Re: XDP socket rings, and LKMM litmus tests

On Tue, Mar 02, 2021 at 09:24:04PM +0100, Björn Töpel wrote:

On Tue, 2 Mar 2021 at 20:57, Paul E. McKenney [off-list ref] wrote:

quoted

On Tue, Mar 02, 2021 at 07:46:27PM +0100, Björn Töpel wrote:

[...]

quoted

Before digging in too deeply, does the following simplification
still capture your intent?

Thanks for having a look, Paul!

quoted

P0(int *prod, int *cons, int *data)
{
    int p;
    int cond = 0;

    p = READ_ONCE(*prod);
    if (p == READ_ONCE(*cons))
            cond = 1;

With this, yes!

quoted

    if (cond) {
        smp_mb();
        WRITE_ONCE(*data, 1);
        smp_wmb();
        WRITE_ONCE(*prod, p ^ 1);
    }
}

P1(int *prod, int *cons, int *data)
{
    int c;
    int d = -1;
    int cond = 0;

    c = READ_ONCE(*cons);
    if (READ_ONCE(*prod) == c)
            cond = 1;

Hmm, this would not be the correct state transition.

c==1 && p==1 would set cond to 1, right?

I would agree with:
  c = READ_ONCE(*cons);
  if (READ_ONCE(*prod) != c)

Right you are!

With that, it looks to me like LKMM is OK with removing the smp_mb().
My guess is that the issue is that LKMM confines the effect of control
dependencies to a single "if" statement, hence my reworking of your
original.

							Thanx, Paul

quoted

    if (cond == 1) {
        smp_rmb();
        d = READ_ONCE(*data);
        smp_mb();
        WRITE_ONCE(*cons, c ^ 1);
    }
}

                                                        Thanx, Paul

[...]

Björn