Hi Hillf,
Unfortunately, above mem barriers don't help. The issue shows up
within 1 minute ...
Hillf Danton <hdan...@sina.com> 于2020年8月27日周四 下午8:58写道：
>
>
> On Thu, 27 Aug 2020 14:56:31 +0800 Kehuan Feng wrote:
> >
> > > Lets see if TCQ_F_NOLOC is making fq_codel different in your testing.
> >
> > I assume you meant disabling NOLOCK for pfifo_fast.
> >
> > Here is the modification,
> >
> > --- ./net/sched/sch_generic.c.orig 2020-08-24 22:02:04.589830751 +0800
> > +++ ./net/sched/sch_generic.c 2020-08-27 10:17:10.148977195 +0800
> > @@ -792,7 +792,7 @@
> > .dump =3D pfifo_fast_dump,
> > .change_tx_queue_len =3D pfifo_fast_change_tx_queue_len,
> > .owner =3D THIS_MODULE,
> > - .static_flags =3D TCQ_F_NOLOCK | TCQ_F_CPUSTATS,
> > + .static_flags =3D TCQ_F_CPUSTATS,
> >
> > The issue never happen again with it for over 3 hours stressing. And I
> > restarted the test for two times. No any surprising. Quite stable...
>
> Jaw off. That is great news and I'm failing again to explain the test
> result wrt the difference TCQ_F_NOLOCK can make in running qdisc.
>
> Nothing comes into mind other than two mem barriers though only one is
> needed...
>
> --- a/net/core/dev.c
> +++ b/net/core/dev.c
> @@ -3040,6 +3040,7 @@ static void __netif_reschedule(struct Qd
>
> void __netif_schedule(struct Qdisc *q)
> {
> + smp_mb__before_atomic();
> if (!test_and_set_bit(__QDISC_STATE_SCHED, &q->state))
> __netif_reschedule(q);
> }
> @@ -4899,6 +4900,7 @@ static __latent_entropy void net_tx_acti
> */
> smp_mb__before_atomic();
> clear_bit(__QDISC_STATE_SCHED, &q->state);
> + smp_mb__after_atomic();
> qdisc_run(q);
> if (root_lock)
> spin_unlock(root_lock);
>
