Re: [PATCH 2/7] atomics/treewide: rework ordering barriers

From: Mark Rutland
Date: Tue Jun 05 2018 - 09:28:11 EST


On Tue, Jun 05, 2018 at 02:16:23PM +0200, Peter Zijlstra wrote:
> On Tue, May 29, 2018 at 07:07:41PM +0100, Mark Rutland wrote:
> > +#ifndef __atomic_mb__after_acquire
> > +#define __atomic_mb__after_acquire smp_mb__after_atomic
> > +#endif
> > +
> > +#ifndef __atomic_mb__before_release
> > +#define __atomic_mb__before_release smp_mb__before_atomic
> > +#endif
> > +
> > +#ifndef __atomic_mb__before_fence
> > +#define __atomic_mb__before_fence smp_mb__before_atomic
> > +#endif
> > +
> > +#ifndef __atomic_mb__after_fence
> > +#define __atomic_mb__after_fence smp_mb__after_atomic
> > +#endif
>
> I really _really_ dislike those names.. because they imply providing an
> MB before/after something else.
>
> But that is exactly what they do not.
>
> How about:
>
> __atomic_acquire_fence
> __atomic_release_fence
>
> for the acquire/release things,

Sure, those sound fine to me.

> and simply using smp_mb__{before,after}_atomic for the full fence, its
> exactly what they were made for.

The snag is arch/alpha, whare we have:

/*
* To ensure dependency ordering is preserved for the _relaxed and
* _release atomics, an smp_read_barrier_depends() is unconditionally
* inserted into the _relaxed variants, which are used to build the
* barriered versions. To avoid redundant back-to-back fences, we can
* define the _acquire and _fence versions explicitly.
*/
#define __atomic_op_acquire(op, args...) op##_relaxed(args)
#define __atomic_op_fence __atomic_op_release

... where alpha's smp_read_barrier_depends() is the same as
smp_mb_after_atomic().

Since alpha's non-value-returning atomics do not have the
smp_read_barrier_depends(), I can't just define an empty
smp_mb_after_atomic().

Thoughts?

Thanks,
Mark.