Re: [RFC][PATCH] mips: Fix arch_spin_unlock()

[Will Deacon]

On Tue, Feb 02, 2016 at 11:55:57AM -0800, Linus Torvalds wrote:
> On Tue, Feb 2, 2016 at 11:30 AM, Will Deacon <will.deacon@xxxxxxx> wrote:
> >
> > FWIW, and this is by no means conclusive, I hacked that up quickly and
> > ran hackbench a few times on the nearest idle arm64 system. The results
> > were consistently ~4% slower using acquire for rcu_dereference.
> 
> Ok, that's *much* more noticeable than I would have expected. I take
> it that load-acquire is really really slow on current arm64
> implementations.

See my reply to Ingo, but it seems a bunch of this was down to rebooting
the system between runs and hackbench being particularly susceptible to
that.

> Just out of interest, is store-release slow too? Because that should
> be easy to make fast.

There's a slight gotcha with arm64's store-release instruction in that
it's RCsc and therefore orders against a subsequent load-acquire. That's
not to say you can't make it fast, but it's potentially more involved
than posting a flag in a store buffer (or whatever you were envisaging :)

Measuring store-release is much more difficult, because you can't replace
it with a dependency or the like, only other barrier constructs.

Will