kernel-ark/mm/slab.c

4161 lines
108 KiB
C
Raw Normal View History

/*
* linux/mm/slab.c
* Written by Mark Hemment, 1996/97.
* (markhe@nextd.demon.co.uk)
*
* kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
*
* Major cleanup, different bufctl logic, per-cpu arrays
* (c) 2000 Manfred Spraul
*
* Cleanup, make the head arrays unconditional, preparation for NUMA
* (c) 2002 Manfred Spraul
*
* An implementation of the Slab Allocator as described in outline in;
* UNIX Internals: The New Frontiers by Uresh Vahalia
* Pub: Prentice Hall ISBN 0-13-101908-2
* or with a little more detail in;
* The Slab Allocator: An Object-Caching Kernel Memory Allocator
* Jeff Bonwick (Sun Microsystems).
* Presented at: USENIX Summer 1994 Technical Conference
*
* The memory is organized in caches, one cache for each object type.
* (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
* Each cache consists out of many slabs (they are small (usually one
* page long) and always contiguous), and each slab contains multiple
* initialized objects.
*
* This means, that your constructor is used only for newly allocated
* slabs and you must pass objects with the same intializations to
* kmem_cache_free.
*
* Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
* normal). If you need a special memory type, then must create a new
* cache for that memory type.
*
* In order to reduce fragmentation, the slabs are sorted in 3 groups:
* full slabs with 0 free objects
* partial slabs
* empty slabs with no allocated objects
*
* If partial slabs exist, then new allocations come from these slabs,
* otherwise from empty slabs or new slabs are allocated.
*
* kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
* during kmem_cache_destroy(). The caller must prevent concurrent allocs.
*
* Each cache has a short per-cpu head array, most allocs
* and frees go into that array, and if that array overflows, then 1/2
* of the entries in the array are given back into the global cache.
* The head array is strictly LIFO and should improve the cache hit rates.
* On SMP, it additionally reduces the spinlock operations.
*
* The c_cpuarray may not be read with enabled local interrupts -
* it's changed with a smp_call_function().
*
* SMP synchronization:
* constructors and destructors are called without any locking.
* Several members in struct kmem_cache and struct slab never change, they
* are accessed without any locking.
* The per-cpu arrays are never accessed from the wrong cpu, no locking,
* and local interrupts are disabled so slab code is preempt-safe.
* The non-constant members are protected with a per-cache irq spinlock.
*
* Many thanks to Mark Hemment, who wrote another per-cpu slab patch
* in 2000 - many ideas in the current implementation are derived from
* his patch.
*
* Further notes from the original documentation:
*
* 11 April '97. Started multi-threading - markhe
* The global cache-chain is protected by the mutex 'cache_chain_mutex'.
* The sem is only needed when accessing/extending the cache-chain, which
* can never happen inside an interrupt (kmem_cache_create(),
* kmem_cache_shrink() and kmem_cache_reap()).
*
* At present, each engine can be growing a cache. This should be blocked.
*
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
* 15 March 2005. NUMA slab allocator.
* Shai Fultheim <shai@scalex86.org>.
* Shobhit Dayal <shobhit@calsoftinc.com>
* Alok N Kataria <alokk@calsoftinc.com>
* Christoph Lameter <christoph@lameter.com>
*
* Modified the slab allocator to be node aware on NUMA systems.
* Each node has its own list of partial, free and full slabs.
* All object allocations for a node occur from node specific slab lists.
*/
#include <linux/config.h>
#include <linux/slab.h>
#include <linux/mm.h>
#include <linux/swap.h>
#include <linux/cache.h>
#include <linux/interrupt.h>
#include <linux/init.h>
#include <linux/compiler.h>
[PATCH] cpuset memory spread slab cache implementation Provide the slab cache infrastructure to support cpuset memory spreading. See the previous patches, cpuset_mem_spread, for an explanation of cpuset memory spreading. This patch provides a slab cache SLAB_MEM_SPREAD flag. If set in the kmem_cache_create() call defining a slab cache, then any task marked with the process state flag PF_MEMSPREAD will spread memory page allocations for that cache over all the allowed nodes, instead of preferring the local (faulting) node. On systems not configured with CONFIG_NUMA, this results in no change to the page allocation code path for slab caches. On systems with cpusets configured in the kernel, but the "memory_spread" cpuset option not enabled for the current tasks cpuset, this adds a call to a cpuset routine and failed bit test of the processor state flag PF_SPREAD_SLAB. For tasks so marked, a second inline test is done for the slab cache flag SLAB_MEM_SPREAD, and if that is set and if the allocation is not in_interrupt(), this adds a call to to a cpuset routine that computes which of the tasks mems_allowed nodes should be preferred for this allocation. ==> This patch adds another hook into the performance critical code path to allocating objects from the slab cache, in the ____cache_alloc() chunk, below. The next patch optimizes this hook, reducing the impact of the combined mempolicy plus memory spreading hooks on this critical code path to a single check against the tasks task_struct flags word. This patch provides the generic slab flags and logic needed to apply memory spreading to a particular slab. A subsequent patch will mark a few specific slab caches for this placement policy. Signed-off-by: Paul Jackson <pj@sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-03-24 11:16:07 +00:00
#include <linux/cpuset.h>
#include <linux/seq_file.h>
#include <linux/notifier.h>
#include <linux/kallsyms.h>
#include <linux/cpu.h>
#include <linux/sysctl.h>
#include <linux/module.h>
#include <linux/rcupdate.h>
#include <linux/string.h>
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
#include <linux/nodemask.h>
[PATCH] NUMA policies in the slab allocator V2 This patch fixes a regression in 2.6.14 against 2.6.13 that causes an imbalance in memory allocation during bootup. The slab allocator in 2.6.13 is not numa aware and simply calls alloc_pages(). This means that memory policies may control the behavior of alloc_pages(). During bootup the memory policy is set to MPOL_INTERLEAVE resulting in the spreading out of allocations during bootup over all available nodes. The slab allocator in 2.6.13 has only a single list of slab pages. As a result the per cpu slab cache and the spinlock controlled page lists may contain slab entries from off node memory. The slab allocator in 2.6.13 makes no effort to discern the locality of an entry on its lists. The NUMA aware slab allocator in 2.6.14 controls locality of the slab pages explicitly by calling alloc_pages_node(). The NUMA slab allocator manages slab entries by having lists of available slab pages for each node. The per cpu slab cache can only contain slab entries associated with the node local to the processor. This guarantees that the default allocation mode of the slab allocator always assigns local memory if available. Setting MPOL_INTERLEAVE as a default policy during bootup has no effect anymore. In 2.6.14 all node unspecific slab allocations are performed on the boot processor. This means that most of key data structures are allocated on one node. Most processors will have to refer to these structures making the boot node a potential bottleneck. This may reduce performance and cause unnecessary memory pressure on the boot node. This patch implements NUMA policies in the slab layer. There is the need of explicit application of NUMA memory policies by the slab allcator itself since the NUMA slab allocator does no longer let the page_allocator control locality. The check for policies is made directly at the beginning of __cache_alloc using current->mempolicy. The memory policy is already frequently checked by the page allocator (alloc_page_vma() and alloc_page_current()). So it is highly likely that the cacheline is present. For MPOL_INTERLEAVE kmalloc() will spread out each request to one node after another so that an equal distribution of allocations can be obtained during bootup. It is not possible to push the policy check to lower layers of the NUMA slab allocator since the per cpu caches are now only containing slab entries from the current node. If the policy says that the local node is not to be preferred or forbidden then there is no point in checking the slab cache or local list of slab pages. The allocation better be directed immediately to the lists containing slab entries for the allowed set of nodes. This way of applying policy also fixes another strange behavior in 2.6.13. alloc_pages() is controlled by the memory allocation policy of the current process. It could therefore be that one process is running with MPOL_INTERLEAVE and would f.e. obtain a new page following that policy since no slab entries are in the lists anymore. A page can typically be used for multiple slab entries but lets say that the current process is only using one. The other entries are then added to the slab lists. These are now non local entries in the slab lists despite of the possible availability of local pages that would provide faster access and increase the performance of the application. Another process without MPOL_INTERLEAVE may now run and expect a local slab entry from kmalloc(). However, there are still these free slab entries from the off node page obtained from the other process via MPOL_INTERLEAVE in the cache. The process will then get an off node slab entry although other slab entries may be available that are local to that process. This means that the policy if one process may contaminate the locality of the slab caches for other processes. This patch in effect insures that a per process policy is followed for the allocation of slab entries and that there cannot be a memory policy influence from one process to another. A process with default policy will always get a local slab entry if one is available. And the process using memory policies will get its memory arranged as requested. Off-node slab allocation will require the use of spinlocks and will make the use of per cpu caches not possible. A process using memory policies to redirect allocations offnode will have to cope with additional lock overhead in addition to the latency added by the need to access a remote slab entry. Changes V1->V2 - Remove #ifdef CONFIG_NUMA by moving forward declaration into prior #ifdef CONFIG_NUMA section. - Give the function determining the node number to use a saner name. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-01-19 01:42:36 +00:00
#include <linux/mempolicy.h>
#include <linux/mutex.h>
#include <asm/uaccess.h>
#include <asm/cacheflush.h>
#include <asm/tlbflush.h>
#include <asm/page.h>
/*
* DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
* SLAB_RED_ZONE & SLAB_POISON.
* 0 for faster, smaller code (especially in the critical paths).
*
* STATS - 1 to collect stats for /proc/slabinfo.
* 0 for faster, smaller code (especially in the critical paths).
*
* FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
*/
#ifdef CONFIG_DEBUG_SLAB
#define DEBUG 1
#define STATS 1
#define FORCED_DEBUG 1
#else
#define DEBUG 0
#define STATS 0
#define FORCED_DEBUG 0
#endif
/* Shouldn't this be in a header file somewhere? */
#define BYTES_PER_WORD sizeof(void *)
#ifndef cache_line_size
#define cache_line_size() L1_CACHE_BYTES
#endif
#ifndef ARCH_KMALLOC_MINALIGN
/*
* Enforce a minimum alignment for the kmalloc caches.
* Usually, the kmalloc caches are cache_line_size() aligned, except when
* DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
* Some archs want to perform DMA into kmalloc caches and need a guaranteed
* alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
* Note that this flag disables some debug features.
*/
#define ARCH_KMALLOC_MINALIGN 0
#endif
#ifndef ARCH_SLAB_MINALIGN
/*
* Enforce a minimum alignment for all caches.
* Intended for archs that get misalignment faults even for BYTES_PER_WORD
* aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
* If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
* some debug features.
*/
#define ARCH_SLAB_MINALIGN 0
#endif
#ifndef ARCH_KMALLOC_FLAGS
#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
#endif
/* Legal flag mask for kmem_cache_create(). */
#if DEBUG
# define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
SLAB_POISON | SLAB_HWCACHE_ALIGN | \
SLAB_CACHE_DMA | \
SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
[PATCH] cpuset memory spread slab cache implementation Provide the slab cache infrastructure to support cpuset memory spreading. See the previous patches, cpuset_mem_spread, for an explanation of cpuset memory spreading. This patch provides a slab cache SLAB_MEM_SPREAD flag. If set in the kmem_cache_create() call defining a slab cache, then any task marked with the process state flag PF_MEMSPREAD will spread memory page allocations for that cache over all the allowed nodes, instead of preferring the local (faulting) node. On systems not configured with CONFIG_NUMA, this results in no change to the page allocation code path for slab caches. On systems with cpusets configured in the kernel, but the "memory_spread" cpuset option not enabled for the current tasks cpuset, this adds a call to a cpuset routine and failed bit test of the processor state flag PF_SPREAD_SLAB. For tasks so marked, a second inline test is done for the slab cache flag SLAB_MEM_SPREAD, and if that is set and if the allocation is not in_interrupt(), this adds a call to to a cpuset routine that computes which of the tasks mems_allowed nodes should be preferred for this allocation. ==> This patch adds another hook into the performance critical code path to allocating objects from the slab cache, in the ____cache_alloc() chunk, below. The next patch optimizes this hook, reducing the impact of the combined mempolicy plus memory spreading hooks on this critical code path to a single check against the tasks task_struct flags word. This patch provides the generic slab flags and logic needed to apply memory spreading to a particular slab. A subsequent patch will mark a few specific slab caches for this placement policy. Signed-off-by: Paul Jackson <pj@sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-03-24 11:16:07 +00:00
SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
#else
# define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
[PATCH] cpuset memory spread slab cache implementation Provide the slab cache infrastructure to support cpuset memory spreading. See the previous patches, cpuset_mem_spread, for an explanation of cpuset memory spreading. This patch provides a slab cache SLAB_MEM_SPREAD flag. If set in the kmem_cache_create() call defining a slab cache, then any task marked with the process state flag PF_MEMSPREAD will spread memory page allocations for that cache over all the allowed nodes, instead of preferring the local (faulting) node. On systems not configured with CONFIG_NUMA, this results in no change to the page allocation code path for slab caches. On systems with cpusets configured in the kernel, but the "memory_spread" cpuset option not enabled for the current tasks cpuset, this adds a call to a cpuset routine and failed bit test of the processor state flag PF_SPREAD_SLAB. For tasks so marked, a second inline test is done for the slab cache flag SLAB_MEM_SPREAD, and if that is set and if the allocation is not in_interrupt(), this adds a call to to a cpuset routine that computes which of the tasks mems_allowed nodes should be preferred for this allocation. ==> This patch adds another hook into the performance critical code path to allocating objects from the slab cache, in the ____cache_alloc() chunk, below. The next patch optimizes this hook, reducing the impact of the combined mempolicy plus memory spreading hooks on this critical code path to a single check against the tasks task_struct flags word. This patch provides the generic slab flags and logic needed to apply memory spreading to a particular slab. A subsequent patch will mark a few specific slab caches for this placement policy. Signed-off-by: Paul Jackson <pj@sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-03-24 11:16:07 +00:00
SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
#endif
/*
* kmem_bufctl_t:
*
* Bufctl's are used for linking objs within a slab
* linked offsets.
*
* This implementation relies on "struct page" for locating the cache &
* slab an object belongs to.
* This allows the bufctl structure to be small (one int), but limits
* the number of objects a slab (not a cache) can contain when off-slab
* bufctls are used. The limit is the size of the largest general cache
* that does not use off-slab slabs.
* For 32bit archs with 4 kB pages, is this 56.
* This is not serious, as it is only for large objects, when it is unwise
* to have too many per slab.
* Note: This limit can be raised by introducing a general cache whose size
* is less than 512 (PAGE_SIZE<<3), but greater than 256.
*/
typedef unsigned int kmem_bufctl_t;
#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
[PATCH] slab: implement /proc/slab_allocators Implement /proc/slab_allocators. It produces output like: idr_layer_cache: 80 idr_pre_get+0x33/0x4e buffer_head: 2555 alloc_buffer_head+0x20/0x75 mm_struct: 9 mm_alloc+0x1e/0x42 mm_struct: 20 dup_mm+0x36/0x370 vm_area_struct: 384 dup_mm+0x18f/0x370 vm_area_struct: 151 do_mmap_pgoff+0x2e0/0x7c3 vm_area_struct: 1 split_vma+0x5a/0x10e vm_area_struct: 11 do_brk+0x206/0x2e2 vm_area_struct: 2 copy_vma+0xda/0x142 vm_area_struct: 9 setup_arg_pages+0x99/0x214 fs_cache: 8 copy_fs_struct+0x21/0x133 fs_cache: 29 copy_process+0xf38/0x10e3 files_cache: 30 alloc_files+0x1b/0xcf signal_cache: 81 copy_process+0xbaa/0x10e3 sighand_cache: 77 copy_process+0xe65/0x10e3 sighand_cache: 1 de_thread+0x4d/0x5f8 anon_vma: 241 anon_vma_prepare+0xd9/0xf3 size-2048: 1 add_sect_attrs+0x5f/0x145 size-2048: 2 journal_init_revoke+0x99/0x302 size-2048: 2 journal_init_revoke+0x137/0x302 size-2048: 2 journal_init_inode+0xf9/0x1c4 Cc: Manfred Spraul <manfred@colorfullife.com> Cc: Alexander Nyberg <alexn@telia.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: Ravikiran Thirumalai <kiran@scalex86.org> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk> DESC slab-leaks3-locking-fix EDESC From: Andrew Morton <akpm@osdl.org> Update for slab-remove-cachep-spinlock.patch Cc: Al Viro <viro@ftp.linux.org.uk> Cc: Manfred Spraul <manfred@colorfullife.com> Cc: Alexander Nyberg <alexn@telia.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: Ravikiran Thirumalai <kiran@scalex86.org> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-03-25 11:06:39 +00:00
#define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
/* Max number of objs-per-slab for caches which use off-slab slabs.
* Needed to avoid a possible looping condition in cache_grow().
*/
static unsigned long offslab_limit;
/*
* struct slab
*
* Manages the objs in a slab. Placed either at the beginning of mem allocated
* for a slab, or allocated from an general cache.
* Slabs are chained into three list: fully used, partial, fully free slabs.
*/
struct slab {
struct list_head list;
unsigned long colouroff;
void *s_mem; /* including colour offset */
unsigned int inuse; /* num of objs active in slab */
kmem_bufctl_t free;
unsigned short nodeid;
};
/*
* struct slab_rcu
*
* slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
* arrange for kmem_freepages to be called via RCU. This is useful if
* we need to approach a kernel structure obliquely, from its address
* obtained without the usual locking. We can lock the structure to
* stabilize it and check it's still at the given address, only if we
* can be sure that the memory has not been meanwhile reused for some
* other kind of object (which our subsystem's lock might corrupt).
*
* rcu_read_lock before reading the address, then rcu_read_unlock after
* taking the spinlock within the structure expected at that address.
*
* We assume struct slab_rcu can overlay struct slab when destroying.
*/
struct slab_rcu {
struct rcu_head head;
struct kmem_cache *cachep;
void *addr;
};
/*
* struct array_cache
*
* Purpose:
* - LIFO ordering, to hand out cache-warm objects from _alloc
* - reduce the number of linked list operations
* - reduce spinlock operations
*
* The limit is stored in the per-cpu structure to reduce the data cache
* footprint.
*
*/
struct array_cache {
unsigned int avail;
unsigned int limit;
unsigned int batchcount;
unsigned int touched;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
spinlock_t lock;
void *entry[0]; /*
* Must have this definition in here for the proper
* alignment of array_cache. Also simplifies accessing
* the entries.
* [0] is for gcc 2.95. It should really be [].
*/
};
/*
* bootstrap: The caches do not work without cpuarrays anymore, but the
* cpuarrays are allocated from the generic caches...
*/
#define BOOT_CPUCACHE_ENTRIES 1
struct arraycache_init {
struct array_cache cache;
void *entries[BOOT_CPUCACHE_ENTRIES];
};
/*
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
* The slab lists for all objects.
*/
struct kmem_list3 {
struct list_head slabs_partial; /* partial list first, better asm code */
struct list_head slabs_full;
struct list_head slabs_free;
unsigned long free_objects;
unsigned int free_limit;
unsigned int colour_next; /* Per-node cache coloring */
spinlock_t list_lock;
struct array_cache *shared; /* shared per node */
struct array_cache **alien; /* on other nodes */
unsigned long next_reap; /* updated without locking */
int free_touched; /* updated without locking */
};
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
/*
* Need this for bootstrapping a per node allocator.
*/
#define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
#define CACHE_CACHE 0
#define SIZE_AC 1
#define SIZE_L3 (1 + MAX_NUMNODES)
/*
* This function must be completely optimized away if a constant is passed to
* it. Mostly the same as what is in linux/slab.h except it returns an index.
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
*/
static __always_inline int index_of(const size_t size)
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
{
extern void __bad_size(void);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
if (__builtin_constant_p(size)) {
int i = 0;
#define CACHE(x) \
if (size <=x) \
return i; \
else \
i++;
#include "linux/kmalloc_sizes.h"
#undef CACHE
__bad_size();
} else
__bad_size();
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
return 0;
}
#define INDEX_AC index_of(sizeof(struct arraycache_init))
#define INDEX_L3 index_of(sizeof(struct kmem_list3))
static void kmem_list3_init(struct kmem_list3 *parent)
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
{
INIT_LIST_HEAD(&parent->slabs_full);
INIT_LIST_HEAD(&parent->slabs_partial);
INIT_LIST_HEAD(&parent->slabs_free);
parent->shared = NULL;
parent->alien = NULL;
parent->colour_next = 0;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
spin_lock_init(&parent->list_lock);
parent->free_objects = 0;
parent->free_touched = 0;
}
#define MAKE_LIST(cachep, listp, slab, nodeid) \
do { \
INIT_LIST_HEAD(listp); \
list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
} while (0)
#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
do { \
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
} while (0)
/*
* struct kmem_cache
*
* manages a cache.
*/
struct kmem_cache {
/* 1) per-cpu data, touched during every alloc/free */
struct array_cache *array[NR_CPUS];
/* 2) Cache tunables. Protected by cache_chain_mutex */
unsigned int batchcount;
unsigned int limit;
unsigned int shared;
unsigned int buffer_size;
/* 3) touched by every alloc & free from the backend */
struct kmem_list3 *nodelists[MAX_NUMNODES];
unsigned int flags; /* constant flags */
unsigned int num; /* # of objs per slab */
/* 4) cache_grow/shrink */
/* order of pgs per slab (2^n) */
unsigned int gfporder;
/* force GFP flags, e.g. GFP_DMA */
gfp_t gfpflags;
size_t colour; /* cache colouring range */
unsigned int colour_off; /* colour offset */
struct kmem_cache *slabp_cache;
unsigned int slab_size;
unsigned int dflags; /* dynamic flags */
/* constructor func */
void (*ctor) (void *, struct kmem_cache *, unsigned long);
/* de-constructor func */
void (*dtor) (void *, struct kmem_cache *, unsigned long);
/* 5) cache creation/removal */
const char *name;
struct list_head next;
/* 6) statistics */
#if STATS
unsigned long num_active;
unsigned long num_allocations;
unsigned long high_mark;
unsigned long grown;
unsigned long reaped;
unsigned long errors;
unsigned long max_freeable;
unsigned long node_allocs;
unsigned long node_frees;
unsigned long node_overflow;
atomic_t allochit;
atomic_t allocmiss;
atomic_t freehit;
atomic_t freemiss;
#endif
#if DEBUG
/*
* If debugging is enabled, then the allocator can add additional
* fields and/or padding to every object. buffer_size contains the total
* object size including these internal fields, the following two
* variables contain the offset to the user object and its size.
*/
int obj_offset;
int obj_size;
#endif
};
#define CFLGS_OFF_SLAB (0x80000000UL)
#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
#define BATCHREFILL_LIMIT 16
/*
* Optimization question: fewer reaps means less probability for unnessary
* cpucache drain/refill cycles.
*
2005-11-08 15:44:08 +00:00
* OTOH the cpuarrays can contain lots of objects,
* which could lock up otherwise freeable slabs.
*/
#define REAPTIMEOUT_CPUC (2*HZ)
#define REAPTIMEOUT_LIST3 (4*HZ)
#if STATS
#define STATS_INC_ACTIVE(x) ((x)->num_active++)
#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
#define STATS_INC_GROWN(x) ((x)->grown++)
#define STATS_INC_REAPED(x) ((x)->reaped++)
#define STATS_SET_HIGH(x) \
do { \
if ((x)->num_active > (x)->high_mark) \
(x)->high_mark = (x)->num_active; \
} while (0)
#define STATS_INC_ERR(x) ((x)->errors++)
#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
#define STATS_SET_FREEABLE(x, i) \
do { \
if ((x)->max_freeable < i) \
(x)->max_freeable = i; \
} while (0)
#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
#else
#define STATS_INC_ACTIVE(x) do { } while (0)
#define STATS_DEC_ACTIVE(x) do { } while (0)
#define STATS_INC_ALLOCED(x) do { } while (0)
#define STATS_INC_GROWN(x) do { } while (0)
#define STATS_INC_REAPED(x) do { } while (0)
#define STATS_SET_HIGH(x) do { } while (0)
#define STATS_INC_ERR(x) do { } while (0)
#define STATS_INC_NODEALLOCS(x) do { } while (0)
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
#define STATS_INC_NODEFREES(x) do { } while (0)
#define STATS_INC_ACOVERFLOW(x) do { } while (0)
#define STATS_SET_FREEABLE(x, i) do { } while (0)
#define STATS_INC_ALLOCHIT(x) do { } while (0)
#define STATS_INC_ALLOCMISS(x) do { } while (0)
#define STATS_INC_FREEHIT(x) do { } while (0)
#define STATS_INC_FREEMISS(x) do { } while (0)
#endif
#if DEBUG
/*
* Magic nums for obj red zoning.
* Placed in the first word before and the first word after an obj.
*/
#define RED_INACTIVE 0x5A2CF071UL /* when obj is inactive */
#define RED_ACTIVE 0x170FC2A5UL /* when obj is active */
/* ...and for poisoning */
#define POISON_INUSE 0x5a /* for use-uninitialised poisoning */
#define POISON_FREE 0x6b /* for use-after-free poisoning */
#define POISON_END 0xa5 /* end-byte of poisoning */
/*
* memory layout of objects:
* 0 : objp
* 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
* the end of an object is aligned with the end of the real
* allocation. Catches writes behind the end of the allocation.
* cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
* redzone word.
* cachep->obj_offset: The real object.
* cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
* cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
* [BYTES_PER_WORD long]
*/
static int obj_offset(struct kmem_cache *cachep)
{
return cachep->obj_offset;
}
static int obj_size(struct kmem_cache *cachep)
{
return cachep->obj_size;
}
static unsigned long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
{
BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD);
}
static unsigned long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
{
BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
if (cachep->flags & SLAB_STORE_USER)
return (unsigned long *)(objp + cachep->buffer_size -
2 * BYTES_PER_WORD);
return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD);
}
static void **dbg_userword(struct kmem_cache *cachep, void *objp)
{
BUG_ON(!(cachep->flags & SLAB_STORE_USER));
return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
}
#else
#define obj_offset(x) 0
#define obj_size(cachep) (cachep->buffer_size)
#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
#endif
/*
* Maximum size of an obj (in 2^order pages) and absolute limit for the gfp
* order.
*/
#if defined(CONFIG_LARGE_ALLOCS)
#define MAX_OBJ_ORDER 13 /* up to 32Mb */
#define MAX_GFP_ORDER 13 /* up to 32Mb */
#elif defined(CONFIG_MMU)
#define MAX_OBJ_ORDER 5 /* 32 pages */
#define MAX_GFP_ORDER 5 /* 32 pages */
#else
#define MAX_OBJ_ORDER 8 /* up to 1Mb */
#define MAX_GFP_ORDER 8 /* up to 1Mb */
#endif
/*
* Do not go above this order unless 0 objects fit into the slab.
*/
#define BREAK_GFP_ORDER_HI 1
#define BREAK_GFP_ORDER_LO 0
static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
/*
* Functions for storing/retrieving the cachep and or slab from the page
* allocator. These are used to find the slab an obj belongs to. With kfree(),
* these are used to find the cache which an obj belongs to.
*/
static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
{
page->lru.next = (struct list_head *)cache;
}
static inline struct kmem_cache *page_get_cache(struct page *page)
{
if (unlikely(PageCompound(page)))
page = (struct page *)page_private(page);
return (struct kmem_cache *)page->lru.next;
}
static inline void page_set_slab(struct page *page, struct slab *slab)
{
page->lru.prev = (struct list_head *)slab;
}
static inline struct slab *page_get_slab(struct page *page)
{
if (unlikely(PageCompound(page)))
page = (struct page *)page_private(page);
return (struct slab *)page->lru.prev;
}
static inline struct kmem_cache *virt_to_cache(const void *obj)
{
struct page *page = virt_to_page(obj);
return page_get_cache(page);
}
static inline struct slab *virt_to_slab(const void *obj)
{
struct page *page = virt_to_page(obj);
return page_get_slab(page);
}
static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
unsigned int idx)
{
return slab->s_mem + cache->buffer_size * idx;
}
static inline unsigned int obj_to_index(struct kmem_cache *cache,
struct slab *slab, void *obj)
{
return (unsigned)(obj - slab->s_mem) / cache->buffer_size;
}
/*
* These are the default caches for kmalloc. Custom caches can have other sizes.
*/
struct cache_sizes malloc_sizes[] = {
#define CACHE(x) { .cs_size = (x) },
#include <linux/kmalloc_sizes.h>
CACHE(ULONG_MAX)
#undef CACHE
};
EXPORT_SYMBOL(malloc_sizes);
/* Must match cache_sizes above. Out of line to keep cache footprint low. */
struct cache_names {
char *name;
char *name_dma;
};
static struct cache_names __initdata cache_names[] = {
#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
#include <linux/kmalloc_sizes.h>
{NULL,}
#undef CACHE
};
static struct arraycache_init initarray_cache __initdata =
{ {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
static struct arraycache_init initarray_generic =
{ {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
/* internal cache of cache description objs */
static struct kmem_cache cache_cache = {
.batchcount = 1,
.limit = BOOT_CPUCACHE_ENTRIES,
.shared = 1,
.buffer_size = sizeof(struct kmem_cache),
.name = "kmem_cache",
#if DEBUG
.obj_size = sizeof(struct kmem_cache),
#endif
};
/* Guard access to the cache-chain. */
static DEFINE_MUTEX(cache_chain_mutex);
static struct list_head cache_chain;
/*
* vm_enough_memory() looks at this to determine how many slab-allocated pages
* are possibly freeable under pressure
*
* SLAB_RECLAIM_ACCOUNT turns this on per-slab
*/
atomic_t slab_reclaim_pages;
/*
* chicken and egg problem: delay the per-cpu array allocation
* until the general caches are up.
*/
static enum {
NONE,
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
PARTIAL_AC,
PARTIAL_L3,
FULL
} g_cpucache_up;
static DEFINE_PER_CPU(struct work_struct, reap_work);
static void free_block(struct kmem_cache *cachep, void **objpp, int len,
int node);
static void enable_cpucache(struct kmem_cache *cachep);
static void cache_reap(void *unused);
static int __node_shrink(struct kmem_cache *cachep, int node);
static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
{
return cachep->array[smp_processor_id()];
}
static inline struct kmem_cache *__find_general_cachep(size_t size,
gfp_t gfpflags)
{
struct cache_sizes *csizep = malloc_sizes;
#if DEBUG
/* This happens if someone tries to call
* kmem_cache_create(), or __kmalloc(), before
* the generic caches are initialized.
*/
BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
#endif
while (size > csizep->cs_size)
csizep++;
/*
* Really subtle: The last entry with cs->cs_size==ULONG_MAX
* has cs_{dma,}cachep==NULL. Thus no special case
* for large kmalloc calls required.
*/
if (unlikely(gfpflags & GFP_DMA))
return csizep->cs_dmacachep;
return csizep->cs_cachep;
}
struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
[PATCH] add kmalloc_node, inline cleanup The patch makes the following function calls available to allocate memory on a specific node without changing the basic operation of the slab allocator: kmem_cache_alloc_node(kmem_cache_t *cachep, unsigned int flags, int node); kmalloc_node(size_t size, unsigned int flags, int node); in a similar way to the existing node-blind functions: kmem_cache_alloc(kmem_cache_t *cachep, unsigned int flags); kmalloc(size, flags); kmem_cache_alloc_node was changed to pass flags and the node information through the existing layers of the slab allocator (which lead to some minor rearrangements). The functions at the lowest layer (kmem_getpages, cache_grow) are already node aware. Also __alloc_percpu can call kmalloc_node now. Performance measurements (using the pageset localization patch) yields: w/o patches: Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.97 Wed Mar 30 20:50:43 2005 100 25170.83 91 251.7083 23.12 150.10 Wed Mar 30 20:51:06 2005 200 34601.66 84 173.0083 33.64 294.14 Wed Mar 30 20:51:40 2005 300 37154.47 86 123.8482 46.99 436.56 Wed Mar 30 20:52:28 2005 400 39839.82 80 99.5995 58.43 580.46 Wed Mar 30 20:53:27 2005 500 40036.32 79 80.0726 72.68 728.60 Wed Mar 30 20:54:40 2005 600 44074.21 79 73.4570 79.23 872.10 Wed Mar 30 20:55:59 2005 700 44016.60 78 62.8809 92.56 1015.84 Wed Mar 30 20:57:32 2005 800 40411.05 80 50.5138 115.22 1161.13 Wed Mar 30 20:59:28 2005 900 42298.56 79 46.9984 123.83 1303.42 Wed Mar 30 21:01:33 2005 1000 40955.05 80 40.9551 142.11 1441.92 Wed Mar 30 21:03:55 2005 with pageset localization and slab API patches: Tasks jobs/min jti jobs/min/task real cpu 1 484.19 100 484.1930 12.02 1.98 Wed Mar 30 21:10:18 2005 100 27428.25 92 274.2825 21.22 149.79 Wed Mar 30 21:10:40 2005 200 37228.94 86 186.1447 31.27 293.49 Wed Mar 30 21:11:12 2005 300 41725.42 85 139.0847 41.84 434.10 Wed Mar 30 21:11:54 2005 400 43032.22 82 107.5805 54.10 582.06 Wed Mar 30 21:12:48 2005 500 42211.23 83 84.4225 68.94 722.61 Wed Mar 30 21:13:58 2005 600 40084.49 82 66.8075 87.12 873.11 Wed Mar 30 21:15:25 2005 700 44169.30 79 63.0990 92.24 1008.77 Wed Mar 30 21:16:58 2005 800 43097.94 79 53.8724 108.03 1155.88 Wed Mar 30 21:18:47 2005 900 41846.75 79 46.4964 125.17 1303.38 Wed Mar 30 21:20:52 2005 1000 40247.85 79 40.2478 144.60 1442.21 Wed Mar 30 21:23:17 2005 Signed-off-by: Christoph Lameter <christoph@lameter.com> Signed-off-by: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-05-01 15:58:38 +00:00
{
return __find_general_cachep(size, gfpflags);
}
EXPORT_SYMBOL(kmem_find_general_cachep);
static size_t slab_mgmt_size(size_t nr_objs, size_t align)
{
return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
}
/*
* Calculate the number of objects and left-over bytes for a given buffer size.
*/
static void cache_estimate(unsigned long gfporder, size_t buffer_size,
size_t align, int flags, size_t *left_over,
unsigned int *num)
{
int nr_objs;
size_t mgmt_size;
size_t slab_size = PAGE_SIZE << gfporder;
/*
* The slab management structure can be either off the slab or
* on it. For the latter case, the memory allocated for a
* slab is used for:
*
* - The struct slab
* - One kmem_bufctl_t for each object
* - Padding to respect alignment of @align
* - @buffer_size bytes for each object
*
* If the slab management structure is off the slab, then the
* alignment will already be calculated into the size. Because
* the slabs are all pages aligned, the objects will be at the
* correct alignment when allocated.
*/
if (flags & CFLGS_OFF_SLAB) {
mgmt_size = 0;
nr_objs = slab_size / buffer_size;
if (nr_objs > SLAB_LIMIT)
nr_objs = SLAB_LIMIT;
} else {
/*
* Ignore padding for the initial guess. The padding
* is at most @align-1 bytes, and @buffer_size is at
* least @align. In the worst case, this result will
* be one greater than the number of objects that fit
* into the memory allocation when taking the padding
* into account.
*/
nr_objs = (slab_size - sizeof(struct slab)) /
(buffer_size + sizeof(kmem_bufctl_t));
/*
* This calculated number will be either the right
* amount, or one greater than what we want.
*/
if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
> slab_size)
nr_objs--;
if (nr_objs > SLAB_LIMIT)
nr_objs = SLAB_LIMIT;
mgmt_size = slab_mgmt_size(nr_objs, align);
}
*num = nr_objs;
*left_over = slab_size - nr_objs*buffer_size - mgmt_size;
}
#define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
static void __slab_error(const char *function, struct kmem_cache *cachep,
char *msg)
{
printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
function, cachep->name, msg);
dump_stack();
}
[PATCH] slab: Node rotor for freeing alien caches and remote per cpu pages. The cache reaper currently tries to free all alien caches and all remote per cpu pages in each pass of cache_reap. For a machines with large number of nodes (such as Altix) this may lead to sporadic delays of around ~10ms. Interrupts are disabled while reclaiming creating unacceptable delays. This patch changes that behavior by adding a per cpu reap_node variable. Instead of attempting to free all caches, we free only one alien cache and the per cpu pages from one remote node. That reduces the time spend in cache_reap. However, doing so will lengthen the time it takes to completely drain all remote per cpu pagesets and all alien caches. The time needed will grow with the number of nodes in the system. All caches are drained when they overflow their respective capacity. So the drawback here is only that a bit of memory may be wasted for awhile longer. Details: 1. Rename drain_remote_pages to drain_node_pages to allow the specification of the node to drain of pcp pages. 2. Add additional functions init_reap_node, next_reap_node for NUMA that manage a per cpu reap_node counter. 3. Add a reap_alien function that reaps only from the current reap_node. For us this seems to be a critical issue. Holdoffs of an average of ~7ms cause some HPC benchmarks to slow down significantly. F.e. NAS parallel slows down dramatically. NAS parallel has a 12-16 seconds runtime w/o rotor compared to 5.8 secs with the rotor patches. It gets down to 5.05 secs with the additional interrupt holdoff reductions. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-03-10 01:33:54 +00:00
#ifdef CONFIG_NUMA
/*
* Special reaping functions for NUMA systems called from cache_reap().
* These take care of doing round robin flushing of alien caches (containing
* objects freed on different nodes from which they were allocated) and the
* flushing of remote pcps by calling drain_node_pages.
*/
static DEFINE_PER_CPU(unsigned long, reap_node);
static void init_reap_node(int cpu)
{
int node;
node = next_node(cpu_to_node(cpu), node_online_map);
if (node == MAX_NUMNODES)
node = first_node(node_online_map);
[PATCH] slab: Node rotor for freeing alien caches and remote per cpu pages. The cache reaper currently tries to free all alien caches and all remote per cpu pages in each pass of cache_reap. For a machines with large number of nodes (such as Altix) this may lead to sporadic delays of around ~10ms. Interrupts are disabled while reclaiming creating unacceptable delays. This patch changes that behavior by adding a per cpu reap_node variable. Instead of attempting to free all caches, we free only one alien cache and the per cpu pages from one remote node. That reduces the time spend in cache_reap. However, doing so will lengthen the time it takes to completely drain all remote per cpu pagesets and all alien caches. The time needed will grow with the number of nodes in the system. All caches are drained when they overflow their respective capacity. So the drawback here is only that a bit of memory may be wasted for awhile longer. Details: 1. Rename drain_remote_pages to drain_node_pages to allow the specification of the node to drain of pcp pages. 2. Add additional functions init_reap_node, next_reap_node for NUMA that manage a per cpu reap_node counter. 3. Add a reap_alien function that reaps only from the current reap_node. For us this seems to be a critical issue. Holdoffs of an average of ~7ms cause some HPC benchmarks to slow down significantly. F.e. NAS parallel slows down dramatically. NAS parallel has a 12-16 seconds runtime w/o rotor compared to 5.8 secs with the rotor patches. It gets down to 5.05 secs with the additional interrupt holdoff reductions. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-03-10 01:33:54 +00:00
__get_cpu_var(reap_node) = node;
}
static void next_reap_node(void)
{
int node = __get_cpu_var(reap_node);
/*
* Also drain per cpu pages on remote zones
*/
if (node != numa_node_id())
drain_node_pages(node);
node = next_node(node, node_online_map);
if (unlikely(node >= MAX_NUMNODES))
node = first_node(node_online_map);
__get_cpu_var(reap_node) = node;
}
#else
#define init_reap_node(cpu) do { } while (0)
#define next_reap_node(void) do { } while (0)
#endif
/*
* Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
* via the workqueue/eventd.
* Add the CPU number into the expiration time to minimize the possibility of
* the CPUs getting into lockstep and contending for the global cache chain
* lock.
*/
static void __devinit start_cpu_timer(int cpu)
{
struct work_struct *reap_work = &per_cpu(reap_work, cpu);
/*
* When this gets called from do_initcalls via cpucache_init(),
* init_workqueues() has already run, so keventd will be setup
* at that time.
*/
if (keventd_up() && reap_work->func == NULL) {
[PATCH] slab: Node rotor for freeing alien caches and remote per cpu pages. The cache reaper currently tries to free all alien caches and all remote per cpu pages in each pass of cache_reap. For a machines with large number of nodes (such as Altix) this may lead to sporadic delays of around ~10ms. Interrupts are disabled while reclaiming creating unacceptable delays. This patch changes that behavior by adding a per cpu reap_node variable. Instead of attempting to free all caches, we free only one alien cache and the per cpu pages from one remote node. That reduces the time spend in cache_reap. However, doing so will lengthen the time it takes to completely drain all remote per cpu pagesets and all alien caches. The time needed will grow with the number of nodes in the system. All caches are drained when they overflow their respective capacity. So the drawback here is only that a bit of memory may be wasted for awhile longer. Details: 1. Rename drain_remote_pages to drain_node_pages to allow the specification of the node to drain of pcp pages. 2. Add additional functions init_reap_node, next_reap_node for NUMA that manage a per cpu reap_node counter. 3. Add a reap_alien function that reaps only from the current reap_node. For us this seems to be a critical issue. Holdoffs of an average of ~7ms cause some HPC benchmarks to slow down significantly. F.e. NAS parallel slows down dramatically. NAS parallel has a 12-16 seconds runtime w/o rotor compared to 5.8 secs with the rotor patches. It gets down to 5.05 secs with the additional interrupt holdoff reductions. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-03-10 01:33:54 +00:00
init_reap_node(cpu);
INIT_WORK(reap_work, cache_reap, NULL);
schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
}
}
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
static struct array_cache *alloc_arraycache(int node, int entries,
int batchcount)
{
int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
struct array_cache *nc = NULL;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
nc = kmalloc_node(memsize, GFP_KERNEL, node);
if (nc) {
nc->avail = 0;
nc->limit = entries;
nc->batchcount = batchcount;
nc->touched = 0;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
spin_lock_init(&nc->lock);
}
return nc;
}
/*
* Transfer objects in one arraycache to another.
* Locking must be handled by the caller.
*
* Return the number of entries transferred.
*/
static int transfer_objects(struct array_cache *to,
struct array_cache *from, unsigned int max)
{
/* Figure out how many entries to transfer */
int nr = min(min(from->avail, max), to->limit - to->avail);
if (!nr)
return 0;
memcpy(to->entry + to->avail, from->entry + from->avail -nr,
sizeof(void *) *nr);
from->avail -= nr;
to->avail += nr;
to->touched = 1;
return nr;
}
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
#ifdef CONFIG_NUMA
static void *__cache_alloc_node(struct kmem_cache *, gfp_t, int);
static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
[PATCH] NUMA policies in the slab allocator V2 This patch fixes a regression in 2.6.14 against 2.6.13 that causes an imbalance in memory allocation during bootup. The slab allocator in 2.6.13 is not numa aware and simply calls alloc_pages(). This means that memory policies may control the behavior of alloc_pages(). During bootup the memory policy is set to MPOL_INTERLEAVE resulting in the spreading out of allocations during bootup over all available nodes. The slab allocator in 2.6.13 has only a single list of slab pages. As a result the per cpu slab cache and the spinlock controlled page lists may contain slab entries from off node memory. The slab allocator in 2.6.13 makes no effort to discern the locality of an entry on its lists. The NUMA aware slab allocator in 2.6.14 controls locality of the slab pages explicitly by calling alloc_pages_node(). The NUMA slab allocator manages slab entries by having lists of available slab pages for each node. The per cpu slab cache can only contain slab entries associated with the node local to the processor. This guarantees that the default allocation mode of the slab allocator always assigns local memory if available. Setting MPOL_INTERLEAVE as a default policy during bootup has no effect anymore. In 2.6.14 all node unspecific slab allocations are performed on the boot processor. This means that most of key data structures are allocated on one node. Most processors will have to refer to these structures making the boot node a potential bottleneck. This may reduce performance and cause unnecessary memory pressure on the boot node. This patch implements NUMA policies in the slab layer. There is the need of explicit application of NUMA memory policies by the slab allcator itself since the NUMA slab allocator does no longer let the page_allocator control locality. The check for policies is made directly at the beginning of __cache_alloc using current->mempolicy. The memory policy is already frequently checked by the page allocator (alloc_page_vma() and alloc_page_current()). So it is highly likely that the cacheline is present. For MPOL_INTERLEAVE kmalloc() will spread out each request to one node after another so that an equal distribution of allocations can be obtained during bootup. It is not possible to push the policy check to lower layers of the NUMA slab allocator since the per cpu caches are now only containing slab entries from the current node. If the policy says that the local node is not to be preferred or forbidden then there is no point in checking the slab cache or local list of slab pages. The allocation better be directed immediately to the lists containing slab entries for the allowed set of nodes. This way of applying policy also fixes another strange behavior in 2.6.13. alloc_pages() is controlled by the memory allocation policy of the current process. It could therefore be that one process is running with MPOL_INTERLEAVE and would f.e. obtain a new page following that policy since no slab entries are in the lists anymore. A page can typically be used for multiple slab entries but lets say that the current process is only using one. The other entries are then added to the slab lists. These are now non local entries in the slab lists despite of the possible availability of local pages that would provide faster access and increase the performance of the application. Another process without MPOL_INTERLEAVE may now run and expect a local slab entry from kmalloc(). However, there are still these free slab entries from the off node page obtained from the other process via MPOL_INTERLEAVE in the cache. The process will then get an off node slab entry although other slab entries may be available that are local to that process. This means that the policy if one process may contaminate the locality of the slab caches for other processes. This patch in effect insures that a per process policy is followed for the allocation of slab entries and that there cannot be a memory policy influence from one process to another. A process with default policy will always get a local slab entry if one is available. And the process using memory policies will get its memory arranged as requested. Off-node slab allocation will require the use of spinlocks and will make the use of per cpu caches not possible. A process using memory policies to redirect allocations offnode will have to cope with additional lock overhead in addition to the latency added by the need to access a remote slab entry. Changes V1->V2 - Remove #ifdef CONFIG_NUMA by moving forward declaration into prior #ifdef CONFIG_NUMA section. - Give the function determining the node number to use a saner name. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-01-19 01:42:36 +00:00
static struct array_cache **alloc_alien_cache(int node, int limit)
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
{
struct array_cache **ac_ptr;
int memsize = sizeof(void *) * MAX_NUMNODES;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
int i;
if (limit > 1)
limit = 12;
ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
if (ac_ptr) {
for_each_node(i) {
if (i == node || !node_online(i)) {
ac_ptr[i] = NULL;
continue;
}
ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
if (!ac_ptr[i]) {
for (i--; i <= 0; i--)
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
kfree(ac_ptr[i]);
kfree(ac_ptr);
return NULL;
}
}
}
return ac_ptr;
}
static void free_alien_cache(struct array_cache **ac_ptr)
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
{
int i;
if (!ac_ptr)
return;
for_each_node(i)
kfree(ac_ptr[i]);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
kfree(ac_ptr);
}
static void __drain_alien_cache(struct kmem_cache *cachep,
struct array_cache *ac, int node)
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
{
struct kmem_list3 *rl3 = cachep->nodelists[node];
if (ac->avail) {
spin_lock(&rl3->list_lock);
/*
* Stuff objects into the remote nodes shared array first.
* That way we could avoid the overhead of putting the objects
* into the free lists and getting them back later.
*/
transfer_objects(rl3->shared, ac, ac->limit);
free_block(cachep, ac->entry, ac->avail, node);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
ac->avail = 0;
spin_unlock(&rl3->list_lock);
}
}
[PATCH] slab: Node rotor for freeing alien caches and remote per cpu pages. The cache reaper currently tries to free all alien caches and all remote per cpu pages in each pass of cache_reap. For a machines with large number of nodes (such as Altix) this may lead to sporadic delays of around ~10ms. Interrupts are disabled while reclaiming creating unacceptable delays. This patch changes that behavior by adding a per cpu reap_node variable. Instead of attempting to free all caches, we free only one alien cache and the per cpu pages from one remote node. That reduces the time spend in cache_reap. However, doing so will lengthen the time it takes to completely drain all remote per cpu pagesets and all alien caches. The time needed will grow with the number of nodes in the system. All caches are drained when they overflow their respective capacity. So the drawback here is only that a bit of memory may be wasted for awhile longer. Details: 1. Rename drain_remote_pages to drain_node_pages to allow the specification of the node to drain of pcp pages. 2. Add additional functions init_reap_node, next_reap_node for NUMA that manage a per cpu reap_node counter. 3. Add a reap_alien function that reaps only from the current reap_node. For us this seems to be a critical issue. Holdoffs of an average of ~7ms cause some HPC benchmarks to slow down significantly. F.e. NAS parallel slows down dramatically. NAS parallel has a 12-16 seconds runtime w/o rotor compared to 5.8 secs with the rotor patches. It gets down to 5.05 secs with the additional interrupt holdoff reductions. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-03-10 01:33:54 +00:00
/*
* Called from cache_reap() to regularly drain alien caches round robin.
*/
static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
{
int node = __get_cpu_var(reap_node);
if (l3->alien) {
struct array_cache *ac = l3->alien[node];
if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
[PATCH] slab: Node rotor for freeing alien caches and remote per cpu pages. The cache reaper currently tries to free all alien caches and all remote per cpu pages in each pass of cache_reap. For a machines with large number of nodes (such as Altix) this may lead to sporadic delays of around ~10ms. Interrupts are disabled while reclaiming creating unacceptable delays. This patch changes that behavior by adding a per cpu reap_node variable. Instead of attempting to free all caches, we free only one alien cache and the per cpu pages from one remote node. That reduces the time spend in cache_reap. However, doing so will lengthen the time it takes to completely drain all remote per cpu pagesets and all alien caches. The time needed will grow with the number of nodes in the system. All caches are drained when they overflow their respective capacity. So the drawback here is only that a bit of memory may be wasted for awhile longer. Details: 1. Rename drain_remote_pages to drain_node_pages to allow the specification of the node to drain of pcp pages. 2. Add additional functions init_reap_node, next_reap_node for NUMA that manage a per cpu reap_node counter. 3. Add a reap_alien function that reaps only from the current reap_node. For us this seems to be a critical issue. Holdoffs of an average of ~7ms cause some HPC benchmarks to slow down significantly. F.e. NAS parallel slows down dramatically. NAS parallel has a 12-16 seconds runtime w/o rotor compared to 5.8 secs with the rotor patches. It gets down to 5.05 secs with the additional interrupt holdoff reductions. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-03-10 01:33:54 +00:00
__drain_alien_cache(cachep, ac, node);
spin_unlock_irq(&ac->lock);
}
}
}
static void drain_alien_cache(struct kmem_cache *cachep,
struct array_cache **alien)
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
{
int i = 0;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
struct array_cache *ac;
unsigned long flags;
for_each_online_node(i) {
[PATCH] NUMA slab locking fixes: fix cpu down and up locking This fixes locking and bugs in cpu_down and cpu_up paths of the NUMA slab allocator. Sonny Rao <sonny@burdell.org> reported problems sometime back on POWER5 boxes, when the last cpu on the nodes were being offlined. We could not reproduce the same on x86_64 because the cpumask (node_to_cpumask) was not being updated on cpu down. Since that issue is now fixed, we can reproduce Sonny's problems on x86_64 NUMA, and here is the fix. The problem earlier was on CPU_DOWN, if it was the last cpu on the node to go down, the array_caches (shared, alien) and the kmem_list3 of the node were being freed (kfree) with the kmem_list3 lock held. If the l3 or the array_caches were to come from the same cache being cleared, we hit on badness. This patch cleans up the locking in cpu_up and cpu_down path. We cannot really free l3 on cpu down because, there is no node offlining yet and even though a cpu is not yet up, node local memory can be allocated for it. So l3s are usually allocated at keme_cache_create and destroyed at kmem_cache_destroy. Hence, we don't need cachep->spinlock protection to get to the cachep->nodelist[nodeid] either. Patch survived onlining and offlining on a 4 core 2 node Tyan box with a 4 dbench process running all the time. Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Ravikiran Thirumalai <kiran@scalex86.org> Cc: Christoph Lameter <christoph@lameter.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-02-05 07:27:59 +00:00
ac = alien[i];
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
if (ac) {
spin_lock_irqsave(&ac->lock, flags);
__drain_alien_cache(cachep, ac, i);
spin_unlock_irqrestore(&ac->lock, flags);
}
}
}
#else
[PATCH] NUMA slab locking fixes: fix cpu down and up locking This fixes locking and bugs in cpu_down and cpu_up paths of the NUMA slab allocator. Sonny Rao <sonny@burdell.org> reported problems sometime back on POWER5 boxes, when the last cpu on the nodes were being offlined. We could not reproduce the same on x86_64 because the cpumask (node_to_cpumask) was not being updated on cpu down. Since that issue is now fixed, we can reproduce Sonny's problems on x86_64 NUMA, and here is the fix. The problem earlier was on CPU_DOWN, if it was the last cpu on the node to go down, the array_caches (shared, alien) and the kmem_list3 of the node were being freed (kfree) with the kmem_list3 lock held. If the l3 or the array_caches were to come from the same cache being cleared, we hit on badness. This patch cleans up the locking in cpu_up and cpu_down path. We cannot really free l3 on cpu down because, there is no node offlining yet and even though a cpu is not yet up, node local memory can be allocated for it. So l3s are usually allocated at keme_cache_create and destroyed at kmem_cache_destroy. Hence, we don't need cachep->spinlock protection to get to the cachep->nodelist[nodeid] either. Patch survived onlining and offlining on a 4 core 2 node Tyan box with a 4 dbench process running all the time. Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Ravikiran Thirumalai <kiran@scalex86.org> Cc: Christoph Lameter <christoph@lameter.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-02-05 07:27:59 +00:00
#define drain_alien_cache(cachep, alien) do { } while (0)
[PATCH] slab: Node rotor for freeing alien caches and remote per cpu pages. The cache reaper currently tries to free all alien caches and all remote per cpu pages in each pass of cache_reap. For a machines with large number of nodes (such as Altix) this may lead to sporadic delays of around ~10ms. Interrupts are disabled while reclaiming creating unacceptable delays. This patch changes that behavior by adding a per cpu reap_node variable. Instead of attempting to free all caches, we free only one alien cache and the per cpu pages from one remote node. That reduces the time spend in cache_reap. However, doing so will lengthen the time it takes to completely drain all remote per cpu pagesets and all alien caches. The time needed will grow with the number of nodes in the system. All caches are drained when they overflow their respective capacity. So the drawback here is only that a bit of memory may be wasted for awhile longer. Details: 1. Rename drain_remote_pages to drain_node_pages to allow the specification of the node to drain of pcp pages. 2. Add additional functions init_reap_node, next_reap_node for NUMA that manage a per cpu reap_node counter. 3. Add a reap_alien function that reaps only from the current reap_node. For us this seems to be a critical issue. Holdoffs of an average of ~7ms cause some HPC benchmarks to slow down significantly. F.e. NAS parallel slows down dramatically. NAS parallel has a 12-16 seconds runtime w/o rotor compared to 5.8 secs with the rotor patches. It gets down to 5.05 secs with the additional interrupt holdoff reductions. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-03-10 01:33:54 +00:00
#define reap_alien(cachep, l3) do { } while (0)
[PATCH] NUMA slab locking fixes: fix cpu down and up locking This fixes locking and bugs in cpu_down and cpu_up paths of the NUMA slab allocator. Sonny Rao <sonny@burdell.org> reported problems sometime back on POWER5 boxes, when the last cpu on the nodes were being offlined. We could not reproduce the same on x86_64 because the cpumask (node_to_cpumask) was not being updated on cpu down. Since that issue is now fixed, we can reproduce Sonny's problems on x86_64 NUMA, and here is the fix. The problem earlier was on CPU_DOWN, if it was the last cpu on the node to go down, the array_caches (shared, alien) and the kmem_list3 of the node were being freed (kfree) with the kmem_list3 lock held. If the l3 or the array_caches were to come from the same cache being cleared, we hit on badness. This patch cleans up the locking in cpu_up and cpu_down path. We cannot really free l3 on cpu down because, there is no node offlining yet and even though a cpu is not yet up, node local memory can be allocated for it. So l3s are usually allocated at keme_cache_create and destroyed at kmem_cache_destroy. Hence, we don't need cachep->spinlock protection to get to the cachep->nodelist[nodeid] either. Patch survived onlining and offlining on a 4 core 2 node Tyan box with a 4 dbench process running all the time. Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Ravikiran Thirumalai <kiran@scalex86.org> Cc: Christoph Lameter <christoph@lameter.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-02-05 07:27:59 +00:00
static inline struct array_cache **alloc_alien_cache(int node, int limit)
{
return (struct array_cache **) 0x01020304ul;
}
[PATCH] NUMA slab locking fixes: fix cpu down and up locking This fixes locking and bugs in cpu_down and cpu_up paths of the NUMA slab allocator. Sonny Rao <sonny@burdell.org> reported problems sometime back on POWER5 boxes, when the last cpu on the nodes were being offlined. We could not reproduce the same on x86_64 because the cpumask (node_to_cpumask) was not being updated on cpu down. Since that issue is now fixed, we can reproduce Sonny's problems on x86_64 NUMA, and here is the fix. The problem earlier was on CPU_DOWN, if it was the last cpu on the node to go down, the array_caches (shared, alien) and the kmem_list3 of the node were being freed (kfree) with the kmem_list3 lock held. If the l3 or the array_caches were to come from the same cache being cleared, we hit on badness. This patch cleans up the locking in cpu_up and cpu_down path. We cannot really free l3 on cpu down because, there is no node offlining yet and even though a cpu is not yet up, node local memory can be allocated for it. So l3s are usually allocated at keme_cache_create and destroyed at kmem_cache_destroy. Hence, we don't need cachep->spinlock protection to get to the cachep->nodelist[nodeid] either. Patch survived onlining and offlining on a 4 core 2 node Tyan box with a 4 dbench process running all the time. Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Ravikiran Thirumalai <kiran@scalex86.org> Cc: Christoph Lameter <christoph@lameter.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-02-05 07:27:59 +00:00
static inline void free_alien_cache(struct array_cache **ac_ptr)
{
}
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
#endif
static int cpuup_callback(struct notifier_block *nfb,
unsigned long action, void *hcpu)
{
long cpu = (long)hcpu;
struct kmem_cache *cachep;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
struct kmem_list3 *l3 = NULL;
int node = cpu_to_node(cpu);
int memsize = sizeof(struct kmem_list3);
switch (action) {
case CPU_UP_PREPARE:
mutex_lock(&cache_chain_mutex);
/*
* We need to do this right in the beginning since
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
* alloc_arraycache's are going to use this list.
* kmalloc_node allows us to add the slab to the right
* kmem_list3 and not this cpu's kmem_list3
*/
list_for_each_entry(cachep, &cache_chain, next) {
/*
* Set up the size64 kmemlist for cpu before we can
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
* begin anything. Make sure some other cpu on this
* node has not already allocated this
*/
if (!cachep->nodelists[node]) {
l3 = kmalloc_node(memsize, GFP_KERNEL, node);
if (!l3)
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
goto bad;
kmem_list3_init(l3);
l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
((unsigned long)cachep) % REAPTIMEOUT_LIST3;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
[PATCH] NUMA slab locking fixes: fix cpu down and up locking This fixes locking and bugs in cpu_down and cpu_up paths of the NUMA slab allocator. Sonny Rao <sonny@burdell.org> reported problems sometime back on POWER5 boxes, when the last cpu on the nodes were being offlined. We could not reproduce the same on x86_64 because the cpumask (node_to_cpumask) was not being updated on cpu down. Since that issue is now fixed, we can reproduce Sonny's problems on x86_64 NUMA, and here is the fix. The problem earlier was on CPU_DOWN, if it was the last cpu on the node to go down, the array_caches (shared, alien) and the kmem_list3 of the node were being freed (kfree) with the kmem_list3 lock held. If the l3 or the array_caches were to come from the same cache being cleared, we hit on badness. This patch cleans up the locking in cpu_up and cpu_down path. We cannot really free l3 on cpu down because, there is no node offlining yet and even though a cpu is not yet up, node local memory can be allocated for it. So l3s are usually allocated at keme_cache_create and destroyed at kmem_cache_destroy. Hence, we don't need cachep->spinlock protection to get to the cachep->nodelist[nodeid] either. Patch survived onlining and offlining on a 4 core 2 node Tyan box with a 4 dbench process running all the time. Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Ravikiran Thirumalai <kiran@scalex86.org> Cc: Christoph Lameter <christoph@lameter.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-02-05 07:27:59 +00:00
/*
* The l3s don't come and go as CPUs come and
* go. cache_chain_mutex is sufficient
* protection here.
*/
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
cachep->nodelists[node] = l3;
}
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
spin_lock_irq(&cachep->nodelists[node]->list_lock);
cachep->nodelists[node]->free_limit =
(1 + nr_cpus_node(node)) *
cachep->batchcount + cachep->num;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
spin_unlock_irq(&cachep->nodelists[node]->list_lock);
}
/*
* Now we can go ahead with allocating the shared arrays and
* array caches
*/
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
list_for_each_entry(cachep, &cache_chain, next) {
struct array_cache *nc;
[PATCH] NUMA slab locking fixes: fix cpu down and up locking This fixes locking and bugs in cpu_down and cpu_up paths of the NUMA slab allocator. Sonny Rao <sonny@burdell.org> reported problems sometime back on POWER5 boxes, when the last cpu on the nodes were being offlined. We could not reproduce the same on x86_64 because the cpumask (node_to_cpumask) was not being updated on cpu down. Since that issue is now fixed, we can reproduce Sonny's problems on x86_64 NUMA, and here is the fix. The problem earlier was on CPU_DOWN, if it was the last cpu on the node to go down, the array_caches (shared, alien) and the kmem_list3 of the node were being freed (kfree) with the kmem_list3 lock held. If the l3 or the array_caches were to come from the same cache being cleared, we hit on badness. This patch cleans up the locking in cpu_up and cpu_down path. We cannot really free l3 on cpu down because, there is no node offlining yet and even though a cpu is not yet up, node local memory can be allocated for it. So l3s are usually allocated at keme_cache_create and destroyed at kmem_cache_destroy. Hence, we don't need cachep->spinlock protection to get to the cachep->nodelist[nodeid] either. Patch survived onlining and offlining on a 4 core 2 node Tyan box with a 4 dbench process running all the time. Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Ravikiran Thirumalai <kiran@scalex86.org> Cc: Christoph Lameter <christoph@lameter.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-02-05 07:27:59 +00:00
struct array_cache *shared;
struct array_cache **alien;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
nc = alloc_arraycache(node, cachep->limit,
[PATCH] NUMA slab locking fixes: fix cpu down and up locking This fixes locking and bugs in cpu_down and cpu_up paths of the NUMA slab allocator. Sonny Rao <sonny@burdell.org> reported problems sometime back on POWER5 boxes, when the last cpu on the nodes were being offlined. We could not reproduce the same on x86_64 because the cpumask (node_to_cpumask) was not being updated on cpu down. Since that issue is now fixed, we can reproduce Sonny's problems on x86_64 NUMA, and here is the fix. The problem earlier was on CPU_DOWN, if it was the last cpu on the node to go down, the array_caches (shared, alien) and the kmem_list3 of the node were being freed (kfree) with the kmem_list3 lock held. If the l3 or the array_caches were to come from the same cache being cleared, we hit on badness. This patch cleans up the locking in cpu_up and cpu_down path. We cannot really free l3 on cpu down because, there is no node offlining yet and even though a cpu is not yet up, node local memory can be allocated for it. So l3s are usually allocated at keme_cache_create and destroyed at kmem_cache_destroy. Hence, we don't need cachep->spinlock protection to get to the cachep->nodelist[nodeid] either. Patch survived onlining and offlining on a 4 core 2 node Tyan box with a 4 dbench process running all the time. Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Ravikiran Thirumalai <kiran@scalex86.org> Cc: Christoph Lameter <christoph@lameter.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-02-05 07:27:59 +00:00
cachep->batchcount);
if (!nc)
goto bad;
[PATCH] NUMA slab locking fixes: fix cpu down and up locking This fixes locking and bugs in cpu_down and cpu_up paths of the NUMA slab allocator. Sonny Rao <sonny@burdell.org> reported problems sometime back on POWER5 boxes, when the last cpu on the nodes were being offlined. We could not reproduce the same on x86_64 because the cpumask (node_to_cpumask) was not being updated on cpu down. Since that issue is now fixed, we can reproduce Sonny's problems on x86_64 NUMA, and here is the fix. The problem earlier was on CPU_DOWN, if it was the last cpu on the node to go down, the array_caches (shared, alien) and the kmem_list3 of the node were being freed (kfree) with the kmem_list3 lock held. If the l3 or the array_caches were to come from the same cache being cleared, we hit on badness. This patch cleans up the locking in cpu_up and cpu_down path. We cannot really free l3 on cpu down because, there is no node offlining yet and even though a cpu is not yet up, node local memory can be allocated for it. So l3s are usually allocated at keme_cache_create and destroyed at kmem_cache_destroy. Hence, we don't need cachep->spinlock protection to get to the cachep->nodelist[nodeid] either. Patch survived onlining and offlining on a 4 core 2 node Tyan box with a 4 dbench process running all the time. Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Ravikiran Thirumalai <kiran@scalex86.org> Cc: Christoph Lameter <christoph@lameter.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-02-05 07:27:59 +00:00
shared = alloc_arraycache(node,
cachep->shared * cachep->batchcount,
0xbaadf00d);
if (!shared)
goto bad;
[PATCH] NUMA slab locking fixes: fix cpu down and up locking This fixes locking and bugs in cpu_down and cpu_up paths of the NUMA slab allocator. Sonny Rao <sonny@burdell.org> reported problems sometime back on POWER5 boxes, when the last cpu on the nodes were being offlined. We could not reproduce the same on x86_64 because the cpumask (node_to_cpumask) was not being updated on cpu down. Since that issue is now fixed, we can reproduce Sonny's problems on x86_64 NUMA, and here is the fix. The problem earlier was on CPU_DOWN, if it was the last cpu on the node to go down, the array_caches (shared, alien) and the kmem_list3 of the node were being freed (kfree) with the kmem_list3 lock held. If the l3 or the array_caches were to come from the same cache being cleared, we hit on badness. This patch cleans up the locking in cpu_up and cpu_down path. We cannot really free l3 on cpu down because, there is no node offlining yet and even though a cpu is not yet up, node local memory can be allocated for it. So l3s are usually allocated at keme_cache_create and destroyed at kmem_cache_destroy. Hence, we don't need cachep->spinlock protection to get to the cachep->nodelist[nodeid] either. Patch survived onlining and offlining on a 4 core 2 node Tyan box with a 4 dbench process running all the time. Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Ravikiran Thirumalai <kiran@scalex86.org> Cc: Christoph Lameter <christoph@lameter.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-02-05 07:27:59 +00:00
alien = alloc_alien_cache(node, cachep->limit);
if (!alien)
goto bad;
cachep->array[cpu] = nc;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
l3 = cachep->nodelists[node];
BUG_ON(!l3);
[PATCH] NUMA slab locking fixes: fix cpu down and up locking This fixes locking and bugs in cpu_down and cpu_up paths of the NUMA slab allocator. Sonny Rao <sonny@burdell.org> reported problems sometime back on POWER5 boxes, when the last cpu on the nodes were being offlined. We could not reproduce the same on x86_64 because the cpumask (node_to_cpumask) was not being updated on cpu down. Since that issue is now fixed, we can reproduce Sonny's problems on x86_64 NUMA, and here is the fix. The problem earlier was on CPU_DOWN, if it was the last cpu on the node to go down, the array_caches (shared, alien) and the kmem_list3 of the node were being freed (kfree) with the kmem_list3 lock held. If the l3 or the array_caches were to come from the same cache being cleared, we hit on badness. This patch cleans up the locking in cpu_up and cpu_down path. We cannot really free l3 on cpu down because, there is no node offlining yet and even though a cpu is not yet up, node local memory can be allocated for it. So l3s are usually allocated at keme_cache_create and destroyed at kmem_cache_destroy. Hence, we don't need cachep->spinlock protection to get to the cachep->nodelist[nodeid] either. Patch survived onlining and offlining on a 4 core 2 node Tyan box with a 4 dbench process running all the time. Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Ravikiran Thirumalai <kiran@scalex86.org> Cc: Christoph Lameter <christoph@lameter.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-02-05 07:27:59 +00:00
spin_lock_irq(&l3->list_lock);
if (!l3->shared) {
/*
* We are serialised from CPU_DEAD or
* CPU_UP_CANCELLED by the cpucontrol lock
*/
l3->shared = shared;
shared = NULL;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
}
[PATCH] NUMA slab locking fixes: fix cpu down and up locking This fixes locking and bugs in cpu_down and cpu_up paths of the NUMA slab allocator. Sonny Rao <sonny@burdell.org> reported problems sometime back on POWER5 boxes, when the last cpu on the nodes were being offlined. We could not reproduce the same on x86_64 because the cpumask (node_to_cpumask) was not being updated on cpu down. Since that issue is now fixed, we can reproduce Sonny's problems on x86_64 NUMA, and here is the fix. The problem earlier was on CPU_DOWN, if it was the last cpu on the node to go down, the array_caches (shared, alien) and the kmem_list3 of the node were being freed (kfree) with the kmem_list3 lock held. If the l3 or the array_caches were to come from the same cache being cleared, we hit on badness. This patch cleans up the locking in cpu_up and cpu_down path. We cannot really free l3 on cpu down because, there is no node offlining yet and even though a cpu is not yet up, node local memory can be allocated for it. So l3s are usually allocated at keme_cache_create and destroyed at kmem_cache_destroy. Hence, we don't need cachep->spinlock protection to get to the cachep->nodelist[nodeid] either. Patch survived onlining and offlining on a 4 core 2 node Tyan box with a 4 dbench process running all the time. Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Ravikiran Thirumalai <kiran@scalex86.org> Cc: Christoph Lameter <christoph@lameter.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-02-05 07:27:59 +00:00
#ifdef CONFIG_NUMA
if (!l3->alien) {
l3->alien = alien;
alien = NULL;
}
#endif
spin_unlock_irq(&l3->list_lock);
kfree(shared);
free_alien_cache(alien);
}
mutex_unlock(&cache_chain_mutex);
break;
case CPU_ONLINE:
start_cpu_timer(cpu);
break;
#ifdef CONFIG_HOTPLUG_CPU
case CPU_DEAD:
[PATCH] NUMA slab locking fixes: fix cpu down and up locking This fixes locking and bugs in cpu_down and cpu_up paths of the NUMA slab allocator. Sonny Rao <sonny@burdell.org> reported problems sometime back on POWER5 boxes, when the last cpu on the nodes were being offlined. We could not reproduce the same on x86_64 because the cpumask (node_to_cpumask) was not being updated on cpu down. Since that issue is now fixed, we can reproduce Sonny's problems on x86_64 NUMA, and here is the fix. The problem earlier was on CPU_DOWN, if it was the last cpu on the node to go down, the array_caches (shared, alien) and the kmem_list3 of the node were being freed (kfree) with the kmem_list3 lock held. If the l3 or the array_caches were to come from the same cache being cleared, we hit on badness. This patch cleans up the locking in cpu_up and cpu_down path. We cannot really free l3 on cpu down because, there is no node offlining yet and even though a cpu is not yet up, node local memory can be allocated for it. So l3s are usually allocated at keme_cache_create and destroyed at kmem_cache_destroy. Hence, we don't need cachep->spinlock protection to get to the cachep->nodelist[nodeid] either. Patch survived onlining and offlining on a 4 core 2 node Tyan box with a 4 dbench process running all the time. Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Ravikiran Thirumalai <kiran@scalex86.org> Cc: Christoph Lameter <christoph@lameter.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-02-05 07:27:59 +00:00
/*
* Even if all the cpus of a node are down, we don't free the
* kmem_list3 of any cache. This to avoid a race between
* cpu_down, and a kmalloc allocation from another cpu for
* memory from the node of the cpu going down. The list3
* structure is usually allocated from kmem_cache_create() and
* gets destroyed at kmem_cache_destroy().
*/
/* fall thru */
case CPU_UP_CANCELED:
mutex_lock(&cache_chain_mutex);
list_for_each_entry(cachep, &cache_chain, next) {
struct array_cache *nc;
[PATCH] NUMA slab locking fixes: fix cpu down and up locking This fixes locking and bugs in cpu_down and cpu_up paths of the NUMA slab allocator. Sonny Rao <sonny@burdell.org> reported problems sometime back on POWER5 boxes, when the last cpu on the nodes were being offlined. We could not reproduce the same on x86_64 because the cpumask (node_to_cpumask) was not being updated on cpu down. Since that issue is now fixed, we can reproduce Sonny's problems on x86_64 NUMA, and here is the fix. The problem earlier was on CPU_DOWN, if it was the last cpu on the node to go down, the array_caches (shared, alien) and the kmem_list3 of the node were being freed (kfree) with the kmem_list3 lock held. If the l3 or the array_caches were to come from the same cache being cleared, we hit on badness. This patch cleans up the locking in cpu_up and cpu_down path. We cannot really free l3 on cpu down because, there is no node offlining yet and even though a cpu is not yet up, node local memory can be allocated for it. So l3s are usually allocated at keme_cache_create and destroyed at kmem_cache_destroy. Hence, we don't need cachep->spinlock protection to get to the cachep->nodelist[nodeid] either. Patch survived onlining and offlining on a 4 core 2 node Tyan box with a 4 dbench process running all the time. Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Ravikiran Thirumalai <kiran@scalex86.org> Cc: Christoph Lameter <christoph@lameter.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-02-05 07:27:59 +00:00
struct array_cache *shared;
struct array_cache **alien;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
cpumask_t mask;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
mask = node_to_cpumask(node);
/* cpu is dead; no one can alloc from it. */
nc = cachep->array[cpu];
cachep->array[cpu] = NULL;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
l3 = cachep->nodelists[node];
if (!l3)
[PATCH] NUMA slab locking fixes: fix cpu down and up locking This fixes locking and bugs in cpu_down and cpu_up paths of the NUMA slab allocator. Sonny Rao <sonny@burdell.org> reported problems sometime back on POWER5 boxes, when the last cpu on the nodes were being offlined. We could not reproduce the same on x86_64 because the cpumask (node_to_cpumask) was not being updated on cpu down. Since that issue is now fixed, we can reproduce Sonny's problems on x86_64 NUMA, and here is the fix. The problem earlier was on CPU_DOWN, if it was the last cpu on the node to go down, the array_caches (shared, alien) and the kmem_list3 of the node were being freed (kfree) with the kmem_list3 lock held. If the l3 or the array_caches were to come from the same cache being cleared, we hit on badness. This patch cleans up the locking in cpu_up and cpu_down path. We cannot really free l3 on cpu down because, there is no node offlining yet and even though a cpu is not yet up, node local memory can be allocated for it. So l3s are usually allocated at keme_cache_create and destroyed at kmem_cache_destroy. Hence, we don't need cachep->spinlock protection to get to the cachep->nodelist[nodeid] either. Patch survived onlining and offlining on a 4 core 2 node Tyan box with a 4 dbench process running all the time. Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Ravikiran Thirumalai <kiran@scalex86.org> Cc: Christoph Lameter <christoph@lameter.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-02-05 07:27:59 +00:00
goto free_array_cache;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
spin_lock_irq(&l3->list_lock);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
/* Free limit for this kmem_list3 */
l3->free_limit -= cachep->batchcount;
if (nc)
free_block(cachep, nc->entry, nc->avail, node);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
if (!cpus_empty(mask)) {
spin_unlock_irq(&l3->list_lock);
[PATCH] NUMA slab locking fixes: fix cpu down and up locking This fixes locking and bugs in cpu_down and cpu_up paths of the NUMA slab allocator. Sonny Rao <sonny@burdell.org> reported problems sometime back on POWER5 boxes, when the last cpu on the nodes were being offlined. We could not reproduce the same on x86_64 because the cpumask (node_to_cpumask) was not being updated on cpu down. Since that issue is now fixed, we can reproduce Sonny's problems on x86_64 NUMA, and here is the fix. The problem earlier was on CPU_DOWN, if it was the last cpu on the node to go down, the array_caches (shared, alien) and the kmem_list3 of the node were being freed (kfree) with the kmem_list3 lock held. If the l3 or the array_caches were to come from the same cache being cleared, we hit on badness. This patch cleans up the locking in cpu_up and cpu_down path. We cannot really free l3 on cpu down because, there is no node offlining yet and even though a cpu is not yet up, node local memory can be allocated for it. So l3s are usually allocated at keme_cache_create and destroyed at kmem_cache_destroy. Hence, we don't need cachep->spinlock protection to get to the cachep->nodelist[nodeid] either. Patch survived onlining and offlining on a 4 core 2 node Tyan box with a 4 dbench process running all the time. Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Ravikiran Thirumalai <kiran@scalex86.org> Cc: Christoph Lameter <christoph@lameter.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-02-05 07:27:59 +00:00
goto free_array_cache;
}
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
[PATCH] NUMA slab locking fixes: fix cpu down and up locking This fixes locking and bugs in cpu_down and cpu_up paths of the NUMA slab allocator. Sonny Rao <sonny@burdell.org> reported problems sometime back on POWER5 boxes, when the last cpu on the nodes were being offlined. We could not reproduce the same on x86_64 because the cpumask (node_to_cpumask) was not being updated on cpu down. Since that issue is now fixed, we can reproduce Sonny's problems on x86_64 NUMA, and here is the fix. The problem earlier was on CPU_DOWN, if it was the last cpu on the node to go down, the array_caches (shared, alien) and the kmem_list3 of the node were being freed (kfree) with the kmem_list3 lock held. If the l3 or the array_caches were to come from the same cache being cleared, we hit on badness. This patch cleans up the locking in cpu_up and cpu_down path. We cannot really free l3 on cpu down because, there is no node offlining yet and even though a cpu is not yet up, node local memory can be allocated for it. So l3s are usually allocated at keme_cache_create and destroyed at kmem_cache_destroy. Hence, we don't need cachep->spinlock protection to get to the cachep->nodelist[nodeid] either. Patch survived onlining and offlining on a 4 core 2 node Tyan box with a 4 dbench process running all the time. Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Ravikiran Thirumalai <kiran@scalex86.org> Cc: Christoph Lameter <christoph@lameter.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-02-05 07:27:59 +00:00
shared = l3->shared;
if (shared) {
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
free_block(cachep, l3->shared->entry,
l3->shared->avail, node);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
l3->shared = NULL;
}
[PATCH] NUMA slab locking fixes: fix cpu down and up locking This fixes locking and bugs in cpu_down and cpu_up paths of the NUMA slab allocator. Sonny Rao <sonny@burdell.org> reported problems sometime back on POWER5 boxes, when the last cpu on the nodes were being offlined. We could not reproduce the same on x86_64 because the cpumask (node_to_cpumask) was not being updated on cpu down. Since that issue is now fixed, we can reproduce Sonny's problems on x86_64 NUMA, and here is the fix. The problem earlier was on CPU_DOWN, if it was the last cpu on the node to go down, the array_caches (shared, alien) and the kmem_list3 of the node were being freed (kfree) with the kmem_list3 lock held. If the l3 or the array_caches were to come from the same cache being cleared, we hit on badness. This patch cleans up the locking in cpu_up and cpu_down path. We cannot really free l3 on cpu down because, there is no node offlining yet and even though a cpu is not yet up, node local memory can be allocated for it. So l3s are usually allocated at keme_cache_create and destroyed at kmem_cache_destroy. Hence, we don't need cachep->spinlock protection to get to the cachep->nodelist[nodeid] either. Patch survived onlining and offlining on a 4 core 2 node Tyan box with a 4 dbench process running all the time. Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Ravikiran Thirumalai <kiran@scalex86.org> Cc: Christoph Lameter <christoph@lameter.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-02-05 07:27:59 +00:00
alien = l3->alien;
l3->alien = NULL;
spin_unlock_irq(&l3->list_lock);
kfree(shared);
if (alien) {
drain_alien_cache(cachep, alien);
free_alien_cache(alien);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
}
[PATCH] NUMA slab locking fixes: fix cpu down and up locking This fixes locking and bugs in cpu_down and cpu_up paths of the NUMA slab allocator. Sonny Rao <sonny@burdell.org> reported problems sometime back on POWER5 boxes, when the last cpu on the nodes were being offlined. We could not reproduce the same on x86_64 because the cpumask (node_to_cpumask) was not being updated on cpu down. Since that issue is now fixed, we can reproduce Sonny's problems on x86_64 NUMA, and here is the fix. The problem earlier was on CPU_DOWN, if it was the last cpu on the node to go down, the array_caches (shared, alien) and the kmem_list3 of the node were being freed (kfree) with the kmem_list3 lock held. If the l3 or the array_caches were to come from the same cache being cleared, we hit on badness. This patch cleans up the locking in cpu_up and cpu_down path. We cannot really free l3 on cpu down because, there is no node offlining yet and even though a cpu is not yet up, node local memory can be allocated for it. So l3s are usually allocated at keme_cache_create and destroyed at kmem_cache_destroy. Hence, we don't need cachep->spinlock protection to get to the cachep->nodelist[nodeid] either. Patch survived onlining and offlining on a 4 core 2 node Tyan box with a 4 dbench process running all the time. Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Ravikiran Thirumalai <kiran@scalex86.org> Cc: Christoph Lameter <christoph@lameter.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-02-05 07:27:59 +00:00
free_array_cache:
kfree(nc);
}
[PATCH] NUMA slab locking fixes: fix cpu down and up locking This fixes locking and bugs in cpu_down and cpu_up paths of the NUMA slab allocator. Sonny Rao <sonny@burdell.org> reported problems sometime back on POWER5 boxes, when the last cpu on the nodes were being offlined. We could not reproduce the same on x86_64 because the cpumask (node_to_cpumask) was not being updated on cpu down. Since that issue is now fixed, we can reproduce Sonny's problems on x86_64 NUMA, and here is the fix. The problem earlier was on CPU_DOWN, if it was the last cpu on the node to go down, the array_caches (shared, alien) and the kmem_list3 of the node were being freed (kfree) with the kmem_list3 lock held. If the l3 or the array_caches were to come from the same cache being cleared, we hit on badness. This patch cleans up the locking in cpu_up and cpu_down path. We cannot really free l3 on cpu down because, there is no node offlining yet and even though a cpu is not yet up, node local memory can be allocated for it. So l3s are usually allocated at keme_cache_create and destroyed at kmem_cache_destroy. Hence, we don't need cachep->spinlock protection to get to the cachep->nodelist[nodeid] either. Patch survived onlining and offlining on a 4 core 2 node Tyan box with a 4 dbench process running all the time. Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Ravikiran Thirumalai <kiran@scalex86.org> Cc: Christoph Lameter <christoph@lameter.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-02-05 07:27:59 +00:00
/*
* In the previous loop, all the objects were freed to
* the respective cache's slabs, now we can go ahead and
* shrink each nodelist to its limit.
*/
list_for_each_entry(cachep, &cache_chain, next) {
l3 = cachep->nodelists[node];
if (!l3)
continue;
spin_lock_irq(&l3->list_lock);
/* free slabs belonging to this node */
__node_shrink(cachep, node);
spin_unlock_irq(&l3->list_lock);
}
mutex_unlock(&cache_chain_mutex);
break;
#endif
}
return NOTIFY_OK;
bad:
mutex_unlock(&cache_chain_mutex);
return NOTIFY_BAD;
}
static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 };
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
/*
* swap the static kmem_list3 with kmalloced memory
*/
static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
int nodeid)
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
{
struct kmem_list3 *ptr;
BUG_ON(cachep->nodelists[nodeid] != list);
ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
BUG_ON(!ptr);
local_irq_disable();
memcpy(ptr, list, sizeof(struct kmem_list3));
MAKE_ALL_LISTS(cachep, ptr, nodeid);
cachep->nodelists[nodeid] = ptr;
local_irq_enable();
}
/*
* Initialisation. Called after the page allocator have been initialised and
* before smp_init().
*/
void __init kmem_cache_init(void)
{
size_t left_over;
struct cache_sizes *sizes;
struct cache_names *names;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
int i;
int order;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
for (i = 0; i < NUM_INIT_LISTS; i++) {
kmem_list3_init(&initkmem_list3[i]);
if (i < MAX_NUMNODES)
cache_cache.nodelists[i] = NULL;
}
/*
* Fragmentation resistance on low memory - only use bigger
* page orders on machines with more than 32MB of memory.
*/
if (num_physpages > (32 << 20) >> PAGE_SHIFT)
slab_break_gfp_order = BREAK_GFP_ORDER_HI;
/* Bootstrap is tricky, because several objects are allocated
* from caches that do not exist yet:
* 1) initialize the cache_cache cache: it contains the struct
* kmem_cache structures of all caches, except cache_cache itself:
* cache_cache is statically allocated.
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
* Initially an __init data area is used for the head array and the
* kmem_list3 structures, it's replaced with a kmalloc allocated
* array at the end of the bootstrap.
* 2) Create the first kmalloc cache.
* The struct kmem_cache for the new cache is allocated normally.
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
* An __init data area is used for the head array.
* 3) Create the remaining kmalloc caches, with minimally sized
* head arrays.
* 4) Replace the __init data head arrays for cache_cache and the first
* kmalloc cache with kmalloc allocated arrays.
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
* 5) Replace the __init data for kmem_list3 for cache_cache and
* the other cache's with kmalloc allocated memory.
* 6) Resize the head arrays of the kmalloc caches to their final sizes.
*/
/* 1) create the cache_cache */
INIT_LIST_HEAD(&cache_chain);
list_add(&cache_cache.next, &cache_chain);
cache_cache.colour_off = cache_line_size();
cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE];
cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
cache_line_size());
for (order = 0; order < MAX_ORDER; order++) {
cache_estimate(order, cache_cache.buffer_size,
cache_line_size(), 0, &left_over, &cache_cache.num);
if (cache_cache.num)
break;
}
BUG_ON(!cache_cache.num);
cache_cache.gfporder = order;
cache_cache.colour = left_over / cache_cache.colour_off;
cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
sizeof(struct slab), cache_line_size());
/* 2+3) create the kmalloc caches */
sizes = malloc_sizes;
names = cache_names;
/*
* Initialize the caches that provide memory for the array cache and the
* kmem_list3 structures first. Without this, further allocations will
* bug.
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
*/
sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
sizes[INDEX_AC].cs_size,
ARCH_KMALLOC_MINALIGN,
ARCH_KMALLOC_FLAGS|SLAB_PANIC,
NULL, NULL);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
if (INDEX_AC != INDEX_L3) {
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
sizes[INDEX_L3].cs_cachep =
kmem_cache_create(names[INDEX_L3].name,
sizes[INDEX_L3].cs_size,
ARCH_KMALLOC_MINALIGN,
ARCH_KMALLOC_FLAGS|SLAB_PANIC,
NULL, NULL);
}
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
while (sizes->cs_size != ULONG_MAX) {
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
/*
* For performance, all the general caches are L1 aligned.
* This should be particularly beneficial on SMP boxes, as it
* eliminates "false sharing".
* Note for systems short on memory removing the alignment will
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
* allow tighter packing of the smaller caches.
*/
if (!sizes->cs_cachep) {
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
sizes->cs_cachep = kmem_cache_create(names->name,
sizes->cs_size,
ARCH_KMALLOC_MINALIGN,
ARCH_KMALLOC_FLAGS|SLAB_PANIC,
NULL, NULL);
}
/* Inc off-slab bufctl limit until the ceiling is hit. */
if (!(OFF_SLAB(sizes->cs_cachep))) {
offslab_limit = sizes->cs_size - sizeof(struct slab);
offslab_limit /= sizeof(kmem_bufctl_t);
}
sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
sizes->cs_size,
ARCH_KMALLOC_MINALIGN,
ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
SLAB_PANIC,
NULL, NULL);
sizes++;
names++;
}
/* 4) Replace the bootstrap head arrays */
{
void *ptr;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
local_irq_disable();
BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
memcpy(ptr, cpu_cache_get(&cache_cache),
sizeof(struct arraycache_init));
cache_cache.array[smp_processor_id()] = ptr;
local_irq_enable();
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
local_irq_disable();
BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
!= &initarray_generic.cache);
memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
sizeof(struct arraycache_init));
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
ptr;
local_irq_enable();
}
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
/* 5) Replace the bootstrap kmem_list3's */
{
int node;
/* Replace the static kmem_list3 structures for the boot cpu */
init_list(&cache_cache, &initkmem_list3[CACHE_CACHE],
numa_node_id());
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
for_each_online_node(node) {
init_list(malloc_sizes[INDEX_AC].cs_cachep,
&initkmem_list3[SIZE_AC + node], node);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
if (INDEX_AC != INDEX_L3) {
init_list(malloc_sizes[INDEX_L3].cs_cachep,
&initkmem_list3[SIZE_L3 + node],
node);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
}
}
}
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
/* 6) resize the head arrays to their final sizes */
{
struct kmem_cache *cachep;
mutex_lock(&cache_chain_mutex);
list_for_each_entry(cachep, &cache_chain, next)
enable_cpucache(cachep);
mutex_unlock(&cache_chain_mutex);
}
/* Done! */
g_cpucache_up = FULL;
/*
* Register a cpu startup notifier callback that initializes
* cpu_cache_get for all new cpus
*/
register_cpu_notifier(&cpucache_notifier);
/*
* The reap timers are started later, with a module init call: That part
* of the kernel is not yet operational.
*/
}
static int __init cpucache_init(void)
{
int cpu;
/*
* Register the timers that return unneeded pages to the page allocator
*/
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
for_each_online_cpu(cpu)
start_cpu_timer(cpu);
return 0;
}
__initcall(cpucache_init);
/*
* Interface to system's page allocator. No need to hold the cache-lock.
*
* If we requested dmaable memory, we will get it. Even if we
* did not request dmaable memory, we might get it, but that
* would be relatively rare and ignorable.
*/
static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
{
struct page *page;
void *addr;
int i;
flags |= cachep->gfpflags;
#ifndef CONFIG_MMU
/* nommu uses slab's for process anonymous memory allocations, so
* requires __GFP_COMP to properly refcount higher order allocations"
*/
page = alloc_pages_node(nodeid, (flags | __GFP_COMP), cachep->gfporder);
#else
page = alloc_pages_node(nodeid, flags, cachep->gfporder);
#endif
if (!page)
return NULL;
addr = page_address(page);
i = (1 << cachep->gfporder);
if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
atomic_add(i, &slab_reclaim_pages);
add_page_state(nr_slab, i);
while (i--) {
__SetPageSlab(page);
page++;
}
return addr;
}
/*
* Interface to system's page release.
*/
static void kmem_freepages(struct kmem_cache *cachep, void *addr)
{
unsigned long i = (1 << cachep->gfporder);
struct page *page = virt_to_page(addr);
const unsigned long nr_freed = i;
while (i--) {
BUG_ON(!PageSlab(page));
__ClearPageSlab(page);
page++;
}
sub_page_state(nr_slab, nr_freed);
if (current->reclaim_state)
current->reclaim_state->reclaimed_slab += nr_freed;
free_pages((unsigned long)addr, cachep->gfporder);
if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages);
}
static void kmem_rcu_free(struct rcu_head *head)
{
struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
struct kmem_cache *cachep = slab_rcu->cachep;
kmem_freepages(cachep, slab_rcu->addr);
if (OFF_SLAB(cachep))
kmem_cache_free(cachep->slabp_cache, slab_rcu);
}
#if DEBUG
#ifdef CONFIG_DEBUG_PAGEALLOC
static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
unsigned long caller)
{
int size = obj_size(cachep);
addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
if (size < 5 * sizeof(unsigned long))
return;
*addr++ = 0x12345678;
*addr++ = caller;
*addr++ = smp_processor_id();
size -= 3 * sizeof(unsigned long);
{
unsigned long *sptr = &caller;
unsigned long svalue;
while (!kstack_end(sptr)) {
svalue = *sptr++;
if (kernel_text_address(svalue)) {
*addr++ = svalue;
size -= sizeof(unsigned long);
if (size <= sizeof(unsigned long))
break;
}
}
}
*addr++ = 0x87654321;
}
#endif
static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
{
int size = obj_size(cachep);
addr = &((char *)addr)[obj_offset(cachep)];
memset(addr, val, size);
*(unsigned char *)(addr + size - 1) = POISON_END;
}
static void dump_line(char *data, int offset, int limit)
{
int i;
printk(KERN_ERR "%03x:", offset);
for (i = 0; i < limit; i++)
printk(" %02x", (unsigned char)data[offset + i]);
printk("\n");
}
#endif
#if DEBUG
static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
{
int i, size;
char *realobj;
if (cachep->flags & SLAB_RED_ZONE) {
printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
*dbg_redzone1(cachep, objp),
*dbg_redzone2(cachep, objp));
}
if (cachep->flags & SLAB_STORE_USER) {
printk(KERN_ERR "Last user: [<%p>]",
*dbg_userword(cachep, objp));
print_symbol("(%s)",
(unsigned long)*dbg_userword(cachep, objp));
printk("\n");
}
realobj = (char *)objp + obj_offset(cachep);
size = obj_size(cachep);
for (i = 0; i < size && lines; i += 16, lines--) {
int limit;
limit = 16;
if (i + limit > size)
limit = size - i;
dump_line(realobj, i, limit);
}
}
static void check_poison_obj(struct kmem_cache *cachep, void *objp)
{
char *realobj;
int size, i;
int lines = 0;
realobj = (char *)objp + obj_offset(cachep);
size = obj_size(cachep);
for (i = 0; i < size; i++) {
char exp = POISON_FREE;
if (i == size - 1)
exp = POISON_END;
if (realobj[i] != exp) {
int limit;
/* Mismatch ! */
/* Print header */
if (lines == 0) {
printk(KERN_ERR
"Slab corruption: start=%p, len=%d\n",
realobj, size);
print_objinfo(cachep, objp, 0);
}
/* Hexdump the affected line */
i = (i / 16) * 16;
limit = 16;
if (i + limit > size)
limit = size - i;
dump_line(realobj, i, limit);
i += 16;
lines++;
/* Limit to 5 lines */
if (lines > 5)
break;
}
}
if (lines != 0) {
/* Print some data about the neighboring objects, if they
* exist:
*/
struct slab *slabp = virt_to_slab(objp);
unsigned int objnr;
objnr = obj_to_index(cachep, slabp, objp);
if (objnr) {
objp = index_to_obj(cachep, slabp, objnr - 1);
realobj = (char *)objp + obj_offset(cachep);
printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
realobj, size);
print_objinfo(cachep, objp, 2);
}
if (objnr + 1 < cachep->num) {
objp = index_to_obj(cachep, slabp, objnr + 1);
realobj = (char *)objp + obj_offset(cachep);
printk(KERN_ERR "Next obj: start=%p, len=%d\n",
realobj, size);
print_objinfo(cachep, objp, 2);
}
}
}
#endif
#if DEBUG
/**
* slab_destroy_objs - destroy a slab and its objects
* @cachep: cache pointer being destroyed
* @slabp: slab pointer being destroyed
*
* Call the registered destructor for each object in a slab that is being
* destroyed.
*/
static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
{
int i;
for (i = 0; i < cachep->num; i++) {
void *objp = index_to_obj(cachep, slabp, i);
if (cachep->flags & SLAB_POISON) {
#ifdef CONFIG_DEBUG_PAGEALLOC
if (cachep->buffer_size % PAGE_SIZE == 0 &&
OFF_SLAB(cachep))
kernel_map_pages(virt_to_page(objp),
cachep->buffer_size / PAGE_SIZE, 1);
else
check_poison_obj(cachep, objp);
#else
check_poison_obj(cachep, objp);
#endif
}
if (cachep->flags & SLAB_RED_ZONE) {
if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
slab_error(cachep, "start of a freed object "
"was overwritten");
if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
slab_error(cachep, "end of a freed object "
"was overwritten");
}
if (cachep->dtor && !(cachep->flags & SLAB_POISON))
(cachep->dtor) (objp + obj_offset(cachep), cachep, 0);
}
}
#else
static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
{
if (cachep->dtor) {
int i;
for (i = 0; i < cachep->num; i++) {
void *objp = index_to_obj(cachep, slabp, i);
(cachep->dtor) (objp, cachep, 0);
}
}
}
#endif
/**
* slab_destroy - destroy and release all objects in a slab
* @cachep: cache pointer being destroyed
* @slabp: slab pointer being destroyed
*
* Destroy all the objs in a slab, and release the mem back to the system.
* Before calling the slab must have been unlinked from the cache. The
* cache-lock is not held/needed.
*/
static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
{
void *addr = slabp->s_mem - slabp->colouroff;
slab_destroy_objs(cachep, slabp);
if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
struct slab_rcu *slab_rcu;
slab_rcu = (struct slab_rcu *)slabp;
slab_rcu->cachep = cachep;
slab_rcu->addr = addr;
call_rcu(&slab_rcu->head, kmem_rcu_free);
} else {
kmem_freepages(cachep, addr);
if (OFF_SLAB(cachep))
kmem_cache_free(cachep->slabp_cache, slabp);
}
}
/*
* For setting up all the kmem_list3s for cache whose buffer_size is same as
* size of kmem_list3.
*/
static void set_up_list3s(struct kmem_cache *cachep, int index)
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
{
int node;
for_each_online_node(node) {
cachep->nodelists[node] = &initkmem_list3[index + node];
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
cachep->nodelists[node]->next_reap = jiffies +
REAPTIMEOUT_LIST3 +
((unsigned long)cachep) % REAPTIMEOUT_LIST3;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
}
}
/**
* calculate_slab_order - calculate size (page order) of slabs
* @cachep: pointer to the cache that is being created
* @size: size of objects to be created in this cache.
* @align: required alignment for the objects.
* @flags: slab allocation flags
*
* Also calculates the number of objects per slab.
*
* This could be made much more intelligent. For now, try to avoid using
* high order pages for slabs. When the gfp() functions are more friendly
* towards high-order requests, this should be changed.
*/
static size_t calculate_slab_order(struct kmem_cache *cachep,
size_t size, size_t align, unsigned long flags)
{
size_t left_over = 0;
int gfporder;
for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) {
unsigned int num;
size_t remainder;
cache_estimate(gfporder, size, align, flags, &remainder, &num);
if (!num)
continue;
/* More than offslab_limit objects will cause problems */
if ((flags & CFLGS_OFF_SLAB) && num > offslab_limit)
break;
/* Found something acceptable - save it away */
cachep->num = num;
cachep->gfporder = gfporder;
left_over = remainder;
/*
* A VFS-reclaimable slab tends to have most allocations
* as GFP_NOFS and we really don't want to have to be allocating
* higher-order pages when we are unable to shrink dcache.
*/
if (flags & SLAB_RECLAIM_ACCOUNT)
break;
/*
* Large number of objects is good, but very large slabs are
* currently bad for the gfp()s.
*/
if (gfporder >= slab_break_gfp_order)
break;
/*
* Acceptable internal fragmentation?
*/
if (left_over * 8 <= (PAGE_SIZE << gfporder))
break;
}
return left_over;
}
static void setup_cpu_cache(struct kmem_cache *cachep)
{
if (g_cpucache_up == FULL) {
enable_cpucache(cachep);
return;
}
if (g_cpucache_up == NONE) {
/*
* Note: the first kmem_cache_create must create the cache
* that's used by kmalloc(24), otherwise the creation of
* further caches will BUG().
*/
cachep->array[smp_processor_id()] = &initarray_generic.cache;
/*
* If the cache that's used by kmalloc(sizeof(kmem_list3)) is
* the first cache, then we need to set up all its list3s,
* otherwise the creation of further caches will BUG().
*/
set_up_list3s(cachep, SIZE_AC);
if (INDEX_AC == INDEX_L3)
g_cpucache_up = PARTIAL_L3;
else
g_cpucache_up = PARTIAL_AC;
} else {
cachep->array[smp_processor_id()] =
kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
if (g_cpucache_up == PARTIAL_AC) {
set_up_list3s(cachep, SIZE_L3);
g_cpucache_up = PARTIAL_L3;
} else {
int node;
for_each_online_node(node) {
cachep->nodelists[node] =
kmalloc_node(sizeof(struct kmem_list3),
GFP_KERNEL, node);
BUG_ON(!cachep->nodelists[node]);
kmem_list3_init(cachep->nodelists[node]);
}
}
}
cachep->nodelists[numa_node_id()]->next_reap =
jiffies + REAPTIMEOUT_LIST3 +
((unsigned long)cachep) % REAPTIMEOUT_LIST3;
cpu_cache_get(cachep)->avail = 0;
cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
cpu_cache_get(cachep)->batchcount = 1;
cpu_cache_get(cachep)->touched = 0;
cachep->batchcount = 1;
cachep->limit = BOOT_CPUCACHE_ENTRIES;
}
/**
* kmem_cache_create - Create a cache.
* @name: A string which is used in /proc/slabinfo to identify this cache.
* @size: The size of objects to be created in this cache.
* @align: The required alignment for the objects.
* @flags: SLAB flags
* @ctor: A constructor for the objects.
* @dtor: A destructor for the objects.
*
* Returns a ptr to the cache on success, NULL on failure.
* Cannot be called within a int, but can be interrupted.
* The @ctor is run when new pages are allocated by the cache
* and the @dtor is run before the pages are handed back.
*
* @name must be valid until the cache is destroyed. This implies that
* the module calling this has to destroy the cache before getting unloaded.
*
* The flags are
*
* %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
* to catch references to uninitialised memory.
*
* %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
* for buffer overruns.
*
* %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
* cacheline. This can be beneficial if you're counting cycles as closely
* as davem.
*/
struct kmem_cache *
kmem_cache_create (const char *name, size_t size, size_t align,
unsigned long flags,
void (*ctor)(void*, struct kmem_cache *, unsigned long),
void (*dtor)(void*, struct kmem_cache *, unsigned long))
{
size_t left_over, slab_size, ralign;
struct kmem_cache *cachep = NULL;
struct list_head *p;
/*
* Sanity checks... these are all serious usage bugs.
*/
if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
(size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
name);
BUG();
}
/*
* Prevent CPUs from coming and going.
* lock_cpu_hotplug() nests outside cache_chain_mutex
*/
lock_cpu_hotplug();
mutex_lock(&cache_chain_mutex);
list_for_each(p, &cache_chain) {
struct kmem_cache *pc = list_entry(p, struct kmem_cache, next);
mm_segment_t old_fs = get_fs();
char tmp;
int res;
/*
* This happens when the module gets unloaded and doesn't
* destroy its slab cache and no-one else reuses the vmalloc
* area of the module. Print a warning.
*/
set_fs(KERNEL_DS);
res = __get_user(tmp, pc->name);
set_fs(old_fs);
if (res) {
printk("SLAB: cache with size %d has lost its name\n",
pc->buffer_size);
continue;
}
if (!strcmp(pc->name, name)) {
printk("kmem_cache_create: duplicate cache %s\n", name);
dump_stack();
goto oops;
}
}
#if DEBUG
WARN_ON(strchr(name, ' ')); /* It confuses parsers */
if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
/* No constructor, but inital state check requested */
printk(KERN_ERR "%s: No con, but init state check "
"requested - %s\n", __FUNCTION__, name);
flags &= ~SLAB_DEBUG_INITIAL;
}
#if FORCED_DEBUG
/*
* Enable redzoning and last user accounting, except for caches with
* large objects, if the increased size would increase the object size
* above the next power of two: caches with object sizes just above a
* power of two have a significant amount of internal fragmentation.
*/
if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD))
flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
if (!(flags & SLAB_DESTROY_BY_RCU))
flags |= SLAB_POISON;
#endif
if (flags & SLAB_DESTROY_BY_RCU)
BUG_ON(flags & SLAB_POISON);
#endif
if (flags & SLAB_DESTROY_BY_RCU)
BUG_ON(dtor);
/*
* Always checks flags, a caller might be expecting debug support which
* isn't available.
*/
BUG_ON(flags & ~CREATE_MASK);
/*
* Check that size is in terms of words. This is needed to avoid
* unaligned accesses for some archs when redzoning is used, and makes
* sure any on-slab bufctl's are also correctly aligned.
*/
if (size & (BYTES_PER_WORD - 1)) {
size += (BYTES_PER_WORD - 1);
size &= ~(BYTES_PER_WORD - 1);
}
/* calculate the final buffer alignment: */
/* 1) arch recommendation: can be overridden for debug */
if (flags & SLAB_HWCACHE_ALIGN) {
/*
* Default alignment: as specified by the arch code. Except if
* an object is really small, then squeeze multiple objects into
* one cacheline.
*/
ralign = cache_line_size();
while (size <= ralign / 2)
ralign /= 2;
} else {
ralign = BYTES_PER_WORD;
}
/* 2) arch mandated alignment: disables debug if necessary */
if (ralign < ARCH_SLAB_MINALIGN) {
ralign = ARCH_SLAB_MINALIGN;
if (ralign > BYTES_PER_WORD)
flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
}
/* 3) caller mandated alignment: disables debug if necessary */
if (ralign < align) {
ralign = align;
if (ralign > BYTES_PER_WORD)
flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
}
/*
* 4) Store it. Note that the debug code below can reduce
* the alignment to BYTES_PER_WORD.
*/
align = ralign;
/* Get cache's description obj. */
cachep = kmem_cache_zalloc(&cache_cache, SLAB_KERNEL);
if (!cachep)
goto oops;
#if DEBUG
cachep->obj_size = size;
if (flags & SLAB_RED_ZONE) {
/* redzoning only works with word aligned caches */
align = BYTES_PER_WORD;
/* add space for red zone words */
cachep->obj_offset += BYTES_PER_WORD;
size += 2 * BYTES_PER_WORD;
}
if (flags & SLAB_STORE_USER) {
/* user store requires word alignment and
* one word storage behind the end of the real
* object.
*/
align = BYTES_PER_WORD;
size += BYTES_PER_WORD;
}
#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
&& cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
cachep->obj_offset += PAGE_SIZE - size;
size = PAGE_SIZE;
}
#endif
#endif
/* Determine if the slab management is 'on' or 'off' slab. */
if (size >= (PAGE_SIZE >> 3))
/*
* Size is large, assume best to place the slab management obj
* off-slab (should allow better packing of objs).
*/
flags |= CFLGS_OFF_SLAB;
size = ALIGN(size, align);
left_over = calculate_slab_order(cachep, size, align, flags);
if (!cachep->num) {
printk("kmem_cache_create: couldn't create cache %s.\n", name);
kmem_cache_free(&cache_cache, cachep);
cachep = NULL;
goto oops;
}
slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
+ sizeof(struct slab), align);
/*
* If the slab has been placed off-slab, and we have enough space then
* move it on-slab. This is at the expense of any extra colouring.
*/
if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
flags &= ~CFLGS_OFF_SLAB;
left_over -= slab_size;
}
if (flags & CFLGS_OFF_SLAB) {
/* really off slab. No need for manual alignment */
slab_size =
cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
}
cachep->colour_off = cache_line_size();
/* Offset must be a multiple of the alignment. */
if (cachep->colour_off < align)
cachep->colour_off = align;
cachep->colour = left_over / cachep->colour_off;
cachep->slab_size = slab_size;
cachep->flags = flags;
cachep->gfpflags = 0;
if (flags & SLAB_CACHE_DMA)
cachep->gfpflags |= GFP_DMA;
cachep->buffer_size = size;
if (flags & CFLGS_OFF_SLAB)
cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
cachep->ctor = ctor;
cachep->dtor = dtor;
cachep->name = name;
setup_cpu_cache(cachep);
/* cache setup completed, link it into the list */
list_add(&cachep->next, &cache_chain);
oops:
if (!cachep && (flags & SLAB_PANIC))
panic("kmem_cache_create(): failed to create slab `%s'\n",
name);
mutex_unlock(&cache_chain_mutex);
unlock_cpu_hotplug();
return cachep;
}
EXPORT_SYMBOL(kmem_cache_create);
#if DEBUG
static void check_irq_off(void)
{
BUG_ON(!irqs_disabled());
}
static void check_irq_on(void)
{
BUG_ON(irqs_disabled());
}
static void check_spinlock_acquired(struct kmem_cache *cachep)
{
#ifdef CONFIG_SMP
check_irq_off();
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
#endif
}
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
{
#ifdef CONFIG_SMP
check_irq_off();
assert_spin_locked(&cachep->nodelists[node]->list_lock);
#endif
}
#else
#define check_irq_off() do { } while(0)
#define check_irq_on() do { } while(0)
#define check_spinlock_acquired(x) do { } while(0)
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
#define check_spinlock_acquired_node(x, y) do { } while(0)
#endif
static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
struct array_cache *ac,
int force, int node);
static void do_drain(void *arg)
{
struct kmem_cache *cachep = arg;
struct array_cache *ac;
int node = numa_node_id();
check_irq_off();
ac = cpu_cache_get(cachep);
spin_lock(&cachep->nodelists[node]->list_lock);
free_block(cachep, ac->entry, ac->avail, node);
spin_unlock(&cachep->nodelists[node]->list_lock);
ac->avail = 0;
}
static void drain_cpu_caches(struct kmem_cache *cachep)
{
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
struct kmem_list3 *l3;
int node;
on_each_cpu(do_drain, cachep, 1, 1);
check_irq_on();
for_each_online_node(node) {
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
l3 = cachep->nodelists[node];
if (l3) {
drain_array(cachep, l3, l3->shared, 1, node);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
if (l3->alien)
[PATCH] NUMA slab locking fixes: fix cpu down and up locking This fixes locking and bugs in cpu_down and cpu_up paths of the NUMA slab allocator. Sonny Rao <sonny@burdell.org> reported problems sometime back on POWER5 boxes, when the last cpu on the nodes were being offlined. We could not reproduce the same on x86_64 because the cpumask (node_to_cpumask) was not being updated on cpu down. Since that issue is now fixed, we can reproduce Sonny's problems on x86_64 NUMA, and here is the fix. The problem earlier was on CPU_DOWN, if it was the last cpu on the node to go down, the array_caches (shared, alien) and the kmem_list3 of the node were being freed (kfree) with the kmem_list3 lock held. If the l3 or the array_caches were to come from the same cache being cleared, we hit on badness. This patch cleans up the locking in cpu_up and cpu_down path. We cannot really free l3 on cpu down because, there is no node offlining yet and even though a cpu is not yet up, node local memory can be allocated for it. So l3s are usually allocated at keme_cache_create and destroyed at kmem_cache_destroy. Hence, we don't need cachep->spinlock protection to get to the cachep->nodelist[nodeid] either. Patch survived onlining and offlining on a 4 core 2 node Tyan box with a 4 dbench process running all the time. Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Ravikiran Thirumalai <kiran@scalex86.org> Cc: Christoph Lameter <christoph@lameter.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-02-05 07:27:59 +00:00
drain_alien_cache(cachep, l3->alien);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
}
}
}
static int __node_shrink(struct kmem_cache *cachep, int node)
{
struct slab *slabp;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
struct kmem_list3 *l3 = cachep->nodelists[node];
int ret;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
for (;;) {
struct list_head *p;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
p = l3->slabs_free.prev;
if (p == &l3->slabs_free)
break;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
slabp = list_entry(l3->slabs_free.prev, struct slab, list);
#if DEBUG
BUG_ON(slabp->inuse);
#endif
list_del(&slabp->list);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
l3->free_objects -= cachep->num;
spin_unlock_irq(&l3->list_lock);
slab_destroy(cachep, slabp);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
spin_lock_irq(&l3->list_lock);
}
ret = !list_empty(&l3->slabs_full) || !list_empty(&l3->slabs_partial);
return ret;
}
static int __cache_shrink(struct kmem_cache *cachep)
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
{
int ret = 0, i = 0;
struct kmem_list3 *l3;
drain_cpu_caches(cachep);
check_irq_on();
for_each_online_node(i) {
l3 = cachep->nodelists[i];
if (l3) {
spin_lock_irq(&l3->list_lock);
ret += __node_shrink(cachep, i);
spin_unlock_irq(&l3->list_lock);
}
}
return (ret ? 1 : 0);
}
/**
* kmem_cache_shrink - Shrink a cache.
* @cachep: The cache to shrink.
*
* Releases as many slabs as possible for a cache.
* To help debugging, a zero exit status indicates all slabs were released.
*/
int kmem_cache_shrink(struct kmem_cache *cachep)
{
BUG_ON(!cachep || in_interrupt());
return __cache_shrink(cachep);
}
EXPORT_SYMBOL(kmem_cache_shrink);
/**
* kmem_cache_destroy - delete a cache
* @cachep: the cache to destroy
*
* Remove a struct kmem_cache object from the slab cache.
* Returns 0 on success.
*
* It is expected this function will be called by a module when it is
* unloaded. This will remove the cache completely, and avoid a duplicate
* cache being allocated each time a module is loaded and unloaded, if the
* module doesn't have persistent in-kernel storage across loads and unloads.
*
* The cache must be empty before calling this function.
*
* The caller must guarantee that noone will allocate memory from the cache
* during the kmem_cache_destroy().
*/
int kmem_cache_destroy(struct kmem_cache *cachep)
{
int i;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
struct kmem_list3 *l3;
BUG_ON(!cachep || in_interrupt());
/* Don't let CPUs to come and go */
lock_cpu_hotplug();
/* Find the cache in the chain of caches. */
mutex_lock(&cache_chain_mutex);
/*
* the chain is never empty, cache_cache is never destroyed
*/
list_del(&cachep->next);
mutex_unlock(&cache_chain_mutex);
if (__cache_shrink(cachep)) {
slab_error(cachep, "Can't free all objects");
mutex_lock(&cache_chain_mutex);
list_add(&cachep->next, &cache_chain);
mutex_unlock(&cache_chain_mutex);
unlock_cpu_hotplug();
return 1;
}
if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
synchronize_rcu();
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
for_each_online_cpu(i)
kfree(cachep->array[i]);
/* NUMA: free the list3 structures */
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
for_each_online_node(i) {
l3 = cachep->nodelists[i];
if (l3) {
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
kfree(l3->shared);
free_alien_cache(l3->alien);
kfree(l3);
}
}
kmem_cache_free(&cache_cache, cachep);
unlock_cpu_hotplug();
return 0;
}
EXPORT_SYMBOL(kmem_cache_destroy);
/* Get the memory for a slab management obj. */
static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
int colour_off, gfp_t local_flags,
int nodeid)
{
struct slab *slabp;
if (OFF_SLAB(cachep)) {
/* Slab management obj is off-slab. */
slabp = kmem_cache_alloc_node(cachep->slabp_cache,
local_flags, nodeid);
if (!slabp)
return NULL;
} else {
slabp = objp + colour_off;
colour_off += cachep->slab_size;
}
slabp->inuse = 0;
slabp->colouroff = colour_off;
slabp->s_mem = objp + colour_off;
slabp->nodeid = nodeid;
return slabp;
}
static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
{
return (kmem_bufctl_t *) (slabp + 1);
}
static void cache_init_objs(struct kmem_cache *cachep,
struct slab *slabp, unsigned long ctor_flags)
{
int i;
for (i = 0; i < cachep->num; i++) {
void *objp = index_to_obj(cachep, slabp, i);
#if DEBUG
/* need to poison the objs? */
if (cachep->flags & SLAB_POISON)
poison_obj(cachep, objp, POISON_FREE);
if (cachep->flags & SLAB_STORE_USER)
*dbg_userword(cachep, objp) = NULL;
if (cachep->flags & SLAB_RED_ZONE) {
*dbg_redzone1(cachep, objp) = RED_INACTIVE;
*dbg_redzone2(cachep, objp) = RED_INACTIVE;
}
/*
* Constructors are not allowed to allocate memory from the same
* cache which they are a constructor for. Otherwise, deadlock.
* They must also be threaded.
*/
if (cachep->ctor && !(cachep->flags & SLAB_POISON))
cachep->ctor(objp + obj_offset(cachep), cachep,
ctor_flags);
if (cachep->flags & SLAB_RED_ZONE) {
if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
slab_error(cachep, "constructor overwrote the"
" end of an object");
if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
slab_error(cachep, "constructor overwrote the"
" start of an object");
}
if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
kernel_map_pages(virt_to_page(objp),
cachep->buffer_size / PAGE_SIZE, 0);
#else
if (cachep->ctor)
cachep->ctor(objp, cachep, ctor_flags);
#endif
slab_bufctl(slabp)[i] = i + 1;
}
slab_bufctl(slabp)[i - 1] = BUFCTL_END;
slabp->free = 0;
}
static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
{
if (flags & SLAB_DMA)
BUG_ON(!(cachep->gfpflags & GFP_DMA));
else
BUG_ON(cachep->gfpflags & GFP_DMA);
}
static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
int nodeid)
{
void *objp = index_to_obj(cachep, slabp, slabp->free);
kmem_bufctl_t next;
slabp->inuse++;
next = slab_bufctl(slabp)[slabp->free];
#if DEBUG
slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
WARN_ON(slabp->nodeid != nodeid);
#endif
slabp->free = next;
return objp;
}
static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
void *objp, int nodeid)
{
unsigned int objnr = obj_to_index(cachep, slabp, objp);
#if DEBUG
/* Verify that the slab belongs to the intended node */
WARN_ON(slabp->nodeid != nodeid);
[PATCH] slab: implement /proc/slab_allocators Implement /proc/slab_allocators. It produces output like: idr_layer_cache: 80 idr_pre_get+0x33/0x4e buffer_head: 2555 alloc_buffer_head+0x20/0x75 mm_struct: 9 mm_alloc+0x1e/0x42 mm_struct: 20 dup_mm+0x36/0x370 vm_area_struct: 384 dup_mm+0x18f/0x370 vm_area_struct: 151 do_mmap_pgoff+0x2e0/0x7c3 vm_area_struct: 1 split_vma+0x5a/0x10e vm_area_struct: 11 do_brk+0x206/0x2e2 vm_area_struct: 2 copy_vma+0xda/0x142 vm_area_struct: 9 setup_arg_pages+0x99/0x214 fs_cache: 8 copy_fs_struct+0x21/0x133 fs_cache: 29 copy_process+0xf38/0x10e3 files_cache: 30 alloc_files+0x1b/0xcf signal_cache: 81 copy_process+0xbaa/0x10e3 sighand_cache: 77 copy_process+0xe65/0x10e3 sighand_cache: 1 de_thread+0x4d/0x5f8 anon_vma: 241 anon_vma_prepare+0xd9/0xf3 size-2048: 1 add_sect_attrs+0x5f/0x145 size-2048: 2 journal_init_revoke+0x99/0x302 size-2048: 2 journal_init_revoke+0x137/0x302 size-2048: 2 journal_init_inode+0xf9/0x1c4 Cc: Manfred Spraul <manfred@colorfullife.com> Cc: Alexander Nyberg <alexn@telia.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: Ravikiran Thirumalai <kiran@scalex86.org> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk> DESC slab-leaks3-locking-fix EDESC From: Andrew Morton <akpm@osdl.org> Update for slab-remove-cachep-spinlock.patch Cc: Al Viro <viro@ftp.linux.org.uk> Cc: Manfred Spraul <manfred@colorfullife.com> Cc: Alexander Nyberg <alexn@telia.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: Ravikiran Thirumalai <kiran@scalex86.org> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-03-25 11:06:39 +00:00
if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
printk(KERN_ERR "slab: double free detected in cache "
"'%s', objp %p\n", cachep->name, objp);
BUG();
}
#endif
slab_bufctl(slabp)[objnr] = slabp->free;
slabp->free = objnr;
slabp->inuse--;
}
static void set_slab_attr(struct kmem_cache *cachep, struct slab *slabp,
void *objp)
{
int i;
struct page *page;
/* Nasty!!!!!! I hope this is OK. */
page = virt_to_page(objp);
i = 1;
if (likely(!PageCompound(page)))
i <<= cachep->gfporder;
do {
page_set_cache(page, cachep);
page_set_slab(page, slabp);
page++;
} while (--i);
}
/*
* Grow (by 1) the number of slabs within a cache. This is called by
* kmem_cache_alloc() when there are no active objs left in a cache.
*/
static int cache_grow(struct kmem_cache *cachep, gfp_t flags, int nodeid)
{
struct slab *slabp;
void *objp;
size_t offset;
gfp_t local_flags;
unsigned long ctor_flags;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
struct kmem_list3 *l3;
/*
* Be lazy and only check for valid flags here, keeping it out of the
* critical path in kmem_cache_alloc().
*/
BUG_ON(flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW));
if (flags & SLAB_NO_GROW)
return 0;
ctor_flags = SLAB_CTOR_CONSTRUCTOR;
local_flags = (flags & SLAB_LEVEL_MASK);
if (!(local_flags & __GFP_WAIT))
/*
* Not allowed to sleep. Need to tell a constructor about
* this - it might need to know...
*/
ctor_flags |= SLAB_CTOR_ATOMIC;
/* Take the l3 list lock to change the colour_next on this node */
check_irq_off();
l3 = cachep->nodelists[nodeid];
spin_lock(&l3->list_lock);
/* Get colour for the slab, and cal the next value. */
offset = l3->colour_next;
l3->colour_next++;
if (l3->colour_next >= cachep->colour)
l3->colour_next = 0;
spin_unlock(&l3->list_lock);
offset *= cachep->colour_off;
if (local_flags & __GFP_WAIT)
local_irq_enable();
/*
* The test for missing atomic flag is performed here, rather than
* the more obvious place, simply to reduce the critical path length
* in kmem_cache_alloc(). If a caller is seriously mis-behaving they
* will eventually be caught here (where it matters).
*/
kmem_flagcheck(cachep, flags);
/*
* Get mem for the objs. Attempt to allocate a physical page from
* 'nodeid'.
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
*/
objp = kmem_getpages(cachep, flags, nodeid);
if (!objp)
goto failed;
/* Get slab management. */
slabp = alloc_slabmgmt(cachep, objp, offset, local_flags, nodeid);
if (!slabp)
goto opps1;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
slabp->nodeid = nodeid;
set_slab_attr(cachep, slabp, objp);
cache_init_objs(cachep, slabp, ctor_flags);
if (local_flags & __GFP_WAIT)
local_irq_disable();
check_irq_off();
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
spin_lock(&l3->list_lock);
/* Make slab active. */
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
list_add_tail(&slabp->list, &(l3->slabs_free));
STATS_INC_GROWN(cachep);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
l3->free_objects += cachep->num;
spin_unlock(&l3->list_lock);
return 1;
opps1:
kmem_freepages(cachep, objp);
failed:
if (local_flags & __GFP_WAIT)
local_irq_disable();
return 0;
}
#if DEBUG
/*
* Perform extra freeing checks:
* - detect bad pointers.
* - POISON/RED_ZONE checking
* - destructor calls, for caches with POISON+dtor
*/
static void kfree_debugcheck(const void *objp)
{
struct page *page;
if (!virt_addr_valid(objp)) {
printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
(unsigned long)objp);
BUG();
}
page = virt_to_page(objp);
if (!PageSlab(page)) {
printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n",
(unsigned long)objp);
BUG();
}
}
static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
void *caller)
{
struct page *page;
unsigned int objnr;
struct slab *slabp;
objp -= obj_offset(cachep);
kfree_debugcheck(objp);
page = virt_to_page(objp);
if (page_get_cache(page) != cachep) {
printk(KERN_ERR "mismatch in kmem_cache_free: expected "
"cache %p, got %p\n",
page_get_cache(page), cachep);
printk(KERN_ERR "%p is %s.\n", cachep, cachep->name);
printk(KERN_ERR "%p is %s.\n", page_get_cache(page),
page_get_cache(page)->name);
WARN_ON(1);
}
slabp = page_get_slab(page);
if (cachep->flags & SLAB_RED_ZONE) {
if (*dbg_redzone1(cachep, objp) != RED_ACTIVE ||
*dbg_redzone2(cachep, objp) != RED_ACTIVE) {
slab_error(cachep, "double free, or memory outside"
" object was overwritten");
printk(KERN_ERR "%p: redzone 1:0x%lx, "
"redzone 2:0x%lx.\n",
objp, *dbg_redzone1(cachep, objp),
*dbg_redzone2(cachep, objp));
}
*dbg_redzone1(cachep, objp) = RED_INACTIVE;
*dbg_redzone2(cachep, objp) = RED_INACTIVE;
}
if (cachep->flags & SLAB_STORE_USER)
*dbg_userword(cachep, objp) = caller;
objnr = obj_to_index(cachep, slabp, objp);
BUG_ON(objnr >= cachep->num);
BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
if (cachep->flags & SLAB_DEBUG_INITIAL) {
/*
* Need to call the slab's constructor so the caller can
* perform a verify of its state (debugging). Called without
* the cache-lock held.
*/
cachep->ctor(objp + obj_offset(cachep),
cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY);
}
if (cachep->flags & SLAB_POISON && cachep->dtor) {
/* we want to cache poison the object,
* call the destruction callback
*/
cachep->dtor(objp + obj_offset(cachep), cachep, 0);
}
[PATCH] slab: implement /proc/slab_allocators Implement /proc/slab_allocators. It produces output like: idr_layer_cache: 80 idr_pre_get+0x33/0x4e buffer_head: 2555 alloc_buffer_head+0x20/0x75 mm_struct: 9 mm_alloc+0x1e/0x42 mm_struct: 20 dup_mm+0x36/0x370 vm_area_struct: 384 dup_mm+0x18f/0x370 vm_area_struct: 151 do_mmap_pgoff+0x2e0/0x7c3 vm_area_struct: 1 split_vma+0x5a/0x10e vm_area_struct: 11 do_brk+0x206/0x2e2 vm_area_struct: 2 copy_vma+0xda/0x142 vm_area_struct: 9 setup_arg_pages+0x99/0x214 fs_cache: 8 copy_fs_struct+0x21/0x133 fs_cache: 29 copy_process+0xf38/0x10e3 files_cache: 30 alloc_files+0x1b/0xcf signal_cache: 81 copy_process+0xbaa/0x10e3 sighand_cache: 77 copy_process+0xe65/0x10e3 sighand_cache: 1 de_thread+0x4d/0x5f8 anon_vma: 241 anon_vma_prepare+0xd9/0xf3 size-2048: 1 add_sect_attrs+0x5f/0x145 size-2048: 2 journal_init_revoke+0x99/0x302 size-2048: 2 journal_init_revoke+0x137/0x302 size-2048: 2 journal_init_inode+0xf9/0x1c4 Cc: Manfred Spraul <manfred@colorfullife.com> Cc: Alexander Nyberg <alexn@telia.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: Ravikiran Thirumalai <kiran@scalex86.org> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk> DESC slab-leaks3-locking-fix EDESC From: Andrew Morton <akpm@osdl.org> Update for slab-remove-cachep-spinlock.patch Cc: Al Viro <viro@ftp.linux.org.uk> Cc: Manfred Spraul <manfred@colorfullife.com> Cc: Alexander Nyberg <alexn@telia.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: Ravikiran Thirumalai <kiran@scalex86.org> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-03-25 11:06:39 +00:00
#ifdef CONFIG_DEBUG_SLAB_LEAK
slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
#endif
if (cachep->flags & SLAB_POISON) {
#ifdef CONFIG_DEBUG_PAGEALLOC
if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
store_stackinfo(cachep, objp, (unsigned long)caller);
kernel_map_pages(virt_to_page(objp),
cachep->buffer_size / PAGE_SIZE, 0);
} else {
poison_obj(cachep, objp, POISON_FREE);
}
#else
poison_obj(cachep, objp, POISON_FREE);
#endif
}
return objp;
}
static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
{
kmem_bufctl_t i;
int entries = 0;
/* Check slab's freelist to see if this obj is there. */
for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
entries++;
if (entries > cachep->num || i >= cachep->num)
goto bad;
}
if (entries != cachep->num - slabp->inuse) {
bad:
printk(KERN_ERR "slab: Internal list corruption detected in "
"cache '%s'(%d), slabp %p(%d). Hexdump:\n",
cachep->name, cachep->num, slabp, slabp->inuse);
for (i = 0;
i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
i++) {
if (i % 16 == 0)
printk("\n%03x:", i);
printk(" %02x", ((unsigned char *)slabp)[i]);
}
printk("\n");
BUG();
}
}
#else
#define kfree_debugcheck(x) do { } while(0)
#define cache_free_debugcheck(x,objp,z) (objp)
#define check_slabp(x,y) do { } while(0)
#endif
static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
{
int batchcount;
struct kmem_list3 *l3;
struct array_cache *ac;
check_irq_off();
ac = cpu_cache_get(cachep);
retry:
batchcount = ac->batchcount;
if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
/*
* If there was little recent activity on this cache, then
* perform only a partial refill. Otherwise we could generate
* refill bouncing.
*/
batchcount = BATCHREFILL_LIMIT;
}
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
l3 = cachep->nodelists[numa_node_id()];
BUG_ON(ac->avail > 0 || !l3);
spin_lock(&l3->list_lock);
/* See if we can refill from the shared array */
if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
goto alloc_done;
while (batchcount > 0) {
struct list_head *entry;
struct slab *slabp;
/* Get slab alloc is to come from. */
entry = l3->slabs_partial.next;
if (entry == &l3->slabs_partial) {
l3->free_touched = 1;
entry = l3->slabs_free.next;
if (entry == &l3->slabs_free)
goto must_grow;
}
slabp = list_entry(entry, struct slab, list);
check_slabp(cachep, slabp);
check_spinlock_acquired(cachep);
while (slabp->inuse < cachep->num && batchcount--) {
STATS_INC_ALLOCED(cachep);
STATS_INC_ACTIVE(cachep);
STATS_SET_HIGH(cachep);
ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
numa_node_id());
}
check_slabp(cachep, slabp);
/* move slabp to correct slabp list: */
list_del(&slabp->list);
if (slabp->free == BUFCTL_END)
list_add(&slabp->list, &l3->slabs_full);
else
list_add(&slabp->list, &l3->slabs_partial);
}
must_grow:
l3->free_objects -= ac->avail;
alloc_done:
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
spin_unlock(&l3->list_lock);
if (unlikely(!ac->avail)) {
int x;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
x = cache_grow(cachep, flags, numa_node_id());
/* cache_grow can reenable interrupts, then ac could change. */
ac = cpu_cache_get(cachep);
if (!x && ac->avail == 0) /* no objects in sight? abort */
return NULL;
if (!ac->avail) /* objects refilled by interrupt? */
goto retry;
}
ac->touched = 1;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
return ac->entry[--ac->avail];
}
static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
gfp_t flags)
{
might_sleep_if(flags & __GFP_WAIT);
#if DEBUG
kmem_flagcheck(cachep, flags);
#endif
}
#if DEBUG
static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
gfp_t flags, void *objp, void *caller)
{
if (!objp)
return objp;
if (cachep->flags & SLAB_POISON) {
#ifdef CONFIG_DEBUG_PAGEALLOC
if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
kernel_map_pages(virt_to_page(objp),
cachep->buffer_size / PAGE_SIZE, 1);
else
check_poison_obj(cachep, objp);
#else
check_poison_obj(cachep, objp);
#endif
poison_obj(cachep, objp, POISON_INUSE);
}
if (cachep->flags & SLAB_STORE_USER)
*dbg_userword(cachep, objp) = caller;
if (cachep->flags & SLAB_RED_ZONE) {
if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
*dbg_redzone2(cachep, objp) != RED_INACTIVE) {
slab_error(cachep, "double free, or memory outside"
" object was overwritten");
printk(KERN_ERR
"%p: redzone 1:0x%lx, redzone 2:0x%lx\n",
objp, *dbg_redzone1(cachep, objp),
*dbg_redzone2(cachep, objp));
}
*dbg_redzone1(cachep, objp) = RED_ACTIVE;
*dbg_redzone2(cachep, objp) = RED_ACTIVE;
}
[PATCH] slab: implement /proc/slab_allocators Implement /proc/slab_allocators. It produces output like: idr_layer_cache: 80 idr_pre_get+0x33/0x4e buffer_head: 2555 alloc_buffer_head+0x20/0x75 mm_struct: 9 mm_alloc+0x1e/0x42 mm_struct: 20 dup_mm+0x36/0x370 vm_area_struct: 384 dup_mm+0x18f/0x370 vm_area_struct: 151 do_mmap_pgoff+0x2e0/0x7c3 vm_area_struct: 1 split_vma+0x5a/0x10e vm_area_struct: 11 do_brk+0x206/0x2e2 vm_area_struct: 2 copy_vma+0xda/0x142 vm_area_struct: 9 setup_arg_pages+0x99/0x214 fs_cache: 8 copy_fs_struct+0x21/0x133 fs_cache: 29 copy_process+0xf38/0x10e3 files_cache: 30 alloc_files+0x1b/0xcf signal_cache: 81 copy_process+0xbaa/0x10e3 sighand_cache: 77 copy_process+0xe65/0x10e3 sighand_cache: 1 de_thread+0x4d/0x5f8 anon_vma: 241 anon_vma_prepare+0xd9/0xf3 size-2048: 1 add_sect_attrs+0x5f/0x145 size-2048: 2 journal_init_revoke+0x99/0x302 size-2048: 2 journal_init_revoke+0x137/0x302 size-2048: 2 journal_init_inode+0xf9/0x1c4 Cc: Manfred Spraul <manfred@colorfullife.com> Cc: Alexander Nyberg <alexn@telia.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: Ravikiran Thirumalai <kiran@scalex86.org> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk> DESC slab-leaks3-locking-fix EDESC From: Andrew Morton <akpm@osdl.org> Update for slab-remove-cachep-spinlock.patch Cc: Al Viro <viro@ftp.linux.org.uk> Cc: Manfred Spraul <manfred@colorfullife.com> Cc: Alexander Nyberg <alexn@telia.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: Ravikiran Thirumalai <kiran@scalex86.org> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-03-25 11:06:39 +00:00
#ifdef CONFIG_DEBUG_SLAB_LEAK
{
struct slab *slabp;
unsigned objnr;
slabp = page_get_slab(virt_to_page(objp));
objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
}
#endif
objp += obj_offset(cachep);
if (cachep->ctor && cachep->flags & SLAB_POISON) {
unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
if (!(flags & __GFP_WAIT))
ctor_flags |= SLAB_CTOR_ATOMIC;
cachep->ctor(objp, cachep, ctor_flags);
}
return objp;
}
#else
#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
#endif
static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
{
void *objp;
struct array_cache *ac;
[PATCH] NUMA policies in the slab allocator V2 This patch fixes a regression in 2.6.14 against 2.6.13 that causes an imbalance in memory allocation during bootup. The slab allocator in 2.6.13 is not numa aware and simply calls alloc_pages(). This means that memory policies may control the behavior of alloc_pages(). During bootup the memory policy is set to MPOL_INTERLEAVE resulting in the spreading out of allocations during bootup over all available nodes. The slab allocator in 2.6.13 has only a single list of slab pages. As a result the per cpu slab cache and the spinlock controlled page lists may contain slab entries from off node memory. The slab allocator in 2.6.13 makes no effort to discern the locality of an entry on its lists. The NUMA aware slab allocator in 2.6.14 controls locality of the slab pages explicitly by calling alloc_pages_node(). The NUMA slab allocator manages slab entries by having lists of available slab pages for each node. The per cpu slab cache can only contain slab entries associated with the node local to the processor. This guarantees that the default allocation mode of the slab allocator always assigns local memory if available. Setting MPOL_INTERLEAVE as a default policy during bootup has no effect anymore. In 2.6.14 all node unspecific slab allocations are performed on the boot processor. This means that most of key data structures are allocated on one node. Most processors will have to refer to these structures making the boot node a potential bottleneck. This may reduce performance and cause unnecessary memory pressure on the boot node. This patch implements NUMA policies in the slab layer. There is the need of explicit application of NUMA memory policies by the slab allcator itself since the NUMA slab allocator does no longer let the page_allocator control locality. The check for policies is made directly at the beginning of __cache_alloc using current->mempolicy. The memory policy is already frequently checked by the page allocator (alloc_page_vma() and alloc_page_current()). So it is highly likely that the cacheline is present. For MPOL_INTERLEAVE kmalloc() will spread out each request to one node after another so that an equal distribution of allocations can be obtained during bootup. It is not possible to push the policy check to lower layers of the NUMA slab allocator since the per cpu caches are now only containing slab entries from the current node. If the policy says that the local node is not to be preferred or forbidden then there is no point in checking the slab cache or local list of slab pages. The allocation better be directed immediately to the lists containing slab entries for the allowed set of nodes. This way of applying policy also fixes another strange behavior in 2.6.13. alloc_pages() is controlled by the memory allocation policy of the current process. It could therefore be that one process is running with MPOL_INTERLEAVE and would f.e. obtain a new page following that policy since no slab entries are in the lists anymore. A page can typically be used for multiple slab entries but lets say that the current process is only using one. The other entries are then added to the slab lists. These are now non local entries in the slab lists despite of the possible availability of local pages that would provide faster access and increase the performance of the application. Another process without MPOL_INTERLEAVE may now run and expect a local slab entry from kmalloc(). However, there are still these free slab entries from the off node page obtained from the other process via MPOL_INTERLEAVE in the cache. The process will then get an off node slab entry although other slab entries may be available that are local to that process. This means that the policy if one process may contaminate the locality of the slab caches for other processes. This patch in effect insures that a per process policy is followed for the allocation of slab entries and that there cannot be a memory policy influence from one process to another. A process with default policy will always get a local slab entry if one is available. And the process using memory policies will get its memory arranged as requested. Off-node slab allocation will require the use of spinlocks and will make the use of per cpu caches not possible. A process using memory policies to redirect allocations offnode will have to cope with additional lock overhead in addition to the latency added by the need to access a remote slab entry. Changes V1->V2 - Remove #ifdef CONFIG_NUMA by moving forward declaration into prior #ifdef CONFIG_NUMA section. - Give the function determining the node number to use a saner name. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-01-19 01:42:36 +00:00
#ifdef CONFIG_NUMA
if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
objp = alternate_node_alloc(cachep, flags);
if (objp != NULL)
return objp;
[PATCH] NUMA policies in the slab allocator V2 This patch fixes a regression in 2.6.14 against 2.6.13 that causes an imbalance in memory allocation during bootup. The slab allocator in 2.6.13 is not numa aware and simply calls alloc_pages(). This means that memory policies may control the behavior of alloc_pages(). During bootup the memory policy is set to MPOL_INTERLEAVE resulting in the spreading out of allocations during bootup over all available nodes. The slab allocator in 2.6.13 has only a single list of slab pages. As a result the per cpu slab cache and the spinlock controlled page lists may contain slab entries from off node memory. The slab allocator in 2.6.13 makes no effort to discern the locality of an entry on its lists. The NUMA aware slab allocator in 2.6.14 controls locality of the slab pages explicitly by calling alloc_pages_node(). The NUMA slab allocator manages slab entries by having lists of available slab pages for each node. The per cpu slab cache can only contain slab entries associated with the node local to the processor. This guarantees that the default allocation mode of the slab allocator always assigns local memory if available. Setting MPOL_INTERLEAVE as a default policy during bootup has no effect anymore. In 2.6.14 all node unspecific slab allocations are performed on the boot processor. This means that most of key data structures are allocated on one node. Most processors will have to refer to these structures making the boot node a potential bottleneck. This may reduce performance and cause unnecessary memory pressure on the boot node. This patch implements NUMA policies in the slab layer. There is the need of explicit application of NUMA memory policies by the slab allcator itself since the NUMA slab allocator does no longer let the page_allocator control locality. The check for policies is made directly at the beginning of __cache_alloc using current->mempolicy. The memory policy is already frequently checked by the page allocator (alloc_page_vma() and alloc_page_current()). So it is highly likely that the cacheline is present. For MPOL_INTERLEAVE kmalloc() will spread out each request to one node after another so that an equal distribution of allocations can be obtained during bootup. It is not possible to push the policy check to lower layers of the NUMA slab allocator since the per cpu caches are now only containing slab entries from the current node. If the policy says that the local node is not to be preferred or forbidden then there is no point in checking the slab cache or local list of slab pages. The allocation better be directed immediately to the lists containing slab entries for the allowed set of nodes. This way of applying policy also fixes another strange behavior in 2.6.13. alloc_pages() is controlled by the memory allocation policy of the current process. It could therefore be that one process is running with MPOL_INTERLEAVE and would f.e. obtain a new page following that policy since no slab entries are in the lists anymore. A page can typically be used for multiple slab entries but lets say that the current process is only using one. The other entries are then added to the slab lists. These are now non local entries in the slab lists despite of the possible availability of local pages that would provide faster access and increase the performance of the application. Another process without MPOL_INTERLEAVE may now run and expect a local slab entry from kmalloc(). However, there are still these free slab entries from the off node page obtained from the other process via MPOL_INTERLEAVE in the cache. The process will then get an off node slab entry although other slab entries may be available that are local to that process. This means that the policy if one process may contaminate the locality of the slab caches for other processes. This patch in effect insures that a per process policy is followed for the allocation of slab entries and that there cannot be a memory policy influence from one process to another. A process with default policy will always get a local slab entry if one is available. And the process using memory policies will get its memory arranged as requested. Off-node slab allocation will require the use of spinlocks and will make the use of per cpu caches not possible. A process using memory policies to redirect allocations offnode will have to cope with additional lock overhead in addition to the latency added by the need to access a remote slab entry. Changes V1->V2 - Remove #ifdef CONFIG_NUMA by moving forward declaration into prior #ifdef CONFIG_NUMA section. - Give the function determining the node number to use a saner name. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-01-19 01:42:36 +00:00
}
#endif
check_irq_off();
ac = cpu_cache_get(cachep);
if (likely(ac->avail)) {
STATS_INC_ALLOCHIT(cachep);
ac->touched = 1;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
objp = ac->entry[--ac->avail];
} else {
STATS_INC_ALLOCMISS(cachep);
objp = cache_alloc_refill(cachep, flags);
}
return objp;
}
static __always_inline void *__cache_alloc(struct kmem_cache *cachep,
gfp_t flags, void *caller)
{
unsigned long save_flags;
void *objp;
cache_alloc_debugcheck_before(cachep, flags);
local_irq_save(save_flags);
objp = ____cache_alloc(cachep, flags);
local_irq_restore(save_flags);
objp = cache_alloc_debugcheck_after(cachep, flags, objp,
caller);
prefetchw(objp);
return objp;
}
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
#ifdef CONFIG_NUMA
/*
* Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
*
* If we are in_interrupt, then process context, including cpusets and
* mempolicy, may not apply and should not be used for allocation policy.
*/
static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
{
int nid_alloc, nid_here;
if (in_interrupt())
return NULL;
nid_alloc = nid_here = numa_node_id();
if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
nid_alloc = cpuset_mem_spread_node();
else if (current->mempolicy)
nid_alloc = slab_node(current->mempolicy);
if (nid_alloc != nid_here)
return __cache_alloc_node(cachep, flags, nid_alloc);
return NULL;
}
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
/*
* A interface to enable slab creation on nodeid
*/
static void *__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
int nodeid)
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
{
struct list_head *entry;
struct slab *slabp;
struct kmem_list3 *l3;
void *obj;
int x;
l3 = cachep->nodelists[nodeid];
BUG_ON(!l3);
retry:
check_irq_off();
spin_lock(&l3->list_lock);
entry = l3->slabs_partial.next;
if (entry == &l3->slabs_partial) {
l3->free_touched = 1;
entry = l3->slabs_free.next;
if (entry == &l3->slabs_free)
goto must_grow;
}
slabp = list_entry(entry, struct slab, list);
check_spinlock_acquired_node(cachep, nodeid);
check_slabp(cachep, slabp);
STATS_INC_NODEALLOCS(cachep);
STATS_INC_ACTIVE(cachep);
STATS_SET_HIGH(cachep);
BUG_ON(slabp->inuse == cachep->num);
obj = slab_get_obj(cachep, slabp, nodeid);
check_slabp(cachep, slabp);
l3->free_objects--;
/* move slabp to correct slabp list: */
list_del(&slabp->list);
if (slabp->free == BUFCTL_END)
list_add(&slabp->list, &l3->slabs_full);
else
list_add(&slabp->list, &l3->slabs_partial);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
spin_unlock(&l3->list_lock);
goto done;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
must_grow:
spin_unlock(&l3->list_lock);
x = cache_grow(cachep, flags, nodeid);
if (!x)
return NULL;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
goto retry;
done:
return obj;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
}
#endif
/*
* Caller needs to acquire correct kmem_list's list_lock
*/
static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
int node)
{
int i;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
struct kmem_list3 *l3;
for (i = 0; i < nr_objects; i++) {
void *objp = objpp[i];
struct slab *slabp;
slabp = virt_to_slab(objp);
l3 = cachep->nodelists[node];
list_del(&slabp->list);
check_spinlock_acquired_node(cachep, node);
check_slabp(cachep, slabp);
slab_put_obj(cachep, slabp, objp, node);
STATS_DEC_ACTIVE(cachep);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
l3->free_objects++;
check_slabp(cachep, slabp);
/* fixup slab chains */
if (slabp->inuse == 0) {
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
if (l3->free_objects > l3->free_limit) {
l3->free_objects -= cachep->num;
slab_destroy(cachep, slabp);
} else {
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
list_add(&slabp->list, &l3->slabs_free);
}
} else {
/* Unconditionally move a slab to the end of the
* partial list on free - maximum time for the
* other objects to be freed, too.
*/
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
list_add_tail(&slabp->list, &l3->slabs_partial);
}
}
}
static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
{
int batchcount;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
struct kmem_list3 *l3;
int node = numa_node_id();
batchcount = ac->batchcount;
#if DEBUG
BUG_ON(!batchcount || batchcount > ac->avail);
#endif
check_irq_off();
l3 = cachep->nodelists[node];
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
spin_lock(&l3->list_lock);
if (l3->shared) {
struct array_cache *shared_array = l3->shared;
int max = shared_array->limit - shared_array->avail;
if (max) {
if (batchcount > max)
batchcount = max;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
memcpy(&(shared_array->entry[shared_array->avail]),
ac->entry, sizeof(void *) * batchcount);
shared_array->avail += batchcount;
goto free_done;
}
}
free_block(cachep, ac->entry, batchcount, node);
free_done:
#if STATS
{
int i = 0;
struct list_head *p;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
p = l3->slabs_free.next;
while (p != &(l3->slabs_free)) {
struct slab *slabp;
slabp = list_entry(p, struct slab, list);
BUG_ON(slabp->inuse);
i++;
p = p->next;
}
STATS_SET_FREEABLE(cachep, i);
}
#endif
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
spin_unlock(&l3->list_lock);
ac->avail -= batchcount;
memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
}
/*
* Release an obj back to its cache. If the obj has a constructed state, it must
* be in this state _before_ it is released. Called with disabled ints.
*/
static inline void __cache_free(struct kmem_cache *cachep, void *objp)
{
struct array_cache *ac = cpu_cache_get(cachep);
check_irq_off();
objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
/* Make sure we are not freeing a object from another
* node to the array cache on this cpu.
*/
#ifdef CONFIG_NUMA
{
struct slab *slabp;
slabp = virt_to_slab(objp);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
if (unlikely(slabp->nodeid != numa_node_id())) {
struct array_cache *alien = NULL;
int nodeid = slabp->nodeid;
struct kmem_list3 *l3;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
l3 = cachep->nodelists[numa_node_id()];
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
STATS_INC_NODEFREES(cachep);
if (l3->alien && l3->alien[nodeid]) {
alien = l3->alien[nodeid];
spin_lock(&alien->lock);
if (unlikely(alien->avail == alien->limit)) {
STATS_INC_ACOVERFLOW(cachep);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
__drain_alien_cache(cachep,
alien, nodeid);
}
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
alien->entry[alien->avail++] = objp;
spin_unlock(&alien->lock);
} else {
spin_lock(&(cachep->nodelists[nodeid])->
list_lock);
free_block(cachep, &objp, 1, nodeid);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
spin_unlock(&(cachep->nodelists[nodeid])->
list_lock);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
}
return;
}
}
#endif
if (likely(ac->avail < ac->limit)) {
STATS_INC_FREEHIT(cachep);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
ac->entry[ac->avail++] = objp;
return;
} else {
STATS_INC_FREEMISS(cachep);
cache_flusharray(cachep, ac);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
ac->entry[ac->avail++] = objp;
}
}
/**
* kmem_cache_alloc - Allocate an object
* @cachep: The cache to allocate from.
* @flags: See kmalloc().
*
* Allocate an object from this cache. The flags are only relevant
* if the cache has no available objects.
*/
void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
{
return __cache_alloc(cachep, flags, __builtin_return_address(0));
}
EXPORT_SYMBOL(kmem_cache_alloc);
/**
* kmem_cache_alloc - Allocate an object. The memory is set to zero.
* @cache: The cache to allocate from.
* @flags: See kmalloc().
*
* Allocate an object from this cache and set the allocated memory to zero.
* The flags are only relevant if the cache has no available objects.
*/
void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags)
{
void *ret = __cache_alloc(cache, flags, __builtin_return_address(0));
if (ret)
memset(ret, 0, obj_size(cache));
return ret;
}
EXPORT_SYMBOL(kmem_cache_zalloc);
/**
* kmem_ptr_validate - check if an untrusted pointer might
* be a slab entry.
* @cachep: the cache we're checking against
* @ptr: pointer to validate
*
* This verifies that the untrusted pointer looks sane:
* it is _not_ a guarantee that the pointer is actually
* part of the slab cache in question, but it at least
* validates that the pointer can be dereferenced and
* looks half-way sane.
*
* Currently only used for dentry validation.
*/
int fastcall kmem_ptr_validate(struct kmem_cache *cachep, void *ptr)
{
unsigned long addr = (unsigned long)ptr;
unsigned long min_addr = PAGE_OFFSET;
unsigned long align_mask = BYTES_PER_WORD - 1;
unsigned long size = cachep->buffer_size;
struct page *page;
if (unlikely(addr < min_addr))
goto out;
if (unlikely(addr > (unsigned long)high_memory - size))
goto out;
if (unlikely(addr & align_mask))
goto out;
if (unlikely(!kern_addr_valid(addr)))
goto out;
if (unlikely(!kern_addr_valid(addr + size - 1)))
goto out;
page = virt_to_page(ptr);
if (unlikely(!PageSlab(page)))
goto out;
if (unlikely(page_get_cache(page) != cachep))
goto out;
return 1;
out:
return 0;
}
#ifdef CONFIG_NUMA
/**
* kmem_cache_alloc_node - Allocate an object on the specified node
* @cachep: The cache to allocate from.
* @flags: See kmalloc().
* @nodeid: node number of the target node.
*
* Identical to kmem_cache_alloc, except that this function is slow
* and can sleep. And it will allocate memory on the given node, which
* can improve the performance for cpu bound structures.
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
* New and improved: it will now make sure that the object gets
* put on the correct node list so that there is no false sharing.
*/
void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
{
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
unsigned long save_flags;
void *ptr;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
cache_alloc_debugcheck_before(cachep, flags);
local_irq_save(save_flags);
if (nodeid == -1 || nodeid == numa_node_id() ||
!cachep->nodelists[nodeid])
ptr = ____cache_alloc(cachep, flags);
else
ptr = __cache_alloc_node(cachep, flags, nodeid);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
local_irq_restore(save_flags);
ptr = cache_alloc_debugcheck_after(cachep, flags, ptr,
__builtin_return_address(0));
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
return ptr;
}
EXPORT_SYMBOL(kmem_cache_alloc_node);
void *kmalloc_node(size_t size, gfp_t flags, int node)
[PATCH] add kmalloc_node, inline cleanup The patch makes the following function calls available to allocate memory on a specific node without changing the basic operation of the slab allocator: kmem_cache_alloc_node(kmem_cache_t *cachep, unsigned int flags, int node); kmalloc_node(size_t size, unsigned int flags, int node); in a similar way to the existing node-blind functions: kmem_cache_alloc(kmem_cache_t *cachep, unsigned int flags); kmalloc(size, flags); kmem_cache_alloc_node was changed to pass flags and the node information through the existing layers of the slab allocator (which lead to some minor rearrangements). The functions at the lowest layer (kmem_getpages, cache_grow) are already node aware. Also __alloc_percpu can call kmalloc_node now. Performance measurements (using the pageset localization patch) yields: w/o patches: Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.97 Wed Mar 30 20:50:43 2005 100 25170.83 91 251.7083 23.12 150.10 Wed Mar 30 20:51:06 2005 200 34601.66 84 173.0083 33.64 294.14 Wed Mar 30 20:51:40 2005 300 37154.47 86 123.8482 46.99 436.56 Wed Mar 30 20:52:28 2005 400 39839.82 80 99.5995 58.43 580.46 Wed Mar 30 20:53:27 2005 500 40036.32 79 80.0726 72.68 728.60 Wed Mar 30 20:54:40 2005 600 44074.21 79 73.4570 79.23 872.10 Wed Mar 30 20:55:59 2005 700 44016.60 78 62.8809 92.56 1015.84 Wed Mar 30 20:57:32 2005 800 40411.05 80 50.5138 115.22 1161.13 Wed Mar 30 20:59:28 2005 900 42298.56 79 46.9984 123.83 1303.42 Wed Mar 30 21:01:33 2005 1000 40955.05 80 40.9551 142.11 1441.92 Wed Mar 30 21:03:55 2005 with pageset localization and slab API patches: Tasks jobs/min jti jobs/min/task real cpu 1 484.19 100 484.1930 12.02 1.98 Wed Mar 30 21:10:18 2005 100 27428.25 92 274.2825 21.22 149.79 Wed Mar 30 21:10:40 2005 200 37228.94 86 186.1447 31.27 293.49 Wed Mar 30 21:11:12 2005 300 41725.42 85 139.0847 41.84 434.10 Wed Mar 30 21:11:54 2005 400 43032.22 82 107.5805 54.10 582.06 Wed Mar 30 21:12:48 2005 500 42211.23 83 84.4225 68.94 722.61 Wed Mar 30 21:13:58 2005 600 40084.49 82 66.8075 87.12 873.11 Wed Mar 30 21:15:25 2005 700 44169.30 79 63.0990 92.24 1008.77 Wed Mar 30 21:16:58 2005 800 43097.94 79 53.8724 108.03 1155.88 Wed Mar 30 21:18:47 2005 900 41846.75 79 46.4964 125.17 1303.38 Wed Mar 30 21:20:52 2005 1000 40247.85 79 40.2478 144.60 1442.21 Wed Mar 30 21:23:17 2005 Signed-off-by: Christoph Lameter <christoph@lameter.com> Signed-off-by: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-05-01 15:58:38 +00:00
{
struct kmem_cache *cachep;
[PATCH] add kmalloc_node, inline cleanup The patch makes the following function calls available to allocate memory on a specific node without changing the basic operation of the slab allocator: kmem_cache_alloc_node(kmem_cache_t *cachep, unsigned int flags, int node); kmalloc_node(size_t size, unsigned int flags, int node); in a similar way to the existing node-blind functions: kmem_cache_alloc(kmem_cache_t *cachep, unsigned int flags); kmalloc(size, flags); kmem_cache_alloc_node was changed to pass flags and the node information through the existing layers of the slab allocator (which lead to some minor rearrangements). The functions at the lowest layer (kmem_getpages, cache_grow) are already node aware. Also __alloc_percpu can call kmalloc_node now. Performance measurements (using the pageset localization patch) yields: w/o patches: Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.97 Wed Mar 30 20:50:43 2005 100 25170.83 91 251.7083 23.12 150.10 Wed Mar 30 20:51:06 2005 200 34601.66 84 173.0083 33.64 294.14 Wed Mar 30 20:51:40 2005 300 37154.47 86 123.8482 46.99 436.56 Wed Mar 30 20:52:28 2005 400 39839.82 80 99.5995 58.43 580.46 Wed Mar 30 20:53:27 2005 500 40036.32 79 80.0726 72.68 728.60 Wed Mar 30 20:54:40 2005 600 44074.21 79 73.4570 79.23 872.10 Wed Mar 30 20:55:59 2005 700 44016.60 78 62.8809 92.56 1015.84 Wed Mar 30 20:57:32 2005 800 40411.05 80 50.5138 115.22 1161.13 Wed Mar 30 20:59:28 2005 900 42298.56 79 46.9984 123.83 1303.42 Wed Mar 30 21:01:33 2005 1000 40955.05 80 40.9551 142.11 1441.92 Wed Mar 30 21:03:55 2005 with pageset localization and slab API patches: Tasks jobs/min jti jobs/min/task real cpu 1 484.19 100 484.1930 12.02 1.98 Wed Mar 30 21:10:18 2005 100 27428.25 92 274.2825 21.22 149.79 Wed Mar 30 21:10:40 2005 200 37228.94 86 186.1447 31.27 293.49 Wed Mar 30 21:11:12 2005 300 41725.42 85 139.0847 41.84 434.10 Wed Mar 30 21:11:54 2005 400 43032.22 82 107.5805 54.10 582.06 Wed Mar 30 21:12:48 2005 500 42211.23 83 84.4225 68.94 722.61 Wed Mar 30 21:13:58 2005 600 40084.49 82 66.8075 87.12 873.11 Wed Mar 30 21:15:25 2005 700 44169.30 79 63.0990 92.24 1008.77 Wed Mar 30 21:16:58 2005 800 43097.94 79 53.8724 108.03 1155.88 Wed Mar 30 21:18:47 2005 900 41846.75 79 46.4964 125.17 1303.38 Wed Mar 30 21:20:52 2005 1000 40247.85 79 40.2478 144.60 1442.21 Wed Mar 30 21:23:17 2005 Signed-off-by: Christoph Lameter <christoph@lameter.com> Signed-off-by: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-05-01 15:58:38 +00:00
cachep = kmem_find_general_cachep(size, flags);
if (unlikely(cachep == NULL))
return NULL;
return kmem_cache_alloc_node(cachep, flags, node);
}
EXPORT_SYMBOL(kmalloc_node);
#endif
/**
* kmalloc - allocate memory
* @size: how many bytes of memory are required.
* @flags: the type of memory to allocate.
* @caller: function caller for debug tracking of the caller
*
* kmalloc is the normal method of allocating memory
* in the kernel.
*
* The @flags argument may be one of:
*
* %GFP_USER - Allocate memory on behalf of user. May sleep.
*
* %GFP_KERNEL - Allocate normal kernel ram. May sleep.
*
* %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers.
*
* Additionally, the %GFP_DMA flag may be set to indicate the memory
* must be suitable for DMA. This can mean different things on different
* platforms. For example, on i386, it means that the memory must come
* from the first 16MB.
*/
static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
void *caller)
{
struct kmem_cache *cachep;
[PATCH] add kmalloc_node, inline cleanup The patch makes the following function calls available to allocate memory on a specific node without changing the basic operation of the slab allocator: kmem_cache_alloc_node(kmem_cache_t *cachep, unsigned int flags, int node); kmalloc_node(size_t size, unsigned int flags, int node); in a similar way to the existing node-blind functions: kmem_cache_alloc(kmem_cache_t *cachep, unsigned int flags); kmalloc(size, flags); kmem_cache_alloc_node was changed to pass flags and the node information through the existing layers of the slab allocator (which lead to some minor rearrangements). The functions at the lowest layer (kmem_getpages, cache_grow) are already node aware. Also __alloc_percpu can call kmalloc_node now. Performance measurements (using the pageset localization patch) yields: w/o patches: Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.97 Wed Mar 30 20:50:43 2005 100 25170.83 91 251.7083 23.12 150.10 Wed Mar 30 20:51:06 2005 200 34601.66 84 173.0083 33.64 294.14 Wed Mar 30 20:51:40 2005 300 37154.47 86 123.8482 46.99 436.56 Wed Mar 30 20:52:28 2005 400 39839.82 80 99.5995 58.43 580.46 Wed Mar 30 20:53:27 2005 500 40036.32 79 80.0726 72.68 728.60 Wed Mar 30 20:54:40 2005 600 44074.21 79 73.4570 79.23 872.10 Wed Mar 30 20:55:59 2005 700 44016.60 78 62.8809 92.56 1015.84 Wed Mar 30 20:57:32 2005 800 40411.05 80 50.5138 115.22 1161.13 Wed Mar 30 20:59:28 2005 900 42298.56 79 46.9984 123.83 1303.42 Wed Mar 30 21:01:33 2005 1000 40955.05 80 40.9551 142.11 1441.92 Wed Mar 30 21:03:55 2005 with pageset localization and slab API patches: Tasks jobs/min jti jobs/min/task real cpu 1 484.19 100 484.1930 12.02 1.98 Wed Mar 30 21:10:18 2005 100 27428.25 92 274.2825 21.22 149.79 Wed Mar 30 21:10:40 2005 200 37228.94 86 186.1447 31.27 293.49 Wed Mar 30 21:11:12 2005 300 41725.42 85 139.0847 41.84 434.10 Wed Mar 30 21:11:54 2005 400 43032.22 82 107.5805 54.10 582.06 Wed Mar 30 21:12:48 2005 500 42211.23 83 84.4225 68.94 722.61 Wed Mar 30 21:13:58 2005 600 40084.49 82 66.8075 87.12 873.11 Wed Mar 30 21:15:25 2005 700 44169.30 79 63.0990 92.24 1008.77 Wed Mar 30 21:16:58 2005 800 43097.94 79 53.8724 108.03 1155.88 Wed Mar 30 21:18:47 2005 900 41846.75 79 46.4964 125.17 1303.38 Wed Mar 30 21:20:52 2005 1000 40247.85 79 40.2478 144.60 1442.21 Wed Mar 30 21:23:17 2005 Signed-off-by: Christoph Lameter <christoph@lameter.com> Signed-off-by: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-05-01 15:58:38 +00:00
/* If you want to save a few bytes .text space: replace
* __ with kmem_.
* Then kmalloc uses the uninlined functions instead of the inline
* functions.
*/
cachep = __find_general_cachep(size, flags);
if (unlikely(cachep == NULL))
return NULL;
return __cache_alloc(cachep, flags, caller);
}
void *__kmalloc(size_t size, gfp_t flags)
{
[PATCH] slab: implement /proc/slab_allocators Implement /proc/slab_allocators. It produces output like: idr_layer_cache: 80 idr_pre_get+0x33/0x4e buffer_head: 2555 alloc_buffer_head+0x20/0x75 mm_struct: 9 mm_alloc+0x1e/0x42 mm_struct: 20 dup_mm+0x36/0x370 vm_area_struct: 384 dup_mm+0x18f/0x370 vm_area_struct: 151 do_mmap_pgoff+0x2e0/0x7c3 vm_area_struct: 1 split_vma+0x5a/0x10e vm_area_struct: 11 do_brk+0x206/0x2e2 vm_area_struct: 2 copy_vma+0xda/0x142 vm_area_struct: 9 setup_arg_pages+0x99/0x214 fs_cache: 8 copy_fs_struct+0x21/0x133 fs_cache: 29 copy_process+0xf38/0x10e3 files_cache: 30 alloc_files+0x1b/0xcf signal_cache: 81 copy_process+0xbaa/0x10e3 sighand_cache: 77 copy_process+0xe65/0x10e3 sighand_cache: 1 de_thread+0x4d/0x5f8 anon_vma: 241 anon_vma_prepare+0xd9/0xf3 size-2048: 1 add_sect_attrs+0x5f/0x145 size-2048: 2 journal_init_revoke+0x99/0x302 size-2048: 2 journal_init_revoke+0x137/0x302 size-2048: 2 journal_init_inode+0xf9/0x1c4 Cc: Manfred Spraul <manfred@colorfullife.com> Cc: Alexander Nyberg <alexn@telia.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: Ravikiran Thirumalai <kiran@scalex86.org> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk> DESC slab-leaks3-locking-fix EDESC From: Andrew Morton <akpm@osdl.org> Update for slab-remove-cachep-spinlock.patch Cc: Al Viro <viro@ftp.linux.org.uk> Cc: Manfred Spraul <manfred@colorfullife.com> Cc: Alexander Nyberg <alexn@telia.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: Ravikiran Thirumalai <kiran@scalex86.org> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-03-25 11:06:39 +00:00
#ifndef CONFIG_DEBUG_SLAB
return __do_kmalloc(size, flags, NULL);
[PATCH] slab: implement /proc/slab_allocators Implement /proc/slab_allocators. It produces output like: idr_layer_cache: 80 idr_pre_get+0x33/0x4e buffer_head: 2555 alloc_buffer_head+0x20/0x75 mm_struct: 9 mm_alloc+0x1e/0x42 mm_struct: 20 dup_mm+0x36/0x370 vm_area_struct: 384 dup_mm+0x18f/0x370 vm_area_struct: 151 do_mmap_pgoff+0x2e0/0x7c3 vm_area_struct: 1 split_vma+0x5a/0x10e vm_area_struct: 11 do_brk+0x206/0x2e2 vm_area_struct: 2 copy_vma+0xda/0x142 vm_area_struct: 9 setup_arg_pages+0x99/0x214 fs_cache: 8 copy_fs_struct+0x21/0x133 fs_cache: 29 copy_process+0xf38/0x10e3 files_cache: 30 alloc_files+0x1b/0xcf signal_cache: 81 copy_process+0xbaa/0x10e3 sighand_cache: 77 copy_process+0xe65/0x10e3 sighand_cache: 1 de_thread+0x4d/0x5f8 anon_vma: 241 anon_vma_prepare+0xd9/0xf3 size-2048: 1 add_sect_attrs+0x5f/0x145 size-2048: 2 journal_init_revoke+0x99/0x302 size-2048: 2 journal_init_revoke+0x137/0x302 size-2048: 2 journal_init_inode+0xf9/0x1c4 Cc: Manfred Spraul <manfred@colorfullife.com> Cc: Alexander Nyberg <alexn@telia.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: Ravikiran Thirumalai <kiran@scalex86.org> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk> DESC slab-leaks3-locking-fix EDESC From: Andrew Morton <akpm@osdl.org> Update for slab-remove-cachep-spinlock.patch Cc: Al Viro <viro@ftp.linux.org.uk> Cc: Manfred Spraul <manfred@colorfullife.com> Cc: Alexander Nyberg <alexn@telia.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: Ravikiran Thirumalai <kiran@scalex86.org> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-03-25 11:06:39 +00:00
#else
return __do_kmalloc(size, flags, __builtin_return_address(0));
#endif
}
EXPORT_SYMBOL(__kmalloc);
[PATCH] slab: implement /proc/slab_allocators Implement /proc/slab_allocators. It produces output like: idr_layer_cache: 80 idr_pre_get+0x33/0x4e buffer_head: 2555 alloc_buffer_head+0x20/0x75 mm_struct: 9 mm_alloc+0x1e/0x42 mm_struct: 20 dup_mm+0x36/0x370 vm_area_struct: 384 dup_mm+0x18f/0x370 vm_area_struct: 151 do_mmap_pgoff+0x2e0/0x7c3 vm_area_struct: 1 split_vma+0x5a/0x10e vm_area_struct: 11 do_brk+0x206/0x2e2 vm_area_struct: 2 copy_vma+0xda/0x142 vm_area_struct: 9 setup_arg_pages+0x99/0x214 fs_cache: 8 copy_fs_struct+0x21/0x133 fs_cache: 29 copy_process+0xf38/0x10e3 files_cache: 30 alloc_files+0x1b/0xcf signal_cache: 81 copy_process+0xbaa/0x10e3 sighand_cache: 77 copy_process+0xe65/0x10e3 sighand_cache: 1 de_thread+0x4d/0x5f8 anon_vma: 241 anon_vma_prepare+0xd9/0xf3 size-2048: 1 add_sect_attrs+0x5f/0x145 size-2048: 2 journal_init_revoke+0x99/0x302 size-2048: 2 journal_init_revoke+0x137/0x302 size-2048: 2 journal_init_inode+0xf9/0x1c4 Cc: Manfred Spraul <manfred@colorfullife.com> Cc: Alexander Nyberg <alexn@telia.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: Ravikiran Thirumalai <kiran@scalex86.org> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk> DESC slab-leaks3-locking-fix EDESC From: Andrew Morton <akpm@osdl.org> Update for slab-remove-cachep-spinlock.patch Cc: Al Viro <viro@ftp.linux.org.uk> Cc: Manfred Spraul <manfred@colorfullife.com> Cc: Alexander Nyberg <alexn@telia.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: Ravikiran Thirumalai <kiran@scalex86.org> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-03-25 11:06:39 +00:00
#ifdef CONFIG_DEBUG_SLAB
void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
{
return __do_kmalloc(size, flags, caller);
}
EXPORT_SYMBOL(__kmalloc_track_caller);
#endif
#ifdef CONFIG_SMP
/**
* __alloc_percpu - allocate one copy of the object for every present
* cpu in the system, zeroing them.
* Objects should be dereferenced using the per_cpu_ptr macro only.
*
* @size: how many bytes of memory are required.
*/
void *__alloc_percpu(size_t size)
{
int i;
struct percpu_data *pdata = kmalloc(sizeof(*pdata), GFP_KERNEL);
if (!pdata)
return NULL;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
/*
* Cannot use for_each_online_cpu since a cpu may come online
* and we have no way of figuring out how to fix the array
* that we have allocated then....
*/
for_each_possible_cpu(i) {
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
int node = cpu_to_node(i);
if (node_online(node))
pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node);
else
pdata->ptrs[i] = kmalloc(size, GFP_KERNEL);
if (!pdata->ptrs[i])
goto unwind_oom;
memset(pdata->ptrs[i], 0, size);
}
/* Catch derefs w/o wrappers */
return (void *)(~(unsigned long)pdata);
unwind_oom:
while (--i >= 0) {
if (!cpu_possible(i))
continue;
kfree(pdata->ptrs[i]);
}
kfree(pdata);
return NULL;
}
EXPORT_SYMBOL(__alloc_percpu);
#endif
/**
* kmem_cache_free - Deallocate an object
* @cachep: The cache the allocation was from.
* @objp: The previously allocated object.
*
* Free an object which was previously allocated from this
* cache.
*/
void kmem_cache_free(struct kmem_cache *cachep, void *objp)
{
unsigned long flags;
local_irq_save(flags);
__cache_free(cachep, objp);
local_irq_restore(flags);
}
EXPORT_SYMBOL(kmem_cache_free);
/**
* kfree - free previously allocated memory
* @objp: pointer returned by kmalloc.
*
* If @objp is NULL, no operation is performed.
*
* Don't free memory not originally allocated by kmalloc()
* or you will run into trouble.
*/
void kfree(const void *objp)
{
struct kmem_cache *c;
unsigned long flags;
if (unlikely(!objp))
return;
local_irq_save(flags);
kfree_debugcheck(objp);
c = virt_to_cache(objp);
mutex_debug_check_no_locks_freed(objp, obj_size(c));
__cache_free(c, (void *)objp);
local_irq_restore(flags);
}
EXPORT_SYMBOL(kfree);
#ifdef CONFIG_SMP
/**
* free_percpu - free previously allocated percpu memory
* @objp: pointer returned by alloc_percpu.
*
* Don't free memory not originally allocated by alloc_percpu()
* The complemented objp is to check for that.
*/
void free_percpu(const void *objp)
{
int i;
struct percpu_data *p = (struct percpu_data *)(~(unsigned long)objp);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
/*
* We allocate for all cpus so we cannot use for online cpu here.
*/
for_each_possible_cpu(i)
kfree(p->ptrs[i]);
kfree(p);
}
EXPORT_SYMBOL(free_percpu);
#endif
unsigned int kmem_cache_size(struct kmem_cache *cachep)
{
return obj_size(cachep);
}
EXPORT_SYMBOL(kmem_cache_size);
const char *kmem_cache_name(struct kmem_cache *cachep)
{
return cachep->name;
}
EXPORT_SYMBOL_GPL(kmem_cache_name);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
/*
* This initializes kmem_list3 or resizes varioius caches for all nodes.
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
*/
static int alloc_kmemlist(struct kmem_cache *cachep)
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
{
int node;
struct kmem_list3 *l3;
struct array_cache *new_shared;
struct array_cache **new_alien;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
for_each_online_node(node) {
new_alien = alloc_alien_cache(node, cachep->limit);
if (!new_alien)
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
goto fail;
new_shared = alloc_arraycache(node,
cachep->shared*cachep->batchcount,
0xbaadf00d);
if (!new_shared) {
free_alien_cache(new_alien);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
goto fail;
}
l3 = cachep->nodelists[node];
if (l3) {
struct array_cache *shared = l3->shared;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
spin_lock_irq(&l3->list_lock);
if (shared)
free_block(cachep, shared->entry,
shared->avail, node);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
l3->shared = new_shared;
if (!l3->alien) {
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
l3->alien = new_alien;
new_alien = NULL;
}
l3->free_limit = (1 + nr_cpus_node(node)) *
cachep->batchcount + cachep->num;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
spin_unlock_irq(&l3->list_lock);
kfree(shared);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
free_alien_cache(new_alien);
continue;
}
l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
if (!l3) {
free_alien_cache(new_alien);
kfree(new_shared);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
goto fail;
}
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
kmem_list3_init(l3);
l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
((unsigned long)cachep) % REAPTIMEOUT_LIST3;
l3->shared = new_shared;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
l3->alien = new_alien;
l3->free_limit = (1 + nr_cpus_node(node)) *
cachep->batchcount + cachep->num;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
cachep->nodelists[node] = l3;
}
return 0;
fail:
if (!cachep->next.next) {
/* Cache is not active yet. Roll back what we did */
node--;
while (node >= 0) {
if (cachep->nodelists[node]) {
l3 = cachep->nodelists[node];
kfree(l3->shared);
free_alien_cache(l3->alien);
kfree(l3);
cachep->nodelists[node] = NULL;
}
node--;
}
}
return -ENOMEM;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
}
struct ccupdate_struct {
struct kmem_cache *cachep;
struct array_cache *new[NR_CPUS];
};
static void do_ccupdate_local(void *info)
{
struct ccupdate_struct *new = info;
struct array_cache *old;
check_irq_off();
old = cpu_cache_get(new->cachep);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
new->new[smp_processor_id()] = old;
}
/* Always called with the cache_chain_mutex held */
static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
int batchcount, int shared)
{
struct ccupdate_struct new;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
int i, err;
memset(&new.new, 0, sizeof(new.new));
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
for_each_online_cpu(i) {
new.new[i] = alloc_arraycache(cpu_to_node(i), limit,
batchcount);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
if (!new.new[i]) {
for (i--; i >= 0; i--)
kfree(new.new[i]);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
return -ENOMEM;
}
}
new.cachep = cachep;
on_each_cpu(do_ccupdate_local, (void *)&new, 1, 1);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
check_irq_on();
cachep->batchcount = batchcount;
cachep->limit = limit;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
cachep->shared = shared;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
for_each_online_cpu(i) {
struct array_cache *ccold = new.new[i];
if (!ccold)
continue;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
kfree(ccold);
}
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
err = alloc_kmemlist(cachep);
if (err) {
printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n",
cachep->name, -err);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
BUG();
}
return 0;
}
/* Called with cache_chain_mutex held always */
static void enable_cpucache(struct kmem_cache *cachep)
{
int err;
int limit, shared;
/*
* The head array serves three purposes:
* - create a LIFO ordering, i.e. return objects that are cache-warm
* - reduce the number of spinlock operations.
* - reduce the number of linked list operations on the slab and
* bufctl chains: array operations are cheaper.
* The numbers are guessed, we should auto-tune as described by
* Bonwick.
*/
if (cachep->buffer_size > 131072)
limit = 1;
else if (cachep->buffer_size > PAGE_SIZE)
limit = 8;
else if (cachep->buffer_size > 1024)
limit = 24;
else if (cachep->buffer_size > 256)
limit = 54;
else
limit = 120;
/*
* CPU bound tasks (e.g. network routing) can exhibit cpu bound
* allocation behaviour: Most allocs on one cpu, most free operations
* on another cpu. For these cases, an efficient object passing between
* cpus is necessary. This is provided by a shared array. The array
* replaces Bonwick's magazine layer.
* On uniprocessor, it's functionally equivalent (but less efficient)
* to a larger limit. Thus disabled by default.
*/
shared = 0;
#ifdef CONFIG_SMP
if (cachep->buffer_size <= PAGE_SIZE)
shared = 8;
#endif
#if DEBUG
/*
* With debugging enabled, large batchcount lead to excessively long
* periods with disabled local interrupts. Limit the batchcount
*/
if (limit > 32)
limit = 32;
#endif
err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
if (err)
printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
cachep->name, -err);
}
/*
* Drain an array if it contains any elements taking the l3 lock only if
* necessary. Note that the l3 listlock also protects the array_cache
* if drain_array() is used on the shared array.
*/
void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
struct array_cache *ac, int force, int node)
{
int tofree;
if (!ac || !ac->avail)
return;
if (ac->touched && !force) {
ac->touched = 0;
} else {
spin_lock_irq(&l3->list_lock);
if (ac->avail) {
tofree = force ? ac->avail : (ac->limit + 4) / 5;
if (tofree > ac->avail)
tofree = (ac->avail + 1) / 2;
free_block(cachep, ac->entry, tofree, node);
ac->avail -= tofree;
memmove(ac->entry, &(ac->entry[tofree]),
sizeof(void *) * ac->avail);
}
spin_unlock_irq(&l3->list_lock);
}
}
/**
* cache_reap - Reclaim memory from caches.
* @unused: unused parameter
*
* Called from workqueue/eventd every few seconds.
* Purpose:
* - clear the per-cpu caches for this CPU.
* - return freeable pages to the main free memory pool.
*
* If we cannot acquire the cache chain mutex then just give up - we'll try
* again on the next iteration.
*/
static void cache_reap(void *unused)
{
struct list_head *walk;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
struct kmem_list3 *l3;
int node = numa_node_id();
if (!mutex_trylock(&cache_chain_mutex)) {
/* Give up. Setup the next iteration. */
schedule_delayed_work(&__get_cpu_var(reap_work),
REAPTIMEOUT_CPUC);
return;
}
list_for_each(walk, &cache_chain) {
struct kmem_cache *searchp;
struct list_head *p;
int tofree;
struct slab *slabp;
searchp = list_entry(walk, struct kmem_cache, next);
check_irq_on();
/*
* We only take the l3 lock if absolutely necessary and we
* have established with reasonable certainty that
* we can do some work if the lock was obtained.
*/
l3 = searchp->nodelists[node];
[PATCH] slab: Node rotor for freeing alien caches and remote per cpu pages. The cache reaper currently tries to free all alien caches and all remote per cpu pages in each pass of cache_reap. For a machines with large number of nodes (such as Altix) this may lead to sporadic delays of around ~10ms. Interrupts are disabled while reclaiming creating unacceptable delays. This patch changes that behavior by adding a per cpu reap_node variable. Instead of attempting to free all caches, we free only one alien cache and the per cpu pages from one remote node. That reduces the time spend in cache_reap. However, doing so will lengthen the time it takes to completely drain all remote per cpu pagesets and all alien caches. The time needed will grow with the number of nodes in the system. All caches are drained when they overflow their respective capacity. So the drawback here is only that a bit of memory may be wasted for awhile longer. Details: 1. Rename drain_remote_pages to drain_node_pages to allow the specification of the node to drain of pcp pages. 2. Add additional functions init_reap_node, next_reap_node for NUMA that manage a per cpu reap_node counter. 3. Add a reap_alien function that reaps only from the current reap_node. For us this seems to be a critical issue. Holdoffs of an average of ~7ms cause some HPC benchmarks to slow down significantly. F.e. NAS parallel slows down dramatically. NAS parallel has a 12-16 seconds runtime w/o rotor compared to 5.8 secs with the rotor patches. It gets down to 5.05 secs with the additional interrupt holdoff reductions. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-03-10 01:33:54 +00:00
reap_alien(searchp, l3);
drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
/*
* These are racy checks but it does not matter
* if we skip one check or scan twice.
*/
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
if (time_after(l3->next_reap, jiffies))
goto next;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
drain_array(searchp, l3, l3->shared, 0, node);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
if (l3->free_touched) {
l3->free_touched = 0;
goto next;
}
tofree = (l3->free_limit + 5 * searchp->num - 1) /
(5 * searchp->num);
do {
/*
* Do not lock if there are no free blocks.
*/
if (list_empty(&l3->slabs_free))
break;
spin_lock_irq(&l3->list_lock);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
p = l3->slabs_free.next;
if (p == &(l3->slabs_free)) {
spin_unlock_irq(&l3->list_lock);
break;
}
slabp = list_entry(p, struct slab, list);
BUG_ON(slabp->inuse);
list_del(&slabp->list);
STATS_INC_REAPED(searchp);
/*
* Safe to drop the lock. The slab is no longer linked
* to the cache. searchp cannot disappear, we hold
* cache_chain_lock
*/
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
l3->free_objects -= searchp->num;
spin_unlock_irq(&l3->list_lock);
slab_destroy(searchp, slabp);
} while (--tofree > 0);
next:
cond_resched();
}
check_irq_on();
mutex_unlock(&cache_chain_mutex);
[PATCH] slab: Node rotor for freeing alien caches and remote per cpu pages. The cache reaper currently tries to free all alien caches and all remote per cpu pages in each pass of cache_reap. For a machines with large number of nodes (such as Altix) this may lead to sporadic delays of around ~10ms. Interrupts are disabled while reclaiming creating unacceptable delays. This patch changes that behavior by adding a per cpu reap_node variable. Instead of attempting to free all caches, we free only one alien cache and the per cpu pages from one remote node. That reduces the time spend in cache_reap. However, doing so will lengthen the time it takes to completely drain all remote per cpu pagesets and all alien caches. The time needed will grow with the number of nodes in the system. All caches are drained when they overflow their respective capacity. So the drawback here is only that a bit of memory may be wasted for awhile longer. Details: 1. Rename drain_remote_pages to drain_node_pages to allow the specification of the node to drain of pcp pages. 2. Add additional functions init_reap_node, next_reap_node for NUMA that manage a per cpu reap_node counter. 3. Add a reap_alien function that reaps only from the current reap_node. For us this seems to be a critical issue. Holdoffs of an average of ~7ms cause some HPC benchmarks to slow down significantly. F.e. NAS parallel slows down dramatically. NAS parallel has a 12-16 seconds runtime w/o rotor compared to 5.8 secs with the rotor patches. It gets down to 5.05 secs with the additional interrupt holdoff reductions. Signed-off-by: Christoph Lameter <clameter@sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-03-10 01:33:54 +00:00
next_reap_node();
/* Set up the next iteration */
schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
}
#ifdef CONFIG_PROC_FS
static void print_slabinfo_header(struct seq_file *m)
{
/*
* Output format version, so at least we can change it
* without _too_ many complaints.
*/
#if STATS
seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
#else
seq_puts(m, "slabinfo - version: 2.1\n");
#endif
seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
"<objperslab> <pagesperslab>");
seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
#if STATS
seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
"<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
#endif
seq_putc(m, '\n');
}
static void *s_start(struct seq_file *m, loff_t *pos)
{
loff_t n = *pos;
struct list_head *p;
mutex_lock(&cache_chain_mutex);
if (!n)
print_slabinfo_header(m);
p = cache_chain.next;
while (n--) {
p = p->next;
if (p == &cache_chain)
return NULL;
}
return list_entry(p, struct kmem_cache, next);
}
static void *s_next(struct seq_file *m, void *p, loff_t *pos)
{
struct kmem_cache *cachep = p;
++*pos;
return cachep->next.next == &cache_chain ?
NULL : list_entry(cachep->next.next, struct kmem_cache, next);
}
static void s_stop(struct seq_file *m, void *p)
{
mutex_unlock(&cache_chain_mutex);
}
static int s_show(struct seq_file *m, void *p)
{
struct kmem_cache *cachep = p;
struct list_head *q;
struct slab *slabp;
unsigned long active_objs;
unsigned long num_objs;
unsigned long active_slabs = 0;
unsigned long num_slabs, free_objects = 0, shared_avail = 0;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
const char *name;
char *error = NULL;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
int node;
struct kmem_list3 *l3;
active_objs = 0;
num_slabs = 0;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
for_each_online_node(node) {
l3 = cachep->nodelists[node];
if (!l3)
continue;
check_irq_on();
spin_lock_irq(&l3->list_lock);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
list_for_each(q, &l3->slabs_full) {
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
slabp = list_entry(q, struct slab, list);
if (slabp->inuse != cachep->num && !error)
error = "slabs_full accounting error";
active_objs += cachep->num;
active_slabs++;
}
list_for_each(q, &l3->slabs_partial) {
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
slabp = list_entry(q, struct slab, list);
if (slabp->inuse == cachep->num && !error)
error = "slabs_partial inuse accounting error";
if (!slabp->inuse && !error)
error = "slabs_partial/inuse accounting error";
active_objs += slabp->inuse;
active_slabs++;
}
list_for_each(q, &l3->slabs_free) {
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
slabp = list_entry(q, struct slab, list);
if (slabp->inuse && !error)
error = "slabs_free/inuse accounting error";
num_slabs++;
}
free_objects += l3->free_objects;
[PATCH] NUMA slab locking fixes: fix cpu down and up locking This fixes locking and bugs in cpu_down and cpu_up paths of the NUMA slab allocator. Sonny Rao <sonny@burdell.org> reported problems sometime back on POWER5 boxes, when the last cpu on the nodes were being offlined. We could not reproduce the same on x86_64 because the cpumask (node_to_cpumask) was not being updated on cpu down. Since that issue is now fixed, we can reproduce Sonny's problems on x86_64 NUMA, and here is the fix. The problem earlier was on CPU_DOWN, if it was the last cpu on the node to go down, the array_caches (shared, alien) and the kmem_list3 of the node were being freed (kfree) with the kmem_list3 lock held. If the l3 or the array_caches were to come from the same cache being cleared, we hit on badness. This patch cleans up the locking in cpu_up and cpu_down path. We cannot really free l3 on cpu down because, there is no node offlining yet and even though a cpu is not yet up, node local memory can be allocated for it. So l3s are usually allocated at keme_cache_create and destroyed at kmem_cache_destroy. Hence, we don't need cachep->spinlock protection to get to the cachep->nodelist[nodeid] either. Patch survived onlining and offlining on a 4 core 2 node Tyan box with a 4 dbench process running all the time. Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Ravikiran Thirumalai <kiran@scalex86.org> Cc: Christoph Lameter <christoph@lameter.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-02-05 07:27:59 +00:00
if (l3->shared)
shared_avail += l3->shared->avail;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
spin_unlock_irq(&l3->list_lock);
}
num_slabs += active_slabs;
num_objs = num_slabs * cachep->num;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
if (num_objs - active_objs != free_objects && !error)
error = "free_objects accounting error";
name = cachep->name;
if (error)
printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
name, active_objs, num_objs, cachep->buffer_size,
cachep->num, (1 << cachep->gfporder));
seq_printf(m, " : tunables %4u %4u %4u",
cachep->limit, cachep->batchcount, cachep->shared);
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
seq_printf(m, " : slabdata %6lu %6lu %6lu",
active_slabs, num_slabs, shared_avail);
#if STATS
{ /* list3 stats */
unsigned long high = cachep->high_mark;
unsigned long allocs = cachep->num_allocations;
unsigned long grown = cachep->grown;
unsigned long reaped = cachep->reaped;
unsigned long errors = cachep->errors;
unsigned long max_freeable = cachep->max_freeable;
unsigned long node_allocs = cachep->node_allocs;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
unsigned long node_frees = cachep->node_frees;
unsigned long overflows = cachep->node_overflow;
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
%4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
reaped, errors, max_freeable, node_allocs,
node_frees, overflows);
}
/* cpu stats */
{
unsigned long allochit = atomic_read(&cachep->allochit);
unsigned long allocmiss = atomic_read(&cachep->allocmiss);
unsigned long freehit = atomic_read(&cachep->freehit);
unsigned long freemiss = atomic_read(&cachep->freemiss);
seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
allochit, allocmiss, freehit, freemiss);
}
#endif
seq_putc(m, '\n');
return 0;
}
/*
* slabinfo_op - iterator that generates /proc/slabinfo
*
* Output layout:
* cache-name
* num-active-objs
* total-objs
* object size
* num-active-slabs
* total-slabs
* num-pages-per-slab
* + further values on SMP and with statistics enabled
*/
struct seq_operations slabinfo_op = {
.start = s_start,
.next = s_next,
.stop = s_stop,
.show = s_show,
};
#define MAX_SLABINFO_WRITE 128
/**
* slabinfo_write - Tuning for the slab allocator
* @file: unused
* @buffer: user buffer
* @count: data length
* @ppos: unused
*/
ssize_t slabinfo_write(struct file *file, const char __user * buffer,
size_t count, loff_t *ppos)
{
char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
int limit, batchcount, shared, res;
struct list_head *p;
if (count > MAX_SLABINFO_WRITE)
return -EINVAL;
if (copy_from_user(&kbuf, buffer, count))
return -EFAULT;
kbuf[MAX_SLABINFO_WRITE] = '\0';
tmp = strchr(kbuf, ' ');
if (!tmp)
return -EINVAL;
*tmp = '\0';
tmp++;
if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
return -EINVAL;
/* Find the cache in the chain of caches. */
mutex_lock(&cache_chain_mutex);
res = -EINVAL;
list_for_each(p, &cache_chain) {
struct kmem_cache *cachep;
cachep = list_entry(p, struct kmem_cache, next);
if (!strcmp(cachep->name, kbuf)) {
if (limit < 1 || batchcount < 1 ||
batchcount > limit || shared < 0) {
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
res = 0;
} else {
[PATCH] Numa-aware slab allocator V5 The NUMA API change that introduced kmalloc_node was accepted for 2.6.12-rc3. Now it is possible to do slab allocations on a node to localize memory structures. This API was used by the pageset localization patch and the block layer localization patch now in mm. The existing kmalloc_node is slow since it simply searches through all pages of the slab to find a page that is on the node requested. The two patches do a one time allocation of slab structures at initialization and therefore the speed of kmalloc node does not matter. This patch allows kmalloc_node to be as fast as kmalloc by introducing node specific page lists for partial, free and full slabs. Slab allocation improves in a NUMA system so that we are seeing a performance gain in AIM7 of about 5% with this patch alone. More NUMA localizations are possible if kmalloc_node operates in an fast way like kmalloc. Test run on a 32p systems with 32G Ram. w/o patch Tasks jobs/min jti jobs/min/task real cpu 1 485.36 100 485.3640 11.99 1.91 Sat Apr 30 14:01:51 2005 100 26582.63 88 265.8263 21.89 144.96 Sat Apr 30 14:02:14 2005 200 29866.83 81 149.3342 38.97 286.08 Sat Apr 30 14:02:53 2005 300 33127.16 78 110.4239 52.71 426.54 Sat Apr 30 14:03:46 2005 400 34889.47 80 87.2237 66.72 568.90 Sat Apr 30 14:04:53 2005 500 35654.34 76 71.3087 81.62 714.55 Sat Apr 30 14:06:15 2005 600 36460.83 75 60.7681 95.77 853.42 Sat Apr 30 14:07:51 2005 700 35957.00 75 51.3671 113.30 990.67 Sat Apr 30 14:09:45 2005 800 33380.65 73 41.7258 139.48 1140.86 Sat Apr 30 14:12:05 2005 900 35095.01 76 38.9945 149.25 1281.30 Sat Apr 30 14:14:35 2005 1000 36094.37 74 36.0944 161.24 1419.66 Sat Apr 30 14:17:17 2005 w/patch Tasks jobs/min jti jobs/min/task real cpu 1 484.27 100 484.2736 12.02 1.93 Sat Apr 30 15:59:45 2005 100 28262.03 90 282.6203 20.59 143.57 Sat Apr 30 16:00:06 2005 200 32246.45 82 161.2322 36.10 282.89 Sat Apr 30 16:00:42 2005 300 37945.80 83 126.4860 46.01 418.75 Sat Apr 30 16:01:28 2005 400 40000.69 81 100.0017 58.20 561.48 Sat Apr 30 16:02:27 2005 500 40976.10 78 81.9522 71.02 696.95 Sat Apr 30 16:03:38 2005 600 41121.54 78 68.5359 84.92 834.86 Sat Apr 30 16:05:04 2005 700 44052.77 78 62.9325 92.48 971.53 Sat Apr 30 16:06:37 2005 800 41066.89 79 51.3336 113.38 1111.15 Sat Apr 30 16:08:31 2005 900 38918.77 79 43.2431 134.59 1252.57 Sat Apr 30 16:10:46 2005 1000 41842.21 76 41.8422 139.09 1392.33 Sat Apr 30 16:13:05 2005 These are measurement taken directly after boot and show a greater improvement than 5%. However, the performance improvements become less over time if the AIM7 runs are repeated and settle down at around 5%. Links to earlier discussions: http://marc.theaimsgroup.com/?t=111094594500003&r=1&w=2 http://marc.theaimsgroup.com/?t=111603406600002&r=1&w=2 Changelog V4-V5: - alloc_arraycache and alloc_aliencache take node parameter instead of cpu - fix initialization so that nodes without cpus are properly handled. - simplify code in kmem_cache_init - patch against Andrews temp mm3 release - Add Shai to credits - fallback to __cache_alloc from __cache_alloc_node if the node's cache is not available yet. Changelog V3-V4: - Patch against 2.6.12-rc5-mm1 - Cleanup patch integrated - More and better use of for_each_node and for_each_cpu - GCC 2.95 fix (do not use [] use [0]) - Correct determination of INDEX_AC - Remove hack to cause an error on platforms that have no CONFIG_NUMA but nodes. - Remove list3_data and list3_data_ptr macros for better readability Changelog V2-V3: - Made to patch against 2.6.12-rc4-mm1 - Revised bootstrap mechanism so that larger size kmem_list3 structs can be supported. Do a generic solution so that the right slab can be found for the internal structs. - use for_each_online_node Changelog V1-V2: - Batching for freeing of wrong-node objects (alien caches) - Locking changes and NUMA #ifdefs as requested by Manfred Signed-off-by: Alok N Kataria <alokk@calsoftinc.com> Signed-off-by: Shobhit Dayal <shobhit@calsoftinc.com> Signed-off-by: Shai Fultheim <Shai@Scalex86.org> Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-09-09 20:03:32 +00:00
res = do_tune_cpucache(cachep, limit,
batchcount, shared);
}
break;
}
}
mutex_unlock(&cache_chain_mutex);
if (res >= 0)
res = count;
return res;
}
[PATCH] slab: implement /proc/slab_allocators Implement /proc/slab_allocators. It produces output like: idr_layer_cache: 80 idr_pre_get+0x33/0x4e buffer_head: 2555 alloc_buffer_head+0x20/0x75 mm_struct: 9 mm_alloc+0x1e/0x42 mm_struct: 20 dup_mm+0x36/0x370 vm_area_struct: 384 dup_mm+0x18f/0x370 vm_area_struct: 151 do_mmap_pgoff+0x2e0/0x7c3 vm_area_struct: 1 split_vma+0x5a/0x10e vm_area_struct: 11 do_brk+0x206/0x2e2 vm_area_struct: 2 copy_vma+0xda/0x142 vm_area_struct: 9 setup_arg_pages+0x99/0x214 fs_cache: 8 copy_fs_struct+0x21/0x133 fs_cache: 29 copy_process+0xf38/0x10e3 files_cache: 30 alloc_files+0x1b/0xcf signal_cache: 81 copy_process+0xbaa/0x10e3 sighand_cache: 77 copy_process+0xe65/0x10e3 sighand_cache: 1 de_thread+0x4d/0x5f8 anon_vma: 241 anon_vma_prepare+0xd9/0xf3 size-2048: 1 add_sect_attrs+0x5f/0x145 size-2048: 2 journal_init_revoke+0x99/0x302 size-2048: 2 journal_init_revoke+0x137/0x302 size-2048: 2 journal_init_inode+0xf9/0x1c4 Cc: Manfred Spraul <manfred@colorfullife.com> Cc: Alexander Nyberg <alexn@telia.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: Ravikiran Thirumalai <kiran@scalex86.org> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk> DESC slab-leaks3-locking-fix EDESC From: Andrew Morton <akpm@osdl.org> Update for slab-remove-cachep-spinlock.patch Cc: Al Viro <viro@ftp.linux.org.uk> Cc: Manfred Spraul <manfred@colorfullife.com> Cc: Alexander Nyberg <alexn@telia.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Cc: Christoph Lameter <clameter@engr.sgi.com> Cc: Ravikiran Thirumalai <kiran@scalex86.org> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-03-25 11:06:39 +00:00
#ifdef CONFIG_DEBUG_SLAB_LEAK
static void *leaks_start(struct seq_file *m, loff_t *pos)
{
loff_t n = *pos;
struct list_head *p;
mutex_lock(&cache_chain_mutex);
p = cache_chain.next;
while (n--) {
p = p->next;
if (p == &cache_chain)
return NULL;
}
return list_entry(p, struct kmem_cache, next);
}
static inline int add_caller(unsigned long *n, unsigned long v)
{
unsigned long *p;
int l;
if (!v)
return 1;
l = n[1];
p = n + 2;
while (l) {
int i = l/2;
unsigned long *q = p + 2 * i;
if (*q == v) {
q[1]++;
return 1;
}
if (*q > v) {
l = i;
} else {
p = q + 2;
l -= i + 1;
}
}
if (++n[1] == n[0])
return 0;
memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
p[0] = v;
p[1] = 1;
return 1;
}
static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
{
void *p;
int i;
if (n[0] == n[1])
return;
for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
continue;
if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
return;
}
}
static void show_symbol(struct seq_file *m, unsigned long address)
{
#ifdef CONFIG_KALLSYMS
char *modname;
const char *name;
unsigned long offset, size;
char namebuf[KSYM_NAME_LEN+1];
name = kallsyms_lookup(address, &size, &offset, &modname, namebuf);
if (name) {
seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
if (modname)
seq_printf(m, " [%s]", modname);
return;
}
#endif
seq_printf(m, "%p", (void *)address);
}
static int leaks_show(struct seq_file *m, void *p)
{
struct kmem_cache *cachep = p;
struct list_head *q;
struct slab *slabp;
struct kmem_list3 *l3;
const char *name;
unsigned long *n = m->private;
int node;
int i;
if (!(cachep->flags & SLAB_STORE_USER))
return 0;
if (!(cachep->flags & SLAB_RED_ZONE))
return 0;
/* OK, we can do it */
n[1] = 0;
for_each_online_node(node) {
l3 = cachep->nodelists[node];
if (!l3)
continue;
check_irq_on();
spin_lock_irq(&l3->list_lock);
list_for_each(q, &l3->slabs_full) {
slabp = list_entry(q, struct slab, list);
handle_slab(n, cachep, slabp);
}
list_for_each(q, &l3->slabs_partial) {
slabp = list_entry(q, struct slab, list);
handle_slab(n, cachep, slabp);
}
spin_unlock_irq(&l3->list_lock);
}
name = cachep->name;
if (n[0] == n[1]) {
/* Increase the buffer size */
mutex_unlock(&cache_chain_mutex);
m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
if (!m->private) {
/* Too bad, we are really out */
m->private = n;
mutex_lock(&cache_chain_mutex);
return -ENOMEM;
}
*(unsigned long *)m->private = n[0] * 2;
kfree(n);
mutex_lock(&cache_chain_mutex);
/* Now make sure this entry will be retried */
m->count = m->size;
return 0;
}
for (i = 0; i < n[1]; i++) {
seq_printf(m, "%s: %lu ", name, n[2*i+3]);
show_symbol(m, n[2*i+2]);
seq_putc(m, '\n');
}
return 0;
}
struct seq_operations slabstats_op = {
.start = leaks_start,
.next = s_next,
.stop = s_stop,
.show = leaks_show,
};
#endif
#endif
/**
* ksize - get the actual amount of memory allocated for a given object
* @objp: Pointer to the object
*
* kmalloc may internally round up allocations and return more memory
* than requested. ksize() can be used to determine the actual amount of
* memory allocated. The caller may use this additional memory, even though
* a smaller amount of memory was initially specified with the kmalloc call.
* The caller must guarantee that objp points to a valid object previously
* allocated with either kmalloc() or kmem_cache_alloc(). The object
* must not be freed during the duration of the call.
*/
unsigned int ksize(const void *objp)
{
if (unlikely(objp == NULL))
return 0;
return obj_size(virt_to_cache(objp));
}