[PATCH v4 0/2] locking/rwsem: optimize rwsem_wakeup()

[PATCH v4 0/2] locking/rwsem: optimize rwsem_wakeup()

[Waiman Long]

 v3->v4:
   - Break out the active writer check into a separate patch and move
     it from __rwsem_do_wake() to rwsem_wake().
   - Use smp_rmb() instead of the incorrect smp_mb__after_atomic() as
     suggested by PeterZ.

 v2->v3:
   - Fix errors in commit log.

 v1->v2:
   - Add a memory barrier before calling spin_trylock for proper memory
     ordering.

This patch set aims to reduce spinlock contention in the wait_lock
due to excessive activity in the rwsem_wake code path. This, in turn,
reduces up_write/up_read latency and improve performance when the
rwsem is heavily contended.

On an 8-socket Westmere-EX server (80 cores, HT off), running AIM7's
high_systime workload (1000 users) on a vanilla 4.0 kernel produced
the following perf profile for spinlock contention:

  9.23%    reaim  [kernel.kallsyms]    [k] _raw_spin_lock_irqsave
              |--97.39%-- rwsem_wake
              |--0.69%-- try_to_wake_up
              |--0.52%-- release_pages
               --1.40%-- [...]

  1.70%    reaim  [kernel.kallsyms]    [k] _raw_spin_lock_irq
              |--96.61%-- rwsem_down_write_failed
              |--2.03%-- __schedule
              |--0.50%-- run_timer_softirq
               --0.86%-- [...]

Here the contended rwsems are the mmap_sem (mm_struct) and the
i_mmap_rwsem (address_space) with mostly write locking.  With a
patched 4.0 kernel, the perf profile became:

  1.87%    reaim  [kernel.kallsyms]    [k] _raw_spin_lock_irqsave
              |--87.64%-- rwsem_wake
              |--2.80%-- release_pages
              |--2.56%-- try_to_wake_up
              |--1.10%-- __wake_up
              |--1.06%-- pagevec_lru_move_fn
              |--0.93%-- prepare_to_wait_exclusive
              |--0.71%-- free_pid
              |--0.58%-- get_page_from_freelist
              |--0.57%-- add_device_randomness
               --2.04%-- [...]

  0.80%    reaim  [kernel.kallsyms]    [k] _raw_spin_lock_irq
              |--92.49%-- rwsem_down_write_failed
              |--4.24%-- __schedule
              |--1.37%-- run_timer_softirq
               --1.91%-- [...]

The table below shows the % improvement in throughput (1100-2000 users)
in the various AIM7's workloads:

	Workload	% increase in throughput
	--------	------------------------
	custom			 3.8%
	five-sec		 3.5%
	fserver			 4.1%
	high_systime		22.2%
	shared			 2.1%
	short			10.1%

Waiman Long (2):
  locking/rwsem: reduce spinlock contention in wakeup after
    up_read/up_write
  locking/rwsem: check for active writer before wakeup

 include/linux/osq_lock.h    |    5 +++
 kernel/locking/rwsem-xadd.c |   65 +++++++++++++++++++++++++++++++++++++++++-
 2 files changed, 68 insertions(+), 2 deletions(-)

[PATCH v4 1/2] locking/rwsem: reduce spinlock contention in wakeup after up_read/up_write

[Waiman Long]

In up_write()/up_read(), rwsem_wake() will be called whenever it
detects that some writers/readers are waiting. The rwsem_wake()
function will take the wait_lock and call __rwsem_do_wake() to do the
real wakeup.  For a heavily contended rwsem, doing a spin_lock() on
wait_lock will cause further contention on the heavily contended rwsem
cacheline resulting in delay in the completion of the up_read/up_write
operations.

This patch makes the wait_lock taking and the call to __rwsem_do_wake()
optional if at least one spinning writer is present. The spinning
writer will be able to take the rwsem and call rwsem_wake() later
when it calls up_write(). With the presence of a spinning writer,
rwsem_wake() will now try to acquire the lock using trylock. If that
fails, it will just quit.

Signed-off-by: Waiman Long <Waiman.Long@hp.com>
Suggested-by: Peter Zijlstra (Intel) <peterz@infradead.org>
---
 include/linux/osq_lock.h    |    5 ++++
 kernel/locking/rwsem-xadd.c |   44 +++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 49 insertions(+), 0 deletions(-)

diff --git a/include/linux/osq_lock.h b/include/linux/osq_lock.h
index 3a6490e..703ea5c 100644
--- a/include/linux/osq_lock.h
+++ b/include/linux/osq_lock.h
@@ -32,4 +32,9 @@ static inline void osq_lock_init(struct optimistic_spin_queue *lock)
 extern bool osq_lock(struct optimistic_spin_queue *lock);
 extern void osq_unlock(struct optimistic_spin_queue *lock);
 
+static inline bool osq_is_locked(struct optimistic_spin_queue *lock)
+{
+	return atomic_read(&lock->tail) != OSQ_UNLOCKED_VAL;
+}
+
 #endif
diff --git a/kernel/locking/rwsem-xadd.c b/kernel/locking/rwsem-xadd.c
index 2f7cc40..2bb25e2 100644
--- a/kernel/locking/rwsem-xadd.c
+++ b/kernel/locking/rwsem-xadd.c
@@ -391,11 +391,24 @@ done:
 	return taken;
 }
 
+/*
+ * Return true if the rwsem has active spinner
+ */
+static inline bool rwsem_has_spinner(struct rw_semaphore *sem)
+{
+	return osq_is_locked(&sem->osq);
+}
+
 #else
 static bool rwsem_optimistic_spin(struct rw_semaphore *sem)
 {
 	return false;
 }
+
+static inline bool rwsem_has_spinner(struct rw_semaphore *sem)
+{
+	return false;
+}
 #endif
 
 /*
@@ -478,7 +491,38 @@ struct rw_semaphore *rwsem_wake(struct rw_semaphore *sem)
 {
 	unsigned long flags;
 
+	/*
+	 * If a spinner is present, it is not necessary to do the wakeup.
+	 * Try to do wakeup only if the trylock succeeds to minimize
+	 * spinlock contention which may introduce too much delay in the
+	 * unlock operation.
+	 *
+	 *    spinning writer		up_write/up_read caller
+	 *    ---------------		-----------------------
+	 * [S]   osq_unlock()		[L]   osq
+	 *	 MB			      RMB
+	 * [RmW] rwsem_try_write_lock() [RmW] spin_trylock(wait_lock)
+	 *
+	 * Here, it is important to make sure that there won't be a missed
+	 * wakeup while the rwsem is free and the only spinning writer goes
+	 * to sleep without taking the rwsem. Even when the spinning writer
+	 * is just going to break out of the waiting loop, it will still do
+	 * a trylock in rwsem_down_write_failed() before sleeping. IOW, if
+	 * rwsem_has_spinner() is true, it will guarantee at least one
+	 * trylock attempt on the rwsem later on.
+	 */
+	if (rwsem_has_spinner(sem)) {
+		/*
+		 * The smp_rmb() here is to make sure that the spinner
+		 * state is consulted before reading the wait_lock.
+		 */
+		smp_rmb();
+		if (!raw_spin_trylock_irqsave(&sem->wait_lock, flags))
+			return sem;
+		goto locked;
+	}
 	raw_spin_lock_irqsave(&sem->wait_lock, flags);
+locked:
 
 	/* do nothing if list empty */
 	if (!list_empty(&sem->wait_list))
-- 
1.7.1

[PATCH v4 2/2] locking/rwsem: check for active writer before wakeup

[Waiman Long]

On a highly contended rwsem, spinlock contention due to the slow
rwsem_wake() call can be a significant portion of the total CPU cycles
used. With writer lock stealing and writer optimistic spinning, there
is also a chance that the lock may have been stolen by the time that
the wait_lock is acquired.

This patch adds a low cost checking code after acquiring the wait_lock
to look for active writer. The presence of an active writer will
abort the wakeup operation.

Signed-off-by: Waiman Long <Waiman.Long@hp.com>
---
 kernel/locking/rwsem-xadd.c |   21 +++++++++++++++++++--
 1 files changed, 19 insertions(+), 2 deletions(-)

diff --git a/kernel/locking/rwsem-xadd.c b/kernel/locking/rwsem-xadd.c
index 2bb25e2..815f0cc 100644
--- a/kernel/locking/rwsem-xadd.c
+++ b/kernel/locking/rwsem-xadd.c
@@ -399,6 +399,15 @@ static inline bool rwsem_has_spinner(struct rw_semaphore *sem)
 	return osq_is_locked(&sem->osq);
 }
 
+/*
+ * Return true if there is an active writer by checking the owner field which
+ * should be set if there is one.
+ */
+static inline bool rwsem_has_active_writer(struct rw_semaphore *sem)
+{
+	return READ_ONCE(sem->owner) != NULL;
+}
+
 #else
 static bool rwsem_optimistic_spin(struct rw_semaphore *sem)
 {
@@ -409,6 +418,11 @@ static inline bool rwsem_has_spinner(struct rw_semaphore *sem)
 {
 	return false;
 }
+
+static inline bool rwsem_has_active_writer(struct rw_semaphore *sem)
+{
+	return false;	/* Assume it has no active writer */
+}
 #endif
 
 /*
@@ -524,8 +538,11 @@ struct rw_semaphore *rwsem_wake(struct rw_semaphore *sem)
 	raw_spin_lock_irqsave(&sem->wait_lock, flags);
 locked:
 
-	/* do nothing if list empty */
-	if (!list_empty(&sem->wait_list))
+	/*
+	 * Do nothing if list empty or the lock has just been stolen by a
+	 * writer after a possibly long wait in getting the wait_lock.
+	 */
+	if (!list_empty(&sem->wait_list) && !rwsem_has_active_writer(sem))
 		sem = __rwsem_do_wake(sem, RWSEM_WAKE_ANY);
 
 	raw_spin_unlock_irqrestore(&sem->wait_lock, flags);
-- 
1.7.1

Re: [PATCH v4 1/2] locking/rwsem: reduce spinlock contention in wakeup after up_read/up_write

[Jason Low]

On Thu, 2015-04-30 at 17:12 -0400, Waiman Long wrote:
> In up_write()/up_read(), rwsem_wake() will be called whenever it
> detects that some writers/readers are waiting. The rwsem_wake()
> function will take the wait_lock and call __rwsem_do_wake() to do the
> real wakeup.  For a heavily contended rwsem, doing a spin_lock() on
> wait_lock will cause further contention on the heavily contended rwsem
> cacheline resulting in delay in the completion of the up_read/up_write
> operations.
> 
> This patch makes the wait_lock taking and the call to __rwsem_do_wake()
> optional if at least one spinning writer is present. The spinning
> writer will be able to take the rwsem and call rwsem_wake() later
> when it calls up_write(). With the presence of a spinning writer,
> rwsem_wake() will now try to acquire the lock using trylock. If that
> fails, it will just quit.
> 
> Signed-off-by: Waiman Long <Waiman.Long@hp.com>
> Suggested-by: Peter Zijlstra (Intel) <peterz@infradead.org>

Acked-by: Jason Low <jason.low2@hp.com>

Re: [PATCH v4 1/2] locking/rwsem: reduce spinlock contention in wakeup after up_read/up_write

[Peter Zijlstra]

On Thu, Apr 30, 2015 at 05:12:16PM -0400, Waiman Long wrote:
> In up_write()/up_read(), rwsem_wake() will be called whenever it
> detects that some writers/readers are waiting. The rwsem_wake()
> function will take the wait_lock and call __rwsem_do_wake() to do the
> real wakeup.  For a heavily contended rwsem, doing a spin_lock() on
> wait_lock will cause further contention on the heavily contended rwsem
> cacheline resulting in delay in the completion of the up_read/up_write
> operations.
> 
> This patch makes the wait_lock taking and the call to __rwsem_do_wake()
> optional if at least one spinning writer is present. The spinning
> writer will be able to take the rwsem and call rwsem_wake() later
> when it calls up_write(). With the presence of a spinning writer,
> rwsem_wake() will now try to acquire the lock using trylock. If that
> fails, it will just quit.
> 
> Signed-off-by: Waiman Long <Waiman.Long@hp.com>
> Suggested-by: Peter Zijlstra (Intel) <peterz@infradead.org>
> ---

Thanks!

Re: [PATCH v4 1/2] locking/rwsem: reduce spinlock contention in wakeup after up_read/up_write

[Davidlohr Bueso]

On Thu, 2015-04-30 at 17:12 -0400, Waiman Long wrote:
> In up_write()/up_read(), rwsem_wake() will be called whenever it
> detects that some writers/readers are waiting. The rwsem_wake()
> function will take the wait_lock and call __rwsem_do_wake() to do the
> real wakeup.  For a heavily contended rwsem, doing a spin_lock() on
> wait_lock will cause further contention on the heavily contended rwsem
> cacheline resulting in delay in the completion of the up_read/up_write
> operations.
> 
> This patch makes the wait_lock taking and the call to __rwsem_do_wake()
> optional if at least one spinning writer is present. The spinning
> writer will be able to take the rwsem and call rwsem_wake() later
> when it calls up_write(). With the presence of a spinning writer,
> rwsem_wake() will now try to acquire the lock using trylock. If that
> fails, it will just quit.
> 
> Signed-off-by: Waiman Long <Waiman.Long@hp.com>
> Suggested-by: Peter Zijlstra (Intel) <peterz@infradead.org>

Reviewed-by: Davidlohr Bueso <dave@stgolabs.net>

Re: [PATCH v4 1/2] locking/rwsem: reduce spinlock contention in wakeup after up_read/up_write

[Davidlohr Bueso]

On Wed, 2015-05-06 at 04:18 -0700, Davidlohr Bueso wrote:
> Reviewed-by: Davidlohr Bueso <dave@stgolabs.net>

A nit, but it would be useful if the benchmark/perf numbers were also in
this changelog, for future references.

Thanks,
Davidlohr

[tip:locking/core] locking/rwsem: Reduce spinlock contention in wakeup after up_read()/up_write()

[tip-bot for Waiman Long]

Commit-ID:  59aabfc7e959f5f213e4e5cc7567ab4934da2adf
Gitweb:     http://git.kernel.org/tip/59aabfc7e959f5f213e4e5cc7567ab4934da2adf
Author:     Waiman Long <Waiman.Long@hp.com>
AuthorDate: Thu, 30 Apr 2015 17:12:16 -0400
Committer:  Ingo Molnar <mingo@kernel.org>
CommitDate: Fri, 8 May 2015 12:27:59 +0200

locking/rwsem: Reduce spinlock contention in wakeup after up_read()/up_write()

In up_write()/up_read(), rwsem_wake() will be called whenever it
detects that some writers/readers are waiting. The rwsem_wake()
function will take the wait_lock and call __rwsem_do_wake() to do the
real wakeup.  For a heavily contended rwsem, doing a spin_lock() on
wait_lock will cause further contention on the heavily contended rwsem
cacheline resulting in delay in the completion of the up_read/up_write
operations.

This patch makes the wait_lock taking and the call to __rwsem_do_wake()
optional if at least one spinning writer is present. The spinning
writer will be able to take the rwsem and call rwsem_wake() later
when it calls up_write(). With the presence of a spinning writer,
rwsem_wake() will now try to acquire the lock using trylock. If that
fails, it will just quit.

Suggested-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Waiman Long <Waiman.Long@hp.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Davidlohr Bueso <dave@stgolabs.net>
Acked-by: Jason Low <jason.low2@hp.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Douglas Hatch <doug.hatch@hp.com>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Scott J Norton <scott.norton@hp.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/1430428337-16802-2-git-send-email-Waiman.Long@hp.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
---
 include/linux/osq_lock.h    |  5 +++++
 kernel/locking/rwsem-xadd.c | 44 ++++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 49 insertions(+)

diff --git a/include/linux/osq_lock.h b/include/linux/osq_lock.h
index 3a6490e..703ea5c 100644
--- a/include/linux/osq_lock.h
+++ b/include/linux/osq_lock.h
@@ -32,4 +32,9 @@ static inline void osq_lock_init(struct optimistic_spin_queue *lock)
 extern bool osq_lock(struct optimistic_spin_queue *lock);
 extern void osq_unlock(struct optimistic_spin_queue *lock);
 
+static inline bool osq_is_locked(struct optimistic_spin_queue *lock)
+{
+	return atomic_read(&lock->tail) != OSQ_UNLOCKED_VAL;
+}
+
 #endif
diff --git a/kernel/locking/rwsem-xadd.c b/kernel/locking/rwsem-xadd.c
index 3417d01..0f18971 100644
--- a/kernel/locking/rwsem-xadd.c
+++ b/kernel/locking/rwsem-xadd.c
@@ -409,11 +409,24 @@ done:
 	return taken;
 }
 
+/*
+ * Return true if the rwsem has active spinner
+ */
+static inline bool rwsem_has_spinner(struct rw_semaphore *sem)
+{
+	return osq_is_locked(&sem->osq);
+}
+
 #else
 static bool rwsem_optimistic_spin(struct rw_semaphore *sem)
 {
 	return false;
 }
+
+static inline bool rwsem_has_spinner(struct rw_semaphore *sem)
+{
+	return false;
+}
 #endif
 
 /*
@@ -496,7 +509,38 @@ struct rw_semaphore *rwsem_wake(struct rw_semaphore *sem)
 {
 	unsigned long flags;
 
+	/*
+	 * If a spinner is present, it is not necessary to do the wakeup.
+	 * Try to do wakeup only if the trylock succeeds to minimize
+	 * spinlock contention which may introduce too much delay in the
+	 * unlock operation.
+	 *
+	 *    spinning writer		up_write/up_read caller
+	 *    ---------------		-----------------------
+	 * [S]   osq_unlock()		[L]   osq
+	 *	 MB			      RMB
+	 * [RmW] rwsem_try_write_lock() [RmW] spin_trylock(wait_lock)
+	 *
+	 * Here, it is important to make sure that there won't be a missed
+	 * wakeup while the rwsem is free and the only spinning writer goes
+	 * to sleep without taking the rwsem. Even when the spinning writer
+	 * is just going to break out of the waiting loop, it will still do
+	 * a trylock in rwsem_down_write_failed() before sleeping. IOW, if
+	 * rwsem_has_spinner() is true, it will guarantee at least one
+	 * trylock attempt on the rwsem later on.
+	 */
+	if (rwsem_has_spinner(sem)) {
+		/*
+		 * The smp_rmb() here is to make sure that the spinner
+		 * state is consulted before reading the wait_lock.
+		 */
+		smp_rmb();
+		if (!raw_spin_trylock_irqsave(&sem->wait_lock, flags))
+			return sem;
+		goto locked;
+	}
 	raw_spin_lock_irqsave(&sem->wait_lock, flags);
+locked:
 
 	/* do nothing if list empty */
 	if (!list_empty(&sem->wait_list))