On Fri, Feb 14, 2020 at 09:12:13AM +0100, Andrea Parri wrote:
> > @@ -0,0 +1,24 @@
> > +C Atomic-set-observable-to-RMW
> > +
> > +(*
> > + * Result: Never
> > + *
> > + * Test of the result of atomic_set() must be observable to atomic RMWs.
> > + *)
> > +
> > +{
> > + atomic_t v = ATOMIC_INIT(1);
> > +}
> > +
> > +P0(atomic_t *v)
> > +{
> > + (void)atomic_add_unless(v,1,0);
>
> We blacklisted this primitive some time ago, cf. section "LIMITATIONS",
> entry (6b) in tools/memory-model/README; the discussion was here:
>
> https://lkml.kernel.org/r/20180829211053.20531-3-paulmck@xxxxxxxxxxxxxxxxxx
>
And in an email replying to that email, you just tried and seemed
atomic_add_unless() works ;-)
> but unfortunately I can't remember other details at the moment: maybe
> it is just a matter of or the proper time to update that section.
>
I spend a few time looking into the changes in herd, the dependency
problem seems to be as follow:
For atomic_add_unless(ptr, a, u), the return value (true or false)
depends on both *ptr and u, this is different than other atomic RMW,
whose return value only depends on *ptr. Considering the following
litmus test:
C atomic_add_unless-dependency
{
int y = 1;
}
P0(int *x, int *y, int *z)
{
int r0;
int r1;
int r2;
r0 = READ_ONCE(*x);
if (atomic_add_unless(y, 2, r0))
WRITE_ONCE(*z, 42);
else
WRITE_ONCE(*z, 1);
}
P1(int *x, int *y, int *z)
{
int r0;
r0 = smp_load_acquire(z);
WRITE_ONCE(*x, 1);
}
exists
(1:r0 = 1 /\ 0:r0 = 1)
, the exist-clause will never trigger, however if we replace
"atomic_add_unless(y, 2, r0)" with "atomic_add_unless(y, 2, 1)", the
write on *z and the read from *x on CPU 0 are not ordered, so we could
observe the exist-clause triggered.
I just tried with the latest herd, and herd can work out this
dependency. So I think we are good now and can change the limitation
section in the document. But I will wait for Luc's input for this. Luc,
did I get this correct? Is there any other limitation on
atomic_add_unless() now?
Regards,
Boqun
> Thanks,
> Andrea
