Re: [PATCH for-3.3] mempool: clean up and document synchronization and memory barrier usage

[Tejun Heo <tj@kernel.org>]

Hello,

On Wed, Dec 21, 2011 at 06:40:58PM +0100, Oleg Nesterov wrote:
> > The first paragraph is saying that at that point full barrier (for
> > both stores and loads) is necessary at that point and the second
> > paragraph is a bit confusing but the last sentence seems to say that
> > only loads after the unlock can creep above unlock,
> 
> Probably, this is because the comment tries to explain the possible
> reordering with the subsequent "if (condition)" check, so it only
> mentions loads.

Ah, I see.

> > Anyways, yeah, you're right.  We need a smp_wmb() before returning but
> > I think the comment on top of prepare_to_wait() is misleading.
> 
> Hmm. I am not sure I understand... Although almost everything written
> in English looks misleading to me ;)

Amen. :) I missed the context there, so please forget about it.

> > Great, thanks.  I'll wait a bit for futher comments and repost w/
> > smp_wmb() added.
> 
> Well. This is almost off-topic, but perhaps we can add
> smp_mb__after_unlock() ? We already have smp_mb__after_lock.
> Afaics prepare_to_wait() could use it.
> 
> I am not talking about perfomance issues, just I think the code
> will be more understandable.

Hmmm... maybe.  I really don't know.

Thanks.

-- 
tejun