kernel_optimize_test/kernel/locking/mutex.c

1035 lines
26 KiB
C
Raw Normal View History

/*
* kernel/locking/mutex.c
*
* Mutexes: blocking mutual exclusion locks
*
* Started by Ingo Molnar:
*
* Copyright (C) 2004, 2005, 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
*
* Many thanks to Arjan van de Ven, Thomas Gleixner, Steven Rostedt and
* David Howells for suggestions and improvements.
*
mutex: implement adaptive spinning Change mutex contention behaviour such that it will sometimes busy wait on acquisition - moving its behaviour closer to that of spinlocks. This concept got ported to mainline from the -rt tree, where it was originally implemented for rtmutexes by Steven Rostedt, based on work by Gregory Haskins. Testing with Ingo's test-mutex application (http://lkml.org/lkml/2006/1/8/50) gave a 345% boost for VFS scalability on my testbox: # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 296604 # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 85870 The key criteria for the busy wait is that the lock owner has to be running on a (different) cpu. The idea is that as long as the owner is running, there is a fair chance it'll release the lock soon, and thus we'll be better off spinning instead of blocking/scheduling. Since regular mutexes (as opposed to rtmutexes) do not atomically track the owner, we add the owner in a non-atomic fashion and deal with the races in the slowpath. Furthermore, to ease the testing of the performance impact of this new code, there is means to disable this behaviour runtime (without having to reboot the system), when scheduler debugging is enabled (CONFIG_SCHED_DEBUG=y), by issuing the following command: # echo NO_OWNER_SPIN > /debug/sched_features This command re-enables spinning again (this is also the default): # echo OWNER_SPIN > /debug/sched_features Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-01-12 21:01:47 +08:00
* - Adaptive spinning for mutexes by Peter Zijlstra. (Ported to mainline
* from the -rt tree, where it was originally implemented for rtmutexes
* by Steven Rostedt, based on work by Gregory Haskins, Peter Morreale
* and Sven Dietrich.
*
locking/Documentation: Move locking related docs into Documentation/locking/ Specifically: Documentation/locking/lockdep-design.txt Documentation/locking/lockstat.txt Documentation/locking/mutex-design.txt Documentation/locking/rt-mutex-design.txt Documentation/locking/rt-mutex.txt Documentation/locking/spinlocks.txt Documentation/locking/ww-mutex-design.txt Signed-off-by: Davidlohr Bueso <davidlohr@hp.com> Acked-by: Randy Dunlap <rdunlap@infradead.org> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Cc: jason.low2@hp.com Cc: aswin@hp.com Cc: Alexei Starovoitov <ast@plumgrid.com> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Chris Mason <clm@fb.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: David Airlie <airlied@linux.ie> Cc: Davidlohr Bueso <davidlohr@hp.com> Cc: David S. Miller <davem@davemloft.net> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: Jason Low <jason.low2@hp.com> Cc: Josef Bacik <jbacik@fusionio.com> Cc: Kees Cook <keescook@chromium.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Lubomir Rintel <lkundrak@v3.sk> Cc: Masanari Iida <standby24x7@gmail.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Randy Dunlap <rdunlap@infradead.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Vineet Gupta <vgupta@synopsys.com> Cc: fengguang.wu@intel.com Link: http://lkml.kernel.org/r/1406752916-3341-6-git-send-email-davidlohr@hp.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-07-31 04:41:55 +08:00
* Also see Documentation/locking/mutex-design.txt.
*/
#include <linux/mutex.h>
#include <linux/ww_mutex.h>
#include <linux/sched.h>
#include <linux/sched/rt.h>
#include <linux/export.h>
#include <linux/spinlock.h>
#include <linux/interrupt.h>
#include <linux/debug_locks.h>
#include <linux/osq_lock.h>
#ifdef CONFIG_DEBUG_MUTEXES
# include "mutex-debug.h"
#else
# include "mutex.h"
#endif
void
__mutex_init(struct mutex *lock, const char *name, struct lock_class_key *key)
{
locking/mutex: Rework mutex::owner The current mutex implementation has an atomic lock word and a non-atomic owner field. This disparity leads to a number of issues with the current mutex code as it means that we can have a locked mutex without an explicit owner (because the owner field has not been set, or already cleared). This leads to a number of weird corner cases, esp. between the optimistic spinning and debug code. Where the optimistic spinning code needs the owner field updated inside the lock region, the debug code is more relaxed because the whole lock is serialized by the wait_lock. Also, the spinning code itself has a few corner cases where we need to deal with a held lock without an owner field. Furthermore, it becomes even more of a problem when trying to fix starvation cases in the current code. We end up stacking special case on special case. To solve this rework the basic mutex implementation to be a single atomic word that contains the owner and uses the low bits for extra state. This matches how PI futexes and rt_mutex already work. By having the owner an integral part of the lock state a lot of the problems dissapear and we get a better option to deal with starvation cases, direct owner handoff. Changing the basic mutex does however invalidate all the arch specific mutex code; this patch leaves that unused in-place, a later patch will remove that. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Will Deacon <will.deacon@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 19:36:04 +08:00
atomic_long_set(&lock->owner, 0);
spin_lock_init(&lock->wait_lock);
INIT_LIST_HEAD(&lock->wait_list);
mutex: Queue mutex spinners with MCS lock to reduce cacheline contention The current mutex spinning code (with MUTEX_SPIN_ON_OWNER option turned on) allow multiple tasks to spin on a single mutex concurrently. A potential problem with the current approach is that when the mutex becomes available, all the spinning tasks will try to acquire the mutex more or less simultaneously. As a result, there will be a lot of cacheline bouncing especially on systems with a large number of CPUs. This patch tries to reduce this kind of contention by putting the mutex spinners into a queue so that only the first one in the queue will try to acquire the mutex. This will reduce contention and allow all the tasks to move forward faster. The queuing of mutex spinners is done using an MCS lock based implementation which will further reduce contention on the mutex cacheline than a similar ticket spinlock based implementation. This patch will add a new field into the mutex data structure for holding the MCS lock. This expands the mutex size by 8 bytes for 64-bit system and 4 bytes for 32-bit system. This overhead will be avoid if the MUTEX_SPIN_ON_OWNER option is turned off. The following table shows the jobs per minute (JPM) scalability data on an 8-node 80-core Westmere box with a 3.7.10 kernel. The numactl command is used to restrict the running of the fserver workloads to 1/2/4/8 nodes with hyperthreading off. +-----------------+-----------+-----------+-------------+----------+ | Configuration | Mean JPM | Mean JPM | Mean JPM | % Change | | | w/o patch | patch 1 | patches 1&2 | 1->1&2 | +-----------------+------------------------------------------------+ | | User Range 1100 - 2000 | +-----------------+------------------------------------------------+ | 8 nodes, HT off | 227972 | 227237 | 305043 | +34.2% | | 4 nodes, HT off | 393503 | 381558 | 394650 | +3.4% | | 2 nodes, HT off | 334957 | 325240 | 338853 | +4.2% | | 1 node , HT off | 198141 | 197972 | 198075 | +0.1% | +-----------------+------------------------------------------------+ | | User Range 200 - 1000 | +-----------------+------------------------------------------------+ | 8 nodes, HT off | 282325 | 312870 | 332185 | +6.2% | | 4 nodes, HT off | 390698 | 378279 | 393419 | +4.0% | | 2 nodes, HT off | 336986 | 326543 | 340260 | +4.2% | | 1 node , HT off | 197588 | 197622 | 197582 | 0.0% | +-----------------+-----------+-----------+-------------+----------+ At low user range 10-100, the JPM differences were within +/-1%. So they are not that interesting. The fserver workload uses mutex spinning extensively. With just the mutex change in the first patch, there is no noticeable change in performance. Rather, there is a slight drop in performance. This mutex spinning patch more than recovers the lost performance and show a significant increase of +30% at high user load with the full 8 nodes. Similar improvements were also seen in a 3.8 kernel. The table below shows the %time spent by different kernel functions as reported by perf when running the fserver workload at 1500 users with all 8 nodes. +-----------------------+-----------+---------+-------------+ | Function | % time | % time | % time | | | w/o patch | patch 1 | patches 1&2 | +-----------------------+-----------+---------+-------------+ | __read_lock_failed | 34.96% | 34.91% | 29.14% | | __write_lock_failed | 10.14% | 10.68% | 7.51% | | mutex_spin_on_owner | 3.62% | 3.42% | 2.33% | | mspin_lock | N/A | N/A | 9.90% | | __mutex_lock_slowpath | 1.46% | 0.81% | 0.14% | | _raw_spin_lock | 2.25% | 2.50% | 1.10% | +-----------------------+-----------+---------+-------------+ The fserver workload for an 8-node system is dominated by the contention in the read/write lock. Mutex contention also plays a role. With the first patch only, mutex contention is down (as shown by the __mutex_lock_slowpath figure) which help a little bit. We saw only a few percents improvement with that. By applying patch 2 as well, the single mutex_spin_on_owner figure is now split out into an additional mspin_lock figure. The time increases from 3.42% to 11.23%. It shows a great reduction in contention among the spinners leading to a 30% improvement. The time ratio 9.9/2.33=4.3 indicates that there are on average 4+ spinners waiting in the spin_lock loop for each spinner in the mutex_spin_on_owner loop. Contention in other locking functions also go down by quite a lot. The table below shows the performance change of both patches 1 & 2 over patch 1 alone in other AIM7 workloads (at 8 nodes, hyperthreading off). +--------------+---------------+----------------+-----------------+ | Workload | mean % change | mean % change | mean % change | | | 10-100 users | 200-1000 users | 1100-2000 users | +--------------+---------------+----------------+-----------------+ | alltests | 0.0% | -0.8% | +0.6% | | five_sec | -0.3% | +0.8% | +0.8% | | high_systime | +0.4% | +2.4% | +2.1% | | new_fserver | +0.1% | +14.1% | +34.2% | | shared | -0.5% | -0.3% | -0.4% | | short | -1.7% | -9.8% | -8.3% | +--------------+---------------+----------------+-----------------+ The short workload is the only one that shows a decline in performance probably due to the spinner locking and queuing overhead. Signed-off-by: Waiman Long <Waiman.Long@hp.com> Reviewed-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Chandramouleeswaran Aswin <aswin@hp.com> Cc: Norton Scott J <scott.norton@hp.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: David Howells <dhowells@redhat.com> Cc: Dave Jones <davej@redhat.com> Cc: Clark Williams <williams@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1366226594-5506-4-git-send-email-Waiman.Long@hp.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-04-18 03:23:13 +08:00
#ifdef CONFIG_MUTEX_SPIN_ON_OWNER
osq_lock_init(&lock->osq);
mutex: Queue mutex spinners with MCS lock to reduce cacheline contention The current mutex spinning code (with MUTEX_SPIN_ON_OWNER option turned on) allow multiple tasks to spin on a single mutex concurrently. A potential problem with the current approach is that when the mutex becomes available, all the spinning tasks will try to acquire the mutex more or less simultaneously. As a result, there will be a lot of cacheline bouncing especially on systems with a large number of CPUs. This patch tries to reduce this kind of contention by putting the mutex spinners into a queue so that only the first one in the queue will try to acquire the mutex. This will reduce contention and allow all the tasks to move forward faster. The queuing of mutex spinners is done using an MCS lock based implementation which will further reduce contention on the mutex cacheline than a similar ticket spinlock based implementation. This patch will add a new field into the mutex data structure for holding the MCS lock. This expands the mutex size by 8 bytes for 64-bit system and 4 bytes for 32-bit system. This overhead will be avoid if the MUTEX_SPIN_ON_OWNER option is turned off. The following table shows the jobs per minute (JPM) scalability data on an 8-node 80-core Westmere box with a 3.7.10 kernel. The numactl command is used to restrict the running of the fserver workloads to 1/2/4/8 nodes with hyperthreading off. +-----------------+-----------+-----------+-------------+----------+ | Configuration | Mean JPM | Mean JPM | Mean JPM | % Change | | | w/o patch | patch 1 | patches 1&2 | 1->1&2 | +-----------------+------------------------------------------------+ | | User Range 1100 - 2000 | +-----------------+------------------------------------------------+ | 8 nodes, HT off | 227972 | 227237 | 305043 | +34.2% | | 4 nodes, HT off | 393503 | 381558 | 394650 | +3.4% | | 2 nodes, HT off | 334957 | 325240 | 338853 | +4.2% | | 1 node , HT off | 198141 | 197972 | 198075 | +0.1% | +-----------------+------------------------------------------------+ | | User Range 200 - 1000 | +-----------------+------------------------------------------------+ | 8 nodes, HT off | 282325 | 312870 | 332185 | +6.2% | | 4 nodes, HT off | 390698 | 378279 | 393419 | +4.0% | | 2 nodes, HT off | 336986 | 326543 | 340260 | +4.2% | | 1 node , HT off | 197588 | 197622 | 197582 | 0.0% | +-----------------+-----------+-----------+-------------+----------+ At low user range 10-100, the JPM differences were within +/-1%. So they are not that interesting. The fserver workload uses mutex spinning extensively. With just the mutex change in the first patch, there is no noticeable change in performance. Rather, there is a slight drop in performance. This mutex spinning patch more than recovers the lost performance and show a significant increase of +30% at high user load with the full 8 nodes. Similar improvements were also seen in a 3.8 kernel. The table below shows the %time spent by different kernel functions as reported by perf when running the fserver workload at 1500 users with all 8 nodes. +-----------------------+-----------+---------+-------------+ | Function | % time | % time | % time | | | w/o patch | patch 1 | patches 1&2 | +-----------------------+-----------+---------+-------------+ | __read_lock_failed | 34.96% | 34.91% | 29.14% | | __write_lock_failed | 10.14% | 10.68% | 7.51% | | mutex_spin_on_owner | 3.62% | 3.42% | 2.33% | | mspin_lock | N/A | N/A | 9.90% | | __mutex_lock_slowpath | 1.46% | 0.81% | 0.14% | | _raw_spin_lock | 2.25% | 2.50% | 1.10% | +-----------------------+-----------+---------+-------------+ The fserver workload for an 8-node system is dominated by the contention in the read/write lock. Mutex contention also plays a role. With the first patch only, mutex contention is down (as shown by the __mutex_lock_slowpath figure) which help a little bit. We saw only a few percents improvement with that. By applying patch 2 as well, the single mutex_spin_on_owner figure is now split out into an additional mspin_lock figure. The time increases from 3.42% to 11.23%. It shows a great reduction in contention among the spinners leading to a 30% improvement. The time ratio 9.9/2.33=4.3 indicates that there are on average 4+ spinners waiting in the spin_lock loop for each spinner in the mutex_spin_on_owner loop. Contention in other locking functions also go down by quite a lot. The table below shows the performance change of both patches 1 & 2 over patch 1 alone in other AIM7 workloads (at 8 nodes, hyperthreading off). +--------------+---------------+----------------+-----------------+ | Workload | mean % change | mean % change | mean % change | | | 10-100 users | 200-1000 users | 1100-2000 users | +--------------+---------------+----------------+-----------------+ | alltests | 0.0% | -0.8% | +0.6% | | five_sec | -0.3% | +0.8% | +0.8% | | high_systime | +0.4% | +2.4% | +2.1% | | new_fserver | +0.1% | +14.1% | +34.2% | | shared | -0.5% | -0.3% | -0.4% | | short | -1.7% | -9.8% | -8.3% | +--------------+---------------+----------------+-----------------+ The short workload is the only one that shows a decline in performance probably due to the spinner locking and queuing overhead. Signed-off-by: Waiman Long <Waiman.Long@hp.com> Reviewed-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Chandramouleeswaran Aswin <aswin@hp.com> Cc: Norton Scott J <scott.norton@hp.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: David Howells <dhowells@redhat.com> Cc: Dave Jones <davej@redhat.com> Cc: Clark Williams <williams@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1366226594-5506-4-git-send-email-Waiman.Long@hp.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-04-18 03:23:13 +08:00
#endif
debug_mutex_init(lock, name, key);
}
EXPORT_SYMBOL(__mutex_init);
locking/mutex: Rework mutex::owner The current mutex implementation has an atomic lock word and a non-atomic owner field. This disparity leads to a number of issues with the current mutex code as it means that we can have a locked mutex without an explicit owner (because the owner field has not been set, or already cleared). This leads to a number of weird corner cases, esp. between the optimistic spinning and debug code. Where the optimistic spinning code needs the owner field updated inside the lock region, the debug code is more relaxed because the whole lock is serialized by the wait_lock. Also, the spinning code itself has a few corner cases where we need to deal with a held lock without an owner field. Furthermore, it becomes even more of a problem when trying to fix starvation cases in the current code. We end up stacking special case on special case. To solve this rework the basic mutex implementation to be a single atomic word that contains the owner and uses the low bits for extra state. This matches how PI futexes and rt_mutex already work. By having the owner an integral part of the lock state a lot of the problems dissapear and we get a better option to deal with starvation cases, direct owner handoff. Changing the basic mutex does however invalidate all the arch specific mutex code; this patch leaves that unused in-place, a later patch will remove that. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Will Deacon <will.deacon@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 19:36:04 +08:00
/*
* @owner: contains: 'struct task_struct *' to the current lock owner,
* NULL means not owned. Since task_struct pointers are aligned at
* ARCH_MIN_TASKALIGN (which is at least sizeof(void *)), we have low
* bits to store extra state.
*
* Bit0 indicates a non-empty waiter list; unlock must issue a wakeup.
locking/mutex: Add lock handoff to avoid starvation Implement lock handoff to avoid lock starvation. Lock starvation is possible because mutex_lock() allows lock stealing, where a running (or optimistic spinning) task beats the woken waiter to the acquire. Lock stealing is an important performance optimization because waiting for a waiter to wake up and get runtime can take a significant time, during which everyboy would stall on the lock. The down-side is of course that it allows for starvation. This patch has the waiter requesting a handoff if it fails to acquire the lock upon waking. This re-introduces some of the wait time, because once we do a handoff we have to wait for the waiter to wake up again. A future patch will add a round of optimistic spinning to attempt to alleviate this penalty, but if that turns out to not be enough, we can add a counter and only request handoff after multiple failed wakeups. There are a few tricky implementation details: - accepting a handoff must only be done in the wait-loop. Since the handoff condition is owner == current, it can easily cause recursive locking trouble. - accepting the handoff must be careful to provide the ACQUIRE semantics. - having the HANDOFF bit set on unlock requires care, we must not clear the owner. - we must be careful to not leave HANDOFF set after we've acquired the lock. The tricky scenario is setting the HANDOFF bit on an unlocked mutex. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Waiman Long <Waiman.Long@hpe.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 20:40:16 +08:00
* Bit1 indicates unlock needs to hand the lock to the top-waiter
locking/mutex: Rework mutex::owner The current mutex implementation has an atomic lock word and a non-atomic owner field. This disparity leads to a number of issues with the current mutex code as it means that we can have a locked mutex without an explicit owner (because the owner field has not been set, or already cleared). This leads to a number of weird corner cases, esp. between the optimistic spinning and debug code. Where the optimistic spinning code needs the owner field updated inside the lock region, the debug code is more relaxed because the whole lock is serialized by the wait_lock. Also, the spinning code itself has a few corner cases where we need to deal with a held lock without an owner field. Furthermore, it becomes even more of a problem when trying to fix starvation cases in the current code. We end up stacking special case on special case. To solve this rework the basic mutex implementation to be a single atomic word that contains the owner and uses the low bits for extra state. This matches how PI futexes and rt_mutex already work. By having the owner an integral part of the lock state a lot of the problems dissapear and we get a better option to deal with starvation cases, direct owner handoff. Changing the basic mutex does however invalidate all the arch specific mutex code; this patch leaves that unused in-place, a later patch will remove that. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Will Deacon <will.deacon@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 19:36:04 +08:00
*/
#define MUTEX_FLAG_WAITERS 0x01
locking/mutex: Add lock handoff to avoid starvation Implement lock handoff to avoid lock starvation. Lock starvation is possible because mutex_lock() allows lock stealing, where a running (or optimistic spinning) task beats the woken waiter to the acquire. Lock stealing is an important performance optimization because waiting for a waiter to wake up and get runtime can take a significant time, during which everyboy would stall on the lock. The down-side is of course that it allows for starvation. This patch has the waiter requesting a handoff if it fails to acquire the lock upon waking. This re-introduces some of the wait time, because once we do a handoff we have to wait for the waiter to wake up again. A future patch will add a round of optimistic spinning to attempt to alleviate this penalty, but if that turns out to not be enough, we can add a counter and only request handoff after multiple failed wakeups. There are a few tricky implementation details: - accepting a handoff must only be done in the wait-loop. Since the handoff condition is owner == current, it can easily cause recursive locking trouble. - accepting the handoff must be careful to provide the ACQUIRE semantics. - having the HANDOFF bit set on unlock requires care, we must not clear the owner. - we must be careful to not leave HANDOFF set after we've acquired the lock. The tricky scenario is setting the HANDOFF bit on an unlocked mutex. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Waiman Long <Waiman.Long@hpe.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 20:40:16 +08:00
#define MUTEX_FLAG_HANDOFF 0x02
locking/mutex: Rework mutex::owner The current mutex implementation has an atomic lock word and a non-atomic owner field. This disparity leads to a number of issues with the current mutex code as it means that we can have a locked mutex without an explicit owner (because the owner field has not been set, or already cleared). This leads to a number of weird corner cases, esp. between the optimistic spinning and debug code. Where the optimistic spinning code needs the owner field updated inside the lock region, the debug code is more relaxed because the whole lock is serialized by the wait_lock. Also, the spinning code itself has a few corner cases where we need to deal with a held lock without an owner field. Furthermore, it becomes even more of a problem when trying to fix starvation cases in the current code. We end up stacking special case on special case. To solve this rework the basic mutex implementation to be a single atomic word that contains the owner and uses the low bits for extra state. This matches how PI futexes and rt_mutex already work. By having the owner an integral part of the lock state a lot of the problems dissapear and we get a better option to deal with starvation cases, direct owner handoff. Changing the basic mutex does however invalidate all the arch specific mutex code; this patch leaves that unused in-place, a later patch will remove that. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Will Deacon <will.deacon@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 19:36:04 +08:00
#define MUTEX_FLAGS 0x03
static inline struct task_struct *__owner_task(unsigned long owner)
{
return (struct task_struct *)(owner & ~MUTEX_FLAGS);
}
static inline unsigned long __owner_flags(unsigned long owner)
{
return owner & MUTEX_FLAGS;
}
/*
* Actual trylock that will work on any unlocked state.
locking/mutex: Add lock handoff to avoid starvation Implement lock handoff to avoid lock starvation. Lock starvation is possible because mutex_lock() allows lock stealing, where a running (or optimistic spinning) task beats the woken waiter to the acquire. Lock stealing is an important performance optimization because waiting for a waiter to wake up and get runtime can take a significant time, during which everyboy would stall on the lock. The down-side is of course that it allows for starvation. This patch has the waiter requesting a handoff if it fails to acquire the lock upon waking. This re-introduces some of the wait time, because once we do a handoff we have to wait for the waiter to wake up again. A future patch will add a round of optimistic spinning to attempt to alleviate this penalty, but if that turns out to not be enough, we can add a counter and only request handoff after multiple failed wakeups. There are a few tricky implementation details: - accepting a handoff must only be done in the wait-loop. Since the handoff condition is owner == current, it can easily cause recursive locking trouble. - accepting the handoff must be careful to provide the ACQUIRE semantics. - having the HANDOFF bit set on unlock requires care, we must not clear the owner. - we must be careful to not leave HANDOFF set after we've acquired the lock. The tricky scenario is setting the HANDOFF bit on an unlocked mutex. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Waiman Long <Waiman.Long@hpe.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 20:40:16 +08:00
*
* When setting the owner field, we must preserve the low flag bits.
*
* Be careful with @handoff, only set that in a wait-loop (where you set
* HANDOFF) to avoid recursive lock attempts.
locking/mutex: Rework mutex::owner The current mutex implementation has an atomic lock word and a non-atomic owner field. This disparity leads to a number of issues with the current mutex code as it means that we can have a locked mutex without an explicit owner (because the owner field has not been set, or already cleared). This leads to a number of weird corner cases, esp. between the optimistic spinning and debug code. Where the optimistic spinning code needs the owner field updated inside the lock region, the debug code is more relaxed because the whole lock is serialized by the wait_lock. Also, the spinning code itself has a few corner cases where we need to deal with a held lock without an owner field. Furthermore, it becomes even more of a problem when trying to fix starvation cases in the current code. We end up stacking special case on special case. To solve this rework the basic mutex implementation to be a single atomic word that contains the owner and uses the low bits for extra state. This matches how PI futexes and rt_mutex already work. By having the owner an integral part of the lock state a lot of the problems dissapear and we get a better option to deal with starvation cases, direct owner handoff. Changing the basic mutex does however invalidate all the arch specific mutex code; this patch leaves that unused in-place, a later patch will remove that. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Will Deacon <will.deacon@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 19:36:04 +08:00
*/
locking/mutex: Add lock handoff to avoid starvation Implement lock handoff to avoid lock starvation. Lock starvation is possible because mutex_lock() allows lock stealing, where a running (or optimistic spinning) task beats the woken waiter to the acquire. Lock stealing is an important performance optimization because waiting for a waiter to wake up and get runtime can take a significant time, during which everyboy would stall on the lock. The down-side is of course that it allows for starvation. This patch has the waiter requesting a handoff if it fails to acquire the lock upon waking. This re-introduces some of the wait time, because once we do a handoff we have to wait for the waiter to wake up again. A future patch will add a round of optimistic spinning to attempt to alleviate this penalty, but if that turns out to not be enough, we can add a counter and only request handoff after multiple failed wakeups. There are a few tricky implementation details: - accepting a handoff must only be done in the wait-loop. Since the handoff condition is owner == current, it can easily cause recursive locking trouble. - accepting the handoff must be careful to provide the ACQUIRE semantics. - having the HANDOFF bit set on unlock requires care, we must not clear the owner. - we must be careful to not leave HANDOFF set after we've acquired the lock. The tricky scenario is setting the HANDOFF bit on an unlocked mutex. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Waiman Long <Waiman.Long@hpe.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 20:40:16 +08:00
static inline bool __mutex_trylock(struct mutex *lock, const bool handoff)
locking/mutex: Rework mutex::owner The current mutex implementation has an atomic lock word and a non-atomic owner field. This disparity leads to a number of issues with the current mutex code as it means that we can have a locked mutex without an explicit owner (because the owner field has not been set, or already cleared). This leads to a number of weird corner cases, esp. between the optimistic spinning and debug code. Where the optimistic spinning code needs the owner field updated inside the lock region, the debug code is more relaxed because the whole lock is serialized by the wait_lock. Also, the spinning code itself has a few corner cases where we need to deal with a held lock without an owner field. Furthermore, it becomes even more of a problem when trying to fix starvation cases in the current code. We end up stacking special case on special case. To solve this rework the basic mutex implementation to be a single atomic word that contains the owner and uses the low bits for extra state. This matches how PI futexes and rt_mutex already work. By having the owner an integral part of the lock state a lot of the problems dissapear and we get a better option to deal with starvation cases, direct owner handoff. Changing the basic mutex does however invalidate all the arch specific mutex code; this patch leaves that unused in-place, a later patch will remove that. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Will Deacon <will.deacon@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 19:36:04 +08:00
{
unsigned long owner, curr = (unsigned long)current;
owner = atomic_long_read(&lock->owner);
for (;;) { /* must loop, can race against a flag */
locking/mutex: Add lock handoff to avoid starvation Implement lock handoff to avoid lock starvation. Lock starvation is possible because mutex_lock() allows lock stealing, where a running (or optimistic spinning) task beats the woken waiter to the acquire. Lock stealing is an important performance optimization because waiting for a waiter to wake up and get runtime can take a significant time, during which everyboy would stall on the lock. The down-side is of course that it allows for starvation. This patch has the waiter requesting a handoff if it fails to acquire the lock upon waking. This re-introduces some of the wait time, because once we do a handoff we have to wait for the waiter to wake up again. A future patch will add a round of optimistic spinning to attempt to alleviate this penalty, but if that turns out to not be enough, we can add a counter and only request handoff after multiple failed wakeups. There are a few tricky implementation details: - accepting a handoff must only be done in the wait-loop. Since the handoff condition is owner == current, it can easily cause recursive locking trouble. - accepting the handoff must be careful to provide the ACQUIRE semantics. - having the HANDOFF bit set on unlock requires care, we must not clear the owner. - we must be careful to not leave HANDOFF set after we've acquired the lock. The tricky scenario is setting the HANDOFF bit on an unlocked mutex. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Waiman Long <Waiman.Long@hpe.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 20:40:16 +08:00
unsigned long old, flags = __owner_flags(owner);
if (__owner_task(owner)) {
if (handoff && unlikely(__owner_task(owner) == current)) {
/*
* Provide ACQUIRE semantics for the lock-handoff.
*
* We cannot easily use load-acquire here, since
* the actual load is a failed cmpxchg, which
* doesn't imply any barriers.
*
* Also, this is a fairly unlikely scenario, and
* this contains the cost.
*/
smp_mb(); /* ACQUIRE */
return true;
}
locking/mutex: Rework mutex::owner The current mutex implementation has an atomic lock word and a non-atomic owner field. This disparity leads to a number of issues with the current mutex code as it means that we can have a locked mutex without an explicit owner (because the owner field has not been set, or already cleared). This leads to a number of weird corner cases, esp. between the optimistic spinning and debug code. Where the optimistic spinning code needs the owner field updated inside the lock region, the debug code is more relaxed because the whole lock is serialized by the wait_lock. Also, the spinning code itself has a few corner cases where we need to deal with a held lock without an owner field. Furthermore, it becomes even more of a problem when trying to fix starvation cases in the current code. We end up stacking special case on special case. To solve this rework the basic mutex implementation to be a single atomic word that contains the owner and uses the low bits for extra state. This matches how PI futexes and rt_mutex already work. By having the owner an integral part of the lock state a lot of the problems dissapear and we get a better option to deal with starvation cases, direct owner handoff. Changing the basic mutex does however invalidate all the arch specific mutex code; this patch leaves that unused in-place, a later patch will remove that. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Will Deacon <will.deacon@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 19:36:04 +08:00
return false;
locking/mutex: Add lock handoff to avoid starvation Implement lock handoff to avoid lock starvation. Lock starvation is possible because mutex_lock() allows lock stealing, where a running (or optimistic spinning) task beats the woken waiter to the acquire. Lock stealing is an important performance optimization because waiting for a waiter to wake up and get runtime can take a significant time, during which everyboy would stall on the lock. The down-side is of course that it allows for starvation. This patch has the waiter requesting a handoff if it fails to acquire the lock upon waking. This re-introduces some of the wait time, because once we do a handoff we have to wait for the waiter to wake up again. A future patch will add a round of optimistic spinning to attempt to alleviate this penalty, but if that turns out to not be enough, we can add a counter and only request handoff after multiple failed wakeups. There are a few tricky implementation details: - accepting a handoff must only be done in the wait-loop. Since the handoff condition is owner == current, it can easily cause recursive locking trouble. - accepting the handoff must be careful to provide the ACQUIRE semantics. - having the HANDOFF bit set on unlock requires care, we must not clear the owner. - we must be careful to not leave HANDOFF set after we've acquired the lock. The tricky scenario is setting the HANDOFF bit on an unlocked mutex. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Waiman Long <Waiman.Long@hpe.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 20:40:16 +08:00
}
/*
* We set the HANDOFF bit, we must make sure it doesn't live
* past the point where we acquire it. This would be possible
* if we (accidentally) set the bit on an unlocked mutex.
*/
if (handoff)
flags &= ~MUTEX_FLAG_HANDOFF;
locking/mutex: Rework mutex::owner The current mutex implementation has an atomic lock word and a non-atomic owner field. This disparity leads to a number of issues with the current mutex code as it means that we can have a locked mutex without an explicit owner (because the owner field has not been set, or already cleared). This leads to a number of weird corner cases, esp. between the optimistic spinning and debug code. Where the optimistic spinning code needs the owner field updated inside the lock region, the debug code is more relaxed because the whole lock is serialized by the wait_lock. Also, the spinning code itself has a few corner cases where we need to deal with a held lock without an owner field. Furthermore, it becomes even more of a problem when trying to fix starvation cases in the current code. We end up stacking special case on special case. To solve this rework the basic mutex implementation to be a single atomic word that contains the owner and uses the low bits for extra state. This matches how PI futexes and rt_mutex already work. By having the owner an integral part of the lock state a lot of the problems dissapear and we get a better option to deal with starvation cases, direct owner handoff. Changing the basic mutex does however invalidate all the arch specific mutex code; this patch leaves that unused in-place, a later patch will remove that. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Will Deacon <will.deacon@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 19:36:04 +08:00
locking/mutex: Add lock handoff to avoid starvation Implement lock handoff to avoid lock starvation. Lock starvation is possible because mutex_lock() allows lock stealing, where a running (or optimistic spinning) task beats the woken waiter to the acquire. Lock stealing is an important performance optimization because waiting for a waiter to wake up and get runtime can take a significant time, during which everyboy would stall on the lock. The down-side is of course that it allows for starvation. This patch has the waiter requesting a handoff if it fails to acquire the lock upon waking. This re-introduces some of the wait time, because once we do a handoff we have to wait for the waiter to wake up again. A future patch will add a round of optimistic spinning to attempt to alleviate this penalty, but if that turns out to not be enough, we can add a counter and only request handoff after multiple failed wakeups. There are a few tricky implementation details: - accepting a handoff must only be done in the wait-loop. Since the handoff condition is owner == current, it can easily cause recursive locking trouble. - accepting the handoff must be careful to provide the ACQUIRE semantics. - having the HANDOFF bit set on unlock requires care, we must not clear the owner. - we must be careful to not leave HANDOFF set after we've acquired the lock. The tricky scenario is setting the HANDOFF bit on an unlocked mutex. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Waiman Long <Waiman.Long@hpe.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 20:40:16 +08:00
old = atomic_long_cmpxchg_acquire(&lock->owner, owner, curr | flags);
locking/mutex: Rework mutex::owner The current mutex implementation has an atomic lock word and a non-atomic owner field. This disparity leads to a number of issues with the current mutex code as it means that we can have a locked mutex without an explicit owner (because the owner field has not been set, or already cleared). This leads to a number of weird corner cases, esp. between the optimistic spinning and debug code. Where the optimistic spinning code needs the owner field updated inside the lock region, the debug code is more relaxed because the whole lock is serialized by the wait_lock. Also, the spinning code itself has a few corner cases where we need to deal with a held lock without an owner field. Furthermore, it becomes even more of a problem when trying to fix starvation cases in the current code. We end up stacking special case on special case. To solve this rework the basic mutex implementation to be a single atomic word that contains the owner and uses the low bits for extra state. This matches how PI futexes and rt_mutex already work. By having the owner an integral part of the lock state a lot of the problems dissapear and we get a better option to deal with starvation cases, direct owner handoff. Changing the basic mutex does however invalidate all the arch specific mutex code; this patch leaves that unused in-place, a later patch will remove that. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Will Deacon <will.deacon@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 19:36:04 +08:00
if (old == owner)
return true;
owner = old;
}
}
#ifndef CONFIG_DEBUG_LOCK_ALLOC
/*
* Lockdep annotations are contained to the slow paths for simplicity.
* There is nothing that would stop spreading the lockdep annotations outwards
* except more code.
*/
/*
* Optimistic trylock that only works in the uncontended case. Make sure to
* follow with a __mutex_trylock() before failing.
*/
static __always_inline bool __mutex_trylock_fast(struct mutex *lock)
{
unsigned long curr = (unsigned long)current;
if (!atomic_long_cmpxchg_acquire(&lock->owner, 0UL, curr))
return true;
return false;
}
static __always_inline bool __mutex_unlock_fast(struct mutex *lock)
{
unsigned long curr = (unsigned long)current;
if (atomic_long_cmpxchg_release(&lock->owner, curr, 0UL) == curr)
return true;
return false;
}
#endif
static inline void __mutex_set_flag(struct mutex *lock, unsigned long flag)
{
atomic_long_or(flag, &lock->owner);
}
static inline void __mutex_clear_flag(struct mutex *lock, unsigned long flag)
{
atomic_long_andnot(flag, &lock->owner);
}
locking/mutex: Add lock handoff to avoid starvation Implement lock handoff to avoid lock starvation. Lock starvation is possible because mutex_lock() allows lock stealing, where a running (or optimistic spinning) task beats the woken waiter to the acquire. Lock stealing is an important performance optimization because waiting for a waiter to wake up and get runtime can take a significant time, during which everyboy would stall on the lock. The down-side is of course that it allows for starvation. This patch has the waiter requesting a handoff if it fails to acquire the lock upon waking. This re-introduces some of the wait time, because once we do a handoff we have to wait for the waiter to wake up again. A future patch will add a round of optimistic spinning to attempt to alleviate this penalty, but if that turns out to not be enough, we can add a counter and only request handoff after multiple failed wakeups. There are a few tricky implementation details: - accepting a handoff must only be done in the wait-loop. Since the handoff condition is owner == current, it can easily cause recursive locking trouble. - accepting the handoff must be careful to provide the ACQUIRE semantics. - having the HANDOFF bit set on unlock requires care, we must not clear the owner. - we must be careful to not leave HANDOFF set after we've acquired the lock. The tricky scenario is setting the HANDOFF bit on an unlocked mutex. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Waiman Long <Waiman.Long@hpe.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 20:40:16 +08:00
static inline bool __mutex_waiter_is_first(struct mutex *lock, struct mutex_waiter *waiter)
{
return list_first_entry(&lock->wait_list, struct mutex_waiter, list) == waiter;
}
/*
* Give up ownership to a specific task, when @task = NULL, this is equivalent
* to a regular unlock. Clears HANDOFF, preserves WAITERS. Provides RELEASE
* semantics like a regular unlock, the __mutex_trylock() provides matching
* ACQUIRE semantics for the handoff.
*/
static void __mutex_handoff(struct mutex *lock, struct task_struct *task)
{
unsigned long owner = atomic_long_read(&lock->owner);
for (;;) {
unsigned long old, new;
#ifdef CONFIG_DEBUG_MUTEXES
DEBUG_LOCKS_WARN_ON(__owner_task(owner) != current);
#endif
new = (owner & MUTEX_FLAG_WAITERS);
new |= (unsigned long)task;
old = atomic_long_cmpxchg_release(&lock->owner, owner, new);
if (old == owner)
break;
owner = old;
}
}
#ifndef CONFIG_DEBUG_LOCK_ALLOC
/*
* We split the mutex lock/unlock logic into separate fastpath and
* slowpath functions, to reduce the register pressure on the fastpath.
* We also put the fastpath first in the kernel image, to make sure the
* branch is predicted by the CPU as default-untaken.
*/
locking/mutex: Rework mutex::owner The current mutex implementation has an atomic lock word and a non-atomic owner field. This disparity leads to a number of issues with the current mutex code as it means that we can have a locked mutex without an explicit owner (because the owner field has not been set, or already cleared). This leads to a number of weird corner cases, esp. between the optimistic spinning and debug code. Where the optimistic spinning code needs the owner field updated inside the lock region, the debug code is more relaxed because the whole lock is serialized by the wait_lock. Also, the spinning code itself has a few corner cases where we need to deal with a held lock without an owner field. Furthermore, it becomes even more of a problem when trying to fix starvation cases in the current code. We end up stacking special case on special case. To solve this rework the basic mutex implementation to be a single atomic word that contains the owner and uses the low bits for extra state. This matches how PI futexes and rt_mutex already work. By having the owner an integral part of the lock state a lot of the problems dissapear and we get a better option to deal with starvation cases, direct owner handoff. Changing the basic mutex does however invalidate all the arch specific mutex code; this patch leaves that unused in-place, a later patch will remove that. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Will Deacon <will.deacon@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 19:36:04 +08:00
static void __sched __mutex_lock_slowpath(struct mutex *lock);
/**
* mutex_lock - acquire the mutex
* @lock: the mutex to be acquired
*
* Lock the mutex exclusively for this task. If the mutex is not
* available right now, it will sleep until it can get it.
*
* The mutex must later on be released by the same task that
* acquired it. Recursive locking is not allowed. The task
* may not exit without first unlocking the mutex. Also, kernel
* memory where the mutex resides must not be freed with
* the mutex still locked. The mutex must first be initialized
* (or statically defined) before it can be locked. memset()-ing
* the mutex to 0 is not allowed.
*
* ( The CONFIG_DEBUG_MUTEXES .config option turns on debugging
* checks that will enforce the restrictions and will also do
* deadlock debugging. )
*
* This function is similar to (but not equivalent to) down().
*/
void __sched mutex_lock(struct mutex *lock)
{
might_sleep();
locking/mutex: Rework mutex::owner The current mutex implementation has an atomic lock word and a non-atomic owner field. This disparity leads to a number of issues with the current mutex code as it means that we can have a locked mutex without an explicit owner (because the owner field has not been set, or already cleared). This leads to a number of weird corner cases, esp. between the optimistic spinning and debug code. Where the optimistic spinning code needs the owner field updated inside the lock region, the debug code is more relaxed because the whole lock is serialized by the wait_lock. Also, the spinning code itself has a few corner cases where we need to deal with a held lock without an owner field. Furthermore, it becomes even more of a problem when trying to fix starvation cases in the current code. We end up stacking special case on special case. To solve this rework the basic mutex implementation to be a single atomic word that contains the owner and uses the low bits for extra state. This matches how PI futexes and rt_mutex already work. By having the owner an integral part of the lock state a lot of the problems dissapear and we get a better option to deal with starvation cases, direct owner handoff. Changing the basic mutex does however invalidate all the arch specific mutex code; this patch leaves that unused in-place, a later patch will remove that. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Will Deacon <will.deacon@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 19:36:04 +08:00
if (!__mutex_trylock_fast(lock))
__mutex_lock_slowpath(lock);
}
EXPORT_SYMBOL(mutex_lock);
#endif
static __always_inline void ww_mutex_lock_acquired(struct ww_mutex *ww,
struct ww_acquire_ctx *ww_ctx)
{
#ifdef CONFIG_DEBUG_MUTEXES
/*
* If this WARN_ON triggers, you used ww_mutex_lock to acquire,
* but released with a normal mutex_unlock in this call.
*
* This should never happen, always use ww_mutex_unlock.
*/
DEBUG_LOCKS_WARN_ON(ww->ctx);
/*
* Not quite done after calling ww_acquire_done() ?
*/
DEBUG_LOCKS_WARN_ON(ww_ctx->done_acquire);
if (ww_ctx->contending_lock) {
/*
* After -EDEADLK you tried to
* acquire a different ww_mutex? Bad!
*/
DEBUG_LOCKS_WARN_ON(ww_ctx->contending_lock != ww);
/*
* You called ww_mutex_lock after receiving -EDEADLK,
* but 'forgot' to unlock everything else first?
*/
DEBUG_LOCKS_WARN_ON(ww_ctx->acquired > 0);
ww_ctx->contending_lock = NULL;
}
/*
* Naughty, using a different class will lead to undefined behavior!
*/
DEBUG_LOCKS_WARN_ON(ww_ctx->ww_class != ww->ww_class);
#endif
ww_ctx->acquired++;
}
/*
* After acquiring lock with fastpath or when we lost out in contested
* slowpath, set ctx and wake up any waiters so they can recheck.
*/
static __always_inline void
ww_mutex_set_context_fastpath(struct ww_mutex *lock,
struct ww_acquire_ctx *ctx)
{
unsigned long flags;
struct mutex_waiter *cur;
ww_mutex_lock_acquired(lock, ctx);
lock->ctx = ctx;
/*
* The lock->ctx update should be visible on all cores before
* the atomic read is done, otherwise contended waiters might be
* missed. The contended waiters will either see ww_ctx == NULL
* and keep spinning, or it will acquire wait_lock, add itself
* to waiter list and sleep.
*/
smp_mb(); /* ^^^ */
/*
* Check if lock is contended, if not there is nobody to wake up
*/
locking/mutex: Rework mutex::owner The current mutex implementation has an atomic lock word and a non-atomic owner field. This disparity leads to a number of issues with the current mutex code as it means that we can have a locked mutex without an explicit owner (because the owner field has not been set, or already cleared). This leads to a number of weird corner cases, esp. between the optimistic spinning and debug code. Where the optimistic spinning code needs the owner field updated inside the lock region, the debug code is more relaxed because the whole lock is serialized by the wait_lock. Also, the spinning code itself has a few corner cases where we need to deal with a held lock without an owner field. Furthermore, it becomes even more of a problem when trying to fix starvation cases in the current code. We end up stacking special case on special case. To solve this rework the basic mutex implementation to be a single atomic word that contains the owner and uses the low bits for extra state. This matches how PI futexes and rt_mutex already work. By having the owner an integral part of the lock state a lot of the problems dissapear and we get a better option to deal with starvation cases, direct owner handoff. Changing the basic mutex does however invalidate all the arch specific mutex code; this patch leaves that unused in-place, a later patch will remove that. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Will Deacon <will.deacon@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 19:36:04 +08:00
if (likely(!(atomic_long_read(&lock->base.owner) & MUTEX_FLAG_WAITERS)))
return;
/*
* Uh oh, we raced in fastpath, wake up everyone in this case,
* so they can see the new lock->ctx.
*/
spin_lock_mutex(&lock->base.wait_lock, flags);
list_for_each_entry(cur, &lock->base.wait_list, list) {
debug_mutex_wake_waiter(&lock->base, cur);
wake_up_process(cur->task);
}
spin_unlock_mutex(&lock->base.wait_lock, flags);
}
/*
* After acquiring lock in the slowpath set ctx and wake up any
* waiters so they can recheck.
*
* Callers must hold the mutex wait_lock.
*/
static __always_inline void
ww_mutex_set_context_slowpath(struct ww_mutex *lock,
struct ww_acquire_ctx *ctx)
{
struct mutex_waiter *cur;
ww_mutex_lock_acquired(lock, ctx);
lock->ctx = ctx;
/*
* Give any possible sleeping processes the chance to wake up,
* so they can recheck if they have to back off.
*/
list_for_each_entry(cur, &lock->base.wait_list, list) {
debug_mutex_wake_waiter(&lock->base, cur);
wake_up_process(cur->task);
}
}
#ifdef CONFIG_MUTEX_SPIN_ON_OWNER
/*
* Look out! "owner" is an entirely speculative pointer
* access and not reliable.
*/
static noinline
bool mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
{
bool ret = true;
rcu_read_lock();
locking/mutex: Rework mutex::owner The current mutex implementation has an atomic lock word and a non-atomic owner field. This disparity leads to a number of issues with the current mutex code as it means that we can have a locked mutex without an explicit owner (because the owner field has not been set, or already cleared). This leads to a number of weird corner cases, esp. between the optimistic spinning and debug code. Where the optimistic spinning code needs the owner field updated inside the lock region, the debug code is more relaxed because the whole lock is serialized by the wait_lock. Also, the spinning code itself has a few corner cases where we need to deal with a held lock without an owner field. Furthermore, it becomes even more of a problem when trying to fix starvation cases in the current code. We end up stacking special case on special case. To solve this rework the basic mutex implementation to be a single atomic word that contains the owner and uses the low bits for extra state. This matches how PI futexes and rt_mutex already work. By having the owner an integral part of the lock state a lot of the problems dissapear and we get a better option to deal with starvation cases, direct owner handoff. Changing the basic mutex does however invalidate all the arch specific mutex code; this patch leaves that unused in-place, a later patch will remove that. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Will Deacon <will.deacon@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 19:36:04 +08:00
while (__mutex_owner(lock) == owner) {
/*
* Ensure we emit the owner->on_cpu, dereference _after_
* checking lock->owner still matches owner. If that fails,
* owner might point to freed memory. If it still matches,
* the rcu_read_lock() ensures the memory stays valid.
*/
barrier();
if (!owner->on_cpu || need_resched()) {
ret = false;
break;
}
arch, locking: Ciao arch_mutex_cpu_relax() The arch_mutex_cpu_relax() function, introduced by 34b133f, is hacky and ugly. It was added a few years ago to address the fact that common cpu_relax() calls include yielding on s390, and thus impact the optimistic spinning functionality of mutexes. Nowadays we use this function well beyond mutexes: rwsem, qrwlock, mcs and lockref. Since the macro that defines the call is in the mutex header, any users must include mutex.h and the naming is misleading as well. This patch (i) renames the call to cpu_relax_lowlatency ("relax, but only if you can do it with very low latency") and (ii) defines it in each arch's asm/processor.h local header, just like for regular cpu_relax functions. On all archs, except s390, cpu_relax_lowlatency is simply cpu_relax, and thus we can take it out of mutex.h. While this can seem redundant, I believe it is a good choice as it allows us to move out arch specific logic from generic locking primitives and enables future(?) archs to transparently define it, similarly to System Z. Signed-off-by: Davidlohr Bueso <davidlohr@hp.com> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Anton Blanchard <anton@samba.org> Cc: Aurelien Jacquiot <a-jacquiot@ti.com> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Bharat Bhushan <r65777@freescale.com> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Chen Liqin <liqin.linux@gmail.com> Cc: Chris Metcalf <cmetcalf@tilera.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Chris Zankel <chris@zankel.net> Cc: David Howells <dhowells@redhat.com> Cc: David S. Miller <davem@davemloft.net> Cc: Deepthi Dharwar <deepthi@linux.vnet.ibm.com> Cc: Dominik Dingel <dingel@linux.vnet.ibm.com> Cc: Fenghua Yu <fenghua.yu@intel.com> Cc: Geert Uytterhoeven <geert@linux-m68k.org> Cc: Guan Xuetao <gxt@mprc.pku.edu.cn> Cc: Haavard Skinnemoen <hskinnemoen@gmail.com> Cc: Hans-Christian Egtvedt <egtvedt@samfundet.no> Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: Helge Deller <deller@gmx.de> Cc: Hirokazu Takata <takata@linux-m32r.org> Cc: Ivan Kokshaysky <ink@jurassic.park.msu.ru> Cc: James E.J. Bottomley <jejb@parisc-linux.org> Cc: James Hogan <james.hogan@imgtec.com> Cc: Jason Wang <jasowang@redhat.com> Cc: Jesper Nilsson <jesper.nilsson@axis.com> Cc: Joe Perches <joe@perches.com> Cc: Jonas Bonn <jonas@southpole.se> Cc: Joseph Myers <joseph@codesourcery.com> Cc: Kees Cook <keescook@chromium.org> Cc: Koichi Yasutake <yasutake.koichi@jp.panasonic.com> Cc: Lennox Wu <lennox.wu@gmail.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mark Salter <msalter@redhat.com> Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Matt Turner <mattst88@gmail.com> Cc: Max Filippov <jcmvbkbc@gmail.com> Cc: Michael Neuling <mikey@neuling.org> Cc: Michal Simek <monstr@monstr.eu> Cc: Mikael Starvik <starvik@axis.com> Cc: Nicolas Pitre <nico@linaro.org> Cc: Paolo Bonzini <pbonzini@redhat.com> Cc: Paul Burton <paul.burton@imgtec.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Paul Mackerras <paulus@samba.org> Cc: Qais Yousef <qais.yousef@imgtec.com> Cc: Qiaowei Ren <qiaowei.ren@intel.com> Cc: Rafael Wysocki <rafael.j.wysocki@intel.com> Cc: Ralf Baechle <ralf@linux-mips.org> Cc: Richard Henderson <rth@twiddle.net> Cc: Richard Kuo <rkuo@codeaurora.org> Cc: Russell King <linux@arm.linux.org.uk> Cc: Steven Miao <realmz6@gmail.com> Cc: Steven Rostedt <srostedt@redhat.com> Cc: Stratos Karafotis <stratosk@semaphore.gr> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Tony Luck <tony.luck@intel.com> Cc: Vasily Kulikov <segoon@openwall.com> Cc: Vineet Gupta <vgupta@synopsys.com> Cc: Vineet Gupta <Vineet.Gupta1@synopsys.com> Cc: Waiman Long <Waiman.Long@hp.com> Cc: Will Deacon <will.deacon@arm.com> Cc: Wolfram Sang <wsa@the-dreams.de> Cc: adi-buildroot-devel@lists.sourceforge.net Cc: linux390@de.ibm.com Cc: linux-alpha@vger.kernel.org Cc: linux-am33-list@redhat.com Cc: linux-arm-kernel@lists.infradead.org Cc: linux-c6x-dev@linux-c6x.org Cc: linux-cris-kernel@axis.com Cc: linux-hexagon@vger.kernel.org Cc: linux-ia64@vger.kernel.org Cc: linux@lists.openrisc.net Cc: linux-m32r-ja@ml.linux-m32r.org Cc: linux-m32r@ml.linux-m32r.org Cc: linux-m68k@lists.linux-m68k.org Cc: linux-metag@vger.kernel.org Cc: linux-mips@linux-mips.org Cc: linux-parisc@vger.kernel.org Cc: linuxppc-dev@lists.ozlabs.org Cc: linux-s390@vger.kernel.org Cc: linux-sh@vger.kernel.org Cc: linux-xtensa@linux-xtensa.org Cc: sparclinux@vger.kernel.org Link: http://lkml.kernel.org/r/1404079773.2619.4.camel@buesod1.americas.hpqcorp.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2014-06-30 06:09:33 +08:00
cpu_relax_lowlatency();
}
rcu_read_unlock();
return ret;
}
mutex: Queue mutex spinners with MCS lock to reduce cacheline contention The current mutex spinning code (with MUTEX_SPIN_ON_OWNER option turned on) allow multiple tasks to spin on a single mutex concurrently. A potential problem with the current approach is that when the mutex becomes available, all the spinning tasks will try to acquire the mutex more or less simultaneously. As a result, there will be a lot of cacheline bouncing especially on systems with a large number of CPUs. This patch tries to reduce this kind of contention by putting the mutex spinners into a queue so that only the first one in the queue will try to acquire the mutex. This will reduce contention and allow all the tasks to move forward faster. The queuing of mutex spinners is done using an MCS lock based implementation which will further reduce contention on the mutex cacheline than a similar ticket spinlock based implementation. This patch will add a new field into the mutex data structure for holding the MCS lock. This expands the mutex size by 8 bytes for 64-bit system and 4 bytes for 32-bit system. This overhead will be avoid if the MUTEX_SPIN_ON_OWNER option is turned off. The following table shows the jobs per minute (JPM) scalability data on an 8-node 80-core Westmere box with a 3.7.10 kernel. The numactl command is used to restrict the running of the fserver workloads to 1/2/4/8 nodes with hyperthreading off. +-----------------+-----------+-----------+-------------+----------+ | Configuration | Mean JPM | Mean JPM | Mean JPM | % Change | | | w/o patch | patch 1 | patches 1&2 | 1->1&2 | +-----------------+------------------------------------------------+ | | User Range 1100 - 2000 | +-----------------+------------------------------------------------+ | 8 nodes, HT off | 227972 | 227237 | 305043 | +34.2% | | 4 nodes, HT off | 393503 | 381558 | 394650 | +3.4% | | 2 nodes, HT off | 334957 | 325240 | 338853 | +4.2% | | 1 node , HT off | 198141 | 197972 | 198075 | +0.1% | +-----------------+------------------------------------------------+ | | User Range 200 - 1000 | +-----------------+------------------------------------------------+ | 8 nodes, HT off | 282325 | 312870 | 332185 | +6.2% | | 4 nodes, HT off | 390698 | 378279 | 393419 | +4.0% | | 2 nodes, HT off | 336986 | 326543 | 340260 | +4.2% | | 1 node , HT off | 197588 | 197622 | 197582 | 0.0% | +-----------------+-----------+-----------+-------------+----------+ At low user range 10-100, the JPM differences were within +/-1%. So they are not that interesting. The fserver workload uses mutex spinning extensively. With just the mutex change in the first patch, there is no noticeable change in performance. Rather, there is a slight drop in performance. This mutex spinning patch more than recovers the lost performance and show a significant increase of +30% at high user load with the full 8 nodes. Similar improvements were also seen in a 3.8 kernel. The table below shows the %time spent by different kernel functions as reported by perf when running the fserver workload at 1500 users with all 8 nodes. +-----------------------+-----------+---------+-------------+ | Function | % time | % time | % time | | | w/o patch | patch 1 | patches 1&2 | +-----------------------+-----------+---------+-------------+ | __read_lock_failed | 34.96% | 34.91% | 29.14% | | __write_lock_failed | 10.14% | 10.68% | 7.51% | | mutex_spin_on_owner | 3.62% | 3.42% | 2.33% | | mspin_lock | N/A | N/A | 9.90% | | __mutex_lock_slowpath | 1.46% | 0.81% | 0.14% | | _raw_spin_lock | 2.25% | 2.50% | 1.10% | +-----------------------+-----------+---------+-------------+ The fserver workload for an 8-node system is dominated by the contention in the read/write lock. Mutex contention also plays a role. With the first patch only, mutex contention is down (as shown by the __mutex_lock_slowpath figure) which help a little bit. We saw only a few percents improvement with that. By applying patch 2 as well, the single mutex_spin_on_owner figure is now split out into an additional mspin_lock figure. The time increases from 3.42% to 11.23%. It shows a great reduction in contention among the spinners leading to a 30% improvement. The time ratio 9.9/2.33=4.3 indicates that there are on average 4+ spinners waiting in the spin_lock loop for each spinner in the mutex_spin_on_owner loop. Contention in other locking functions also go down by quite a lot. The table below shows the performance change of both patches 1 & 2 over patch 1 alone in other AIM7 workloads (at 8 nodes, hyperthreading off). +--------------+---------------+----------------+-----------------+ | Workload | mean % change | mean % change | mean % change | | | 10-100 users | 200-1000 users | 1100-2000 users | +--------------+---------------+----------------+-----------------+ | alltests | 0.0% | -0.8% | +0.6% | | five_sec | -0.3% | +0.8% | +0.8% | | high_systime | +0.4% | +2.4% | +2.1% | | new_fserver | +0.1% | +14.1% | +34.2% | | shared | -0.5% | -0.3% | -0.4% | | short | -1.7% | -9.8% | -8.3% | +--------------+---------------+----------------+-----------------+ The short workload is the only one that shows a decline in performance probably due to the spinner locking and queuing overhead. Signed-off-by: Waiman Long <Waiman.Long@hp.com> Reviewed-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Chandramouleeswaran Aswin <aswin@hp.com> Cc: Norton Scott J <scott.norton@hp.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: David Howells <dhowells@redhat.com> Cc: Dave Jones <davej@redhat.com> Cc: Clark Williams <williams@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1366226594-5506-4-git-send-email-Waiman.Long@hp.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-04-18 03:23:13 +08:00
/*
* Initial check for entering the mutex spinning loop
*/
static inline int mutex_can_spin_on_owner(struct mutex *lock)
{
struct task_struct *owner;
mutex: Queue mutex spinners with MCS lock to reduce cacheline contention The current mutex spinning code (with MUTEX_SPIN_ON_OWNER option turned on) allow multiple tasks to spin on a single mutex concurrently. A potential problem with the current approach is that when the mutex becomes available, all the spinning tasks will try to acquire the mutex more or less simultaneously. As a result, there will be a lot of cacheline bouncing especially on systems with a large number of CPUs. This patch tries to reduce this kind of contention by putting the mutex spinners into a queue so that only the first one in the queue will try to acquire the mutex. This will reduce contention and allow all the tasks to move forward faster. The queuing of mutex spinners is done using an MCS lock based implementation which will further reduce contention on the mutex cacheline than a similar ticket spinlock based implementation. This patch will add a new field into the mutex data structure for holding the MCS lock. This expands the mutex size by 8 bytes for 64-bit system and 4 bytes for 32-bit system. This overhead will be avoid if the MUTEX_SPIN_ON_OWNER option is turned off. The following table shows the jobs per minute (JPM) scalability data on an 8-node 80-core Westmere box with a 3.7.10 kernel. The numactl command is used to restrict the running of the fserver workloads to 1/2/4/8 nodes with hyperthreading off. +-----------------+-----------+-----------+-------------+----------+ | Configuration | Mean JPM | Mean JPM | Mean JPM | % Change | | | w/o patch | patch 1 | patches 1&2 | 1->1&2 | +-----------------+------------------------------------------------+ | | User Range 1100 - 2000 | +-----------------+------------------------------------------------+ | 8 nodes, HT off | 227972 | 227237 | 305043 | +34.2% | | 4 nodes, HT off | 393503 | 381558 | 394650 | +3.4% | | 2 nodes, HT off | 334957 | 325240 | 338853 | +4.2% | | 1 node , HT off | 198141 | 197972 | 198075 | +0.1% | +-----------------+------------------------------------------------+ | | User Range 200 - 1000 | +-----------------+------------------------------------------------+ | 8 nodes, HT off | 282325 | 312870 | 332185 | +6.2% | | 4 nodes, HT off | 390698 | 378279 | 393419 | +4.0% | | 2 nodes, HT off | 336986 | 326543 | 340260 | +4.2% | | 1 node , HT off | 197588 | 197622 | 197582 | 0.0% | +-----------------+-----------+-----------+-------------+----------+ At low user range 10-100, the JPM differences were within +/-1%. So they are not that interesting. The fserver workload uses mutex spinning extensively. With just the mutex change in the first patch, there is no noticeable change in performance. Rather, there is a slight drop in performance. This mutex spinning patch more than recovers the lost performance and show a significant increase of +30% at high user load with the full 8 nodes. Similar improvements were also seen in a 3.8 kernel. The table below shows the %time spent by different kernel functions as reported by perf when running the fserver workload at 1500 users with all 8 nodes. +-----------------------+-----------+---------+-------------+ | Function | % time | % time | % time | | | w/o patch | patch 1 | patches 1&2 | +-----------------------+-----------+---------+-------------+ | __read_lock_failed | 34.96% | 34.91% | 29.14% | | __write_lock_failed | 10.14% | 10.68% | 7.51% | | mutex_spin_on_owner | 3.62% | 3.42% | 2.33% | | mspin_lock | N/A | N/A | 9.90% | | __mutex_lock_slowpath | 1.46% | 0.81% | 0.14% | | _raw_spin_lock | 2.25% | 2.50% | 1.10% | +-----------------------+-----------+---------+-------------+ The fserver workload for an 8-node system is dominated by the contention in the read/write lock. Mutex contention also plays a role. With the first patch only, mutex contention is down (as shown by the __mutex_lock_slowpath figure) which help a little bit. We saw only a few percents improvement with that. By applying patch 2 as well, the single mutex_spin_on_owner figure is now split out into an additional mspin_lock figure. The time increases from 3.42% to 11.23%. It shows a great reduction in contention among the spinners leading to a 30% improvement. The time ratio 9.9/2.33=4.3 indicates that there are on average 4+ spinners waiting in the spin_lock loop for each spinner in the mutex_spin_on_owner loop. Contention in other locking functions also go down by quite a lot. The table below shows the performance change of both patches 1 & 2 over patch 1 alone in other AIM7 workloads (at 8 nodes, hyperthreading off). +--------------+---------------+----------------+-----------------+ | Workload | mean % change | mean % change | mean % change | | | 10-100 users | 200-1000 users | 1100-2000 users | +--------------+---------------+----------------+-----------------+ | alltests | 0.0% | -0.8% | +0.6% | | five_sec | -0.3% | +0.8% | +0.8% | | high_systime | +0.4% | +2.4% | +2.1% | | new_fserver | +0.1% | +14.1% | +34.2% | | shared | -0.5% | -0.3% | -0.4% | | short | -1.7% | -9.8% | -8.3% | +--------------+---------------+----------------+-----------------+ The short workload is the only one that shows a decline in performance probably due to the spinner locking and queuing overhead. Signed-off-by: Waiman Long <Waiman.Long@hp.com> Reviewed-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Chandramouleeswaran Aswin <aswin@hp.com> Cc: Norton Scott J <scott.norton@hp.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: David Howells <dhowells@redhat.com> Cc: Dave Jones <davej@redhat.com> Cc: Clark Williams <williams@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1366226594-5506-4-git-send-email-Waiman.Long@hp.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-04-18 03:23:13 +08:00
int retval = 1;
if (need_resched())
return 0;
mutex: Queue mutex spinners with MCS lock to reduce cacheline contention The current mutex spinning code (with MUTEX_SPIN_ON_OWNER option turned on) allow multiple tasks to spin on a single mutex concurrently. A potential problem with the current approach is that when the mutex becomes available, all the spinning tasks will try to acquire the mutex more or less simultaneously. As a result, there will be a lot of cacheline bouncing especially on systems with a large number of CPUs. This patch tries to reduce this kind of contention by putting the mutex spinners into a queue so that only the first one in the queue will try to acquire the mutex. This will reduce contention and allow all the tasks to move forward faster. The queuing of mutex spinners is done using an MCS lock based implementation which will further reduce contention on the mutex cacheline than a similar ticket spinlock based implementation. This patch will add a new field into the mutex data structure for holding the MCS lock. This expands the mutex size by 8 bytes for 64-bit system and 4 bytes for 32-bit system. This overhead will be avoid if the MUTEX_SPIN_ON_OWNER option is turned off. The following table shows the jobs per minute (JPM) scalability data on an 8-node 80-core Westmere box with a 3.7.10 kernel. The numactl command is used to restrict the running of the fserver workloads to 1/2/4/8 nodes with hyperthreading off. +-----------------+-----------+-----------+-------------+----------+ | Configuration | Mean JPM | Mean JPM | Mean JPM | % Change | | | w/o patch | patch 1 | patches 1&2 | 1->1&2 | +-----------------+------------------------------------------------+ | | User Range 1100 - 2000 | +-----------------+------------------------------------------------+ | 8 nodes, HT off | 227972 | 227237 | 305043 | +34.2% | | 4 nodes, HT off | 393503 | 381558 | 394650 | +3.4% | | 2 nodes, HT off | 334957 | 325240 | 338853 | +4.2% | | 1 node , HT off | 198141 | 197972 | 198075 | +0.1% | +-----------------+------------------------------------------------+ | | User Range 200 - 1000 | +-----------------+------------------------------------------------+ | 8 nodes, HT off | 282325 | 312870 | 332185 | +6.2% | | 4 nodes, HT off | 390698 | 378279 | 393419 | +4.0% | | 2 nodes, HT off | 336986 | 326543 | 340260 | +4.2% | | 1 node , HT off | 197588 | 197622 | 197582 | 0.0% | +-----------------+-----------+-----------+-------------+----------+ At low user range 10-100, the JPM differences were within +/-1%. So they are not that interesting. The fserver workload uses mutex spinning extensively. With just the mutex change in the first patch, there is no noticeable change in performance. Rather, there is a slight drop in performance. This mutex spinning patch more than recovers the lost performance and show a significant increase of +30% at high user load with the full 8 nodes. Similar improvements were also seen in a 3.8 kernel. The table below shows the %time spent by different kernel functions as reported by perf when running the fserver workload at 1500 users with all 8 nodes. +-----------------------+-----------+---------+-------------+ | Function | % time | % time | % time | | | w/o patch | patch 1 | patches 1&2 | +-----------------------+-----------+---------+-------------+ | __read_lock_failed | 34.96% | 34.91% | 29.14% | | __write_lock_failed | 10.14% | 10.68% | 7.51% | | mutex_spin_on_owner | 3.62% | 3.42% | 2.33% | | mspin_lock | N/A | N/A | 9.90% | | __mutex_lock_slowpath | 1.46% | 0.81% | 0.14% | | _raw_spin_lock | 2.25% | 2.50% | 1.10% | +-----------------------+-----------+---------+-------------+ The fserver workload for an 8-node system is dominated by the contention in the read/write lock. Mutex contention also plays a role. With the first patch only, mutex contention is down (as shown by the __mutex_lock_slowpath figure) which help a little bit. We saw only a few percents improvement with that. By applying patch 2 as well, the single mutex_spin_on_owner figure is now split out into an additional mspin_lock figure. The time increases from 3.42% to 11.23%. It shows a great reduction in contention among the spinners leading to a 30% improvement. The time ratio 9.9/2.33=4.3 indicates that there are on average 4+ spinners waiting in the spin_lock loop for each spinner in the mutex_spin_on_owner loop. Contention in other locking functions also go down by quite a lot. The table below shows the performance change of both patches 1 & 2 over patch 1 alone in other AIM7 workloads (at 8 nodes, hyperthreading off). +--------------+---------------+----------------+-----------------+ | Workload | mean % change | mean % change | mean % change | | | 10-100 users | 200-1000 users | 1100-2000 users | +--------------+---------------+----------------+-----------------+ | alltests | 0.0% | -0.8% | +0.6% | | five_sec | -0.3% | +0.8% | +0.8% | | high_systime | +0.4% | +2.4% | +2.1% | | new_fserver | +0.1% | +14.1% | +34.2% | | shared | -0.5% | -0.3% | -0.4% | | short | -1.7% | -9.8% | -8.3% | +--------------+---------------+----------------+-----------------+ The short workload is the only one that shows a decline in performance probably due to the spinner locking and queuing overhead. Signed-off-by: Waiman Long <Waiman.Long@hp.com> Reviewed-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Chandramouleeswaran Aswin <aswin@hp.com> Cc: Norton Scott J <scott.norton@hp.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: David Howells <dhowells@redhat.com> Cc: Dave Jones <davej@redhat.com> Cc: Clark Williams <williams@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1366226594-5506-4-git-send-email-Waiman.Long@hp.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-04-18 03:23:13 +08:00
rcu_read_lock();
locking/mutex: Rework mutex::owner The current mutex implementation has an atomic lock word and a non-atomic owner field. This disparity leads to a number of issues with the current mutex code as it means that we can have a locked mutex without an explicit owner (because the owner field has not been set, or already cleared). This leads to a number of weird corner cases, esp. between the optimistic spinning and debug code. Where the optimistic spinning code needs the owner field updated inside the lock region, the debug code is more relaxed because the whole lock is serialized by the wait_lock. Also, the spinning code itself has a few corner cases where we need to deal with a held lock without an owner field. Furthermore, it becomes even more of a problem when trying to fix starvation cases in the current code. We end up stacking special case on special case. To solve this rework the basic mutex implementation to be a single atomic word that contains the owner and uses the low bits for extra state. This matches how PI futexes and rt_mutex already work. By having the owner an integral part of the lock state a lot of the problems dissapear and we get a better option to deal with starvation cases, direct owner handoff. Changing the basic mutex does however invalidate all the arch specific mutex code; this patch leaves that unused in-place, a later patch will remove that. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Will Deacon <will.deacon@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 19:36:04 +08:00
owner = __mutex_owner(lock);
if (owner)
retval = owner->on_cpu;
mutex: Queue mutex spinners with MCS lock to reduce cacheline contention The current mutex spinning code (with MUTEX_SPIN_ON_OWNER option turned on) allow multiple tasks to spin on a single mutex concurrently. A potential problem with the current approach is that when the mutex becomes available, all the spinning tasks will try to acquire the mutex more or less simultaneously. As a result, there will be a lot of cacheline bouncing especially on systems with a large number of CPUs. This patch tries to reduce this kind of contention by putting the mutex spinners into a queue so that only the first one in the queue will try to acquire the mutex. This will reduce contention and allow all the tasks to move forward faster. The queuing of mutex spinners is done using an MCS lock based implementation which will further reduce contention on the mutex cacheline than a similar ticket spinlock based implementation. This patch will add a new field into the mutex data structure for holding the MCS lock. This expands the mutex size by 8 bytes for 64-bit system and 4 bytes for 32-bit system. This overhead will be avoid if the MUTEX_SPIN_ON_OWNER option is turned off. The following table shows the jobs per minute (JPM) scalability data on an 8-node 80-core Westmere box with a 3.7.10 kernel. The numactl command is used to restrict the running of the fserver workloads to 1/2/4/8 nodes with hyperthreading off. +-----------------+-----------+-----------+-------------+----------+ | Configuration | Mean JPM | Mean JPM | Mean JPM | % Change | | | w/o patch | patch 1 | patches 1&2 | 1->1&2 | +-----------------+------------------------------------------------+ | | User Range 1100 - 2000 | +-----------------+------------------------------------------------+ | 8 nodes, HT off | 227972 | 227237 | 305043 | +34.2% | | 4 nodes, HT off | 393503 | 381558 | 394650 | +3.4% | | 2 nodes, HT off | 334957 | 325240 | 338853 | +4.2% | | 1 node , HT off | 198141 | 197972 | 198075 | +0.1% | +-----------------+------------------------------------------------+ | | User Range 200 - 1000 | +-----------------+------------------------------------------------+ | 8 nodes, HT off | 282325 | 312870 | 332185 | +6.2% | | 4 nodes, HT off | 390698 | 378279 | 393419 | +4.0% | | 2 nodes, HT off | 336986 | 326543 | 340260 | +4.2% | | 1 node , HT off | 197588 | 197622 | 197582 | 0.0% | +-----------------+-----------+-----------+-------------+----------+ At low user range 10-100, the JPM differences were within +/-1%. So they are not that interesting. The fserver workload uses mutex spinning extensively. With just the mutex change in the first patch, there is no noticeable change in performance. Rather, there is a slight drop in performance. This mutex spinning patch more than recovers the lost performance and show a significant increase of +30% at high user load with the full 8 nodes. Similar improvements were also seen in a 3.8 kernel. The table below shows the %time spent by different kernel functions as reported by perf when running the fserver workload at 1500 users with all 8 nodes. +-----------------------+-----------+---------+-------------+ | Function | % time | % time | % time | | | w/o patch | patch 1 | patches 1&2 | +-----------------------+-----------+---------+-------------+ | __read_lock_failed | 34.96% | 34.91% | 29.14% | | __write_lock_failed | 10.14% | 10.68% | 7.51% | | mutex_spin_on_owner | 3.62% | 3.42% | 2.33% | | mspin_lock | N/A | N/A | 9.90% | | __mutex_lock_slowpath | 1.46% | 0.81% | 0.14% | | _raw_spin_lock | 2.25% | 2.50% | 1.10% | +-----------------------+-----------+---------+-------------+ The fserver workload for an 8-node system is dominated by the contention in the read/write lock. Mutex contention also plays a role. With the first patch only, mutex contention is down (as shown by the __mutex_lock_slowpath figure) which help a little bit. We saw only a few percents improvement with that. By applying patch 2 as well, the single mutex_spin_on_owner figure is now split out into an additional mspin_lock figure. The time increases from 3.42% to 11.23%. It shows a great reduction in contention among the spinners leading to a 30% improvement. The time ratio 9.9/2.33=4.3 indicates that there are on average 4+ spinners waiting in the spin_lock loop for each spinner in the mutex_spin_on_owner loop. Contention in other locking functions also go down by quite a lot. The table below shows the performance change of both patches 1 & 2 over patch 1 alone in other AIM7 workloads (at 8 nodes, hyperthreading off). +--------------+---------------+----------------+-----------------+ | Workload | mean % change | mean % change | mean % change | | | 10-100 users | 200-1000 users | 1100-2000 users | +--------------+---------------+----------------+-----------------+ | alltests | 0.0% | -0.8% | +0.6% | | five_sec | -0.3% | +0.8% | +0.8% | | high_systime | +0.4% | +2.4% | +2.1% | | new_fserver | +0.1% | +14.1% | +34.2% | | shared | -0.5% | -0.3% | -0.4% | | short | -1.7% | -9.8% | -8.3% | +--------------+---------------+----------------+-----------------+ The short workload is the only one that shows a decline in performance probably due to the spinner locking and queuing overhead. Signed-off-by: Waiman Long <Waiman.Long@hp.com> Reviewed-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Chandramouleeswaran Aswin <aswin@hp.com> Cc: Norton Scott J <scott.norton@hp.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: David Howells <dhowells@redhat.com> Cc: Dave Jones <davej@redhat.com> Cc: Clark Williams <williams@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1366226594-5506-4-git-send-email-Waiman.Long@hp.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-04-18 03:23:13 +08:00
rcu_read_unlock();
locking/mutex: Rework mutex::owner The current mutex implementation has an atomic lock word and a non-atomic owner field. This disparity leads to a number of issues with the current mutex code as it means that we can have a locked mutex without an explicit owner (because the owner field has not been set, or already cleared). This leads to a number of weird corner cases, esp. between the optimistic spinning and debug code. Where the optimistic spinning code needs the owner field updated inside the lock region, the debug code is more relaxed because the whole lock is serialized by the wait_lock. Also, the spinning code itself has a few corner cases where we need to deal with a held lock without an owner field. Furthermore, it becomes even more of a problem when trying to fix starvation cases in the current code. We end up stacking special case on special case. To solve this rework the basic mutex implementation to be a single atomic word that contains the owner and uses the low bits for extra state. This matches how PI futexes and rt_mutex already work. By having the owner an integral part of the lock state a lot of the problems dissapear and we get a better option to deal with starvation cases, direct owner handoff. Changing the basic mutex does however invalidate all the arch specific mutex code; this patch leaves that unused in-place, a later patch will remove that. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Will Deacon <will.deacon@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 19:36:04 +08:00
mutex: Queue mutex spinners with MCS lock to reduce cacheline contention The current mutex spinning code (with MUTEX_SPIN_ON_OWNER option turned on) allow multiple tasks to spin on a single mutex concurrently. A potential problem with the current approach is that when the mutex becomes available, all the spinning tasks will try to acquire the mutex more or less simultaneously. As a result, there will be a lot of cacheline bouncing especially on systems with a large number of CPUs. This patch tries to reduce this kind of contention by putting the mutex spinners into a queue so that only the first one in the queue will try to acquire the mutex. This will reduce contention and allow all the tasks to move forward faster. The queuing of mutex spinners is done using an MCS lock based implementation which will further reduce contention on the mutex cacheline than a similar ticket spinlock based implementation. This patch will add a new field into the mutex data structure for holding the MCS lock. This expands the mutex size by 8 bytes for 64-bit system and 4 bytes for 32-bit system. This overhead will be avoid if the MUTEX_SPIN_ON_OWNER option is turned off. The following table shows the jobs per minute (JPM) scalability data on an 8-node 80-core Westmere box with a 3.7.10 kernel. The numactl command is used to restrict the running of the fserver workloads to 1/2/4/8 nodes with hyperthreading off. +-----------------+-----------+-----------+-------------+----------+ | Configuration | Mean JPM | Mean JPM | Mean JPM | % Change | | | w/o patch | patch 1 | patches 1&2 | 1->1&2 | +-----------------+------------------------------------------------+ | | User Range 1100 - 2000 | +-----------------+------------------------------------------------+ | 8 nodes, HT off | 227972 | 227237 | 305043 | +34.2% | | 4 nodes, HT off | 393503 | 381558 | 394650 | +3.4% | | 2 nodes, HT off | 334957 | 325240 | 338853 | +4.2% | | 1 node , HT off | 198141 | 197972 | 198075 | +0.1% | +-----------------+------------------------------------------------+ | | User Range 200 - 1000 | +-----------------+------------------------------------------------+ | 8 nodes, HT off | 282325 | 312870 | 332185 | +6.2% | | 4 nodes, HT off | 390698 | 378279 | 393419 | +4.0% | | 2 nodes, HT off | 336986 | 326543 | 340260 | +4.2% | | 1 node , HT off | 197588 | 197622 | 197582 | 0.0% | +-----------------+-----------+-----------+-------------+----------+ At low user range 10-100, the JPM differences were within +/-1%. So they are not that interesting. The fserver workload uses mutex spinning extensively. With just the mutex change in the first patch, there is no noticeable change in performance. Rather, there is a slight drop in performance. This mutex spinning patch more than recovers the lost performance and show a significant increase of +30% at high user load with the full 8 nodes. Similar improvements were also seen in a 3.8 kernel. The table below shows the %time spent by different kernel functions as reported by perf when running the fserver workload at 1500 users with all 8 nodes. +-----------------------+-----------+---------+-------------+ | Function | % time | % time | % time | | | w/o patch | patch 1 | patches 1&2 | +-----------------------+-----------+---------+-------------+ | __read_lock_failed | 34.96% | 34.91% | 29.14% | | __write_lock_failed | 10.14% | 10.68% | 7.51% | | mutex_spin_on_owner | 3.62% | 3.42% | 2.33% | | mspin_lock | N/A | N/A | 9.90% | | __mutex_lock_slowpath | 1.46% | 0.81% | 0.14% | | _raw_spin_lock | 2.25% | 2.50% | 1.10% | +-----------------------+-----------+---------+-------------+ The fserver workload for an 8-node system is dominated by the contention in the read/write lock. Mutex contention also plays a role. With the first patch only, mutex contention is down (as shown by the __mutex_lock_slowpath figure) which help a little bit. We saw only a few percents improvement with that. By applying patch 2 as well, the single mutex_spin_on_owner figure is now split out into an additional mspin_lock figure. The time increases from 3.42% to 11.23%. It shows a great reduction in contention among the spinners leading to a 30% improvement. The time ratio 9.9/2.33=4.3 indicates that there are on average 4+ spinners waiting in the spin_lock loop for each spinner in the mutex_spin_on_owner loop. Contention in other locking functions also go down by quite a lot. The table below shows the performance change of both patches 1 & 2 over patch 1 alone in other AIM7 workloads (at 8 nodes, hyperthreading off). +--------------+---------------+----------------+-----------------+ | Workload | mean % change | mean % change | mean % change | | | 10-100 users | 200-1000 users | 1100-2000 users | +--------------+---------------+----------------+-----------------+ | alltests | 0.0% | -0.8% | +0.6% | | five_sec | -0.3% | +0.8% | +0.8% | | high_systime | +0.4% | +2.4% | +2.1% | | new_fserver | +0.1% | +14.1% | +34.2% | | shared | -0.5% | -0.3% | -0.4% | | short | -1.7% | -9.8% | -8.3% | +--------------+---------------+----------------+-----------------+ The short workload is the only one that shows a decline in performance probably due to the spinner locking and queuing overhead. Signed-off-by: Waiman Long <Waiman.Long@hp.com> Reviewed-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Chandramouleeswaran Aswin <aswin@hp.com> Cc: Norton Scott J <scott.norton@hp.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: David Howells <dhowells@redhat.com> Cc: Dave Jones <davej@redhat.com> Cc: Clark Williams <williams@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1366226594-5506-4-git-send-email-Waiman.Long@hp.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-04-18 03:23:13 +08:00
/*
locking/mutex: Rework mutex::owner The current mutex implementation has an atomic lock word and a non-atomic owner field. This disparity leads to a number of issues with the current mutex code as it means that we can have a locked mutex without an explicit owner (because the owner field has not been set, or already cleared). This leads to a number of weird corner cases, esp. between the optimistic spinning and debug code. Where the optimistic spinning code needs the owner field updated inside the lock region, the debug code is more relaxed because the whole lock is serialized by the wait_lock. Also, the spinning code itself has a few corner cases where we need to deal with a held lock without an owner field. Furthermore, it becomes even more of a problem when trying to fix starvation cases in the current code. We end up stacking special case on special case. To solve this rework the basic mutex implementation to be a single atomic word that contains the owner and uses the low bits for extra state. This matches how PI futexes and rt_mutex already work. By having the owner an integral part of the lock state a lot of the problems dissapear and we get a better option to deal with starvation cases, direct owner handoff. Changing the basic mutex does however invalidate all the arch specific mutex code; this patch leaves that unused in-place, a later patch will remove that. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Will Deacon <will.deacon@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 19:36:04 +08:00
* If lock->owner is not set, the mutex has been released. Return true
* such that we'll trylock in the spin path, which is a faster option
* than the blocking slow path.
mutex: Queue mutex spinners with MCS lock to reduce cacheline contention The current mutex spinning code (with MUTEX_SPIN_ON_OWNER option turned on) allow multiple tasks to spin on a single mutex concurrently. A potential problem with the current approach is that when the mutex becomes available, all the spinning tasks will try to acquire the mutex more or less simultaneously. As a result, there will be a lot of cacheline bouncing especially on systems with a large number of CPUs. This patch tries to reduce this kind of contention by putting the mutex spinners into a queue so that only the first one in the queue will try to acquire the mutex. This will reduce contention and allow all the tasks to move forward faster. The queuing of mutex spinners is done using an MCS lock based implementation which will further reduce contention on the mutex cacheline than a similar ticket spinlock based implementation. This patch will add a new field into the mutex data structure for holding the MCS lock. This expands the mutex size by 8 bytes for 64-bit system and 4 bytes for 32-bit system. This overhead will be avoid if the MUTEX_SPIN_ON_OWNER option is turned off. The following table shows the jobs per minute (JPM) scalability data on an 8-node 80-core Westmere box with a 3.7.10 kernel. The numactl command is used to restrict the running of the fserver workloads to 1/2/4/8 nodes with hyperthreading off. +-----------------+-----------+-----------+-------------+----------+ | Configuration | Mean JPM | Mean JPM | Mean JPM | % Change | | | w/o patch | patch 1 | patches 1&2 | 1->1&2 | +-----------------+------------------------------------------------+ | | User Range 1100 - 2000 | +-----------------+------------------------------------------------+ | 8 nodes, HT off | 227972 | 227237 | 305043 | +34.2% | | 4 nodes, HT off | 393503 | 381558 | 394650 | +3.4% | | 2 nodes, HT off | 334957 | 325240 | 338853 | +4.2% | | 1 node , HT off | 198141 | 197972 | 198075 | +0.1% | +-----------------+------------------------------------------------+ | | User Range 200 - 1000 | +-----------------+------------------------------------------------+ | 8 nodes, HT off | 282325 | 312870 | 332185 | +6.2% | | 4 nodes, HT off | 390698 | 378279 | 393419 | +4.0% | | 2 nodes, HT off | 336986 | 326543 | 340260 | +4.2% | | 1 node , HT off | 197588 | 197622 | 197582 | 0.0% | +-----------------+-----------+-----------+-------------+----------+ At low user range 10-100, the JPM differences were within +/-1%. So they are not that interesting. The fserver workload uses mutex spinning extensively. With just the mutex change in the first patch, there is no noticeable change in performance. Rather, there is a slight drop in performance. This mutex spinning patch more than recovers the lost performance and show a significant increase of +30% at high user load with the full 8 nodes. Similar improvements were also seen in a 3.8 kernel. The table below shows the %time spent by different kernel functions as reported by perf when running the fserver workload at 1500 users with all 8 nodes. +-----------------------+-----------+---------+-------------+ | Function | % time | % time | % time | | | w/o patch | patch 1 | patches 1&2 | +-----------------------+-----------+---------+-------------+ | __read_lock_failed | 34.96% | 34.91% | 29.14% | | __write_lock_failed | 10.14% | 10.68% | 7.51% | | mutex_spin_on_owner | 3.62% | 3.42% | 2.33% | | mspin_lock | N/A | N/A | 9.90% | | __mutex_lock_slowpath | 1.46% | 0.81% | 0.14% | | _raw_spin_lock | 2.25% | 2.50% | 1.10% | +-----------------------+-----------+---------+-------------+ The fserver workload for an 8-node system is dominated by the contention in the read/write lock. Mutex contention also plays a role. With the first patch only, mutex contention is down (as shown by the __mutex_lock_slowpath figure) which help a little bit. We saw only a few percents improvement with that. By applying patch 2 as well, the single mutex_spin_on_owner figure is now split out into an additional mspin_lock figure. The time increases from 3.42% to 11.23%. It shows a great reduction in contention among the spinners leading to a 30% improvement. The time ratio 9.9/2.33=4.3 indicates that there are on average 4+ spinners waiting in the spin_lock loop for each spinner in the mutex_spin_on_owner loop. Contention in other locking functions also go down by quite a lot. The table below shows the performance change of both patches 1 & 2 over patch 1 alone in other AIM7 workloads (at 8 nodes, hyperthreading off). +--------------+---------------+----------------+-----------------+ | Workload | mean % change | mean % change | mean % change | | | 10-100 users | 200-1000 users | 1100-2000 users | +--------------+---------------+----------------+-----------------+ | alltests | 0.0% | -0.8% | +0.6% | | five_sec | -0.3% | +0.8% | +0.8% | | high_systime | +0.4% | +2.4% | +2.1% | | new_fserver | +0.1% | +14.1% | +34.2% | | shared | -0.5% | -0.3% | -0.4% | | short | -1.7% | -9.8% | -8.3% | +--------------+---------------+----------------+-----------------+ The short workload is the only one that shows a decline in performance probably due to the spinner locking and queuing overhead. Signed-off-by: Waiman Long <Waiman.Long@hp.com> Reviewed-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Chandramouleeswaran Aswin <aswin@hp.com> Cc: Norton Scott J <scott.norton@hp.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: David Howells <dhowells@redhat.com> Cc: Dave Jones <davej@redhat.com> Cc: Clark Williams <williams@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1366226594-5506-4-git-send-email-Waiman.Long@hp.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-04-18 03:23:13 +08:00
*/
return retval;
}
/*
* Optimistic spinning.
*
* We try to spin for acquisition when we find that the lock owner
* is currently running on a (different) CPU and while we don't
* need to reschedule. The rationale is that if the lock owner is
* running, it is likely to release the lock soon.
*
* The mutex spinners are queued up using MCS lock so that only one
* spinner can compete for the mutex. However, if mutex spinning isn't
* going to happen, there is no point in going through the lock/unlock
* overhead.
*
* Returns true when the lock was taken, otherwise false, indicating
* that we need to jump to the slowpath and sleep.
*/
static bool mutex_optimistic_spin(struct mutex *lock,
struct ww_acquire_ctx *ww_ctx, const bool use_ww_ctx)
{
struct task_struct *task = current;
if (!mutex_can_spin_on_owner(lock))
goto done;
/*
* In order to avoid a stampede of mutex spinners trying to
* acquire the mutex all at once, the spinners need to take a
* MCS (queued) lock first before spinning on the owner field.
*/
if (!osq_lock(&lock->osq))
goto done;
while (true) {
struct task_struct *owner;
if (use_ww_ctx && ww_ctx->acquired > 0) {
struct ww_mutex *ww;
ww = container_of(lock, struct ww_mutex, base);
/*
* If ww->ctx is set the contents are undefined, only
* by acquiring wait_lock there is a guarantee that
* they are not invalid when reading.
*
* As such, when deadlock detection needs to be
* performed the optimistic spinning cannot be done.
*/
if (READ_ONCE(ww->ctx))
break;
}
/*
* If there's an owner, wait for it to either
* release the lock or go to sleep.
*/
locking/mutex: Rework mutex::owner The current mutex implementation has an atomic lock word and a non-atomic owner field. This disparity leads to a number of issues with the current mutex code as it means that we can have a locked mutex without an explicit owner (because the owner field has not been set, or already cleared). This leads to a number of weird corner cases, esp. between the optimistic spinning and debug code. Where the optimistic spinning code needs the owner field updated inside the lock region, the debug code is more relaxed because the whole lock is serialized by the wait_lock. Also, the spinning code itself has a few corner cases where we need to deal with a held lock without an owner field. Furthermore, it becomes even more of a problem when trying to fix starvation cases in the current code. We end up stacking special case on special case. To solve this rework the basic mutex implementation to be a single atomic word that contains the owner and uses the low bits for extra state. This matches how PI futexes and rt_mutex already work. By having the owner an integral part of the lock state a lot of the problems dissapear and we get a better option to deal with starvation cases, direct owner handoff. Changing the basic mutex does however invalidate all the arch specific mutex code; this patch leaves that unused in-place, a later patch will remove that. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Will Deacon <will.deacon@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 19:36:04 +08:00
owner = __mutex_owner(lock);
if (owner && !mutex_spin_on_owner(lock, owner))
break;
/* Try to acquire the mutex if it is unlocked. */
locking/mutex: Add lock handoff to avoid starvation Implement lock handoff to avoid lock starvation. Lock starvation is possible because mutex_lock() allows lock stealing, where a running (or optimistic spinning) task beats the woken waiter to the acquire. Lock stealing is an important performance optimization because waiting for a waiter to wake up and get runtime can take a significant time, during which everyboy would stall on the lock. The down-side is of course that it allows for starvation. This patch has the waiter requesting a handoff if it fails to acquire the lock upon waking. This re-introduces some of the wait time, because once we do a handoff we have to wait for the waiter to wake up again. A future patch will add a round of optimistic spinning to attempt to alleviate this penalty, but if that turns out to not be enough, we can add a counter and only request handoff after multiple failed wakeups. There are a few tricky implementation details: - accepting a handoff must only be done in the wait-loop. Since the handoff condition is owner == current, it can easily cause recursive locking trouble. - accepting the handoff must be careful to provide the ACQUIRE semantics. - having the HANDOFF bit set on unlock requires care, we must not clear the owner. - we must be careful to not leave HANDOFF set after we've acquired the lock. The tricky scenario is setting the HANDOFF bit on an unlocked mutex. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Waiman Long <Waiman.Long@hpe.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 20:40:16 +08:00
if (__mutex_trylock(lock, false)) {
osq_unlock(&lock->osq);
return true;
}
/*
* The cpu_relax() call is a compiler barrier which forces
* everything in this loop to be re-loaded. We don't need
* memory barriers as we'll eventually observe the right
* values at the cost of a few extra spins.
*/
cpu_relax_lowlatency();
}
osq_unlock(&lock->osq);
done:
/*
* If we fell out of the spin path because of need_resched(),
* reschedule now, before we try-lock the mutex. This avoids getting
* scheduled out right after we obtained the mutex.
*/
if (need_resched()) {
/*
* We _should_ have TASK_RUNNING here, but just in case
* we do not, make it so, otherwise we might get stuck.
*/
__set_current_state(TASK_RUNNING);
schedule_preempt_disabled();
}
return false;
}
#else
static bool mutex_optimistic_spin(struct mutex *lock,
struct ww_acquire_ctx *ww_ctx, const bool use_ww_ctx)
{
return false;
}
#endif
locking/mutex: Rework mutex::owner The current mutex implementation has an atomic lock word and a non-atomic owner field. This disparity leads to a number of issues with the current mutex code as it means that we can have a locked mutex without an explicit owner (because the owner field has not been set, or already cleared). This leads to a number of weird corner cases, esp. between the optimistic spinning and debug code. Where the optimistic spinning code needs the owner field updated inside the lock region, the debug code is more relaxed because the whole lock is serialized by the wait_lock. Also, the spinning code itself has a few corner cases where we need to deal with a held lock without an owner field. Furthermore, it becomes even more of a problem when trying to fix starvation cases in the current code. We end up stacking special case on special case. To solve this rework the basic mutex implementation to be a single atomic word that contains the owner and uses the low bits for extra state. This matches how PI futexes and rt_mutex already work. By having the owner an integral part of the lock state a lot of the problems dissapear and we get a better option to deal with starvation cases, direct owner handoff. Changing the basic mutex does however invalidate all the arch specific mutex code; this patch leaves that unused in-place, a later patch will remove that. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Will Deacon <will.deacon@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 19:36:04 +08:00
static noinline void __sched __mutex_unlock_slowpath(struct mutex *lock, unsigned long ip);
/**
* mutex_unlock - release the mutex
* @lock: the mutex to be released
*
* Unlock a mutex that has been locked by this task previously.
*
* This function must not be used in interrupt context. Unlocking
* of a not locked mutex is not allowed.
*
* This function is similar to (but not equivalent to) up().
*/
void __sched mutex_unlock(struct mutex *lock)
{
locking/mutex: Rework mutex::owner The current mutex implementation has an atomic lock word and a non-atomic owner field. This disparity leads to a number of issues with the current mutex code as it means that we can have a locked mutex without an explicit owner (because the owner field has not been set, or already cleared). This leads to a number of weird corner cases, esp. between the optimistic spinning and debug code. Where the optimistic spinning code needs the owner field updated inside the lock region, the debug code is more relaxed because the whole lock is serialized by the wait_lock. Also, the spinning code itself has a few corner cases where we need to deal with a held lock without an owner field. Furthermore, it becomes even more of a problem when trying to fix starvation cases in the current code. We end up stacking special case on special case. To solve this rework the basic mutex implementation to be a single atomic word that contains the owner and uses the low bits for extra state. This matches how PI futexes and rt_mutex already work. By having the owner an integral part of the lock state a lot of the problems dissapear and we get a better option to deal with starvation cases, direct owner handoff. Changing the basic mutex does however invalidate all the arch specific mutex code; this patch leaves that unused in-place, a later patch will remove that. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Will Deacon <will.deacon@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 19:36:04 +08:00
#ifndef CONFIG_DEBUG_LOCK_ALLOC
if (__mutex_unlock_fast(lock))
return;
mutex: implement adaptive spinning Change mutex contention behaviour such that it will sometimes busy wait on acquisition - moving its behaviour closer to that of spinlocks. This concept got ported to mainline from the -rt tree, where it was originally implemented for rtmutexes by Steven Rostedt, based on work by Gregory Haskins. Testing with Ingo's test-mutex application (http://lkml.org/lkml/2006/1/8/50) gave a 345% boost for VFS scalability on my testbox: # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 296604 # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 85870 The key criteria for the busy wait is that the lock owner has to be running on a (different) cpu. The idea is that as long as the owner is running, there is a fair chance it'll release the lock soon, and thus we'll be better off spinning instead of blocking/scheduling. Since regular mutexes (as opposed to rtmutexes) do not atomically track the owner, we add the owner in a non-atomic fashion and deal with the races in the slowpath. Furthermore, to ease the testing of the performance impact of this new code, there is means to disable this behaviour runtime (without having to reboot the system), when scheduler debugging is enabled (CONFIG_SCHED_DEBUG=y), by issuing the following command: # echo NO_OWNER_SPIN > /debug/sched_features This command re-enables spinning again (this is also the default): # echo OWNER_SPIN > /debug/sched_features Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-01-12 21:01:47 +08:00
#endif
locking/mutex: Rework mutex::owner The current mutex implementation has an atomic lock word and a non-atomic owner field. This disparity leads to a number of issues with the current mutex code as it means that we can have a locked mutex without an explicit owner (because the owner field has not been set, or already cleared). This leads to a number of weird corner cases, esp. between the optimistic spinning and debug code. Where the optimistic spinning code needs the owner field updated inside the lock region, the debug code is more relaxed because the whole lock is serialized by the wait_lock. Also, the spinning code itself has a few corner cases where we need to deal with a held lock without an owner field. Furthermore, it becomes even more of a problem when trying to fix starvation cases in the current code. We end up stacking special case on special case. To solve this rework the basic mutex implementation to be a single atomic word that contains the owner and uses the low bits for extra state. This matches how PI futexes and rt_mutex already work. By having the owner an integral part of the lock state a lot of the problems dissapear and we get a better option to deal with starvation cases, direct owner handoff. Changing the basic mutex does however invalidate all the arch specific mutex code; this patch leaves that unused in-place, a later patch will remove that. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Will Deacon <will.deacon@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 19:36:04 +08:00
__mutex_unlock_slowpath(lock, _RET_IP_);
}
EXPORT_SYMBOL(mutex_unlock);
/**
* ww_mutex_unlock - release the w/w mutex
* @lock: the mutex to be released
*
* Unlock a mutex that has been locked by this task previously with any of the
* ww_mutex_lock* functions (with or without an acquire context). It is
* forbidden to release the locks after releasing the acquire context.
*
* This function must not be used in interrupt context. Unlocking
* of a unlocked mutex is not allowed.
*/
void __sched ww_mutex_unlock(struct ww_mutex *lock)
{
/*
* The unlocking fastpath is the 0->1 transition from 'locked'
* into 'unlocked' state:
*/
if (lock->ctx) {
#ifdef CONFIG_DEBUG_MUTEXES
DEBUG_LOCKS_WARN_ON(!lock->ctx->acquired);
#endif
if (lock->ctx->acquired > 0)
lock->ctx->acquired--;
lock->ctx = NULL;
}
locking/mutex: Rework mutex::owner The current mutex implementation has an atomic lock word and a non-atomic owner field. This disparity leads to a number of issues with the current mutex code as it means that we can have a locked mutex without an explicit owner (because the owner field has not been set, or already cleared). This leads to a number of weird corner cases, esp. between the optimistic spinning and debug code. Where the optimistic spinning code needs the owner field updated inside the lock region, the debug code is more relaxed because the whole lock is serialized by the wait_lock. Also, the spinning code itself has a few corner cases where we need to deal with a held lock without an owner field. Furthermore, it becomes even more of a problem when trying to fix starvation cases in the current code. We end up stacking special case on special case. To solve this rework the basic mutex implementation to be a single atomic word that contains the owner and uses the low bits for extra state. This matches how PI futexes and rt_mutex already work. By having the owner an integral part of the lock state a lot of the problems dissapear and we get a better option to deal with starvation cases, direct owner handoff. Changing the basic mutex does however invalidate all the arch specific mutex code; this patch leaves that unused in-place, a later patch will remove that. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Will Deacon <will.deacon@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 19:36:04 +08:00
mutex_unlock(&lock->base);
}
EXPORT_SYMBOL(ww_mutex_unlock);
static inline int __sched
__ww_mutex_lock_check_stamp(struct mutex *lock, struct ww_acquire_ctx *ctx)
{
struct ww_mutex *ww = container_of(lock, struct ww_mutex, base);
struct ww_acquire_ctx *hold_ctx = READ_ONCE(ww->ctx);
if (!hold_ctx)
return 0;
if (ctx->stamp - hold_ctx->stamp <= LONG_MAX &&
(ctx->stamp != hold_ctx->stamp || ctx > hold_ctx)) {
#ifdef CONFIG_DEBUG_MUTEXES
DEBUG_LOCKS_WARN_ON(ctx->contending_lock);
ctx->contending_lock = ww;
#endif
return -EDEADLK;
}
return 0;
}
/*
* Lock a mutex (possibly interruptible), slowpath:
*/
static __always_inline int __sched
__mutex_lock_common(struct mutex *lock, long state, unsigned int subclass,
struct lockdep_map *nest_lock, unsigned long ip,
struct ww_acquire_ctx *ww_ctx, const bool use_ww_ctx)
{
struct task_struct *task = current;
struct mutex_waiter waiter;
unsigned long flags;
locking/mutex: Add lock handoff to avoid starvation Implement lock handoff to avoid lock starvation. Lock starvation is possible because mutex_lock() allows lock stealing, where a running (or optimistic spinning) task beats the woken waiter to the acquire. Lock stealing is an important performance optimization because waiting for a waiter to wake up and get runtime can take a significant time, during which everyboy would stall on the lock. The down-side is of course that it allows for starvation. This patch has the waiter requesting a handoff if it fails to acquire the lock upon waking. This re-introduces some of the wait time, because once we do a handoff we have to wait for the waiter to wake up again. A future patch will add a round of optimistic spinning to attempt to alleviate this penalty, but if that turns out to not be enough, we can add a counter and only request handoff after multiple failed wakeups. There are a few tricky implementation details: - accepting a handoff must only be done in the wait-loop. Since the handoff condition is owner == current, it can easily cause recursive locking trouble. - accepting the handoff must be careful to provide the ACQUIRE semantics. - having the HANDOFF bit set on unlock requires care, we must not clear the owner. - we must be careful to not leave HANDOFF set after we've acquired the lock. The tricky scenario is setting the HANDOFF bit on an unlocked mutex. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Waiman Long <Waiman.Long@hpe.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 20:40:16 +08:00
bool first = false;
int ret;
if (use_ww_ctx) {
struct ww_mutex *ww = container_of(lock, struct ww_mutex, base);
if (unlikely(ww_ctx == READ_ONCE(ww->ctx)))
return -EALREADY;
}
preempt_disable();
mutex_acquire_nest(&lock->dep_map, subclass, 0, nest_lock, ip);
locking/mutex: Add lock handoff to avoid starvation Implement lock handoff to avoid lock starvation. Lock starvation is possible because mutex_lock() allows lock stealing, where a running (or optimistic spinning) task beats the woken waiter to the acquire. Lock stealing is an important performance optimization because waiting for a waiter to wake up and get runtime can take a significant time, during which everyboy would stall on the lock. The down-side is of course that it allows for starvation. This patch has the waiter requesting a handoff if it fails to acquire the lock upon waking. This re-introduces some of the wait time, because once we do a handoff we have to wait for the waiter to wake up again. A future patch will add a round of optimistic spinning to attempt to alleviate this penalty, but if that turns out to not be enough, we can add a counter and only request handoff after multiple failed wakeups. There are a few tricky implementation details: - accepting a handoff must only be done in the wait-loop. Since the handoff condition is owner == current, it can easily cause recursive locking trouble. - accepting the handoff must be careful to provide the ACQUIRE semantics. - having the HANDOFF bit set on unlock requires care, we must not clear the owner. - we must be careful to not leave HANDOFF set after we've acquired the lock. The tricky scenario is setting the HANDOFF bit on an unlocked mutex. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Waiman Long <Waiman.Long@hpe.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 20:40:16 +08:00
if (__mutex_trylock(lock, false) ||
mutex_optimistic_spin(lock, ww_ctx, use_ww_ctx)) {
/* got the lock, yay! */
locking/mutex: Rework mutex::owner The current mutex implementation has an atomic lock word and a non-atomic owner field. This disparity leads to a number of issues with the current mutex code as it means that we can have a locked mutex without an explicit owner (because the owner field has not been set, or already cleared). This leads to a number of weird corner cases, esp. between the optimistic spinning and debug code. Where the optimistic spinning code needs the owner field updated inside the lock region, the debug code is more relaxed because the whole lock is serialized by the wait_lock. Also, the spinning code itself has a few corner cases where we need to deal with a held lock without an owner field. Furthermore, it becomes even more of a problem when trying to fix starvation cases in the current code. We end up stacking special case on special case. To solve this rework the basic mutex implementation to be a single atomic word that contains the owner and uses the low bits for extra state. This matches how PI futexes and rt_mutex already work. By having the owner an integral part of the lock state a lot of the problems dissapear and we get a better option to deal with starvation cases, direct owner handoff. Changing the basic mutex does however invalidate all the arch specific mutex code; this patch leaves that unused in-place, a later patch will remove that. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Will Deacon <will.deacon@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 19:36:04 +08:00
lock_acquired(&lock->dep_map, ip);
if (use_ww_ctx) {
struct ww_mutex *ww;
ww = container_of(lock, struct ww_mutex, base);
ww_mutex_set_context_fastpath(ww, ww_ctx);
}
preempt_enable();
return 0;
mutex: implement adaptive spinning Change mutex contention behaviour such that it will sometimes busy wait on acquisition - moving its behaviour closer to that of spinlocks. This concept got ported to mainline from the -rt tree, where it was originally implemented for rtmutexes by Steven Rostedt, based on work by Gregory Haskins. Testing with Ingo's test-mutex application (http://lkml.org/lkml/2006/1/8/50) gave a 345% boost for VFS scalability on my testbox: # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 296604 # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 85870 The key criteria for the busy wait is that the lock owner has to be running on a (different) cpu. The idea is that as long as the owner is running, there is a fair chance it'll release the lock soon, and thus we'll be better off spinning instead of blocking/scheduling. Since regular mutexes (as opposed to rtmutexes) do not atomically track the owner, we add the owner in a non-atomic fashion and deal with the races in the slowpath. Furthermore, to ease the testing of the performance impact of this new code, there is means to disable this behaviour runtime (without having to reboot the system), when scheduler debugging is enabled (CONFIG_SCHED_DEBUG=y), by issuing the following command: # echo NO_OWNER_SPIN > /debug/sched_features This command re-enables spinning again (this is also the default): # echo OWNER_SPIN > /debug/sched_features Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-01-12 21:01:47 +08:00
}
spin_lock_mutex(&lock->wait_lock, flags);
/*
locking/mutex: Rework mutex::owner The current mutex implementation has an atomic lock word and a non-atomic owner field. This disparity leads to a number of issues with the current mutex code as it means that we can have a locked mutex without an explicit owner (because the owner field has not been set, or already cleared). This leads to a number of weird corner cases, esp. between the optimistic spinning and debug code. Where the optimistic spinning code needs the owner field updated inside the lock region, the debug code is more relaxed because the whole lock is serialized by the wait_lock. Also, the spinning code itself has a few corner cases where we need to deal with a held lock without an owner field. Furthermore, it becomes even more of a problem when trying to fix starvation cases in the current code. We end up stacking special case on special case. To solve this rework the basic mutex implementation to be a single atomic word that contains the owner and uses the low bits for extra state. This matches how PI futexes and rt_mutex already work. By having the owner an integral part of the lock state a lot of the problems dissapear and we get a better option to deal with starvation cases, direct owner handoff. Changing the basic mutex does however invalidate all the arch specific mutex code; this patch leaves that unused in-place, a later patch will remove that. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Will Deacon <will.deacon@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 19:36:04 +08:00
* After waiting to acquire the wait_lock, try again.
*/
locking/mutex: Add lock handoff to avoid starvation Implement lock handoff to avoid lock starvation. Lock starvation is possible because mutex_lock() allows lock stealing, where a running (or optimistic spinning) task beats the woken waiter to the acquire. Lock stealing is an important performance optimization because waiting for a waiter to wake up and get runtime can take a significant time, during which everyboy would stall on the lock. The down-side is of course that it allows for starvation. This patch has the waiter requesting a handoff if it fails to acquire the lock upon waking. This re-introduces some of the wait time, because once we do a handoff we have to wait for the waiter to wake up again. A future patch will add a round of optimistic spinning to attempt to alleviate this penalty, but if that turns out to not be enough, we can add a counter and only request handoff after multiple failed wakeups. There are a few tricky implementation details: - accepting a handoff must only be done in the wait-loop. Since the handoff condition is owner == current, it can easily cause recursive locking trouble. - accepting the handoff must be careful to provide the ACQUIRE semantics. - having the HANDOFF bit set on unlock requires care, we must not clear the owner. - we must be careful to not leave HANDOFF set after we've acquired the lock. The tricky scenario is setting the HANDOFF bit on an unlocked mutex. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Waiman Long <Waiman.Long@hpe.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 20:40:16 +08:00
if (__mutex_trylock(lock, false))
mutex: Do not unnecessarily deal with waiters Upon entering the slowpath, we immediately attempt to acquire the lock by checking if it is already unlocked. If we are lucky enough that this is the case, then we don't need to deal with any waiter related logic. Furthermore any checks for an empty wait_list are unnecessary as we already know that count is non-negative and hence no one is waiting for the lock. Move the count check and xchg calls to be done before any waiters are setup - including waiter debugging. Upon failure to acquire the lock, the xchg sets the counter to 0, instead of -1 as it was originally. This can be done here since we set it back to -1 right at the beginning of the loop so other waiters are woken up when the lock is released. When tested on a 8-socket (80 core) system against a vanilla 3.10-rc1 kernel, this patch provides some small performance benefits (+2-6%). While it could be considered in the noise level, the average percentages were stable across multiple runs and no performance regressions were seen. Two big winners, for small amounts of users (10-100), were the short and compute workloads had a +19.36% and +%15.76% in jobs per minute. Also change some break statements to 'goto slowpath', which IMO makes a little more intuitive to read. Signed-off-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Acked-by: Rik van Riel <riel@redhat.com> Acked-by: Maarten Lankhorst <maarten.lankhorst@canonical.com> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/1372450398.2106.1.camel@buesod1.americas.hpqcorp.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-06-29 04:13:18 +08:00
goto skip_wait;
debug_mutex_lock_common(lock, &waiter);
debug_mutex_add_waiter(lock, &waiter, task);
/* add waiting tasks to the end of the waitqueue (FIFO): */
list_add_tail(&waiter.list, &lock->wait_list);
waiter.task = task;
locking/mutex: Add lock handoff to avoid starvation Implement lock handoff to avoid lock starvation. Lock starvation is possible because mutex_lock() allows lock stealing, where a running (or optimistic spinning) task beats the woken waiter to the acquire. Lock stealing is an important performance optimization because waiting for a waiter to wake up and get runtime can take a significant time, during which everyboy would stall on the lock. The down-side is of course that it allows for starvation. This patch has the waiter requesting a handoff if it fails to acquire the lock upon waking. This re-introduces some of the wait time, because once we do a handoff we have to wait for the waiter to wake up again. A future patch will add a round of optimistic spinning to attempt to alleviate this penalty, but if that turns out to not be enough, we can add a counter and only request handoff after multiple failed wakeups. There are a few tricky implementation details: - accepting a handoff must only be done in the wait-loop. Since the handoff condition is owner == current, it can easily cause recursive locking trouble. - accepting the handoff must be careful to provide the ACQUIRE semantics. - having the HANDOFF bit set on unlock requires care, we must not clear the owner. - we must be careful to not leave HANDOFF set after we've acquired the lock. The tricky scenario is setting the HANDOFF bit on an unlocked mutex. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Waiman Long <Waiman.Long@hpe.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 20:40:16 +08:00
if (__mutex_waiter_is_first(lock, &waiter))
locking/mutex: Rework mutex::owner The current mutex implementation has an atomic lock word and a non-atomic owner field. This disparity leads to a number of issues with the current mutex code as it means that we can have a locked mutex without an explicit owner (because the owner field has not been set, or already cleared). This leads to a number of weird corner cases, esp. between the optimistic spinning and debug code. Where the optimistic spinning code needs the owner field updated inside the lock region, the debug code is more relaxed because the whole lock is serialized by the wait_lock. Also, the spinning code itself has a few corner cases where we need to deal with a held lock without an owner field. Furthermore, it becomes even more of a problem when trying to fix starvation cases in the current code. We end up stacking special case on special case. To solve this rework the basic mutex implementation to be a single atomic word that contains the owner and uses the low bits for extra state. This matches how PI futexes and rt_mutex already work. By having the owner an integral part of the lock state a lot of the problems dissapear and we get a better option to deal with starvation cases, direct owner handoff. Changing the basic mutex does however invalidate all the arch specific mutex code; this patch leaves that unused in-place, a later patch will remove that. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Will Deacon <will.deacon@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 19:36:04 +08:00
__mutex_set_flag(lock, MUTEX_FLAG_WAITERS);
lock_contended(&lock->dep_map, ip);
set_task_state(task, state);
for (;;) {
/*
* Once we hold wait_lock, we're serialized against
* mutex_unlock() handing the lock off to us, do a trylock
* before testing the error conditions to make sure we pick up
* the handoff.
*/
locking/mutex: Add lock handoff to avoid starvation Implement lock handoff to avoid lock starvation. Lock starvation is possible because mutex_lock() allows lock stealing, where a running (or optimistic spinning) task beats the woken waiter to the acquire. Lock stealing is an important performance optimization because waiting for a waiter to wake up and get runtime can take a significant time, during which everyboy would stall on the lock. The down-side is of course that it allows for starvation. This patch has the waiter requesting a handoff if it fails to acquire the lock upon waking. This re-introduces some of the wait time, because once we do a handoff we have to wait for the waiter to wake up again. A future patch will add a round of optimistic spinning to attempt to alleviate this penalty, but if that turns out to not be enough, we can add a counter and only request handoff after multiple failed wakeups. There are a few tricky implementation details: - accepting a handoff must only be done in the wait-loop. Since the handoff condition is owner == current, it can easily cause recursive locking trouble. - accepting the handoff must be careful to provide the ACQUIRE semantics. - having the HANDOFF bit set on unlock requires care, we must not clear the owner. - we must be careful to not leave HANDOFF set after we've acquired the lock. The tricky scenario is setting the HANDOFF bit on an unlocked mutex. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Waiman Long <Waiman.Long@hpe.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 20:40:16 +08:00
if (__mutex_trylock(lock, first))
goto acquired;
/*
* Check for signals and wound conditions while holding
* wait_lock. This ensures the lock cancellation is ordered
* against mutex_unlock() and wake-ups do not go missing.
*/
if (unlikely(signal_pending_state(state, task))) {
ret = -EINTR;
goto err;
}
if (use_ww_ctx && ww_ctx->acquired > 0) {
ret = __ww_mutex_lock_check_stamp(lock, ww_ctx);
if (ret)
goto err;
}
spin_unlock_mutex(&lock->wait_lock, flags);
schedule_preempt_disabled();
locking/mutex: Add lock handoff to avoid starvation Implement lock handoff to avoid lock starvation. Lock starvation is possible because mutex_lock() allows lock stealing, where a running (or optimistic spinning) task beats the woken waiter to the acquire. Lock stealing is an important performance optimization because waiting for a waiter to wake up and get runtime can take a significant time, during which everyboy would stall on the lock. The down-side is of course that it allows for starvation. This patch has the waiter requesting a handoff if it fails to acquire the lock upon waking. This re-introduces some of the wait time, because once we do a handoff we have to wait for the waiter to wake up again. A future patch will add a round of optimistic spinning to attempt to alleviate this penalty, but if that turns out to not be enough, we can add a counter and only request handoff after multiple failed wakeups. There are a few tricky implementation details: - accepting a handoff must only be done in the wait-loop. Since the handoff condition is owner == current, it can easily cause recursive locking trouble. - accepting the handoff must be careful to provide the ACQUIRE semantics. - having the HANDOFF bit set on unlock requires care, we must not clear the owner. - we must be careful to not leave HANDOFF set after we've acquired the lock. The tricky scenario is setting the HANDOFF bit on an unlocked mutex. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Waiman Long <Waiman.Long@hpe.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 20:40:16 +08:00
if (!first && __mutex_waiter_is_first(lock, &waiter)) {
first = true;
__mutex_set_flag(lock, MUTEX_FLAG_HANDOFF);
}
set_task_state(task, state);
/*
* Here we order against unlock; we must either see it change
* state back to RUNNING and fall through the next schedule(),
* or we must see its unlock and acquire.
*/
if (__mutex_trylock(lock, first))
break;
spin_lock_mutex(&lock->wait_lock, flags);
}
spin_lock_mutex(&lock->wait_lock, flags);
acquired:
__set_task_state(task, TASK_RUNNING);
mutex_remove_waiter(lock, &waiter, task);
mutex: Do not unnecessarily deal with waiters Upon entering the slowpath, we immediately attempt to acquire the lock by checking if it is already unlocked. If we are lucky enough that this is the case, then we don't need to deal with any waiter related logic. Furthermore any checks for an empty wait_list are unnecessary as we already know that count is non-negative and hence no one is waiting for the lock. Move the count check and xchg calls to be done before any waiters are setup - including waiter debugging. Upon failure to acquire the lock, the xchg sets the counter to 0, instead of -1 as it was originally. This can be done here since we set it back to -1 right at the beginning of the loop so other waiters are woken up when the lock is released. When tested on a 8-socket (80 core) system against a vanilla 3.10-rc1 kernel, this patch provides some small performance benefits (+2-6%). While it could be considered in the noise level, the average percentages were stable across multiple runs and no performance regressions were seen. Two big winners, for small amounts of users (10-100), were the short and compute workloads had a +19.36% and +%15.76% in jobs per minute. Also change some break statements to 'goto slowpath', which IMO makes a little more intuitive to read. Signed-off-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Acked-by: Rik van Riel <riel@redhat.com> Acked-by: Maarten Lankhorst <maarten.lankhorst@canonical.com> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/1372450398.2106.1.camel@buesod1.americas.hpqcorp.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-06-29 04:13:18 +08:00
if (likely(list_empty(&lock->wait_list)))
locking/mutex: Add lock handoff to avoid starvation Implement lock handoff to avoid lock starvation. Lock starvation is possible because mutex_lock() allows lock stealing, where a running (or optimistic spinning) task beats the woken waiter to the acquire. Lock stealing is an important performance optimization because waiting for a waiter to wake up and get runtime can take a significant time, during which everyboy would stall on the lock. The down-side is of course that it allows for starvation. This patch has the waiter requesting a handoff if it fails to acquire the lock upon waking. This re-introduces some of the wait time, because once we do a handoff we have to wait for the waiter to wake up again. A future patch will add a round of optimistic spinning to attempt to alleviate this penalty, but if that turns out to not be enough, we can add a counter and only request handoff after multiple failed wakeups. There are a few tricky implementation details: - accepting a handoff must only be done in the wait-loop. Since the handoff condition is owner == current, it can easily cause recursive locking trouble. - accepting the handoff must be careful to provide the ACQUIRE semantics. - having the HANDOFF bit set on unlock requires care, we must not clear the owner. - we must be careful to not leave HANDOFF set after we've acquired the lock. The tricky scenario is setting the HANDOFF bit on an unlocked mutex. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Waiman Long <Waiman.Long@hpe.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 20:40:16 +08:00
__mutex_clear_flag(lock, MUTEX_FLAGS);
locking/mutex: Rework mutex::owner The current mutex implementation has an atomic lock word and a non-atomic owner field. This disparity leads to a number of issues with the current mutex code as it means that we can have a locked mutex without an explicit owner (because the owner field has not been set, or already cleared). This leads to a number of weird corner cases, esp. between the optimistic spinning and debug code. Where the optimistic spinning code needs the owner field updated inside the lock region, the debug code is more relaxed because the whole lock is serialized by the wait_lock. Also, the spinning code itself has a few corner cases where we need to deal with a held lock without an owner field. Furthermore, it becomes even more of a problem when trying to fix starvation cases in the current code. We end up stacking special case on special case. To solve this rework the basic mutex implementation to be a single atomic word that contains the owner and uses the low bits for extra state. This matches how PI futexes and rt_mutex already work. By having the owner an integral part of the lock state a lot of the problems dissapear and we get a better option to deal with starvation cases, direct owner handoff. Changing the basic mutex does however invalidate all the arch specific mutex code; this patch leaves that unused in-place, a later patch will remove that. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Will Deacon <will.deacon@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 19:36:04 +08:00
mutex: Do not unnecessarily deal with waiters Upon entering the slowpath, we immediately attempt to acquire the lock by checking if it is already unlocked. If we are lucky enough that this is the case, then we don't need to deal with any waiter related logic. Furthermore any checks for an empty wait_list are unnecessary as we already know that count is non-negative and hence no one is waiting for the lock. Move the count check and xchg calls to be done before any waiters are setup - including waiter debugging. Upon failure to acquire the lock, the xchg sets the counter to 0, instead of -1 as it was originally. This can be done here since we set it back to -1 right at the beginning of the loop so other waiters are woken up when the lock is released. When tested on a 8-socket (80 core) system against a vanilla 3.10-rc1 kernel, this patch provides some small performance benefits (+2-6%). While it could be considered in the noise level, the average percentages were stable across multiple runs and no performance regressions were seen. Two big winners, for small amounts of users (10-100), were the short and compute workloads had a +19.36% and +%15.76% in jobs per minute. Also change some break statements to 'goto slowpath', which IMO makes a little more intuitive to read. Signed-off-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Acked-by: Rik van Riel <riel@redhat.com> Acked-by: Maarten Lankhorst <maarten.lankhorst@canonical.com> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/1372450398.2106.1.camel@buesod1.americas.hpqcorp.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-06-29 04:13:18 +08:00
debug_mutex_free_waiter(&waiter);
mutex: Do not unnecessarily deal with waiters Upon entering the slowpath, we immediately attempt to acquire the lock by checking if it is already unlocked. If we are lucky enough that this is the case, then we don't need to deal with any waiter related logic. Furthermore any checks for an empty wait_list are unnecessary as we already know that count is non-negative and hence no one is waiting for the lock. Move the count check and xchg calls to be done before any waiters are setup - including waiter debugging. Upon failure to acquire the lock, the xchg sets the counter to 0, instead of -1 as it was originally. This can be done here since we set it back to -1 right at the beginning of the loop so other waiters are woken up when the lock is released. When tested on a 8-socket (80 core) system against a vanilla 3.10-rc1 kernel, this patch provides some small performance benefits (+2-6%). While it could be considered in the noise level, the average percentages were stable across multiple runs and no performance regressions were seen. Two big winners, for small amounts of users (10-100), were the short and compute workloads had a +19.36% and +%15.76% in jobs per minute. Also change some break statements to 'goto slowpath', which IMO makes a little more intuitive to read. Signed-off-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Acked-by: Rik van Riel <riel@redhat.com> Acked-by: Maarten Lankhorst <maarten.lankhorst@canonical.com> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/1372450398.2106.1.camel@buesod1.americas.hpqcorp.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-06-29 04:13:18 +08:00
skip_wait:
/* got the lock - cleanup and rejoice! */
lock_acquired(&lock->dep_map, ip);
if (use_ww_ctx) {
mutex: Do not unnecessarily deal with waiters Upon entering the slowpath, we immediately attempt to acquire the lock by checking if it is already unlocked. If we are lucky enough that this is the case, then we don't need to deal with any waiter related logic. Furthermore any checks for an empty wait_list are unnecessary as we already know that count is non-negative and hence no one is waiting for the lock. Move the count check and xchg calls to be done before any waiters are setup - including waiter debugging. Upon failure to acquire the lock, the xchg sets the counter to 0, instead of -1 as it was originally. This can be done here since we set it back to -1 right at the beginning of the loop so other waiters are woken up when the lock is released. When tested on a 8-socket (80 core) system against a vanilla 3.10-rc1 kernel, this patch provides some small performance benefits (+2-6%). While it could be considered in the noise level, the average percentages were stable across multiple runs and no performance regressions were seen. Two big winners, for small amounts of users (10-100), were the short and compute workloads had a +19.36% and +%15.76% in jobs per minute. Also change some break statements to 'goto slowpath', which IMO makes a little more intuitive to read. Signed-off-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Acked-by: Rik van Riel <riel@redhat.com> Acked-by: Maarten Lankhorst <maarten.lankhorst@canonical.com> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Link: http://lkml.kernel.org/r/1372450398.2106.1.camel@buesod1.americas.hpqcorp.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-06-29 04:13:18 +08:00
struct ww_mutex *ww = container_of(lock, struct ww_mutex, base);
ww_mutex_set_context_slowpath(ww, ww_ctx);
}
spin_unlock_mutex(&lock->wait_lock, flags);
preempt_enable();
return 0;
err:
__set_task_state(task, TASK_RUNNING);
mutex_remove_waiter(lock, &waiter, task);
spin_unlock_mutex(&lock->wait_lock, flags);
debug_mutex_free_waiter(&waiter);
mutex_release(&lock->dep_map, 1, ip);
preempt_enable();
return ret;
}
#ifdef CONFIG_DEBUG_LOCK_ALLOC
void __sched
mutex_lock_nested(struct mutex *lock, unsigned int subclass)
{
might_sleep();
__mutex_lock_common(lock, TASK_UNINTERRUPTIBLE,
subclass, NULL, _RET_IP_, NULL, 0);
}
EXPORT_SYMBOL_GPL(mutex_lock_nested);
void __sched
_mutex_lock_nest_lock(struct mutex *lock, struct lockdep_map *nest)
{
might_sleep();
__mutex_lock_common(lock, TASK_UNINTERRUPTIBLE,
0, nest, _RET_IP_, NULL, 0);
}
EXPORT_SYMBOL_GPL(_mutex_lock_nest_lock);
int __sched
mutex_lock_killable_nested(struct mutex *lock, unsigned int subclass)
{
might_sleep();
return __mutex_lock_common(lock, TASK_KILLABLE,
subclass, NULL, _RET_IP_, NULL, 0);
}
EXPORT_SYMBOL_GPL(mutex_lock_killable_nested);
int __sched
mutex_lock_interruptible_nested(struct mutex *lock, unsigned int subclass)
{
might_sleep();
mutex: implement adaptive spinning Change mutex contention behaviour such that it will sometimes busy wait on acquisition - moving its behaviour closer to that of spinlocks. This concept got ported to mainline from the -rt tree, where it was originally implemented for rtmutexes by Steven Rostedt, based on work by Gregory Haskins. Testing with Ingo's test-mutex application (http://lkml.org/lkml/2006/1/8/50) gave a 345% boost for VFS scalability on my testbox: # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 296604 # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 85870 The key criteria for the busy wait is that the lock owner has to be running on a (different) cpu. The idea is that as long as the owner is running, there is a fair chance it'll release the lock soon, and thus we'll be better off spinning instead of blocking/scheduling. Since regular mutexes (as opposed to rtmutexes) do not atomically track the owner, we add the owner in a non-atomic fashion and deal with the races in the slowpath. Furthermore, to ease the testing of the performance impact of this new code, there is means to disable this behaviour runtime (without having to reboot the system), when scheduler debugging is enabled (CONFIG_SCHED_DEBUG=y), by issuing the following command: # echo NO_OWNER_SPIN > /debug/sched_features This command re-enables spinning again (this is also the default): # echo OWNER_SPIN > /debug/sched_features Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-01-12 21:01:47 +08:00
return __mutex_lock_common(lock, TASK_INTERRUPTIBLE,
subclass, NULL, _RET_IP_, NULL, 0);
}
EXPORT_SYMBOL_GPL(mutex_lock_interruptible_nested);
mutex: Add w/w mutex slowpath debugging Injects EDEADLK conditions at pseudo-random interval, with exponential backoff up to UINT_MAX (to ensure that every lock operation still completes in a reasonable time). This way we can test the wound slowpath even for ww mutex users where contention is never expected, and the ww deadlock avoidance algorithm is only needed for correctness against malicious userspace. An example would be protecting kernel modesetting properties, which thanks to single-threaded X isn't really expected to contend, ever. I've looked into using the CONFIG_FAULT_INJECTION infrastructure, but decided against it for two reasons: - EDEADLK handling is mandatory for ww mutex users and should never affect the outcome of a syscall. This is in contrast to -ENOMEM injection. So fine configurability isn't required. - The fault injection framework only allows to set a simple probability for failure. Now the probability that a ww mutex acquire stage with N locks will never complete (due to too many injected EDEADLK backoffs) is zero. But the expected number of ww_mutex_lock operations for the completely uncontended case would be O(exp(N)). The per-acuiqire ctx exponential backoff solution choosen here only results in O(log N) overhead due to injection and so O(log N * N) lock operations. This way we can fail with high probability (and so have good test coverage even for fancy backoff and lock acquisition paths) without running into patalogical cases. Note that EDEADLK will only ever be injected when we managed to acquire the lock. This prevents any behaviour changes for users which rely on the EALREADY semantics. Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: dri-devel@lists.freedesktop.org Cc: linaro-mm-sig@lists.linaro.org Cc: rostedt@goodmis.org Cc: daniel@ffwll.ch Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/20130620113117.4001.21681.stgit@patser Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-06-20 19:31:17 +08:00
static inline int
ww_mutex_deadlock_injection(struct ww_mutex *lock, struct ww_acquire_ctx *ctx)
{
#ifdef CONFIG_DEBUG_WW_MUTEX_SLOWPATH
unsigned tmp;
if (ctx->deadlock_inject_countdown-- == 0) {
tmp = ctx->deadlock_inject_interval;
if (tmp > UINT_MAX/4)
tmp = UINT_MAX;
else
tmp = tmp*2 + tmp + tmp/2;
ctx->deadlock_inject_interval = tmp;
ctx->deadlock_inject_countdown = tmp;
ctx->contending_lock = lock;
ww_mutex_unlock(lock);
return -EDEADLK;
}
#endif
return 0;
}
int __sched
__ww_mutex_lock(struct ww_mutex *lock, struct ww_acquire_ctx *ctx)
{
mutex: Add w/w mutex slowpath debugging Injects EDEADLK conditions at pseudo-random interval, with exponential backoff up to UINT_MAX (to ensure that every lock operation still completes in a reasonable time). This way we can test the wound slowpath even for ww mutex users where contention is never expected, and the ww deadlock avoidance algorithm is only needed for correctness against malicious userspace. An example would be protecting kernel modesetting properties, which thanks to single-threaded X isn't really expected to contend, ever. I've looked into using the CONFIG_FAULT_INJECTION infrastructure, but decided against it for two reasons: - EDEADLK handling is mandatory for ww mutex users and should never affect the outcome of a syscall. This is in contrast to -ENOMEM injection. So fine configurability isn't required. - The fault injection framework only allows to set a simple probability for failure. Now the probability that a ww mutex acquire stage with N locks will never complete (due to too many injected EDEADLK backoffs) is zero. But the expected number of ww_mutex_lock operations for the completely uncontended case would be O(exp(N)). The per-acuiqire ctx exponential backoff solution choosen here only results in O(log N) overhead due to injection and so O(log N * N) lock operations. This way we can fail with high probability (and so have good test coverage even for fancy backoff and lock acquisition paths) without running into patalogical cases. Note that EDEADLK will only ever be injected when we managed to acquire the lock. This prevents any behaviour changes for users which rely on the EALREADY semantics. Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: dri-devel@lists.freedesktop.org Cc: linaro-mm-sig@lists.linaro.org Cc: rostedt@goodmis.org Cc: daniel@ffwll.ch Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/20130620113117.4001.21681.stgit@patser Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-06-20 19:31:17 +08:00
int ret;
might_sleep();
mutex: Add w/w mutex slowpath debugging Injects EDEADLK conditions at pseudo-random interval, with exponential backoff up to UINT_MAX (to ensure that every lock operation still completes in a reasonable time). This way we can test the wound slowpath even for ww mutex users where contention is never expected, and the ww deadlock avoidance algorithm is only needed for correctness against malicious userspace. An example would be protecting kernel modesetting properties, which thanks to single-threaded X isn't really expected to contend, ever. I've looked into using the CONFIG_FAULT_INJECTION infrastructure, but decided against it for two reasons: - EDEADLK handling is mandatory for ww mutex users and should never affect the outcome of a syscall. This is in contrast to -ENOMEM injection. So fine configurability isn't required. - The fault injection framework only allows to set a simple probability for failure. Now the probability that a ww mutex acquire stage with N locks will never complete (due to too many injected EDEADLK backoffs) is zero. But the expected number of ww_mutex_lock operations for the completely uncontended case would be O(exp(N)). The per-acuiqire ctx exponential backoff solution choosen here only results in O(log N) overhead due to injection and so O(log N * N) lock operations. This way we can fail with high probability (and so have good test coverage even for fancy backoff and lock acquisition paths) without running into patalogical cases. Note that EDEADLK will only ever be injected when we managed to acquire the lock. This prevents any behaviour changes for users which rely on the EALREADY semantics. Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: dri-devel@lists.freedesktop.org Cc: linaro-mm-sig@lists.linaro.org Cc: rostedt@goodmis.org Cc: daniel@ffwll.ch Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/20130620113117.4001.21681.stgit@patser Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-06-20 19:31:17 +08:00
ret = __mutex_lock_common(&lock->base, TASK_UNINTERRUPTIBLE,
0, &ctx->dep_map, _RET_IP_, ctx, 1);
if (!ret && ctx->acquired > 1)
mutex: Add w/w mutex slowpath debugging Injects EDEADLK conditions at pseudo-random interval, with exponential backoff up to UINT_MAX (to ensure that every lock operation still completes in a reasonable time). This way we can test the wound slowpath even for ww mutex users where contention is never expected, and the ww deadlock avoidance algorithm is only needed for correctness against malicious userspace. An example would be protecting kernel modesetting properties, which thanks to single-threaded X isn't really expected to contend, ever. I've looked into using the CONFIG_FAULT_INJECTION infrastructure, but decided against it for two reasons: - EDEADLK handling is mandatory for ww mutex users and should never affect the outcome of a syscall. This is in contrast to -ENOMEM injection. So fine configurability isn't required. - The fault injection framework only allows to set a simple probability for failure. Now the probability that a ww mutex acquire stage with N locks will never complete (due to too many injected EDEADLK backoffs) is zero. But the expected number of ww_mutex_lock operations for the completely uncontended case would be O(exp(N)). The per-acuiqire ctx exponential backoff solution choosen here only results in O(log N) overhead due to injection and so O(log N * N) lock operations. This way we can fail with high probability (and so have good test coverage even for fancy backoff and lock acquisition paths) without running into patalogical cases. Note that EDEADLK will only ever be injected when we managed to acquire the lock. This prevents any behaviour changes for users which rely on the EALREADY semantics. Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: dri-devel@lists.freedesktop.org Cc: linaro-mm-sig@lists.linaro.org Cc: rostedt@goodmis.org Cc: daniel@ffwll.ch Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/20130620113117.4001.21681.stgit@patser Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-06-20 19:31:17 +08:00
return ww_mutex_deadlock_injection(lock, ctx);
return ret;
}
EXPORT_SYMBOL_GPL(__ww_mutex_lock);
int __sched
__ww_mutex_lock_interruptible(struct ww_mutex *lock, struct ww_acquire_ctx *ctx)
{
mutex: Add w/w mutex slowpath debugging Injects EDEADLK conditions at pseudo-random interval, with exponential backoff up to UINT_MAX (to ensure that every lock operation still completes in a reasonable time). This way we can test the wound slowpath even for ww mutex users where contention is never expected, and the ww deadlock avoidance algorithm is only needed for correctness against malicious userspace. An example would be protecting kernel modesetting properties, which thanks to single-threaded X isn't really expected to contend, ever. I've looked into using the CONFIG_FAULT_INJECTION infrastructure, but decided against it for two reasons: - EDEADLK handling is mandatory for ww mutex users and should never affect the outcome of a syscall. This is in contrast to -ENOMEM injection. So fine configurability isn't required. - The fault injection framework only allows to set a simple probability for failure. Now the probability that a ww mutex acquire stage with N locks will never complete (due to too many injected EDEADLK backoffs) is zero. But the expected number of ww_mutex_lock operations for the completely uncontended case would be O(exp(N)). The per-acuiqire ctx exponential backoff solution choosen here only results in O(log N) overhead due to injection and so O(log N * N) lock operations. This way we can fail with high probability (and so have good test coverage even for fancy backoff and lock acquisition paths) without running into patalogical cases. Note that EDEADLK will only ever be injected when we managed to acquire the lock. This prevents any behaviour changes for users which rely on the EALREADY semantics. Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: dri-devel@lists.freedesktop.org Cc: linaro-mm-sig@lists.linaro.org Cc: rostedt@goodmis.org Cc: daniel@ffwll.ch Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/20130620113117.4001.21681.stgit@patser Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-06-20 19:31:17 +08:00
int ret;
might_sleep();
mutex: Add w/w mutex slowpath debugging Injects EDEADLK conditions at pseudo-random interval, with exponential backoff up to UINT_MAX (to ensure that every lock operation still completes in a reasonable time). This way we can test the wound slowpath even for ww mutex users where contention is never expected, and the ww deadlock avoidance algorithm is only needed for correctness against malicious userspace. An example would be protecting kernel modesetting properties, which thanks to single-threaded X isn't really expected to contend, ever. I've looked into using the CONFIG_FAULT_INJECTION infrastructure, but decided against it for two reasons: - EDEADLK handling is mandatory for ww mutex users and should never affect the outcome of a syscall. This is in contrast to -ENOMEM injection. So fine configurability isn't required. - The fault injection framework only allows to set a simple probability for failure. Now the probability that a ww mutex acquire stage with N locks will never complete (due to too many injected EDEADLK backoffs) is zero. But the expected number of ww_mutex_lock operations for the completely uncontended case would be O(exp(N)). The per-acuiqire ctx exponential backoff solution choosen here only results in O(log N) overhead due to injection and so O(log N * N) lock operations. This way we can fail with high probability (and so have good test coverage even for fancy backoff and lock acquisition paths) without running into patalogical cases. Note that EDEADLK will only ever be injected when we managed to acquire the lock. This prevents any behaviour changes for users which rely on the EALREADY semantics. Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: dri-devel@lists.freedesktop.org Cc: linaro-mm-sig@lists.linaro.org Cc: rostedt@goodmis.org Cc: daniel@ffwll.ch Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/20130620113117.4001.21681.stgit@patser Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-06-20 19:31:17 +08:00
ret = __mutex_lock_common(&lock->base, TASK_INTERRUPTIBLE,
0, &ctx->dep_map, _RET_IP_, ctx, 1);
mutex: Add w/w mutex slowpath debugging Injects EDEADLK conditions at pseudo-random interval, with exponential backoff up to UINT_MAX (to ensure that every lock operation still completes in a reasonable time). This way we can test the wound slowpath even for ww mutex users where contention is never expected, and the ww deadlock avoidance algorithm is only needed for correctness against malicious userspace. An example would be protecting kernel modesetting properties, which thanks to single-threaded X isn't really expected to contend, ever. I've looked into using the CONFIG_FAULT_INJECTION infrastructure, but decided against it for two reasons: - EDEADLK handling is mandatory for ww mutex users and should never affect the outcome of a syscall. This is in contrast to -ENOMEM injection. So fine configurability isn't required. - The fault injection framework only allows to set a simple probability for failure. Now the probability that a ww mutex acquire stage with N locks will never complete (due to too many injected EDEADLK backoffs) is zero. But the expected number of ww_mutex_lock operations for the completely uncontended case would be O(exp(N)). The per-acuiqire ctx exponential backoff solution choosen here only results in O(log N) overhead due to injection and so O(log N * N) lock operations. This way we can fail with high probability (and so have good test coverage even for fancy backoff and lock acquisition paths) without running into patalogical cases. Note that EDEADLK will only ever be injected when we managed to acquire the lock. This prevents any behaviour changes for users which rely on the EALREADY semantics. Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: dri-devel@lists.freedesktop.org Cc: linaro-mm-sig@lists.linaro.org Cc: rostedt@goodmis.org Cc: daniel@ffwll.ch Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/20130620113117.4001.21681.stgit@patser Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-06-20 19:31:17 +08:00
if (!ret && ctx->acquired > 1)
mutex: Add w/w mutex slowpath debugging Injects EDEADLK conditions at pseudo-random interval, with exponential backoff up to UINT_MAX (to ensure that every lock operation still completes in a reasonable time). This way we can test the wound slowpath even for ww mutex users where contention is never expected, and the ww deadlock avoidance algorithm is only needed for correctness against malicious userspace. An example would be protecting kernel modesetting properties, which thanks to single-threaded X isn't really expected to contend, ever. I've looked into using the CONFIG_FAULT_INJECTION infrastructure, but decided against it for two reasons: - EDEADLK handling is mandatory for ww mutex users and should never affect the outcome of a syscall. This is in contrast to -ENOMEM injection. So fine configurability isn't required. - The fault injection framework only allows to set a simple probability for failure. Now the probability that a ww mutex acquire stage with N locks will never complete (due to too many injected EDEADLK backoffs) is zero. But the expected number of ww_mutex_lock operations for the completely uncontended case would be O(exp(N)). The per-acuiqire ctx exponential backoff solution choosen here only results in O(log N) overhead due to injection and so O(log N * N) lock operations. This way we can fail with high probability (and so have good test coverage even for fancy backoff and lock acquisition paths) without running into patalogical cases. Note that EDEADLK will only ever be injected when we managed to acquire the lock. This prevents any behaviour changes for users which rely on the EALREADY semantics. Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: dri-devel@lists.freedesktop.org Cc: linaro-mm-sig@lists.linaro.org Cc: rostedt@goodmis.org Cc: daniel@ffwll.ch Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/20130620113117.4001.21681.stgit@patser Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-06-20 19:31:17 +08:00
return ww_mutex_deadlock_injection(lock, ctx);
return ret;
}
EXPORT_SYMBOL_GPL(__ww_mutex_lock_interruptible);
#endif
/*
* Release the lock, slowpath:
*/
locking/mutex: Rework mutex::owner The current mutex implementation has an atomic lock word and a non-atomic owner field. This disparity leads to a number of issues with the current mutex code as it means that we can have a locked mutex without an explicit owner (because the owner field has not been set, or already cleared). This leads to a number of weird corner cases, esp. between the optimistic spinning and debug code. Where the optimistic spinning code needs the owner field updated inside the lock region, the debug code is more relaxed because the whole lock is serialized by the wait_lock. Also, the spinning code itself has a few corner cases where we need to deal with a held lock without an owner field. Furthermore, it becomes even more of a problem when trying to fix starvation cases in the current code. We end up stacking special case on special case. To solve this rework the basic mutex implementation to be a single atomic word that contains the owner and uses the low bits for extra state. This matches how PI futexes and rt_mutex already work. By having the owner an integral part of the lock state a lot of the problems dissapear and we get a better option to deal with starvation cases, direct owner handoff. Changing the basic mutex does however invalidate all the arch specific mutex code; this patch leaves that unused in-place, a later patch will remove that. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Will Deacon <will.deacon@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 19:36:04 +08:00
static noinline void __sched __mutex_unlock_slowpath(struct mutex *lock, unsigned long ip)
{
locking/mutex: Add lock handoff to avoid starvation Implement lock handoff to avoid lock starvation. Lock starvation is possible because mutex_lock() allows lock stealing, where a running (or optimistic spinning) task beats the woken waiter to the acquire. Lock stealing is an important performance optimization because waiting for a waiter to wake up and get runtime can take a significant time, during which everyboy would stall on the lock. The down-side is of course that it allows for starvation. This patch has the waiter requesting a handoff if it fails to acquire the lock upon waking. This re-introduces some of the wait time, because once we do a handoff we have to wait for the waiter to wake up again. A future patch will add a round of optimistic spinning to attempt to alleviate this penalty, but if that turns out to not be enough, we can add a counter and only request handoff after multiple failed wakeups. There are a few tricky implementation details: - accepting a handoff must only be done in the wait-loop. Since the handoff condition is owner == current, it can easily cause recursive locking trouble. - accepting the handoff must be careful to provide the ACQUIRE semantics. - having the HANDOFF bit set on unlock requires care, we must not clear the owner. - we must be careful to not leave HANDOFF set after we've acquired the lock. The tricky scenario is setting the HANDOFF bit on an unlocked mutex. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Waiman Long <Waiman.Long@hpe.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 20:40:16 +08:00
struct task_struct *next = NULL;
locking/mutex: Rework mutex::owner The current mutex implementation has an atomic lock word and a non-atomic owner field. This disparity leads to a number of issues with the current mutex code as it means that we can have a locked mutex without an explicit owner (because the owner field has not been set, or already cleared). This leads to a number of weird corner cases, esp. between the optimistic spinning and debug code. Where the optimistic spinning code needs the owner field updated inside the lock region, the debug code is more relaxed because the whole lock is serialized by the wait_lock. Also, the spinning code itself has a few corner cases where we need to deal with a held lock without an owner field. Furthermore, it becomes even more of a problem when trying to fix starvation cases in the current code. We end up stacking special case on special case. To solve this rework the basic mutex implementation to be a single atomic word that contains the owner and uses the low bits for extra state. This matches how PI futexes and rt_mutex already work. By having the owner an integral part of the lock state a lot of the problems dissapear and we get a better option to deal with starvation cases, direct owner handoff. Changing the basic mutex does however invalidate all the arch specific mutex code; this patch leaves that unused in-place, a later patch will remove that. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Will Deacon <will.deacon@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 19:36:04 +08:00
unsigned long owner, flags;
WAKE_Q(wake_q);
locking/mutex: Rework mutex::owner The current mutex implementation has an atomic lock word and a non-atomic owner field. This disparity leads to a number of issues with the current mutex code as it means that we can have a locked mutex without an explicit owner (because the owner field has not been set, or already cleared). This leads to a number of weird corner cases, esp. between the optimistic spinning and debug code. Where the optimistic spinning code needs the owner field updated inside the lock region, the debug code is more relaxed because the whole lock is serialized by the wait_lock. Also, the spinning code itself has a few corner cases where we need to deal with a held lock without an owner field. Furthermore, it becomes even more of a problem when trying to fix starvation cases in the current code. We end up stacking special case on special case. To solve this rework the basic mutex implementation to be a single atomic word that contains the owner and uses the low bits for extra state. This matches how PI futexes and rt_mutex already work. By having the owner an integral part of the lock state a lot of the problems dissapear and we get a better option to deal with starvation cases, direct owner handoff. Changing the basic mutex does however invalidate all the arch specific mutex code; this patch leaves that unused in-place, a later patch will remove that. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Will Deacon <will.deacon@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 19:36:04 +08:00
mutex_release(&lock->dep_map, 1, ip);
/*
locking/mutex: Add lock handoff to avoid starvation Implement lock handoff to avoid lock starvation. Lock starvation is possible because mutex_lock() allows lock stealing, where a running (or optimistic spinning) task beats the woken waiter to the acquire. Lock stealing is an important performance optimization because waiting for a waiter to wake up and get runtime can take a significant time, during which everyboy would stall on the lock. The down-side is of course that it allows for starvation. This patch has the waiter requesting a handoff if it fails to acquire the lock upon waking. This re-introduces some of the wait time, because once we do a handoff we have to wait for the waiter to wake up again. A future patch will add a round of optimistic spinning to attempt to alleviate this penalty, but if that turns out to not be enough, we can add a counter and only request handoff after multiple failed wakeups. There are a few tricky implementation details: - accepting a handoff must only be done in the wait-loop. Since the handoff condition is owner == current, it can easily cause recursive locking trouble. - accepting the handoff must be careful to provide the ACQUIRE semantics. - having the HANDOFF bit set on unlock requires care, we must not clear the owner. - we must be careful to not leave HANDOFF set after we've acquired the lock. The tricky scenario is setting the HANDOFF bit on an unlocked mutex. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Waiman Long <Waiman.Long@hpe.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 20:40:16 +08:00
* Release the lock before (potentially) taking the spinlock such that
* other contenders can get on with things ASAP.
*
* Except when HANDOFF, in that case we must not clear the owner field,
* but instead set it to the top waiter.
*/
locking/mutex: Add lock handoff to avoid starvation Implement lock handoff to avoid lock starvation. Lock starvation is possible because mutex_lock() allows lock stealing, where a running (or optimistic spinning) task beats the woken waiter to the acquire. Lock stealing is an important performance optimization because waiting for a waiter to wake up and get runtime can take a significant time, during which everyboy would stall on the lock. The down-side is of course that it allows for starvation. This patch has the waiter requesting a handoff if it fails to acquire the lock upon waking. This re-introduces some of the wait time, because once we do a handoff we have to wait for the waiter to wake up again. A future patch will add a round of optimistic spinning to attempt to alleviate this penalty, but if that turns out to not be enough, we can add a counter and only request handoff after multiple failed wakeups. There are a few tricky implementation details: - accepting a handoff must only be done in the wait-loop. Since the handoff condition is owner == current, it can easily cause recursive locking trouble. - accepting the handoff must be careful to provide the ACQUIRE semantics. - having the HANDOFF bit set on unlock requires care, we must not clear the owner. - we must be careful to not leave HANDOFF set after we've acquired the lock. The tricky scenario is setting the HANDOFF bit on an unlocked mutex. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Waiman Long <Waiman.Long@hpe.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 20:40:16 +08:00
owner = atomic_long_read(&lock->owner);
for (;;) {
unsigned long old;
#ifdef CONFIG_DEBUG_MUTEXES
DEBUG_LOCKS_WARN_ON(__owner_task(owner) != current);
#endif
if (owner & MUTEX_FLAG_HANDOFF)
break;
old = atomic_long_cmpxchg_release(&lock->owner, owner,
__owner_flags(owner));
if (old == owner) {
if (owner & MUTEX_FLAG_WAITERS)
break;
return;
}
owner = old;
}
spin_lock_mutex(&lock->wait_lock, flags);
debug_mutex_unlock(lock);
if (!list_empty(&lock->wait_list)) {
/* get the first entry from the wait-list: */
struct mutex_waiter *waiter =
locking/mutex: Add lock handoff to avoid starvation Implement lock handoff to avoid lock starvation. Lock starvation is possible because mutex_lock() allows lock stealing, where a running (or optimistic spinning) task beats the woken waiter to the acquire. Lock stealing is an important performance optimization because waiting for a waiter to wake up and get runtime can take a significant time, during which everyboy would stall on the lock. The down-side is of course that it allows for starvation. This patch has the waiter requesting a handoff if it fails to acquire the lock upon waking. This re-introduces some of the wait time, because once we do a handoff we have to wait for the waiter to wake up again. A future patch will add a round of optimistic spinning to attempt to alleviate this penalty, but if that turns out to not be enough, we can add a counter and only request handoff after multiple failed wakeups. There are a few tricky implementation details: - accepting a handoff must only be done in the wait-loop. Since the handoff condition is owner == current, it can easily cause recursive locking trouble. - accepting the handoff must be careful to provide the ACQUIRE semantics. - having the HANDOFF bit set on unlock requires care, we must not clear the owner. - we must be careful to not leave HANDOFF set after we've acquired the lock. The tricky scenario is setting the HANDOFF bit on an unlocked mutex. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Waiman Long <Waiman.Long@hpe.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 20:40:16 +08:00
list_first_entry(&lock->wait_list,
struct mutex_waiter, list);
next = waiter->task;
debug_mutex_wake_waiter(lock, waiter);
locking/mutex: Add lock handoff to avoid starvation Implement lock handoff to avoid lock starvation. Lock starvation is possible because mutex_lock() allows lock stealing, where a running (or optimistic spinning) task beats the woken waiter to the acquire. Lock stealing is an important performance optimization because waiting for a waiter to wake up and get runtime can take a significant time, during which everyboy would stall on the lock. The down-side is of course that it allows for starvation. This patch has the waiter requesting a handoff if it fails to acquire the lock upon waking. This re-introduces some of the wait time, because once we do a handoff we have to wait for the waiter to wake up again. A future patch will add a round of optimistic spinning to attempt to alleviate this penalty, but if that turns out to not be enough, we can add a counter and only request handoff after multiple failed wakeups. There are a few tricky implementation details: - accepting a handoff must only be done in the wait-loop. Since the handoff condition is owner == current, it can easily cause recursive locking trouble. - accepting the handoff must be careful to provide the ACQUIRE semantics. - having the HANDOFF bit set on unlock requires care, we must not clear the owner. - we must be careful to not leave HANDOFF set after we've acquired the lock. The tricky scenario is setting the HANDOFF bit on an unlocked mutex. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Waiman Long <Waiman.Long@hpe.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 20:40:16 +08:00
wake_q_add(&wake_q, next);
}
locking/mutex: Add lock handoff to avoid starvation Implement lock handoff to avoid lock starvation. Lock starvation is possible because mutex_lock() allows lock stealing, where a running (or optimistic spinning) task beats the woken waiter to the acquire. Lock stealing is an important performance optimization because waiting for a waiter to wake up and get runtime can take a significant time, during which everyboy would stall on the lock. The down-side is of course that it allows for starvation. This patch has the waiter requesting a handoff if it fails to acquire the lock upon waking. This re-introduces some of the wait time, because once we do a handoff we have to wait for the waiter to wake up again. A future patch will add a round of optimistic spinning to attempt to alleviate this penalty, but if that turns out to not be enough, we can add a counter and only request handoff after multiple failed wakeups. There are a few tricky implementation details: - accepting a handoff must only be done in the wait-loop. Since the handoff condition is owner == current, it can easily cause recursive locking trouble. - accepting the handoff must be careful to provide the ACQUIRE semantics. - having the HANDOFF bit set on unlock requires care, we must not clear the owner. - we must be careful to not leave HANDOFF set after we've acquired the lock. The tricky scenario is setting the HANDOFF bit on an unlocked mutex. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Waiman Long <Waiman.Long@hpe.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 20:40:16 +08:00
if (owner & MUTEX_FLAG_HANDOFF)
__mutex_handoff(lock, next);
spin_unlock_mutex(&lock->wait_lock, flags);
locking/mutex: Add lock handoff to avoid starvation Implement lock handoff to avoid lock starvation. Lock starvation is possible because mutex_lock() allows lock stealing, where a running (or optimistic spinning) task beats the woken waiter to the acquire. Lock stealing is an important performance optimization because waiting for a waiter to wake up and get runtime can take a significant time, during which everyboy would stall on the lock. The down-side is of course that it allows for starvation. This patch has the waiter requesting a handoff if it fails to acquire the lock upon waking. This re-introduces some of the wait time, because once we do a handoff we have to wait for the waiter to wake up again. A future patch will add a round of optimistic spinning to attempt to alleviate this penalty, but if that turns out to not be enough, we can add a counter and only request handoff after multiple failed wakeups. There are a few tricky implementation details: - accepting a handoff must only be done in the wait-loop. Since the handoff condition is owner == current, it can easily cause recursive locking trouble. - accepting the handoff must be careful to provide the ACQUIRE semantics. - having the HANDOFF bit set on unlock requires care, we must not clear the owner. - we must be careful to not leave HANDOFF set after we've acquired the lock. The tricky scenario is setting the HANDOFF bit on an unlocked mutex. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Waiman Long <Waiman.Long@hpe.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 20:40:16 +08:00
wake_up_q(&wake_q);
}
#ifndef CONFIG_DEBUG_LOCK_ALLOC
/*
* Here come the less common (and hence less performance-critical) APIs:
* mutex_lock_interruptible() and mutex_trylock().
*/
static noinline int __sched
__mutex_lock_killable_slowpath(struct mutex *lock);
static noinline int __sched
__mutex_lock_interruptible_slowpath(struct mutex *lock);
/**
* mutex_lock_interruptible - acquire the mutex, interruptible
* @lock: the mutex to be acquired
*
* Lock the mutex like mutex_lock(), and return 0 if the mutex has
* been acquired or sleep until the mutex becomes available. If a
* signal arrives while waiting for the lock then this function
* returns -EINTR.
*
* This function is similar to (but not equivalent to) down_interruptible().
*/
int __sched mutex_lock_interruptible(struct mutex *lock)
{
might_sleep();
locking/mutex: Rework mutex::owner The current mutex implementation has an atomic lock word and a non-atomic owner field. This disparity leads to a number of issues with the current mutex code as it means that we can have a locked mutex without an explicit owner (because the owner field has not been set, or already cleared). This leads to a number of weird corner cases, esp. between the optimistic spinning and debug code. Where the optimistic spinning code needs the owner field updated inside the lock region, the debug code is more relaxed because the whole lock is serialized by the wait_lock. Also, the spinning code itself has a few corner cases where we need to deal with a held lock without an owner field. Furthermore, it becomes even more of a problem when trying to fix starvation cases in the current code. We end up stacking special case on special case. To solve this rework the basic mutex implementation to be a single atomic word that contains the owner and uses the low bits for extra state. This matches how PI futexes and rt_mutex already work. By having the owner an integral part of the lock state a lot of the problems dissapear and we get a better option to deal with starvation cases, direct owner handoff. Changing the basic mutex does however invalidate all the arch specific mutex code; this patch leaves that unused in-place, a later patch will remove that. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Will Deacon <will.deacon@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 19:36:04 +08:00
if (__mutex_trylock_fast(lock))
return 0;
locking/mutex: Rework mutex::owner The current mutex implementation has an atomic lock word and a non-atomic owner field. This disparity leads to a number of issues with the current mutex code as it means that we can have a locked mutex without an explicit owner (because the owner field has not been set, or already cleared). This leads to a number of weird corner cases, esp. between the optimistic spinning and debug code. Where the optimistic spinning code needs the owner field updated inside the lock region, the debug code is more relaxed because the whole lock is serialized by the wait_lock. Also, the spinning code itself has a few corner cases where we need to deal with a held lock without an owner field. Furthermore, it becomes even more of a problem when trying to fix starvation cases in the current code. We end up stacking special case on special case. To solve this rework the basic mutex implementation to be a single atomic word that contains the owner and uses the low bits for extra state. This matches how PI futexes and rt_mutex already work. By having the owner an integral part of the lock state a lot of the problems dissapear and we get a better option to deal with starvation cases, direct owner handoff. Changing the basic mutex does however invalidate all the arch specific mutex code; this patch leaves that unused in-place, a later patch will remove that. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Will Deacon <will.deacon@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 19:36:04 +08:00
return __mutex_lock_interruptible_slowpath(lock);
}
EXPORT_SYMBOL(mutex_lock_interruptible);
int __sched mutex_lock_killable(struct mutex *lock)
{
might_sleep();
locking/mutex: Rework mutex::owner The current mutex implementation has an atomic lock word and a non-atomic owner field. This disparity leads to a number of issues with the current mutex code as it means that we can have a locked mutex without an explicit owner (because the owner field has not been set, or already cleared). This leads to a number of weird corner cases, esp. between the optimistic spinning and debug code. Where the optimistic spinning code needs the owner field updated inside the lock region, the debug code is more relaxed because the whole lock is serialized by the wait_lock. Also, the spinning code itself has a few corner cases where we need to deal with a held lock without an owner field. Furthermore, it becomes even more of a problem when trying to fix starvation cases in the current code. We end up stacking special case on special case. To solve this rework the basic mutex implementation to be a single atomic word that contains the owner and uses the low bits for extra state. This matches how PI futexes and rt_mutex already work. By having the owner an integral part of the lock state a lot of the problems dissapear and we get a better option to deal with starvation cases, direct owner handoff. Changing the basic mutex does however invalidate all the arch specific mutex code; this patch leaves that unused in-place, a later patch will remove that. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Will Deacon <will.deacon@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 19:36:04 +08:00
if (__mutex_trylock_fast(lock))
return 0;
locking/mutex: Rework mutex::owner The current mutex implementation has an atomic lock word and a non-atomic owner field. This disparity leads to a number of issues with the current mutex code as it means that we can have a locked mutex without an explicit owner (because the owner field has not been set, or already cleared). This leads to a number of weird corner cases, esp. between the optimistic spinning and debug code. Where the optimistic spinning code needs the owner field updated inside the lock region, the debug code is more relaxed because the whole lock is serialized by the wait_lock. Also, the spinning code itself has a few corner cases where we need to deal with a held lock without an owner field. Furthermore, it becomes even more of a problem when trying to fix starvation cases in the current code. We end up stacking special case on special case. To solve this rework the basic mutex implementation to be a single atomic word that contains the owner and uses the low bits for extra state. This matches how PI futexes and rt_mutex already work. By having the owner an integral part of the lock state a lot of the problems dissapear and we get a better option to deal with starvation cases, direct owner handoff. Changing the basic mutex does however invalidate all the arch specific mutex code; this patch leaves that unused in-place, a later patch will remove that. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Will Deacon <will.deacon@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 19:36:04 +08:00
return __mutex_lock_killable_slowpath(lock);
}
EXPORT_SYMBOL(mutex_lock_killable);
locking/mutex: Rework mutex::owner The current mutex implementation has an atomic lock word and a non-atomic owner field. This disparity leads to a number of issues with the current mutex code as it means that we can have a locked mutex without an explicit owner (because the owner field has not been set, or already cleared). This leads to a number of weird corner cases, esp. between the optimistic spinning and debug code. Where the optimistic spinning code needs the owner field updated inside the lock region, the debug code is more relaxed because the whole lock is serialized by the wait_lock. Also, the spinning code itself has a few corner cases where we need to deal with a held lock without an owner field. Furthermore, it becomes even more of a problem when trying to fix starvation cases in the current code. We end up stacking special case on special case. To solve this rework the basic mutex implementation to be a single atomic word that contains the owner and uses the low bits for extra state. This matches how PI futexes and rt_mutex already work. By having the owner an integral part of the lock state a lot of the problems dissapear and we get a better option to deal with starvation cases, direct owner handoff. Changing the basic mutex does however invalidate all the arch specific mutex code; this patch leaves that unused in-place, a later patch will remove that. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Will Deacon <will.deacon@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 19:36:04 +08:00
static noinline void __sched
__mutex_lock_slowpath(struct mutex *lock)
{
__mutex_lock_common(lock, TASK_UNINTERRUPTIBLE, 0,
NULL, _RET_IP_, NULL, 0);
}
static noinline int __sched
__mutex_lock_killable_slowpath(struct mutex *lock)
{
return __mutex_lock_common(lock, TASK_KILLABLE, 0,
NULL, _RET_IP_, NULL, 0);
}
static noinline int __sched
__mutex_lock_interruptible_slowpath(struct mutex *lock)
{
return __mutex_lock_common(lock, TASK_INTERRUPTIBLE, 0,
NULL, _RET_IP_, NULL, 0);
}
static noinline int __sched
__ww_mutex_lock_slowpath(struct ww_mutex *lock, struct ww_acquire_ctx *ctx)
{
return __mutex_lock_common(&lock->base, TASK_UNINTERRUPTIBLE, 0,
NULL, _RET_IP_, ctx, 1);
}
static noinline int __sched
__ww_mutex_lock_interruptible_slowpath(struct ww_mutex *lock,
struct ww_acquire_ctx *ctx)
{
return __mutex_lock_common(&lock->base, TASK_INTERRUPTIBLE, 0,
NULL, _RET_IP_, ctx, 1);
}
#endif
/**
* mutex_trylock - try to acquire the mutex, without waiting
* @lock: the mutex to be acquired
*
* Try to acquire the mutex atomically. Returns 1 if the mutex
* has been acquired successfully, and 0 on contention.
*
* NOTE: this function follows the spin_trylock() convention, so
* it is negated from the down_trylock() return values! Be careful
* about this when converting semaphore users to mutexes.
*
* This function must not be used in interrupt context. The
* mutex must be released by the same task that acquired it.
*/
int __sched mutex_trylock(struct mutex *lock)
{
locking/mutex: Add lock handoff to avoid starvation Implement lock handoff to avoid lock starvation. Lock starvation is possible because mutex_lock() allows lock stealing, where a running (or optimistic spinning) task beats the woken waiter to the acquire. Lock stealing is an important performance optimization because waiting for a waiter to wake up and get runtime can take a significant time, during which everyboy would stall on the lock. The down-side is of course that it allows for starvation. This patch has the waiter requesting a handoff if it fails to acquire the lock upon waking. This re-introduces some of the wait time, because once we do a handoff we have to wait for the waiter to wake up again. A future patch will add a round of optimistic spinning to attempt to alleviate this penalty, but if that turns out to not be enough, we can add a counter and only request handoff after multiple failed wakeups. There are a few tricky implementation details: - accepting a handoff must only be done in the wait-loop. Since the handoff condition is owner == current, it can easily cause recursive locking trouble. - accepting the handoff must be careful to provide the ACQUIRE semantics. - having the HANDOFF bit set on unlock requires care, we must not clear the owner. - we must be careful to not leave HANDOFF set after we've acquired the lock. The tricky scenario is setting the HANDOFF bit on an unlocked mutex. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Waiman Long <Waiman.Long@hpe.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 20:40:16 +08:00
bool locked = __mutex_trylock(lock, false);
mutex: implement adaptive spinning Change mutex contention behaviour such that it will sometimes busy wait on acquisition - moving its behaviour closer to that of spinlocks. This concept got ported to mainline from the -rt tree, where it was originally implemented for rtmutexes by Steven Rostedt, based on work by Gregory Haskins. Testing with Ingo's test-mutex application (http://lkml.org/lkml/2006/1/8/50) gave a 345% boost for VFS scalability on my testbox: # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 296604 # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 85870 The key criteria for the busy wait is that the lock owner has to be running on a (different) cpu. The idea is that as long as the owner is running, there is a fair chance it'll release the lock soon, and thus we'll be better off spinning instead of blocking/scheduling. Since regular mutexes (as opposed to rtmutexes) do not atomically track the owner, we add the owner in a non-atomic fashion and deal with the races in the slowpath. Furthermore, to ease the testing of the performance impact of this new code, there is means to disable this behaviour runtime (without having to reboot the system), when scheduler debugging is enabled (CONFIG_SCHED_DEBUG=y), by issuing the following command: # echo NO_OWNER_SPIN > /debug/sched_features This command re-enables spinning again (this is also the default): # echo OWNER_SPIN > /debug/sched_features Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-01-12 21:01:47 +08:00
locking/mutex: Rework mutex::owner The current mutex implementation has an atomic lock word and a non-atomic owner field. This disparity leads to a number of issues with the current mutex code as it means that we can have a locked mutex without an explicit owner (because the owner field has not been set, or already cleared). This leads to a number of weird corner cases, esp. between the optimistic spinning and debug code. Where the optimistic spinning code needs the owner field updated inside the lock region, the debug code is more relaxed because the whole lock is serialized by the wait_lock. Also, the spinning code itself has a few corner cases where we need to deal with a held lock without an owner field. Furthermore, it becomes even more of a problem when trying to fix starvation cases in the current code. We end up stacking special case on special case. To solve this rework the basic mutex implementation to be a single atomic word that contains the owner and uses the low bits for extra state. This matches how PI futexes and rt_mutex already work. By having the owner an integral part of the lock state a lot of the problems dissapear and we get a better option to deal with starvation cases, direct owner handoff. Changing the basic mutex does however invalidate all the arch specific mutex code; this patch leaves that unused in-place, a later patch will remove that. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Will Deacon <will.deacon@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 19:36:04 +08:00
if (locked)
mutex_acquire(&lock->dep_map, 0, 1, _RET_IP_);
mutex: implement adaptive spinning Change mutex contention behaviour such that it will sometimes busy wait on acquisition - moving its behaviour closer to that of spinlocks. This concept got ported to mainline from the -rt tree, where it was originally implemented for rtmutexes by Steven Rostedt, based on work by Gregory Haskins. Testing with Ingo's test-mutex application (http://lkml.org/lkml/2006/1/8/50) gave a 345% boost for VFS scalability on my testbox: # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 296604 # ./test-mutex-shm V 16 10 | grep "^avg ops" avg ops/sec: 85870 The key criteria for the busy wait is that the lock owner has to be running on a (different) cpu. The idea is that as long as the owner is running, there is a fair chance it'll release the lock soon, and thus we'll be better off spinning instead of blocking/scheduling. Since regular mutexes (as opposed to rtmutexes) do not atomically track the owner, we add the owner in a non-atomic fashion and deal with the races in the slowpath. Furthermore, to ease the testing of the performance impact of this new code, there is means to disable this behaviour runtime (without having to reboot the system), when scheduler debugging is enabled (CONFIG_SCHED_DEBUG=y), by issuing the following command: # echo NO_OWNER_SPIN > /debug/sched_features This command re-enables spinning again (this is also the default): # echo OWNER_SPIN > /debug/sched_features Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-01-12 21:01:47 +08:00
locking/mutex: Rework mutex::owner The current mutex implementation has an atomic lock word and a non-atomic owner field. This disparity leads to a number of issues with the current mutex code as it means that we can have a locked mutex without an explicit owner (because the owner field has not been set, or already cleared). This leads to a number of weird corner cases, esp. between the optimistic spinning and debug code. Where the optimistic spinning code needs the owner field updated inside the lock region, the debug code is more relaxed because the whole lock is serialized by the wait_lock. Also, the spinning code itself has a few corner cases where we need to deal with a held lock without an owner field. Furthermore, it becomes even more of a problem when trying to fix starvation cases in the current code. We end up stacking special case on special case. To solve this rework the basic mutex implementation to be a single atomic word that contains the owner and uses the low bits for extra state. This matches how PI futexes and rt_mutex already work. By having the owner an integral part of the lock state a lot of the problems dissapear and we get a better option to deal with starvation cases, direct owner handoff. Changing the basic mutex does however invalidate all the arch specific mutex code; this patch leaves that unused in-place, a later patch will remove that. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Will Deacon <will.deacon@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 19:36:04 +08:00
return locked;
}
EXPORT_SYMBOL(mutex_trylock);
#ifndef CONFIG_DEBUG_LOCK_ALLOC
int __sched
__ww_mutex_lock(struct ww_mutex *lock, struct ww_acquire_ctx *ctx)
{
might_sleep();
locking/mutex: Rework mutex::owner The current mutex implementation has an atomic lock word and a non-atomic owner field. This disparity leads to a number of issues with the current mutex code as it means that we can have a locked mutex without an explicit owner (because the owner field has not been set, or already cleared). This leads to a number of weird corner cases, esp. between the optimistic spinning and debug code. Where the optimistic spinning code needs the owner field updated inside the lock region, the debug code is more relaxed because the whole lock is serialized by the wait_lock. Also, the spinning code itself has a few corner cases where we need to deal with a held lock without an owner field. Furthermore, it becomes even more of a problem when trying to fix starvation cases in the current code. We end up stacking special case on special case. To solve this rework the basic mutex implementation to be a single atomic word that contains the owner and uses the low bits for extra state. This matches how PI futexes and rt_mutex already work. By having the owner an integral part of the lock state a lot of the problems dissapear and we get a better option to deal with starvation cases, direct owner handoff. Changing the basic mutex does however invalidate all the arch specific mutex code; this patch leaves that unused in-place, a later patch will remove that. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Will Deacon <will.deacon@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 19:36:04 +08:00
if (__mutex_trylock_fast(&lock->base)) {
ww_mutex_set_context_fastpath(lock, ctx);
locking/mutex: Rework mutex::owner The current mutex implementation has an atomic lock word and a non-atomic owner field. This disparity leads to a number of issues with the current mutex code as it means that we can have a locked mutex without an explicit owner (because the owner field has not been set, or already cleared). This leads to a number of weird corner cases, esp. between the optimistic spinning and debug code. Where the optimistic spinning code needs the owner field updated inside the lock region, the debug code is more relaxed because the whole lock is serialized by the wait_lock. Also, the spinning code itself has a few corner cases where we need to deal with a held lock without an owner field. Furthermore, it becomes even more of a problem when trying to fix starvation cases in the current code. We end up stacking special case on special case. To solve this rework the basic mutex implementation to be a single atomic word that contains the owner and uses the low bits for extra state. This matches how PI futexes and rt_mutex already work. By having the owner an integral part of the lock state a lot of the problems dissapear and we get a better option to deal with starvation cases, direct owner handoff. Changing the basic mutex does however invalidate all the arch specific mutex code; this patch leaves that unused in-place, a later patch will remove that. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Will Deacon <will.deacon@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 19:36:04 +08:00
return 0;
}
return __ww_mutex_lock_slowpath(lock, ctx);
}
EXPORT_SYMBOL(__ww_mutex_lock);
int __sched
__ww_mutex_lock_interruptible(struct ww_mutex *lock, struct ww_acquire_ctx *ctx)
{
might_sleep();
locking/mutex: Rework mutex::owner The current mutex implementation has an atomic lock word and a non-atomic owner field. This disparity leads to a number of issues with the current mutex code as it means that we can have a locked mutex without an explicit owner (because the owner field has not been set, or already cleared). This leads to a number of weird corner cases, esp. between the optimistic spinning and debug code. Where the optimistic spinning code needs the owner field updated inside the lock region, the debug code is more relaxed because the whole lock is serialized by the wait_lock. Also, the spinning code itself has a few corner cases where we need to deal with a held lock without an owner field. Furthermore, it becomes even more of a problem when trying to fix starvation cases in the current code. We end up stacking special case on special case. To solve this rework the basic mutex implementation to be a single atomic word that contains the owner and uses the low bits for extra state. This matches how PI futexes and rt_mutex already work. By having the owner an integral part of the lock state a lot of the problems dissapear and we get a better option to deal with starvation cases, direct owner handoff. Changing the basic mutex does however invalidate all the arch specific mutex code; this patch leaves that unused in-place, a later patch will remove that. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Will Deacon <will.deacon@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 19:36:04 +08:00
if (__mutex_trylock_fast(&lock->base)) {
ww_mutex_set_context_fastpath(lock, ctx);
locking/mutex: Rework mutex::owner The current mutex implementation has an atomic lock word and a non-atomic owner field. This disparity leads to a number of issues with the current mutex code as it means that we can have a locked mutex without an explicit owner (because the owner field has not been set, or already cleared). This leads to a number of weird corner cases, esp. between the optimistic spinning and debug code. Where the optimistic spinning code needs the owner field updated inside the lock region, the debug code is more relaxed because the whole lock is serialized by the wait_lock. Also, the spinning code itself has a few corner cases where we need to deal with a held lock without an owner field. Furthermore, it becomes even more of a problem when trying to fix starvation cases in the current code. We end up stacking special case on special case. To solve this rework the basic mutex implementation to be a single atomic word that contains the owner and uses the low bits for extra state. This matches how PI futexes and rt_mutex already work. By having the owner an integral part of the lock state a lot of the problems dissapear and we get a better option to deal with starvation cases, direct owner handoff. Changing the basic mutex does however invalidate all the arch specific mutex code; this patch leaves that unused in-place, a later patch will remove that. Tested-by: Jason Low <jason.low2@hpe.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Will Deacon <will.deacon@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-23 19:36:04 +08:00
return 0;
}
return __ww_mutex_lock_interruptible_slowpath(lock, ctx);
}
EXPORT_SYMBOL(__ww_mutex_lock_interruptible);
#endif
/**
* atomic_dec_and_mutex_lock - return holding mutex if we dec to 0
* @cnt: the atomic which we are to dec
* @lock: the mutex to return holding if we dec to 0
*
* return true and hold lock if we dec to 0, return false otherwise
*/
int atomic_dec_and_mutex_lock(atomic_t *cnt, struct mutex *lock)
{
/* dec if we can't possibly hit 0 */
if (atomic_add_unless(cnt, -1, 1))
return 0;
/* we might hit 0, so take the lock */
mutex_lock(lock);
if (!atomic_dec_and_test(cnt)) {
/* when we actually did the dec, we didn't hit 0 */
mutex_unlock(lock);
return 0;
}
/* we hit 0, and we hold the lock */
return 1;
}
EXPORT_SYMBOL(atomic_dec_and_mutex_lock);