documentation: atomic_add_unless() doesn't imply mb() on failure
authorOleg Nesterov <oleg@tv-sign.ru>
Sat, 23 Feb 2008 21:03:29 +0000 (00:03 +0300)
committerLinus Torvalds <torvalds@woody.linux-foundation.org>
Sun, 24 Feb 2008 01:52:36 +0000 (17:52 -0800)
commit02c608c1fefe821a4c6fc34c45a0dea3cebf4764
tree8fd360da8bc83dd45ecec5621791baa73ddfac70
parent2dda81ca31dc73e695ff8b83351f7aaefbef192a
documentation: atomic_add_unless() doesn't imply mb() on failure

(sorry for being offtpoic, but while experts are here...)

A "typical" implementation of atomic_add_unless() can return 0 immediately
after the first atomic_read() (before doing cmpxchg). In that case it doesn't
provide any barrier semantics. See include/asm-ia64/atomic.h as an example.

We should either change the implementation, or fix the docs.

Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>
Acked-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Documentation/atomic_ops.txt
Documentation/memory-barriers.txt