)
From: Ingo Molnar <mingo@elte.hu>
This is the latest version of the scheduler cache-hot-auto-tune patch.
The first problem was that detection time scaled with O(N^2), which is
unacceptable on larger SMP and NUMA systems. To solve this:
- I've added a 'domain distance' function, which is used to cache
measurement results. Each distance is only measured once. This means
that e.g. on NUMA distances of 0, 1 and 2 might be measured, on HT
distances 0 and 1, and on SMP distance 0 is measured. The code walks
the domain tree to determine the distance, so it automatically follows
whatever hierarchy an architecture sets up. This cuts down on the boot
time significantly and removes the O(N^2) limit. The only assumption
is that migration costs can be expressed as a function of domain
distance - this covers the overwhelming majority of existing systems,
and is a good guess even for more assymetric systems.
[ People hacking systems that have assymetries that break this
assumption (e.g. different CPU speeds) should experiment a bit with
the cpu_distance() function. Adding a ->migration_distance factor to
the domain structure would be one possible solution - but lets first
see the problem systems, if they exist at all. Lets not overdesign. ]
Another problem was that only a single cache-size was used for measuring
the cost of migration, and most architectures didnt set that variable
up. Furthermore, a single cache-size does not fit NUMA hierarchies with
L3 caches and does not fit HT setups, where different CPUs will often
have different 'effective cache sizes'. To solve this problem:
- Instead of relying on a single cache-size provided by the platform and
sticking to it, the code now auto-detects the 'effective migration
cost' between two measured CPUs, via iterating through a wide range of
cachesizes. The code searches for the maximum migration cost, which
occurs when the working set of the test-workload falls just below the
'effective cache size'. I.e. real-life optimized search is done for
the maximum migration cost, between two real CPUs.
This, amongst other things, has the positive effect hat if e.g. two
CPUs share a L2/L3 cache, a different (and accurate) migration cost
will be found than between two CPUs on the same system that dont share
any caches.
(The reliable measurement of migration costs is tricky - see the source
for details.)
Furthermore i've added various boot-time options to override/tune
migration behavior.
Firstly, there's a blanket override for autodetection:
migration_cost=1000,2000,3000
will override the depth 0/1/2 values with 1msec/2msec/3msec values.
Secondly, there's a global factor that can be used to increase (or
decrease) the autodetected values:
migration_factor=120
will increase the autodetected values by 20%. This option is useful to
tune things in a workload-dependent way - e.g. if a workload is
cache-insensitive then CPU utilization can be maximized by specifying
migration_factor=0.
I've tested the autodetection code quite extensively on x86, on 3
P3/Xeon/2MB, and the autodetected values look pretty good:
Dual Celeron (128K L2 cache):
---------------------
migration cost matrix (max_cache_size: 131072, cpu: 467 MHz):
---------------------
[00] [01]
[00]: - 1.7(1)
[01]: 1.7(1) -
---------------------
cacheflush times [2]: 0.0 (0) 1.7 (1784008)
---------------------
Here the slow memory subsystem dominates system performance, and even
though caches are small, the migration cost is 1.7 msecs.
Dual HT P4 (512K L2 cache):
---------------------
migration cost matrix (max_cache_size: 524288, cpu: 2379 MHz):
---------------------
[00] [01] [02] [03]
[00]: - 0.4(1) 0.0(0) 0.4(1)
[01]: 0.4(1) - 0.4(1) 0.0(0)
[02]: 0.0(0) 0.4(1) - 0.4(1)
[03]: 0.4(1) 0.0(0) 0.4(1) -
---------------------
cacheflush times [2]: 0.0 (33900) 0.4 (448514)
---------------------
Here it can be seen that there is no migration cost between two HT
siblings (CPU#0/2 and CPU#1/3 are separate physical CPUs). A fast memory
system makes inter-physical-CPU migration pretty cheap: 0.4 msecs.
8-way P3/Xeon [2MB L2 cache]:
---------------------
migration cost matrix (max_cache_size: 2097152, cpu: 700 MHz):
---------------------
[00] [01] [02] [03] [04] [05] [06] [07]
[00]: - 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1)
[01]: 19.2(1) - 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1)
[02]: 19.2(1) 19.2(1) - 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1)
[03]: 19.2(1) 19.2(1) 19.2(1) - 19.2(1) 19.2(1) 19.2(1) 19.2(1)
[04]: 19.2(1) 19.2(1) 19.2(1) 19.2(1) - 19.2(1) 19.2(1) 19.2(1)
[05]: 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) - 19.2(1) 19.2(1)
[06]: 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) - 19.2(1)
[07]: 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) -
---------------------
cacheflush times [2]: 0.0 (0) 19.2 (19281756)
---------------------
This one has huge caches and a relatively slow memory subsystem - so the
migration cost is 19 msecs.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Ashok Raj <ashok.raj@intel.com>
Signed-off-by: Ken Chen <kenneth.w.chen@intel.com>
Cc: <wilder@us.ibm.com>
Signed-off-by: John Hawkes <hawkes@sgi.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Add per-arch sched_cacheflush() which is a write-back cacheflush used by
the migration-cost calibration code at bootup time.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Gcc has a tradition of misscompiling the previous construct using the
address of a label as argument to inline assembler. Gas otoh has the
annoying difference between la and dla which are only usable for 32-bit
rsp. 64-bit code, so can't be used without conditional compilation.
The alterantive is switching the assembler to 64-bit code which happens
to work right even for 32-bit code ...
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
local_irq_restore uses di which saves the whole status content, not
just the IE bit resulting in local_irq_restore() to fail. This only
happens if both CONFIG_CPU_MIPSR2 and CONFIG_IRQ_CPU are enabled.
Signed-off-by: Maxime Bizon <mbizon@freebox.fr>
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
dump_regs() is used by a bunch of drivers for their internal stuff;
renamed mips instance (one that is seen in system-wide headers)
to elf_dump_regs()
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
USB OpenHCI host controller on Au1550 only decodes memory addresses from
0x14020000 to 0x1407FFFF according to the databook, which gives 0x60000
(on the prior Au1x00 chips the map size was 1MB).
Signed-off-by: Sergei Shtylyov <sshtylyov@ru.mvista.com>
Acked-by: Jordan Crouse <jordan.crouse@amd.com>
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
mdelay(1) (i.e. udelay(1000)) does not work correctly due to overflow.
1000 * 0x004189374BC6A7f0 = 0x10000000000000180 (>= 2**64)
0x004189374BC6A7ef (0x004189374BC6A7f0 - 1) is OK and it is exactly
same as catchall case (0x8000000000000000UL / (500000 / HZ)).
Signed-off-by: Atsushi Nemoto <anemo@mba.ocn.ne.jp>
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
used_dsp was meant to be used like used_math - but since the FPU context
is small and lazy context switching is a stupid idea on multiprocessors
this idea only got halfway implemented and those bits are were now
breaking ptrace.
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
add the per-arch mutex.h files for the remaining architectures.
We default to asm-generic/mutex-dec.h, because that performs
quite well on most arches. Arches that do not have atomic
decrement/increment instructions should switch to mutex-xchg.h
instead. Arches can also provide their own implementation for
the mutex fastpath primitives.
Signed-off-by: Arjan van de Ven <arjan@infradead.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
add atomic_xchg() to all the architectures. Needed by the new mutex code.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Arjan van de Ven <arjan@infradead.org>
Kill L1_CACHE_SHIFT from all arches. Since L1_CACHE_SHIFT_MAX is not used
anymore with the introduction of INTERNODE_CACHE, kill L1_CACHE_SHIFT_MAX.
Signed-off-by: Ravikiran Thirumalai <kiran@scalex86.org>
Signed-off-by: Shai Fultheim <shai@scalex86.org>
Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Several counters already have the need to use 64 atomic variables on 64 bit
platforms (see mm_counter_t in sched.h). We have to do ugly ifdefs to fall
back to 32 bit atomic on 32 bit platforms.
The VM statistics patch that I am working on will also make more extensive
use of atomic64.
This patch introduces a new type atomic_long_t by providing definitions in
asm-generic/atomic.h that works similar to the c "long" type. Its 32 bits
on 32 bit platforms and 64 bits on 64 bit platforms.
Also cleans up the determination of the mm_counter_t in sched.h.
Signed-off-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Here is the patch to implement madvise(MADV_REMOVE) - which frees up a
given range of pages & its associated backing store. Current
implementation supports only shmfs/tmpfs and other filesystems return
-ENOSYS.
"Some app allocates large tmpfs files, then when some task quits and some
client disconnect, some memory can be released. However the only way to
release tmpfs-swap is to MADV_REMOVE". - Andrea Arcangeli
Databases want to use this feature to drop a section of their bufferpool
(shared memory segments) - without writing back to disk/swap space.
This feature is also useful for supporting hot-plug memory on UML.
Concerns raised by Andrew Morton:
- "We have no plan for holepunching! If we _do_ have such a plan (or
might in the future) then what would the API look like? I think
sys_holepunch(fd, start, len), so we should start out with that."
- Using madvise is very weird, because people will ask "why do I need to
mmap my file before I can stick a hole in it?"
- None of the other madvise operations call into the filesystem in this
manner. A broad question is: is this capability an MM operation or a
filesytem operation? truncate, for example, is a filesystem operation
which sometimes has MM side-effects. madvise is an mm operation and with
this patch, it gains FS side-effects, only they're really, really
significant ones."
Comments:
- Andrea suggested the fs operation too but then it's more efficient to
have it as a mm operation with fs side effects, because they don't
immediatly know fd and physical offset of the range. It's possible to
fixup in userland and to use the fs operation but it's more expensive,
the vmas are already in the kernel and we can use them.
Short term plan & Future Direction:
- We seem to need this interface only for shmfs/tmpfs files in the short
term. We have to add hooks into the filesystem for correctness and
completeness. This is what this patch does.
- In the future, plan is to support both fs and mmap apis also. This
also involves (other) filesystem specific functions to be implemented.
- Current patch doesn't support VM_NONLINEAR - which can be addressed in
the future.
Signed-off-by: Badari Pulavarty <pbadari@us.ibm.com>
Cc: Hugh Dickins <hugh@veritas.com>
Cc: Andrea Arcangeli <andrea@suse.de>
Cc: Michael Kerrisk <mtk-manpages@gmx.net>
Cc: Ulrich Drepper <drepper@redhat.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Changes here include removing all of CONFIG_PM while it is being repeatedly
smacked with a lead pipe, moving the BURSTMODE param to a #define (it should
be defined almost always anyway), fixing the rqsize stuff, pulling ide_ioreg_t,
and general cleanups and whatnot.
Signed-off-by: Jordan Crouse <jordan.crouse@amd.com>
Signed-off-by: Bartlomiej Zolnierkiewicz <bzolnier@gmail.com>
There is no definition for seadint_init() and the unprotected prototype
breaks compilation of assembler files.
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
Introduce an atomic_inc_not_zero operation. Make this a special case of
atomic_add_unless because lockless pagecache actually wants
atomic_inc_not_negativeone due to its offset refcount.
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: "Paul E. McKenney" <paulmck@us.ibm.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Add missing bits to fix D-cache aliasing problem in the PIO IDE driver.
Signed-off-by: Atsushi Nemoto <anemo@mba.ocn.ne.jp>
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
If HZ was 1000, mdelay(2) cause overflow on multiplication in
__udelay. We should define MAX_UDELAY_MS properly to prevent this.
Signed-off-by: Atsushi Nemoto <anemo@mba.ocn.ne.jp>
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
Many RTC routines were not protected against each other, so there are
potential races, for example, ntp-update against /dev/rtc. This patch
fixes them using rtc_lock.
Signed-off-by: Atsushi Nemoto <anemo@mba.ocn.ne.jp>
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
The mips_rtc_lock is no longer needed because RTC operations should be
protected already by other mechanism. (rtc_lock, local_irq_save, etc.)
Also, locking whole rtc_get_time/rtc_set_time should be avoided while
some RTC routines might take very long time (a few seconds).
Signed-off-by: Atsushi Nemoto <anemo@mba.ocn.ne.jp>
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
