[PATCH] synchronize_irq needs a barrier

[Benjamin Herrenschmidt]

On Thu, 2007-10-18 at 22:35 +0800, Herbert Xu wrote:
> Benjamin Herrenschmidt <benh at kernel.crashing.org> wrote:
> > 
> > Note that some kind of read barrier or compiler barrier should be needed
> > regardless, or we are just not sync'ing with anything at all (we may
> > have loaded the value ages ago and thus operate on a totally stale
> > value). I prefer a full barrier to also ensure all previous stores are
> > pushed out.
> 
> We already have a compiler barrier there in the form of cpu_relax.

Isn't it too late ? The barrier should be before the test_bit, to
prevent it from moving up.

Ben.