On Wed, Aug 24, 2016 at 06:52:44PM +0200, Peter Zijlstra wrote:
> On Wed, Aug 24, 2016 at 05:34:12PM +0200, Peter Zijlstra wrote:
> > On Wed, Aug 24, 2016 at 10:56:59AM +0100, Will Deacon wrote:
> > > > + owner = atomic_long_read(&lock->owner);
> > > > + for (;;) {
> > > > + unsigned long old;
> > > > +
> > > > + old = atomic_long_cmpxchg_release(&lock->owner, owner,
> > > > owner & 0x03);
> > > > + if (old == owner)
> > > > + break;
> > > > +
> > > > + owner = old;
> > > > + }
> > >
> > > Can you rewrite this using atomic_long_fetch_and_release?
> >
> > Yes, until patch 3/3.. but now that you mention it I think we can do:
> >
> > owner = atomic_long_read(&lock->owner);
> > if (!(owner & MUTEX_FLAG_HANDOFF))
> > (void)atomic_long_fetch_and_release(MUTEX_FLAGS, &lock->owner);
> >
>
> And of course, x86 would very much like atomic_long_and_release() here,
> such that it can do LOCK ADD instead of a LOCK CMPXCHG loop. But of
> course, we don't have that ...
... yeah, I noticed that. There is a curious use of atomic_and in
linux/atomic.h, but it's packed full of false promises.
Will
