ipc/sem.c: change memory barrier in sem_lock() to smp_rmb()

When I fixed bugs in the sem_lock() logic, I was more conservative than
necessary.  Therefore it is safe to replace the smp_mb() with smp_rmb().
And: With smp_rmb(), semop() syscalls are up to 10% faster.

The race we must protect against is:

	sem->lock is free
	sma->complex_count = 0
	sma->sem_perm.lock held by thread B

thread A:

A: spin_lock(&sem->lock)

			B: sma->complex_count++; (now 1)
			B: spin_unlock(&sma->sem_perm.lock);

A: spin_is_locked(&sma->sem_perm.lock);
A: XXXXX memory barrier
A: if (sma->complex_count == 0)

Thread A must read the increased complex_count value, i.e. the read must
not be reordered with the read of sem_perm.lock done by spin_is_locked().

Since it's about ordering of reads, smp_rmb() is sufficient.

[[email protected]: update sem_lock() comment, from Davidlohr]
Signed-off-by: Manfred Spraul <[email protected]>
Reviewed-by: Davidlohr Bueso <[email protected]>
Acked-by: Rafael Aquini <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>

ipc/sem.c

@@ -326,10 +326,17 @@ static inline int sem_lock(struct sem_array *sma, struct sembuf *sops,
 
 		/* Then check that the global lock is free */
 		if (!spin_is_locked(&sma->sem_perm.lock)) {
-			/* spin_is_locked() is not a memory barrier */
-			smp_mb();
+			/*
+			 * The ipc object lock check must be visible on all
+			 * cores before rechecking the complex count.  Otherwise
+			 * we can race with  another thread that does:
+			 *	complex_count++;
+			 *	spin_unlock(sem_perm.lock);
+			 */
+			smp_rmb();
 
-			/* Now repeat the test of complex_count:
+			/*
+			 * Now repeat the test of complex_count:
 			 * It can't change anymore until we drop sem->lock.
 			 * Thus: if is now 0, then it will stay 0.
 			 */