Re: WARN ON at kernel/sched/deadline.c task_non_contending

[Juri Lelli <juri.lelli@redhat.com>]

Hi,

On 13/03/19 15:49, luca abeni wrote:
> Hi,
> 
> (I added Juri in cc)
> 
> On Tue, 12 Mar 2019 10:03:12 +0800
> "chengjian (D)" <cj.chengjian@huawei.com> wrote:
> [...]
> > diff --git a/kernel/sched/deadline.c b/kernel/sched/deadline.c
> > index 31c050a0d0ce..d73cb033a06d 100644
> > --- a/kernel/sched/deadline.c
> > +++ b/kernel/sched/deadline.c
> > @@ -252,7 +252,6 @@ static void task_non_contending(struct
> > task_struct *p) if (dl_entity_is_special(dl_se))
> >                  return;
> > 
> > -       WARN_ON(hrtimer_active(&dl_se->inactive_timer));
> >          WARN_ON(dl_se->dl_non_contending);
> > 
> >          zerolag_time = dl_se->deadline -
> > @@ -287,7 +286,9 @@ static void task_non_contending(struct
> > task_struct *p) }
> > 
> >          dl_se->dl_non_contending = 1;
> > -       get_task_struct(p);
> > +
> > +       if (!hrtimer_active(&dl_se->inactive_timer));
> > +               get_task_struct(p);
> >          hrtimer_start(timer, ns_to_ktime(zerolag_time),
> > HRTIMER_MODE_REL); }
> 
> After looking at the patch a little bit more and running some tests,
> I suspect this solution might be racy:
> when the timer is already active, (and hrtimer_start() fails), it
> relies on its handler to decrease the running bw (by setting
> dl_non_contending to 1)... But inactive_task_timer() might have
> already checked dl_non_contending, finding it equal to 0 (so, it
> ends up doing nothing and the running bw is not decreased).
> 
> 
> So, I would prefer a different solution. I think this patch should work:
> 
> diff --git a/kernel/sched/deadline.c b/kernel/sched/deadline.c
> index 6a73e41a2016..43901fa3f269 100644
> --- a/kernel/sched/deadline.c
> +++ b/kernel/sched/deadline.c
> @@ -252,7 +252,6 @@ static void task_non_contending(struct task_struct *p)
>  	if (dl_entity_is_special(dl_se))
>  		return;
>  
> -	WARN_ON(hrtimer_active(&dl_se->inactive_timer));
>  	WARN_ON(dl_se->dl_non_contending);
>  
>  	zerolag_time = dl_se->deadline -
> @@ -269,7 +268,7 @@ static void task_non_contending(struct task_struct *p)
>  	 * If the "0-lag time" already passed, decrease the active
>  	 * utilization now, instead of starting a timer
>  	 */
> -	if (zerolag_time < 0) {
> +	if ((zerolag_time < 0) || hrtimer_active(&dl_se->inactive_timer)) {
>  		if (dl_task(p))
>  			sub_running_bw(dl_se, dl_rq);
>  		if (!dl_task(p) || p->state == TASK_DEAD) {
> 
> 
> The idea is that if the timer is active, we leave dl_non_contending set to
> 0 (so that the timer handler does nothing), and we immediately decrease the
> running bw.
> I think this is OK, because this situation can happen only if the task
> blocks, wakes up while the timer handler is running, and then immediately
> blocks again - while the timer handler is still running. So, the "zero lag
> time" cannot be too much in the future.

And if we get here and the handler is running it means that the handler
is spinning on rq->lock waiting the dequeue to release it. So, this
looks safe to me as well.

BTW, I could reproduce with Steve's deadline_test [1], and this seems to
fix it.

Would you mind sending out a proper patch Luca?

Thanks!

- Juri

1 - https://goo.gl/fVbRSu