Re: Possible memory ordering bug in page reclaim?

[Benjamin Herrenschmidt <benh@kernel.crashing.org>]

> Even a write barrier is required on the left side, the read barrier on
> the right side is useless if there is no write barrier on the left side.
> 
> Note that the barrier in atomic_add_negative is useless here because it
> happens way too late, _after_ the count is decremented (not _before_)
> so the decreased count could be already visible to the other cpu.

Not on ppc64. Our atomic*return functions have a write barrier before
the atomic operation. I think there is just too much code out there that
either expects implicit barriers done by atomics (abuse them as locks)
or simply written by people who don't understand those ordering issues
(to be fair, they can be fairly brain damaging).

I agree this is not a good generic solution though.

> Not all archs are like x86 where a barrier happens implicitly both
> before and after the instruction, and the way atomic_add_negative is
> implemented the barrier from a common code point of view is only added
> _after_ the instruction.