Re: [PATCH] ptr_ring: Add barriers to fix NULL-pointer exception

["Michael S. Tsirkin" <mst@redhat.com>]

On Wed, Dec 06, 2017 at 02:08:54PM +0000, Cherian, George wrote:
> > @@ -275,6 +281,13 @@ static inline void *__ptr_ring_consume(struct ptr_ring
> *r)
> > if (ptr)
> > __ptr_ring_discard_one(r);
> >
> > + /*
> > + * This barrier is necessary in order to prevent race condition with
> > + * with __ptr_ring_produce(). Make sure all the elements of ptr is
> > + * in sync with the earlier writes which was done prior to pushing
> > + * it to ring
> > + */
> > + rmb();
> > return ptr;
> > }
> 
> You are trying to synchronise two CPUs so non-smp barriers make no
> sense. wmb/rmb are for synchronising with MMIO.
> 
> What happens when CONFIG_SMP is not set. smp_wmb/rmb becomes compiler barriers
> (atleast for arm64).

And that is because all read and writes always appear in order when done
from the same CPU.

In case of reads, we do not need a barrier at all (except on dec alpha)
because a read through a pointer can't bypass a read of a pointer.

> I guess that is not what we need.

Maybe, but I don't yet see why not.

> An SMP barrier cannot
> replace a mandatory barrier, but a mandatory barrier can replace an SMP
> barrier.

This does imply that you can always replace a weak barrier with a strong
one, but does not mean you should.

> I will try out your patch too and update the results.
> But I would need couple of days time. Sorry for the delay.

Thanks for the testing.

-- 
MST