Xiaohui Xin and some other folks at Intel have been looking into what's
behind the performance hit of paravirt_ops when running native.
It appears that the hit is entirely due to the paravirtualized
spinlocks introduced by:
| commit 8efcbab674
| Date: Mon Jul 7 12:07:51 2008 -0700
|
| paravirt: introduce a "lock-byte" spinlock implementation
The extra call/return in the spinlock path is somehow
causing an increase in the cycles/instruction of somewhere around 2-7%
(seems to vary quite a lot from test to test). The working theory is
that the CPU's pipeline is getting upset about the
call->call->locked-op->return->return, and seems to be failing to
speculate (though I haven't seen anything definitive about the precise
reasons). This doesn't entirely make sense, because the performance
hit is also visible on unlock and other operations which don't involve
locked instructions. But spinlock operations clearly swamp all the
other pvops operations, even though I can't imagine that they're
nearly as common (there's only a .05% increase in instructions
executed).
If I disable just the pv-spinlock calls, my tests show that pvops is
identical to non-pvops performance on native (my measurements show that
it is actually about .1% faster, but Xiaohui shows a .05% slowdown).
Summary of results, averaging 10 runs of the "mmperf" test, using a
no-pvops build as baseline:
nopv Pv-nospin Pv-spin
CPU cycles 100.00% 99.89% 102.18%
instructions 100.00% 100.10% 100.15%
CPI 100.00% 99.79% 102.03%
cache ref 100.00% 100.84% 100.28%
cache miss 100.00% 90.47% 88.56%
cache miss rate 100.00% 89.72% 88.31%
branches 100.00% 99.93% 100.04%
branch miss 100.00% 103.66% 107.72%
branch miss rt 100.00% 103.73% 107.67%
wallclock 100.00% 99.90% 102.20%
The clear effect here is that the 2% increase in CPI is
directly reflected in the final wallclock time.
(The other interesting effect is that the more ops are
out of line calls via pvops, the lower the cache access
and miss rates. Not too surprising, but it suggests that
the non-pvops kernel is over-inlined. On the flipside,
the branch misses go up correspondingly...)
So, what's the fix?
Paravirt patching turns all the pvops calls into direct calls, so
_spin_lock etc do end up having direct calls. For example, the compiler
generated code for paravirtualized _spin_lock is:
<_spin_lock+0>: mov %gs:0xb4c8,%rax
<_spin_lock+9>: incl 0xffffffffffffe044(%rax)
<_spin_lock+15>: callq *0xffffffff805a5b30
<_spin_lock+22>: retq
The indirect call will get patched to:
<_spin_lock+0>: mov %gs:0xb4c8,%rax
<_spin_lock+9>: incl 0xffffffffffffe044(%rax)
<_spin_lock+15>: callq <__ticket_spin_lock>
<_spin_lock+20>: nop; nop /* or whatever 2-byte nop */
<_spin_lock+22>: retq
One possibility is to inline _spin_lock, etc, when building an
optimised kernel (ie, when there's no spinlock/preempt
instrumentation/debugging enabled). That will remove the outer
call/return pair, returning the instruction stream to a single
call/return, which will presumably execute the same as the non-pvops
case. The downsides arel 1) it will replicate the
preempt_disable/enable code at eack lock/unlock callsite; this code is
fairly small, but not nothing; and 2) the spinlock definitions are
already a very heavily tangled mass of #ifdefs and other preprocessor
magic, and making any changes will be non-trivial.
The other obvious answer is to disable pv-spinlocks. Making them a
separate config option is fairly easy, and it would be trivial to
enable them only when Xen is enabled (as the only non-default user).
But it doesn't really address the common case of a distro build which
is going to have Xen support enabled, and leaves the open question of
whether the native performance cost of pv-spinlocks is worth the
performance improvement on a loaded Xen system (10% saving of overall
system CPU when guests block rather than spin). Still it is a
reasonable short-term workaround.
[ Impact: fix pvops performance regression when running native ]
Analysed-by: "Xin Xiaohui" <xiaohui.xin@intel.com>
Analysed-by: "Li Xin" <xin.li@intel.com>
Analysed-by: "Nakajima Jun" <jun.nakajima@intel.com>
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Acked-by: H. Peter Anvin <hpa@zytor.com>
Cc: Nick Piggin <npiggin@suse.de>
Cc: Xen-devel <xen-devel@lists.xensource.com>
LKML-Reference: <4A0B62F7.5030802@goop.org>
[ fixed the help text ]
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: cleanup
Make x86_quirks support more transparent. The highlevel
methods are now named:
extern void x86_quirk_pre_intr_init(void);
extern void x86_quirk_intr_init(void);
extern void x86_quirk_trap_init(void);
extern void x86_quirk_pre_time_init(void);
extern void x86_quirk_time_init(void);
This makes it clear that if some platform extension has to
do something here that it is considered ... weird, and is
discouraged.
Also remove arch_hooks.h and move it into setup.h (and other
header files where appropriate).
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: Catch cases where lazy MMU state is active in a preemtible context
arch_flush_lazy_mmu_cpu() has been changed to disable preemption so
the checks in enter/leave will never trigger. Put the preemtible()
check into arch_flush_lazy_mmu_cpu() to catch such cases.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Impact: avoid access to percpu vars in preempible context
They are intended to be used whenever there's the possibility
that there's some stale state which is going to be overwritten
with a queued update, or to force a state change when we may be
in lazy mode. Either way, we could end up calling it with
preemption enabled, so wrap the functions in their own little
preempt-disable section so they can be safely called in any
context (though preemption should never be enabled if we're actually
in a lazy state).
(Move out of line to avoid #include dependencies.)
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Impact: Optimization
In the native case, pte_val, make_pte, etc are all just identity
functions, so there's no need to clobber a lot of registers over them.
(This changes the 32-bit callee-save calling convention to return both
EAX and EDX so functions can return 64-bit values.)
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
Impact: Optimization
One of the problems with inserting a pile of C calls where previously
there were none is that the register pressure is greatly increased.
The C calling convention says that the caller must expect a certain
set of registers may be trashed by the callee, and that the callee can
use those registers without restriction. This includes the function
argument registers, and several others.
This patch seeks to alleviate this pressure by introducing wrapper
thunks that will do the register saving/restoring, so that the
callsite doesn't need to worry about it, but the callee function can
be conventional compiler-generated code. In many cases (particularly
performance-sensitive cases) the callee will be in assembler anyway,
and need not use the compiler's calling convention.
Standard calling convention is:
arguments return scratch
x86-32 eax edx ecx eax ?
x86-64 rdi rsi rdx rcx rax r8 r9 r10 r11
The thunk preserves all argument and scratch registers. The return
register is not preserved, and is available as a scratch register for
unwrapped callee code (and of course the return value).
Wrapped function pointers are themselves wrapped in a struct
paravirt_callee_save structure, in order to get some warning from the
compiler when functions with mismatched calling conventions are used.
The most common paravirt ops, both statically and dynamically, are
interrupt enable/disable/save/restore, so handle them first. This is
particularly easy since their calls are handled specially anyway.
XXX Deal with VMI. What's their calling convention?
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
Impact: Optimization
Several paravirt ops implementations simply return their arguments,
the most obvious being the make_pte/pte_val class of operations on
native.
On 32-bit, the identity function is literally a no-op, as the calling
convention uses the same registers for the first argument and return.
On 64-bit, it can be implemented with a single "mov".
This patch adds special identity functions for 32 and 64 bit argument,
and machinery to recognize them and replace them with either nops or a
mov as appropriate.
At the moment, the only users for the identity functions are the
pagetable entry conversion functions.
The result is a measureable improvement on pagetable-heavy benchmarks
(2-3%, reducing the pvops overhead from 5 to 2%).
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
pte_flags() was introduced as a new pvop in order to extract just the
flags portion of a pte, which is a potentially cheaper operation than
extracting the page number as well. It turns out this operation is
not needed, because simply using a mask to extract the flags from a
pte is sufficient for all current users.
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
None of the spinlock API is exported GPL, so there's no reason for
pv_lock_ops to be.
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Cc: drago01 <drago01@gmail.com>
ftrace requires certain low-level code, like spinlocks and timestamps,
to be compiled without -pg in order to avoid infinite recursion. This
patch splits out the core paravirt spinlocks and the Xen spinlocks
into separate files which can be compiled without -pg.
Also do xen/time.c while we're about it. As a result, we can now use
ftrace within a Xen domain.
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
LTP testing showed that Xen does not properly implement
sys_modify_ldt(). This patch does the final little bits needed to
make the ldt work properly.
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
(Jeremy said:
rusty: use PTE_MASK
rusty: use PTE_MASK
rusty: use PTE_MASK
When I asked:
jsgf: does that include the NX flag?
He responded eloquently:
rusty: use PTE_MASK
rusty: use PTE_MASK
yes, it's the official constant of masking flags out of ptes
)
Change a15af1c9ea 'x86/paravirt: add
pte_flags to just get pte flags' removed lguest's private pte_flags()
in favor of a generic one.
Unfortunately, the generic one doesn't filter out the non-flags bits:
this results in lguest creating corrupt shadow page tables and blowing
up host memory.
Since noone is supposed to use the pfn part of pte_flags(), it seems
safest to always do the filtering.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Acked-by: Jeremy Fitzhardinge <jeremy@goop.org>
Signed-off-and-morning-tea-spilled-by: Ingo Molnar <mingo@elte.hu>
Use alternatives to select the workaround for the 11AP Pentium erratum
for the affected steppings on the fly rather than build time. Remove the
X86_GOOD_APIC configuration option and replace all the calls to
apic_write_around() with plain apic_write(), protecting accesses to the
ESR as appropriate due to the 3AP Pentium erratum. Remove
apic_read_around() and all its invocations altogether as not needed.
Remove apic_write_atomic() and all its implementing backends. The use of
ASM_OUTPUT2() is not strictly needed for input constraints, but I have
used it for readability's sake.
I had the feeling no one else was brave enough to do it, so I went ahead
and here it is. Verified by checking the generated assembly and tested
with both a 32-bit and a 64-bit configuration, also with the 11AP
"feature" forced on and verified with gdb on /proc/kcore to work as
expected (as an 11AP machines are quite hard to get hands on these days).
Some script complained about the use of "volatile", but apic_write() needs
it for the same reason and is effectively a replacement for writel(), so I
have disregarded it.
I am not sure what the policy wrt defconfig files is, they are generated
and there is risk of a conflict resulting from an unrelated change, so I
have left changes to them out. The option will get removed from them at
the next run.
Some testing with machines other than mine will be needed to avoid some
stupid mistake, but despite its volume, the change is not really that
intrusive, so I am fairly confident that because it works for me, it will
everywhere.
Signed-off-by: Maciej W. Rozycki <macro@linux-mips.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Implement a version of the old spinlock algorithm, in which everyone
spins waiting for a lock byte. In order to be compatible with the
ticket-lock's use of a zero initializer, this uses the convention of
'0' for unlocked and '1' for locked.
This algorithm is much better than ticket locks in a virtual
envionment, because it doesn't interact badly with the vcpu scheduler.
If there are multiple vcpus spinning on a lock and the lock is
released, the next vcpu to be scheduled will take the lock, rather
than cycling around until the next ticketed vcpu gets it.
To use this, you must call paravirt_use_bytelocks() very early, before
any spinlocks have been taken.
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <clameter@linux-foundation.org>
Cc: Petr Tesarik <ptesarik@suse.cz>
Cc: Virtualization <virtualization@lists.linux-foundation.org>
Cc: Xen devel <xen-devel@lists.xensource.com>
Cc: Thomas Friebel <thomas.friebel@amd.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Ticket spinlocks have absolutely ghastly worst-case performance
characteristics in a virtual environment. If there is any contention
for physical CPUs (ie, there are more runnable vcpus than cpus), then
ticket locks can cause the system to end up spending 90+% of its time
spinning.
The problem is that (v)cpus waiting on a ticket spinlock will be
granted access to the lock in strict order they got their tickets. If
the hypervisor scheduler doesn't give the vcpus time in that order,
they will burn timeslices waiting for the scheduler to give the right
vcpu some time. In the worst case it could take O(n^2) vcpu scheduler
timeslices for everyone waiting on the lock to get it, not counting
new cpus trying to take the lock while the log-jam is sorted out.
These hooks allow a paravirt backend to replace the spinlock
implementation.
At the very least, this could revert the implementation back to the
old lock algorithm, which allows the next scheduled vcpu to take the
lock, and has basically fairly good performance.
It also allows the spinlocks to take advantages of the hypervisor
features to make locks more efficient (spin and block, for example).
The cost to native execution is an extra direct call when using a
spinlock function. There's no overhead if CONFIG_PARAVIRT is turned
off.
The lock structure is fixed at a single "unsigned int", initialized to
zero, but the spinlock implementation can use it as it wishes.
Thanks to Thomas Friebel's Xen Summit talk "Preventing Guests from
Spinning Around" for pointing out this problem.
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Christoph Lameter <clameter@linux-foundation.org>
Cc: Petr Tesarik <ptesarik@suse.cz>
Cc: Virtualization <virtualization@lists.linux-foundation.org>
Cc: Xen devel <xen-devel@lists.xensource.com>
Cc: Thomas Friebel <thomas.friebel@amd.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Call paravirt_pagetable_setup_{start,done}
These paravirt_ops functions were not being called on x86_64.
Signed-off-by: Eduardo Habkost <ehabkost@redhat.com>
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Cc: Stephen Tweedie <sct@redhat.com>
Cc: Mark McLoughlin <markmc@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
fix for pv - clean up the namespace there too.
Signed-off-by: Yinghai Lu <yhlu.kernel@gmail.com>
Cc: Suresh Siddha <suresh.b.siddha@intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Introduce basic apic operations which handle the apic programming. This
will be used later to introduce another specific operations for x2apic.
For the perfomance critial accesses like IPI's, EOI etc, we use the
native operations as they are already referenced by different
indirections like genapic, irq_chip etc.
64bit Paravirt ops can also define their apic operations accordingly.
Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com>
Cc: akpm@linux-foundation.org
Cc: arjan@linux.intel.com
Cc: andi@firstfloor.org
Cc: ebiederm@xmission.com
Cc: jbarnes@virtuousgeek.org
Cc: steiner@sgi.com
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Rename the paravirtualized calculate_cpu_khz to calibrate_tsc.
In all cases, we actually calibrate_tsc and use that as the cpu_khz value.
Signed-off-by: Alok N Kataria <akataria@vmware.com>
Signed-off-by: Dan Hecht <dhecht@vmware.com>
Cc: Dan Hecht <dhecht@vmware.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
64-bit Xen pushes a couple of extra words onto an exception frame.
Add a hook to deal with them.
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Cc: xen-devel <xen-devel@lists.xensource.com>
Cc: Stephen Tweedie <sct@redhat.com>
Cc: Eduardo Habkost <ehabkost@redhat.com>
Cc: Mark McLoughlin <markmc@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
In a 64-bit system, we need separate sysret/sysexit operations to
return to a 32-bit userspace.
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citirx.com>
Cc: xen-devel <xen-devel@lists.xensource.com>
Cc: Stephen Tweedie <sct@redhat.com>
Cc: Eduardo Habkost <ehabkost@redhat.com>
Cc: Mark McLoughlin <markmc@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
There's no need to combine restoring the user rsp within the sysret
pvop, so split it out. This makes the pvop's semantics closer to the
machine instruction.
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citirx.com>
Cc: xen-devel <xen-devel@lists.xensource.com>
Cc: Stephen Tweedie <sct@redhat.com>
Cc: Eduardo Habkost <ehabkost@redhat.com>
Cc: Mark McLoughlin <markmc@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Don't conflate sysret and sysexit; they're different instructions with
different semantics, and may be in use at the same time (at least
within the same kernel, depending on whether its an Intel or AMD
system).
sysexit - just return to userspace, does no register restoration of
any kind; must explicitly atomically enable interrupts.
sysret - reloads flags from r11, so no need to explicitly enable
interrupts on 64-bit, responsible for restoring usermode %gs
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citirx.com>
Cc: xen-devel <xen-devel@lists.xensource.com>
Cc: Stephen Tweedie <sct@redhat.com>
Cc: Eduardo Habkost <ehabkost@redhat.com>
Cc: Mark McLoughlin <markmc@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Add hooks which are called at pgd_alloc/free time. The pgd_alloc hook
may return an error code, which if non-zero, causes the pgd allocation
to be failed. The hooks may be used to allocate/free auxillary
per-pgd information.
also fix:
> * Ingo Molnar <mingo@elte.hu> wrote:
>
> include/asm/pgalloc.h: In function ‘paravirt_pgd_free':
> include/asm/pgalloc.h:14: error: parameter name omitted
> arch/x86/kernel/entry_64.S: In file included from
> arch/x86/kernel/traps_64.c:51:include/asm/pgalloc.h: In function ‘paravirt_pgd_free':
> include/asm/pgalloc.h:14: error: parameter name omitted
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Cc: xen-devel <xen-devel@lists.xensource.com>
Cc: Stephen Tweedie <sct@redhat.com>
Cc: Eduardo Habkost <ehabkost@redhat.com>
Cc: Mark McLoughlin <markmc@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
This patch adds paravirt-ops hooks in pv_mmu_ops for ptep_modify_prot_start and
ptep_modify_prot_commit. This allows the hypervisor-specific backends to
implement these in some more efficient way.
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
Acked-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Add pte_flags() to extract the flags from a pte. This is a special
case of pte_val() which is only guaranteed to return the pte's flags
correctly; the page number may be corrupted or missing.
The intent is to allow paravirt implementations to return pte flags
without having to do any translation of the page number (most notably,
Xen).
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Rename (alloc|release)_(pt|pd) to pte/pmd to explicitly match the name
of the appropriate pagetable level structure.
[ x86.git merge work by Mark McLoughlin <markmc@redhat.com> ]
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Mark McLoughlin <markmc@redhat.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
The memory resource is also used for main memory, and we need it to
allocate physical addresses for memory hotplug. Knobbling io space is
enough to get the job done anyway.
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Cc: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
This finally makes paravirt-ops able to compile and boot under x86_64.
Signed-off-by: Eduardo Habkost <ehabkost@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
paravirt_pagetable_setup_{start,done}() are not used (yet) under x86_64,
and native_pagetable_setup_{start,done}() don't exist on x86_64. So they
don't need to be set.
Signed-off-by: Eduardo Habkost <ehabkost@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
This patch fills in the read and write cr8 fields with their
native version.
Signed-off-by: Glauber de Oliveira Costa <gcosta@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
