Re: futex atomic vs ordering constraints

[Will Deacon <will.deacon@arm.com>]

On Tue, Sep 01, 2015 at 05:42:47PM +0100, Peter Zijlstra wrote:
> On Tue, Sep 01, 2015 at 05:31:40PM +0100, Will Deacon wrote:
> > On Wed, Aug 26, 2015 at 07:16:59PM +0100, Peter Zijlstra wrote:
> > > I tried to keep this email short, but failed miserably at this. For
> > > the TL;DR skip to the tail.
> > 
> > [...]
> > 
> > > There are a few options:
> > > 
> > >  1) punt, mandate they're both fully ordered and stop thinking about it
> > > 
> > >  2) make them both fully relaxed, rely on implied barriers and employ
> > >     smp_mb__{before,after}_atomic in key places
> > > 
> > > Given the current state of things and that I don't really think there is
> > > a compelling performance argument to be made for 2, I would suggest we
> > > go with 1.
> > 
> > I'd also go for (1). Since there is a userspace side to this, I'd *really*
> > like to avoid a potential situation on arm64 where the kernel builds its
> > side of the futex using barrier instructions (e.g. treat LDR + smp_mb()
> > as acquire) and userspace builds its side out of native acquire/release
> > instructions and the two end up interacting badly (for example, loss of
> > SC).
> 
> I thought your native acquire/release were RCsc, or is it that in
> combination with the 'fancy' 'full' barrier of stlxr + dmb-ish something
> goes sideways?

Yeah, they don't interact nicely because you can lose the multi-copy
atomicity guarantees you get from using either native acquire/release
everywhere or explicit barriers everywhere. IRIW shows the failure
if you use {DMB; STR} for the writers and LDAR for the readers.

> But yes, unless Thomas has other plans, I'll go ahead and create some
> patches to make sure everybody is fully ordered so we can forget about
> it again.

Sounds good to me!

Will