Re: Broken ARM atomic ops wrt memory barriers (was : [PATCH] Add cmpxchg support for ARMv6+ systems)

[Catalin Marinas <catalin.marinas@arm.com>]

On Mon, 2009-05-25 at 20:34 +0100, Russell King - ARM Linux wrote:
> On Mon, May 25, 2009 at 01:29:55PM -0400, Mathieu Desnoyers wrote:
> > Basically, to make sure we don't forget anything, someone should go
> > through all the atomic_ops.txt document once more and audit all ARM
> > primitives.
> 
> That's easy to say - it took me more than half an hour of reading through
> atomic_ops.txt to work out what was required for things like cmpxchg,
> xchg, etc because it's _very_ waffley and verbose, and directs you to
> other parts of the document.
> 
> For example, for atomic_cmpxchg it directs you to 'cas' but cas doesn't
> really say what the barrier requirements are - reading it leaves me
> to expect that provided spinlocks are serializing themselves, it's
> fine if 'cas' itself isn't.
> 
> Maybe if someone has a few days to translate atomic_ops.txt into a
> succinct set of requirements, and get _that_ reviewed, then we could
> properly audit this stuff.

Documentation/memory-barriers.txt has a section on atomic operations and
it states that barriers should be placed on each side of the cmpxchg and
xchg operations.

-- 
Catalin