locking/mutexes: Unlock the mutex without the wait_lock

When running workloads that have high contention in mutexes on an 8 socket
machine, mutex spinners would often spin for a long time with no lock owner.

The main reason why this is occuring is in __mutex_unlock_common_slowpath(),
if __mutex_slowpath_needs_to_unlock(), then the owner needs to acquire the
mutex->wait_lock before releasing the mutex (setting lock->count to 1). When
the wait_lock is contended, this delays the mutex from being released.
We should be able to release the mutex without holding the wait_lock.

Signed-off-by: Jason Low <[email protected]>
Cc: [email protected]
Cc: [email protected]
Cc: [email protected]
Cc: [email protected]
Cc: [email protected]
Cc: [email protected]
Cc: [email protected]
Cc: [email protected]
Cc: [email protected]
Cc: [email protected]
Cc: [email protected]
Cc: [email protected]
Signed-off-by: Peter Zijlstra <[email protected]>
Link: http://lkml.kernel.org/r/[email protected]
Signed-off-by: Ingo Molnar <[email protected]>

kernel/locking/mutex.c

@@ -671,10 +671,6 @@ __mutex_unlock_common_slowpath(atomic_t *lock_count, int nested)
 	struct mutex *lock = container_of(lock_count, struct mutex, count);
 	unsigned long flags;
 
-	spin_lock_mutex(&lock->wait_lock, flags);
-	mutex_release(&lock->dep_map, nested, _RET_IP_);
-	debug_mutex_unlock(lock);
-
 	/*
 	 * some architectures leave the lock unlocked in the fastpath failure
 	 * case, others need to leave it locked. In the later case we have to
@@ -683,6 +679,10 @@ __mutex_unlock_common_slowpath(atomic_t *lock_count, int nested)
 	if (__mutex_slowpath_needs_to_unlock())
 		atomic_set(&lock->count, 1);
 
+	spin_lock_mutex(&lock->wait_lock, flags);
+	mutex_release(&lock->dep_map, nested, _RET_IP_);
+	debug_mutex_unlock(lock);
+
 	if (!list_empty(&lock->wait_list)) {
 		/* get the first entry from the wait-list: */
 		struct mutex_waiter *waiter =