On Fri, Jul 17, 2020 at 06:02:47PM -0700, Eric Biggers wrote:
> On Fri, Jul 17, 2020 at 01:51:38PM -0400, Alan Stern wrote:
> > On Fri, Jul 17, 2020 at 06:47:50PM +0100, Matthew Wilcox wrote:
> > > On Thu, Jul 16, 2020 at 09:44:27PM -0700, Eric Biggers wrote:
> > ...
> > > > + /* on success, pairs with smp_load_acquire() above and below */
> > > > + if (cmpxchg_release(&foo, NULL, p) != NULL) {
> > >
> > > Why do we have cmpxchg_release() anyway? Under what circumstances is
> > > cmpxchg() useful _without_ having release semantics?
> >
> > To answer just the last question: cmpxchg() is useful for lock
> > acquisition, in which case it needs to have acquire semantics rather
> > than release semantics.
> >
>
> To clarify, there are 4 versions of cmpxchg:
>
> cmpxchg(): does ACQUIRE and RELEASE (on success)
> cmpxchg_acquire(): does ACQUIRE only (on success)
> cmpxchg_release(): does RELEASE only (on success)
> cmpxchg_relaxed(): no barriers
>
> The problem here is that here we need RELEASE on success and ACQUIRE on failure.
> But no version guarantees any barrier on failure.
Why not? Do CPU designers not do load-linked-with-acquire-semantics?
Or is it our fault for not using the appropriate instruction?
> So as far as I can tell, the best we can do is use cmpxchg_release() (or
> cmpxchg() which would be stronger but unnecessary), followed by a separate
> ACQUIRE on failure.
OK, but that detail needs to be hidden behind a higher level primitive,
not exposed to device driver writers.
