OpenCloudOS-Kernel/kernel/semaphore.c

264 lines
7.2 KiB
C
Raw Normal View History

/*
* Copyright (c) 2008 Intel Corporation
* Author: Matthew Wilcox <willy@linux.intel.com>
*
* Distributed under the terms of the GNU GPL, version 2
*
* This file implements counting semaphores.
* A counting semaphore may be acquired 'n' times before sleeping.
* See mutex.c for single-acquisition sleeping locks which enforce
* rules which allow code to be debugged more easily.
*/
/*
* Some notes on the implementation:
*
* The spinlock controls access to the other members of the semaphore.
* down_trylock() and up() can be called from interrupt context, so we
* have to disable interrupts when taking the lock. It turns out various
* parts of the kernel expect to be able to use down() on a semaphore in
* interrupt context when they know it will succeed, so we have to use
* irqsave variants for down(), down_interruptible() and down_killable()
* too.
*
* The ->count variable represents how many more tasks can acquire this
* semaphore. If it's zero, there may be tasks waiting on the wait_list.
*/
#include <linux/compiler.h>
#include <linux/kernel.h>
#include <linux/export.h>
#include <linux/sched.h>
#include <linux/semaphore.h>
#include <linux/spinlock.h>
#include <linux/ftrace.h>
static noinline void __down(struct semaphore *sem);
static noinline int __down_interruptible(struct semaphore *sem);
static noinline int __down_killable(struct semaphore *sem);
static noinline int __down_timeout(struct semaphore *sem, long jiffies);
static noinline void __up(struct semaphore *sem);
/**
* down - acquire the semaphore
* @sem: the semaphore to be acquired
*
* Acquires the semaphore. If no more tasks are allowed to acquire the
* semaphore, calling this function will put the task to sleep until the
* semaphore is released.
*
* Use of this function is deprecated, please use down_interruptible() or
* down_killable() instead.
*/
void down(struct semaphore *sem)
{
unsigned long flags;
raw_spin_lock_irqsave(&sem->lock, flags);
if (likely(sem->count > 0))
sem->count--;
else
__down(sem);
raw_spin_unlock_irqrestore(&sem->lock, flags);
}
EXPORT_SYMBOL(down);
/**
* down_interruptible - acquire the semaphore unless interrupted
* @sem: the semaphore to be acquired
*
* Attempts to acquire the semaphore. If no more tasks are allowed to
* acquire the semaphore, calling this function will put the task to sleep.
* If the sleep is interrupted by a signal, this function will return -EINTR.
* If the semaphore is successfully acquired, this function returns 0.
*/
int down_interruptible(struct semaphore *sem)
{
unsigned long flags;
int result = 0;
raw_spin_lock_irqsave(&sem->lock, flags);
if (likely(sem->count > 0))
semaphore: fix Yanmin Zhang reported: | Comparing with kernel 2.6.25, AIM7 (use tmpfs) has more th | regression under 2.6.26-rc1 on my 8-core stoakley, 16-core tigerton, | and Itanium Montecito. Bisect located the patch below: | | 64ac24e738823161693bf791f87adc802cf529ff is first bad commit | commit 64ac24e738823161693bf791f87adc802cf529ff | Author: Matthew Wilcox <matthew@wil.cx> | Date: Fri Mar 7 21:55:58 2008 -0500 | | Generic semaphore implementation | | After I manually reverted the patch against 2.6.26-rc1 while fixing | lots of conflicts/errors, aim7 regression became less than 2%. i reproduced the AIM7 workload and can confirm Yanmin's findings that -.26-rc1 regresses over .25 - by over 67% here. Looking at the workload i found and fixed what i believe to be the real bug causing the AIM7 regression: it was inefficient wakeup / scheduling / locking behavior of the new generic semaphore code, causing suboptimal performance. The problem comes from the following code. The new semaphore code does this on down(): spin_lock_irqsave(&sem->lock, flags); if (likely(sem->count > 0)) sem->count--; else __down(sem); spin_unlock_irqrestore(&sem->lock, flags); and this on up(): spin_lock_irqsave(&sem->lock, flags); if (likely(list_empty(&sem->wait_list))) sem->count++; else __up(sem); spin_unlock_irqrestore(&sem->lock, flags); where __up() does: list_del(&waiter->list); waiter->up = 1; wake_up_process(waiter->task); and where __down() does this in essence: list_add_tail(&waiter.list, &sem->wait_list); waiter.task = task; waiter.up = 0; for (;;) { [...] spin_unlock_irq(&sem->lock); timeout = schedule_timeout(timeout); spin_lock_irq(&sem->lock); if (waiter.up) return 0; } the fastpath looks good and obvious, but note the following property of the contended path: if there's a task on the ->wait_list, the up() of the current owner will "pass over" ownership to that waiting task, in a wake-one manner, via the waiter->up flag and by removing the waiter from the wait list. That is all and fine in principle, but as implemented in kernel/semaphore.c it also creates a nasty, hidden source of contention! The contention comes from the following property of the new semaphore code: the new owner owns the semaphore exclusively, even if it is not running yet. So if the old owner, even if just a few instructions later, does a down() [lock_kernel()] again, it will be blocked and will have to wait on the new owner to eventually be scheduled (possibly on another CPU)! Or if another task gets to lock_kernel() sooner than the "new owner" scheduled, it will be blocked unnecessarily and for a very long time when there are 2000 tasks running. I.e. the implementation of the new semaphores code does wake-one and lock ownership in a very restrictive way - it does not allow opportunistic re-locking of the lock at all and keeps the scheduler from picking task order intelligently. This kind of scheduling, with 2000 AIM7 processes running, creates awful cross-scheduling between those 2000 tasks, causes reduced parallelism, a throttled runqueue length and a lot of idle time. With increasing number of CPUs it causes an exponentially worse behavior in AIM7, as the chance for a newly woken new-owner task to actually run anytime soon is less and less likely. Note that it takes just a tiny bit of contention for the 'new-semaphore catastrophy' to happen: the wakeup latencies get added to whatever small contention there is, and quickly snowball out of control! I believe Yanmin's findings and numbers support this analysis too. The best fix for this problem is to use the same scheduling logic that the kernel/mutex.c code uses: keep the wake-one behavior (that is OK and wanted because we do not want to over-schedule), but also allow opportunistic locking of the lock even if a wakee is already "in flight". The patch below implements this new logic. With this patch applied the AIM7 regression is largely fixed on my quad testbox: # v2.6.25 vanilla: .................. Tasks Jobs/Min JTI Real CPU Jobs/sec/task 2000 56096.4 91 207.5 789.7 0.4675 2000 55894.4 94 208.2 792.7 0.4658 # v2.6.26-rc1-166-gc0a1811 vanilla: ................................... Tasks Jobs/Min JTI Real CPU Jobs/sec/task 2000 33230.6 83 350.3 784.5 0.2769 2000 31778.1 86 366.3 783.6 0.2648 # v2.6.26-rc1-166-gc0a1811 + semaphore-speedup: ............................................... Tasks Jobs/Min JTI Real CPU Jobs/sec/task 2000 55707.1 92 209.0 795.6 0.4642 2000 55704.4 96 209.0 796.0 0.4642 i.e. a 67% speedup. We are now back to within 1% of the v2.6.25 performance levels and have zero idle time during the test, as expected. Btw., interactivity also improved dramatically with the fix - for example console-switching became almost instantaneous during this workload (which after all is running 2000 tasks at once!), without the patch it was stuck for a minute at times. There's another nice side-effect of this speedup patch, the new generic semaphore code got even smaller: text data bss dec hex filename 1241 0 0 1241 4d9 semaphore.o.before 1207 0 0 1207 4b7 semaphore.o.after (because the waiter.up complication got removed.) Longer-term we should look into using the mutex code for the generic semaphore code as well - but i's not easy due to legacies and it's outside of the scope of v2.6.26 and outside the scope of this patch as well. Bisected-by: "Zhang, Yanmin" <yanmin_zhang@linux.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-05-08 17:53:48 +08:00
sem->count--;
else
result = __down_interruptible(sem);
raw_spin_unlock_irqrestore(&sem->lock, flags);
return result;
}
EXPORT_SYMBOL(down_interruptible);
/**
* down_killable - acquire the semaphore unless killed
* @sem: the semaphore to be acquired
*
* Attempts to acquire the semaphore. If no more tasks are allowed to
* acquire the semaphore, calling this function will put the task to sleep.
* If the sleep is interrupted by a fatal signal, this function will return
* -EINTR. If the semaphore is successfully acquired, this function returns
* 0.
*/
int down_killable(struct semaphore *sem)
{
unsigned long flags;
int result = 0;
raw_spin_lock_irqsave(&sem->lock, flags);
if (likely(sem->count > 0))
semaphore: fix Yanmin Zhang reported: | Comparing with kernel 2.6.25, AIM7 (use tmpfs) has more th | regression under 2.6.26-rc1 on my 8-core stoakley, 16-core tigerton, | and Itanium Montecito. Bisect located the patch below: | | 64ac24e738823161693bf791f87adc802cf529ff is first bad commit | commit 64ac24e738823161693bf791f87adc802cf529ff | Author: Matthew Wilcox <matthew@wil.cx> | Date: Fri Mar 7 21:55:58 2008 -0500 | | Generic semaphore implementation | | After I manually reverted the patch against 2.6.26-rc1 while fixing | lots of conflicts/errors, aim7 regression became less than 2%. i reproduced the AIM7 workload and can confirm Yanmin's findings that -.26-rc1 regresses over .25 - by over 67% here. Looking at the workload i found and fixed what i believe to be the real bug causing the AIM7 regression: it was inefficient wakeup / scheduling / locking behavior of the new generic semaphore code, causing suboptimal performance. The problem comes from the following code. The new semaphore code does this on down(): spin_lock_irqsave(&sem->lock, flags); if (likely(sem->count > 0)) sem->count--; else __down(sem); spin_unlock_irqrestore(&sem->lock, flags); and this on up(): spin_lock_irqsave(&sem->lock, flags); if (likely(list_empty(&sem->wait_list))) sem->count++; else __up(sem); spin_unlock_irqrestore(&sem->lock, flags); where __up() does: list_del(&waiter->list); waiter->up = 1; wake_up_process(waiter->task); and where __down() does this in essence: list_add_tail(&waiter.list, &sem->wait_list); waiter.task = task; waiter.up = 0; for (;;) { [...] spin_unlock_irq(&sem->lock); timeout = schedule_timeout(timeout); spin_lock_irq(&sem->lock); if (waiter.up) return 0; } the fastpath looks good and obvious, but note the following property of the contended path: if there's a task on the ->wait_list, the up() of the current owner will "pass over" ownership to that waiting task, in a wake-one manner, via the waiter->up flag and by removing the waiter from the wait list. That is all and fine in principle, but as implemented in kernel/semaphore.c it also creates a nasty, hidden source of contention! The contention comes from the following property of the new semaphore code: the new owner owns the semaphore exclusively, even if it is not running yet. So if the old owner, even if just a few instructions later, does a down() [lock_kernel()] again, it will be blocked and will have to wait on the new owner to eventually be scheduled (possibly on another CPU)! Or if another task gets to lock_kernel() sooner than the "new owner" scheduled, it will be blocked unnecessarily and for a very long time when there are 2000 tasks running. I.e. the implementation of the new semaphores code does wake-one and lock ownership in a very restrictive way - it does not allow opportunistic re-locking of the lock at all and keeps the scheduler from picking task order intelligently. This kind of scheduling, with 2000 AIM7 processes running, creates awful cross-scheduling between those 2000 tasks, causes reduced parallelism, a throttled runqueue length and a lot of idle time. With increasing number of CPUs it causes an exponentially worse behavior in AIM7, as the chance for a newly woken new-owner task to actually run anytime soon is less and less likely. Note that it takes just a tiny bit of contention for the 'new-semaphore catastrophy' to happen: the wakeup latencies get added to whatever small contention there is, and quickly snowball out of control! I believe Yanmin's findings and numbers support this analysis too. The best fix for this problem is to use the same scheduling logic that the kernel/mutex.c code uses: keep the wake-one behavior (that is OK and wanted because we do not want to over-schedule), but also allow opportunistic locking of the lock even if a wakee is already "in flight". The patch below implements this new logic. With this patch applied the AIM7 regression is largely fixed on my quad testbox: # v2.6.25 vanilla: .................. Tasks Jobs/Min JTI Real CPU Jobs/sec/task 2000 56096.4 91 207.5 789.7 0.4675 2000 55894.4 94 208.2 792.7 0.4658 # v2.6.26-rc1-166-gc0a1811 vanilla: ................................... Tasks Jobs/Min JTI Real CPU Jobs/sec/task 2000 33230.6 83 350.3 784.5 0.2769 2000 31778.1 86 366.3 783.6 0.2648 # v2.6.26-rc1-166-gc0a1811 + semaphore-speedup: ............................................... Tasks Jobs/Min JTI Real CPU Jobs/sec/task 2000 55707.1 92 209.0 795.6 0.4642 2000 55704.4 96 209.0 796.0 0.4642 i.e. a 67% speedup. We are now back to within 1% of the v2.6.25 performance levels and have zero idle time during the test, as expected. Btw., interactivity also improved dramatically with the fix - for example console-switching became almost instantaneous during this workload (which after all is running 2000 tasks at once!), without the patch it was stuck for a minute at times. There's another nice side-effect of this speedup patch, the new generic semaphore code got even smaller: text data bss dec hex filename 1241 0 0 1241 4d9 semaphore.o.before 1207 0 0 1207 4b7 semaphore.o.after (because the waiter.up complication got removed.) Longer-term we should look into using the mutex code for the generic semaphore code as well - but i's not easy due to legacies and it's outside of the scope of v2.6.26 and outside the scope of this patch as well. Bisected-by: "Zhang, Yanmin" <yanmin_zhang@linux.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-05-08 17:53:48 +08:00
sem->count--;
else
result = __down_killable(sem);
raw_spin_unlock_irqrestore(&sem->lock, flags);
return result;
}
EXPORT_SYMBOL(down_killable);
/**
* down_trylock - try to acquire the semaphore, without waiting
* @sem: the semaphore to be acquired
*
* Try to acquire the semaphore atomically. Returns 0 if the semaphore has
* been acquired successfully or 1 if it it cannot be acquired.
*
* NOTE: This return value is inverted from both spin_trylock and
* mutex_trylock! Be careful about this when converting code.
*
* Unlike mutex_trylock, this function can be used from interrupt context,
* and the semaphore can be released by any task or interrupt.
*/
int down_trylock(struct semaphore *sem)
{
unsigned long flags;
int count;
raw_spin_lock_irqsave(&sem->lock, flags);
count = sem->count - 1;
if (likely(count >= 0))
sem->count = count;
raw_spin_unlock_irqrestore(&sem->lock, flags);
return (count < 0);
}
EXPORT_SYMBOL(down_trylock);
/**
* down_timeout - acquire the semaphore within a specified time
* @sem: the semaphore to be acquired
* @jiffies: how long to wait before failing
*
* Attempts to acquire the semaphore. If no more tasks are allowed to
* acquire the semaphore, calling this function will put the task to sleep.
* If the semaphore is not released within the specified number of jiffies,
* this function returns -ETIME. It returns 0 if the semaphore was acquired.
*/
int down_timeout(struct semaphore *sem, long jiffies)
{
unsigned long flags;
int result = 0;
raw_spin_lock_irqsave(&sem->lock, flags);
if (likely(sem->count > 0))
semaphore: fix Yanmin Zhang reported: | Comparing with kernel 2.6.25, AIM7 (use tmpfs) has more th | regression under 2.6.26-rc1 on my 8-core stoakley, 16-core tigerton, | and Itanium Montecito. Bisect located the patch below: | | 64ac24e738823161693bf791f87adc802cf529ff is first bad commit | commit 64ac24e738823161693bf791f87adc802cf529ff | Author: Matthew Wilcox <matthew@wil.cx> | Date: Fri Mar 7 21:55:58 2008 -0500 | | Generic semaphore implementation | | After I manually reverted the patch against 2.6.26-rc1 while fixing | lots of conflicts/errors, aim7 regression became less than 2%. i reproduced the AIM7 workload and can confirm Yanmin's findings that -.26-rc1 regresses over .25 - by over 67% here. Looking at the workload i found and fixed what i believe to be the real bug causing the AIM7 regression: it was inefficient wakeup / scheduling / locking behavior of the new generic semaphore code, causing suboptimal performance. The problem comes from the following code. The new semaphore code does this on down(): spin_lock_irqsave(&sem->lock, flags); if (likely(sem->count > 0)) sem->count--; else __down(sem); spin_unlock_irqrestore(&sem->lock, flags); and this on up(): spin_lock_irqsave(&sem->lock, flags); if (likely(list_empty(&sem->wait_list))) sem->count++; else __up(sem); spin_unlock_irqrestore(&sem->lock, flags); where __up() does: list_del(&waiter->list); waiter->up = 1; wake_up_process(waiter->task); and where __down() does this in essence: list_add_tail(&waiter.list, &sem->wait_list); waiter.task = task; waiter.up = 0; for (;;) { [...] spin_unlock_irq(&sem->lock); timeout = schedule_timeout(timeout); spin_lock_irq(&sem->lock); if (waiter.up) return 0; } the fastpath looks good and obvious, but note the following property of the contended path: if there's a task on the ->wait_list, the up() of the current owner will "pass over" ownership to that waiting task, in a wake-one manner, via the waiter->up flag and by removing the waiter from the wait list. That is all and fine in principle, but as implemented in kernel/semaphore.c it also creates a nasty, hidden source of contention! The contention comes from the following property of the new semaphore code: the new owner owns the semaphore exclusively, even if it is not running yet. So if the old owner, even if just a few instructions later, does a down() [lock_kernel()] again, it will be blocked and will have to wait on the new owner to eventually be scheduled (possibly on another CPU)! Or if another task gets to lock_kernel() sooner than the "new owner" scheduled, it will be blocked unnecessarily and for a very long time when there are 2000 tasks running. I.e. the implementation of the new semaphores code does wake-one and lock ownership in a very restrictive way - it does not allow opportunistic re-locking of the lock at all and keeps the scheduler from picking task order intelligently. This kind of scheduling, with 2000 AIM7 processes running, creates awful cross-scheduling between those 2000 tasks, causes reduced parallelism, a throttled runqueue length and a lot of idle time. With increasing number of CPUs it causes an exponentially worse behavior in AIM7, as the chance for a newly woken new-owner task to actually run anytime soon is less and less likely. Note that it takes just a tiny bit of contention for the 'new-semaphore catastrophy' to happen: the wakeup latencies get added to whatever small contention there is, and quickly snowball out of control! I believe Yanmin's findings and numbers support this analysis too. The best fix for this problem is to use the same scheduling logic that the kernel/mutex.c code uses: keep the wake-one behavior (that is OK and wanted because we do not want to over-schedule), but also allow opportunistic locking of the lock even if a wakee is already "in flight". The patch below implements this new logic. With this patch applied the AIM7 regression is largely fixed on my quad testbox: # v2.6.25 vanilla: .................. Tasks Jobs/Min JTI Real CPU Jobs/sec/task 2000 56096.4 91 207.5 789.7 0.4675 2000 55894.4 94 208.2 792.7 0.4658 # v2.6.26-rc1-166-gc0a1811 vanilla: ................................... Tasks Jobs/Min JTI Real CPU Jobs/sec/task 2000 33230.6 83 350.3 784.5 0.2769 2000 31778.1 86 366.3 783.6 0.2648 # v2.6.26-rc1-166-gc0a1811 + semaphore-speedup: ............................................... Tasks Jobs/Min JTI Real CPU Jobs/sec/task 2000 55707.1 92 209.0 795.6 0.4642 2000 55704.4 96 209.0 796.0 0.4642 i.e. a 67% speedup. We are now back to within 1% of the v2.6.25 performance levels and have zero idle time during the test, as expected. Btw., interactivity also improved dramatically with the fix - for example console-switching became almost instantaneous during this workload (which after all is running 2000 tasks at once!), without the patch it was stuck for a minute at times. There's another nice side-effect of this speedup patch, the new generic semaphore code got even smaller: text data bss dec hex filename 1241 0 0 1241 4d9 semaphore.o.before 1207 0 0 1207 4b7 semaphore.o.after (because the waiter.up complication got removed.) Longer-term we should look into using the mutex code for the generic semaphore code as well - but i's not easy due to legacies and it's outside of the scope of v2.6.26 and outside the scope of this patch as well. Bisected-by: "Zhang, Yanmin" <yanmin_zhang@linux.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-05-08 17:53:48 +08:00
sem->count--;
else
result = __down_timeout(sem, jiffies);
raw_spin_unlock_irqrestore(&sem->lock, flags);
return result;
}
EXPORT_SYMBOL(down_timeout);
/**
* up - release the semaphore
* @sem: the semaphore to release
*
* Release the semaphore. Unlike mutexes, up() may be called from any
* context and even by tasks which have never called down().
*/
void up(struct semaphore *sem)
{
unsigned long flags;
raw_spin_lock_irqsave(&sem->lock, flags);
if (likely(list_empty(&sem->wait_list)))
sem->count++;
else
__up(sem);
raw_spin_unlock_irqrestore(&sem->lock, flags);
}
EXPORT_SYMBOL(up);
/* Functions for the contended case */
struct semaphore_waiter {
struct list_head list;
struct task_struct *task;
int up;
};
/*
* Because this function is inlined, the 'state' parameter will be
* constant, and thus optimised away by the compiler. Likewise the
* 'timeout' parameter for the cases without timeouts.
*/
static inline int __sched __down_common(struct semaphore *sem, long state,
long timeout)
{
struct task_struct *task = current;
struct semaphore_waiter waiter;
semaphore: fix Yanmin Zhang reported: | Comparing with kernel 2.6.25, AIM7 (use tmpfs) has more th | regression under 2.6.26-rc1 on my 8-core stoakley, 16-core tigerton, | and Itanium Montecito. Bisect located the patch below: | | 64ac24e738823161693bf791f87adc802cf529ff is first bad commit | commit 64ac24e738823161693bf791f87adc802cf529ff | Author: Matthew Wilcox <matthew@wil.cx> | Date: Fri Mar 7 21:55:58 2008 -0500 | | Generic semaphore implementation | | After I manually reverted the patch against 2.6.26-rc1 while fixing | lots of conflicts/errors, aim7 regression became less than 2%. i reproduced the AIM7 workload and can confirm Yanmin's findings that -.26-rc1 regresses over .25 - by over 67% here. Looking at the workload i found and fixed what i believe to be the real bug causing the AIM7 regression: it was inefficient wakeup / scheduling / locking behavior of the new generic semaphore code, causing suboptimal performance. The problem comes from the following code. The new semaphore code does this on down(): spin_lock_irqsave(&sem->lock, flags); if (likely(sem->count > 0)) sem->count--; else __down(sem); spin_unlock_irqrestore(&sem->lock, flags); and this on up(): spin_lock_irqsave(&sem->lock, flags); if (likely(list_empty(&sem->wait_list))) sem->count++; else __up(sem); spin_unlock_irqrestore(&sem->lock, flags); where __up() does: list_del(&waiter->list); waiter->up = 1; wake_up_process(waiter->task); and where __down() does this in essence: list_add_tail(&waiter.list, &sem->wait_list); waiter.task = task; waiter.up = 0; for (;;) { [...] spin_unlock_irq(&sem->lock); timeout = schedule_timeout(timeout); spin_lock_irq(&sem->lock); if (waiter.up) return 0; } the fastpath looks good and obvious, but note the following property of the contended path: if there's a task on the ->wait_list, the up() of the current owner will "pass over" ownership to that waiting task, in a wake-one manner, via the waiter->up flag and by removing the waiter from the wait list. That is all and fine in principle, but as implemented in kernel/semaphore.c it also creates a nasty, hidden source of contention! The contention comes from the following property of the new semaphore code: the new owner owns the semaphore exclusively, even if it is not running yet. So if the old owner, even if just a few instructions later, does a down() [lock_kernel()] again, it will be blocked and will have to wait on the new owner to eventually be scheduled (possibly on another CPU)! Or if another task gets to lock_kernel() sooner than the "new owner" scheduled, it will be blocked unnecessarily and for a very long time when there are 2000 tasks running. I.e. the implementation of the new semaphores code does wake-one and lock ownership in a very restrictive way - it does not allow opportunistic re-locking of the lock at all and keeps the scheduler from picking task order intelligently. This kind of scheduling, with 2000 AIM7 processes running, creates awful cross-scheduling between those 2000 tasks, causes reduced parallelism, a throttled runqueue length and a lot of idle time. With increasing number of CPUs it causes an exponentially worse behavior in AIM7, as the chance for a newly woken new-owner task to actually run anytime soon is less and less likely. Note that it takes just a tiny bit of contention for the 'new-semaphore catastrophy' to happen: the wakeup latencies get added to whatever small contention there is, and quickly snowball out of control! I believe Yanmin's findings and numbers support this analysis too. The best fix for this problem is to use the same scheduling logic that the kernel/mutex.c code uses: keep the wake-one behavior (that is OK and wanted because we do not want to over-schedule), but also allow opportunistic locking of the lock even if a wakee is already "in flight". The patch below implements this new logic. With this patch applied the AIM7 regression is largely fixed on my quad testbox: # v2.6.25 vanilla: .................. Tasks Jobs/Min JTI Real CPU Jobs/sec/task 2000 56096.4 91 207.5 789.7 0.4675 2000 55894.4 94 208.2 792.7 0.4658 # v2.6.26-rc1-166-gc0a1811 vanilla: ................................... Tasks Jobs/Min JTI Real CPU Jobs/sec/task 2000 33230.6 83 350.3 784.5 0.2769 2000 31778.1 86 366.3 783.6 0.2648 # v2.6.26-rc1-166-gc0a1811 + semaphore-speedup: ............................................... Tasks Jobs/Min JTI Real CPU Jobs/sec/task 2000 55707.1 92 209.0 795.6 0.4642 2000 55704.4 96 209.0 796.0 0.4642 i.e. a 67% speedup. We are now back to within 1% of the v2.6.25 performance levels and have zero idle time during the test, as expected. Btw., interactivity also improved dramatically with the fix - for example console-switching became almost instantaneous during this workload (which after all is running 2000 tasks at once!), without the patch it was stuck for a minute at times. There's another nice side-effect of this speedup patch, the new generic semaphore code got even smaller: text data bss dec hex filename 1241 0 0 1241 4d9 semaphore.o.before 1207 0 0 1207 4b7 semaphore.o.after (because the waiter.up complication got removed.) Longer-term we should look into using the mutex code for the generic semaphore code as well - but i's not easy due to legacies and it's outside of the scope of v2.6.26 and outside the scope of this patch as well. Bisected-by: "Zhang, Yanmin" <yanmin_zhang@linux.intel.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-05-08 17:53:48 +08:00
list_add_tail(&waiter.list, &sem->wait_list);
waiter.task = task;
waiter.up = 0;
for (;;) {
if (signal_pending_state(state, task))
goto interrupted;
if (timeout <= 0)
goto timed_out;
__set_task_state(task, state);
raw_spin_unlock_irq(&sem->lock);
timeout = schedule_timeout(timeout);
raw_spin_lock_irq(&sem->lock);
if (waiter.up)
return 0;
}
timed_out:
list_del(&waiter.list);
return -ETIME;
interrupted:
list_del(&waiter.list);
return -EINTR;
}
static noinline void __sched __down(struct semaphore *sem)
{
__down_common(sem, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
}
static noinline int __sched __down_interruptible(struct semaphore *sem)
{
return __down_common(sem, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
}
static noinline int __sched __down_killable(struct semaphore *sem)
{
return __down_common(sem, TASK_KILLABLE, MAX_SCHEDULE_TIMEOUT);
}
static noinline int __sched __down_timeout(struct semaphore *sem, long jiffies)
{
return __down_common(sem, TASK_UNINTERRUPTIBLE, jiffies);
}
static noinline void __sched __up(struct semaphore *sem)
{
struct semaphore_waiter *waiter = list_first_entry(&sem->wait_list,
struct semaphore_waiter, list);
list_del(&waiter->list);
waiter->up = 1;
wake_up_process(waiter->task);
}