Correct fences for sys/refcount.h.

The acq barrier in refcount_acquire() has no use, constructor must
ensure that the changes are visible before publication by other means.
Last release must sync/with the constructor and all updaters.

This is based on the refcount/shared_ptr analysis I heard at the Hans
Boehm and Herb Sutter talks about C++ atomics.

Reviewed by:	alc, jhb, markj
Sponsored by:	The FreeBSD Foundation
MFC after:	2 weeks
Differential revision:	https://reviews.freebsd.org/D11270

Author: kostikbel

sys/sys/refcount.h

@@ -50,18 +50,28 @@ refcount_acquire(volatile u_int *count)
 {
 
 	KASSERT(*count < UINT_MAX, ("refcount %p overflowed", count));
-	atomic_add_acq_int(count, 1);	
+	atomic_add_int(count, 1);
 }
 
 static __inline int
 refcount_release(volatile u_int *count)
 {
 	u_int old;
 
-	/* XXX: Should this have a rel membar? */
+	atomic_thread_fence_rel();
 	old = atomic_fetchadd_int(count, -1);
 	KASSERT(old > 0, ("negative refcount %p", count));
-	return (old == 1);
+	if (old > 1)
+		return (0);
+
+	/*
+	 * Last reference.  Signal the user to call the destructor.
+	 *
+	 * Ensure that the destructor sees all updates.  The fence_rel
+	 * at the start of the function synchronized with this fence.
+	 */
+	atomic_thread_fence_acq();
+	return (1);
 }
 
 #endif	/* ! __SYS_REFCOUNT_H__ */