On Thu, Mar 26, 2015 at 02:51:44PM +0000, Peter Zijlstra wrote:
> On Thu, Mar 26, 2015 at 02:41:54PM +0000, Will Deacon wrote:
> > > +++ b/lib/lockref.c
> > > @@ -18,7 +18,8 @@
> > > #define CMPXCHG_LOOP(CODE, SUCCESS) do { \
> > > struct lockref old; \
> > > BUILD_BUG_ON(sizeof(old) != 8); \
> > > - old.lock_count = READ_ONCE(lockref->lock_count); \
> > > + barrier(); \
> > > + old.lock_count = lockref->lock_count; \
> > > while (likely(arch_spin_value_unlocked(old.lock.rlock.raw_lock))) { \
> > > struct lockref new = old, prev = old; \
> > > CODE \
> >
> > Is ACCESS_ONCE actually going away?
>
> I've been arguing for that yes, having two APIs for the 'same' thing is
> confusing at best, and as the comment near the READ_ONCE() thing
> explains, ACCESS_ONCE() has serious, silent, issues.
>
> > It has its problems, but I think it's
> > what we want here and reads better than magic barrier() imo.
>
> Yeah, but its also misleading because we rely on silent fail. Part of
> the ACCESS_ONCE() semantics is that it should avoid split loads, and
> we're here actually relying on emitting just that.
In which case, on the premise that we comment the barrier():
Acked-by: Will Deacon <will.deacon@xxxxxxx>
As an aside, ARMv7 (32-bit) with LPAE *can* emit single-copy atomic 64-bit
memory accesses and we rely on that for things like atomic64_read and
writing ptes. If we see WRITE_ONCE(pte), then we'll have genuine issues
with the way it's currently implemented.
Will
--
To unsubscribe from this list: send the line "unsubscribe linux-next" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
