atomic_add_unless() and mb()

[Oleg Nesterov <oleg@tv-sign.ru>]

Both Documentation/memory-barriers.txt and Documentation/atomic_ops.txt state
that atomic_add_unless() implies smp_mb() on each side of the operation.

Is it true?

include/asm-most_of/atomic.h:

	#define atomic_add_unless(v, a, u)				\
	({								\
		int c, old;						\
		c = atomic_read(v);					\
		for (;;) {						\
			if (unlikely(c == (u)))				\
				break;					\
			old = atomic_cmpxchg((v), c, c + (a));		\
			if (likely(old == c))				\
				break;					\
			c = old;					\
		}							\
		c != (u);						\
	})

This looks like atomic_add_unless() implies mb() only if it returns 1.
Otherwise it could fail (return 0) before the first atomic_cmpxchg(),
but atomic_read() provides a compiler barrier only.

Could you clarify?

Oleg.