Re: [PATCHv2 06/11] atomics/treewide: rework ordering barriers

[Will Deacon]

On Thu, Jul 05, 2018 at 11:12:41AM +0100, Mark Rutland wrote:
> On Wed, Jul 04, 2018 at 06:50:00PM +0100, Will Deacon wrote:
> > On Wed, Jul 04, 2018 at 04:56:19PM +0100, Mark Rutland wrote:
> > > On Wed, Jul 04, 2018 at 04:06:46PM +0100, Will Deacon wrote:
> > > > On Mon, Jun 25, 2018 at 11:59:47AM +0100, Mark Rutland wrote:
> > > > > Currently architectures can override __atomic_op_*() to define the barriers
> > > > > used before/after a relaxed atomic when used to build acquire/release/fence
> > > > > variants.
> > > > > 
> > > > > This has the unfortunate property of requiring the architecture to define the
> > > > > full wrapper for the atomics, rather than just the barriers they care about,
> > > > > and gets in the way of generating atomics which can be easily read.
> > > > > 
> > > > > Instead, this patch has architectures define an optional set of barriers,
> > > > > __atomic_mb_{before,after}_{acquire,release,fence}(), which <linux/atomic.h>
> > > > > uses to build the wrappers.
> > > > 
> > > > Looks like you've renamed these in the patch but not updated the commit
> > > > message.
> > > 
> > > Yup; Peter also pointed that out. In my branch this now looks like:
> > > 
> > > ----
> > > Instead, this patch has architectures define an optional set of barriers:
> > > 
> > > * __atomic_acquire_fence()
> > > * __atomic_release_fence()
> > > * __atomic_pre_fence()
> > > * __atomic_post_fence()
> > > 
> > > ... which <linux/atomic.h> uses to build the wrappers.
> > > ----
> > > 
> > > ... which is hopefully more legible, too!
> > > 
> > > > Also, to add to the bikeshedding, would it worth adding "rmw" in there
> > > > somewhere, e.g. __atomic_post_rmw_fence, since I assume these only
> > > > apply to value-returning stuff?
> > > 
> > > I don't have any opinion there, but I'm also not sure I've parsed your
> > > rationale correctly. I guess a !RMW full-fence op doesn't make sense? Or
> > > that's something we want to avoid in the API?
> > > 
> > > AFAICT, we only use __atomic_{pre,post}_fence() for RMW ops today.
> > 
> > No, I think you're right and my terminology is confused. Leave it as-is
> > for the moment.
> 
> Sure thing.
> 
> Perhaps __atomic_{pre,post}_full_fence() might be better, assuming
> you're trying to avoid people erroneously assuming that
> __atomic_{pre,post}_fence() are like acquire/release fences.

Good idea, I think that's better.

Will