Hi,

As far as I can remember I have implemented atomic_add_unless in herd7.

As to your test, I have first run a slightly modified version of your test
as a kernel module (using klitmus7).

C atomic_add_unless-dependency
{
        atomic_t y = ATOMIC_INIT(1);
}
  P0(int *x, atomic_t *y, int *z)
{
        int r0;
        r0 = READ_ONCE(*x);
        if (atomic_add_unless((atomic_t *)y, 2, r0))
                WRITE_ONCE(*z, 42);
        else
                WRITE_ONCE(*z, 1);
}
  P1(int *x, int *z)
{
        int r0;
        r0 = smp_load_acquire(z);
        WRITE_ONCE(*x, 1);
}
locations [y]
exists
(1:r0 = 1 /\ 0:r0 = 1)


The test is also accepted by herd7, here producing teh same final values
as actual run on a raspberry PI4B.

--Luc

Luc,

Could you have a look at the problem Andrea and I discuss here? It seems
that you have done a few things in herd for atomic_add_unless() in
particular, and based on the experiments of Andrea and me, seems
atomic_add_unless() works correctly. So can you confirm that herd now
can handle atomic_add_unless() or there is still something missing?

Thanks!

Regards,
Boqun

On Fri, Feb 14, 2020 at 06:40:03PM +0800, Boqun Feng wrote:

quoted

On Fri, Feb 14, 2020 at 09:12:13AM +0100, Andrea Parri wrote:

quoted

quoted

@@ -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@linux.vnet.ibm.com

And in an email replying to that email, you just tried and seemed
atomic_add_unless() works ;-)

quoted

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

quoted

Thanks,
  Andrea