Re: [PATCH v1 1/6] drm/lima: fix devfreq refcount imbalance for job timeouts

From: Qiang Yu
Date: Tue Jan 23 2024 - 20:04:25 EST


On Wed, Jan 24, 2024 at 7:19 AM Erico Nunes <nunes.erico@xxxxxxxxx> wrote:
>
> On Fri, Jan 19, 2024 at 2:50 AM Qiang Yu <yuq825@xxxxxxxxx> wrote:
> >
> > On Thu, Jan 18, 2024 at 7:14 PM Erico Nunes <nunes.erico@xxxxxxxxx> wrote:
> > >
> > > On Thu, Jan 18, 2024 at 2:36 AM Qiang Yu <yuq825@xxxxxxxxx> wrote:
> > > >
> > > > So this is caused by same job trigger both done and timeout handling?
> > > > I think a better way to solve this is to make sure only one handler
> > > > (done or timeout) process the job instead of just making lima_pm_idle()
> > > > unique.
> > >
> > > It's not very clear to me how to best ensure that, with the drm_sched
> > > software timeout and the irq happening potentially at the same time.
> > This could be done by stopping scheduler run more job and disable
> > GP/PP interrupt. Then after sync irq, there should be no more new
> > irq gets in when we handling timeout.
> >
> > > I think patch 4 in this series describes and covers the most common
> > > case that this would be hit. So maybe now this patch could be dropped
> > > in favour of just that one.
> > Yes.
>
> After dropping the patch while testing to send v2, I managed to
> reproduce this issue again.
> Looking at a trace it seems that this actually happened with the test workload:
> lima_sched_timedout_job -> (fence by is not signaled and new fence
> check is passed) -> irq happens preempting lima_sched_timedout_job,
> fence is signaled and lima_pm_idle called once at
> lima_sched_pipe_task_done -> lima_sched_timedout_job continues and
> calls lima_pm_idle again
>
> So I think we still need this patch to at least prevent the bug with
> the current software timeout. If we move to the hardware watchdog
> timeout later we might be able to remove it anyway, but that will
> still require separate work and testing.
>
Then you do need to solve the IRQ and reset race by disabling GP/PP
IRQ and sync irq after drm_sched_stop(). I mean you may keep the
patch 4 for spurious timeout, and add a new patch for solving the IRQ
and reset race.

Disable pm idle does not solve the race completely, other reset logic
may also conflict if not sequence the IRQ and reset .