mirror of
https://github.com/torvalds/linux.git
synced 2024-11-21 19:41:42 +00:00
f5fe24ef17
On the x86-64 architecture even a failing cmpxchg grants exclusive access to the cacheline, making it preferable to retry the failed op immediately instead of stalling with the pause instruction. To illustrate the impact, below are benchmark results obtained by running various will-it-scale tests on top of the 6.2-rc3 kernel and Cascade Lake (2 sockets * 24 cores * 2 threads) CPU. All results in ops/s. Note there is some variance in re-runs, but the code is consistently faster when contention is present. open3 ("Same file open/close"): proc stock no-pause 1 805603 814942 (+%1) 2 1054980 1054781 (-0%) 8 1544802 1822858 (+18%) 24 1191064 2199665 (+84%) 48 851582 1469860 (+72%) 96 609481 1427170 (+134%) fstat2 ("Same file fstat"): proc stock no-pause 1 3013872 3047636 (+1%) 2 4284687 4400421 (+2%) 8 3257721 5530156 (+69%) 24 2239819 5466127 (+144%) 48 1701072 5256609 (+209%) 961269157
6649326 (+423%) Additionally, a kernel with a private patch to help access() scalability: access2 ("Same file access"): proc stock patched patched +nopause 24 2378041 2005501 5370335 (-15% / +125%) That is, fixing the problems in access itself *reduces* scalability after the cacheline ping-pong only happens in lockref with the pause instruction. Note that fstat and access benchmarks are not currently integrated into will-it-scale, but interested parties can find them in pull requests to said project. Code at hand has a rather tortured history. First modification showed up in commitd472d9d98b
("lockref: Relax in cmpxchg loop"), written with Itanium in mind. Later it got patched up to use an arch-dependent macro to stop doing it on s390 where it caused a significant regression. Said macro had undergone revisions and was ultimately eliminated later, going back to cpu_relax. While I intended to only remove cpu_relax for x86-64, I got the following comment from Linus: I would actually prefer just removing it entirely and see if somebody else hollers. You have the numbers to prove it hurts on real hardware, and I don't think we have any numbers to the contrary. So I think it's better to trust the numbers and remove it as a failure, than say "let's just remove it on x86-64 and leave everybody else with the potentially broken code" Additionally, Will Deacon (maintainer of the arm64 port, one of the architectures previously benchmarked): So, from the arm64 side of the fence, I'm perfectly happy just removing the cpu_relax() calls from lockref. As such, come back full circle in history and whack it altogether. Signed-off-by: Mateusz Guzik <mjguzik@gmail.com> Link: https://lore.kernel.org/all/CAGudoHHx0Nqg6DE70zAVA75eV-HXfWyhVMWZ-aSeOofkA_=WdA@mail.gmail.com/ Acked-by: Tony Luck <tony.luck@intel.com> # ia64 Acked-by: Nicholas Piggin <npiggin@gmail.com> # powerpc Acked-by: Will Deacon <will@kernel.org> # arm64 Acked-by: Peter Zijlstra <peterz@infradead.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
194 lines
3.9 KiB
C
194 lines
3.9 KiB
C
// SPDX-License-Identifier: GPL-2.0
|
|
#include <linux/export.h>
|
|
#include <linux/lockref.h>
|
|
|
|
#if USE_CMPXCHG_LOCKREF
|
|
|
|
/*
|
|
* Note that the "cmpxchg()" reloads the "old" value for the
|
|
* failure case.
|
|
*/
|
|
#define CMPXCHG_LOOP(CODE, SUCCESS) do { \
|
|
int retry = 100; \
|
|
struct lockref old; \
|
|
BUILD_BUG_ON(sizeof(old) != 8); \
|
|
old.lock_count = READ_ONCE(lockref->lock_count); \
|
|
while (likely(arch_spin_value_unlocked(old.lock.rlock.raw_lock))) { \
|
|
struct lockref new = old; \
|
|
CODE \
|
|
if (likely(try_cmpxchg64_relaxed(&lockref->lock_count, \
|
|
&old.lock_count, \
|
|
new.lock_count))) { \
|
|
SUCCESS; \
|
|
} \
|
|
if (!--retry) \
|
|
break; \
|
|
} \
|
|
} while (0)
|
|
|
|
#else
|
|
|
|
#define CMPXCHG_LOOP(CODE, SUCCESS) do { } while (0)
|
|
|
|
#endif
|
|
|
|
/**
|
|
* lockref_get - Increments reference count unconditionally
|
|
* @lockref: pointer to lockref structure
|
|
*
|
|
* This operation is only valid if you already hold a reference
|
|
* to the object, so you know the count cannot be zero.
|
|
*/
|
|
void lockref_get(struct lockref *lockref)
|
|
{
|
|
CMPXCHG_LOOP(
|
|
new.count++;
|
|
,
|
|
return;
|
|
);
|
|
|
|
spin_lock(&lockref->lock);
|
|
lockref->count++;
|
|
spin_unlock(&lockref->lock);
|
|
}
|
|
EXPORT_SYMBOL(lockref_get);
|
|
|
|
/**
|
|
* lockref_get_not_zero - Increments count unless the count is 0 or dead
|
|
* @lockref: pointer to lockref structure
|
|
* Return: 1 if count updated successfully or 0 if count was zero
|
|
*/
|
|
int lockref_get_not_zero(struct lockref *lockref)
|
|
{
|
|
int retval;
|
|
|
|
CMPXCHG_LOOP(
|
|
new.count++;
|
|
if (old.count <= 0)
|
|
return 0;
|
|
,
|
|
return 1;
|
|
);
|
|
|
|
spin_lock(&lockref->lock);
|
|
retval = 0;
|
|
if (lockref->count > 0) {
|
|
lockref->count++;
|
|
retval = 1;
|
|
}
|
|
spin_unlock(&lockref->lock);
|
|
return retval;
|
|
}
|
|
EXPORT_SYMBOL(lockref_get_not_zero);
|
|
|
|
/**
|
|
* lockref_put_not_zero - Decrements count unless count <= 1 before decrement
|
|
* @lockref: pointer to lockref structure
|
|
* Return: 1 if count updated successfully or 0 if count would become zero
|
|
*/
|
|
int lockref_put_not_zero(struct lockref *lockref)
|
|
{
|
|
int retval;
|
|
|
|
CMPXCHG_LOOP(
|
|
new.count--;
|
|
if (old.count <= 1)
|
|
return 0;
|
|
,
|
|
return 1;
|
|
);
|
|
|
|
spin_lock(&lockref->lock);
|
|
retval = 0;
|
|
if (lockref->count > 1) {
|
|
lockref->count--;
|
|
retval = 1;
|
|
}
|
|
spin_unlock(&lockref->lock);
|
|
return retval;
|
|
}
|
|
EXPORT_SYMBOL(lockref_put_not_zero);
|
|
|
|
/**
|
|
* lockref_put_return - Decrement reference count if possible
|
|
* @lockref: pointer to lockref structure
|
|
*
|
|
* Decrement the reference count and return the new value.
|
|
* If the lockref was dead or locked, return an error.
|
|
*/
|
|
int lockref_put_return(struct lockref *lockref)
|
|
{
|
|
CMPXCHG_LOOP(
|
|
new.count--;
|
|
if (old.count <= 0)
|
|
return -1;
|
|
,
|
|
return new.count;
|
|
);
|
|
return -1;
|
|
}
|
|
EXPORT_SYMBOL(lockref_put_return);
|
|
|
|
/**
|
|
* lockref_put_or_lock - decrements count unless count <= 1 before decrement
|
|
* @lockref: pointer to lockref structure
|
|
* Return: 1 if count updated successfully or 0 if count <= 1 and lock taken
|
|
*/
|
|
int lockref_put_or_lock(struct lockref *lockref)
|
|
{
|
|
CMPXCHG_LOOP(
|
|
new.count--;
|
|
if (old.count <= 1)
|
|
break;
|
|
,
|
|
return 1;
|
|
);
|
|
|
|
spin_lock(&lockref->lock);
|
|
if (lockref->count <= 1)
|
|
return 0;
|
|
lockref->count--;
|
|
spin_unlock(&lockref->lock);
|
|
return 1;
|
|
}
|
|
EXPORT_SYMBOL(lockref_put_or_lock);
|
|
|
|
/**
|
|
* lockref_mark_dead - mark lockref dead
|
|
* @lockref: pointer to lockref structure
|
|
*/
|
|
void lockref_mark_dead(struct lockref *lockref)
|
|
{
|
|
assert_spin_locked(&lockref->lock);
|
|
lockref->count = -128;
|
|
}
|
|
EXPORT_SYMBOL(lockref_mark_dead);
|
|
|
|
/**
|
|
* lockref_get_not_dead - Increments count unless the ref is dead
|
|
* @lockref: pointer to lockref structure
|
|
* Return: 1 if count updated successfully or 0 if lockref was dead
|
|
*/
|
|
int lockref_get_not_dead(struct lockref *lockref)
|
|
{
|
|
int retval;
|
|
|
|
CMPXCHG_LOOP(
|
|
new.count++;
|
|
if (old.count < 0)
|
|
return 0;
|
|
,
|
|
return 1;
|
|
);
|
|
|
|
spin_lock(&lockref->lock);
|
|
retval = 0;
|
|
if (lockref->count >= 0) {
|
|
lockref->count++;
|
|
retval = 1;
|
|
}
|
|
spin_unlock(&lockref->lock);
|
|
return retval;
|
|
}
|
|
EXPORT_SYMBOL(lockref_get_not_dead);
|