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.comSigned-off-by: Ingo Molnar <mingo@kernel.org>

include/linux/mutex.h

@@ -53,6 +53,9 @@ struct mutex {
 #if defined(CONFIG_DEBUG_MUTEXES) || defined(CONFIG_SMP)
 	struct task_struct	*owner;
 #endif
+#ifdef CONFIG_MUTEX_SPIN_ON_OWNER
+	void			*spin_mlock;	/* Spinner MCS lock */
+#endif
 #ifdef CONFIG_DEBUG_MUTEXES
 	const char 		*name;
 	void			*magic;

kernel/mutex.c

@@ -55,6 +55,9 @@ __mutex_init(struct mutex *lock, const char *name, struct lock_class_key *key)
 	spin_lock_init(&lock->wait_lock);
 	INIT_LIST_HEAD(&lock->wait_list);
 	mutex_clear_owner(lock);
+#ifdef CONFIG_MUTEX_SPIN_ON_OWNER
+	lock->spin_mlock = NULL;
+#endif
 
 	debug_mutex_init(lock, name, key);
 }
@@ -107,6 +110,60 @@ EXPORT_SYMBOL(mutex_lock);
 #endif
 
 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
+/*
+ * In order to avoid a stampede of mutex spinners from acquiring the mutex
+ * more or less simultaneously, the spinners need to acquire a MCS lock
+ * first before spinning on the owner field.
+ *
+ * We don't inline mspin_lock() so that perf can correctly account for the
+ * time spent in this lock function.
+ */
+struct mspin_node {
+	struct mspin_node *next ;
+	int		  locked;	/* 1 if lock acquired */
+};
+#define	MLOCK(mutex)	((struct mspin_node **)&((mutex)->spin_mlock))
+
+static noinline
+void mspin_lock(struct mspin_node **lock, struct mspin_node *node)
+{
+	struct mspin_node *prev;
+
+	/* Init node */
+	node->locked = 0;
+	node->next   = NULL;
+
+	prev = xchg(lock, node);
+	if (likely(prev == NULL)) {
+		/* Lock acquired */
+		node->locked = 1;
+		return;
+	}
+	ACCESS_ONCE(prev->next) = node;
+	smp_wmb();
+	/* Wait until the lock holder passes the lock down */
+	while (!ACCESS_ONCE(node->locked))
+		arch_mutex_cpu_relax();
+}
+
+static void mspin_unlock(struct mspin_node **lock, struct mspin_node *node)
+{
+	struct mspin_node *next = ACCESS_ONCE(node->next);
+
+	if (likely(!next)) {
+		/*
+		 * Release the lock by setting it to NULL
+		 */
+		if (cmpxchg(lock, node, NULL) == node)
+			return;
+		/* Wait until the next pointer is set */
+		while (!(next = ACCESS_ONCE(node->next)))
+			arch_mutex_cpu_relax();
+	}
+	ACCESS_ONCE(next->locked) = 1;
+	smp_wmb();
+}
+
 /*
  * Mutex spinning code migrated from kernel/sched/core.c
  */
@@ -150,6 +207,24 @@ int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
 	 */
 	return lock->owner == NULL;
 }
+
+/*
+ * Initial check for entering the mutex spinning loop
+ */
+static inline int mutex_can_spin_on_owner(struct mutex *lock)
+{
+	int retval = 1;
+
+	rcu_read_lock();
+	if (lock->owner)
+		retval = lock->owner->on_cpu;
+	rcu_read_unlock();
+	/*
+	 * if lock->owner is not set, the mutex owner may have just acquired
+	 * it and not set the owner yet or the mutex has been released.
+	 */
+	return retval;
+}
 #endif
 
 static __used noinline void __sched __mutex_unlock_slowpath(atomic_t *lock_count);
@@ -215,26 +290,39 @@ __mutex_lock_common(struct mutex *lock, long state, unsigned int subclass,
 	 *
 	 * We can't do this for DEBUG_MUTEXES because that relies on wait_lock
 	 * to serialize everything.
+	 *
+	 * 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.
 	 */
+	if (!mutex_can_spin_on_owner(lock))
+		goto slowpath;
 
 	for (;;) {
 		struct task_struct *owner;
+		struct mspin_node  node;
 
 		/*
 		 * If there's an owner, wait for it to either
 		 * release the lock or go to sleep.
 		 */
+		mspin_lock(MLOCK(lock), &node);
 		owner = ACCESS_ONCE(lock->owner);
-		if (owner && !mutex_spin_on_owner(lock, owner))
+		if (owner && !mutex_spin_on_owner(lock, owner)) {
+			mspin_unlock(MLOCK(lock), &node);
 			break;
+		}
 
 		if ((atomic_read(&lock->count) == 1) &&
 		    (atomic_cmpxchg(&lock->count, 1, 0) == 1)) {
 			lock_acquired(&lock->dep_map, ip);
 			mutex_set_owner(lock);
+			mspin_unlock(MLOCK(lock), &node);
 			preempt_enable();
 			return 0;
 		}
+		mspin_unlock(MLOCK(lock), &node);
 
 		/*
 		 * When there's no owner, we might have preempted between the
@@ -253,6 +341,7 @@ __mutex_lock_common(struct mutex *lock, long state, unsigned int subclass,
 		 */
 		arch_mutex_cpu_relax();
 	}
+slowpath:
 #endif
 	spin_lock_mutex(&lock->wait_lock, flags);