lkml.org 
[lkml]   [2019]   [Oct]   [4]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: drm_sched with panfrost crash on T820
From
Date
On 04/10/2019 16:34, Koenig, Christian wrote:
> Am 04.10.19 um 17:27 schrieb Steven Price:
>> On 04/10/2019 16:03, Neil Armstrong wrote:
>>> On 04/10/2019 16:53, Grodzovsky, Andrey wrote:
>>>> On 10/3/19 4:34 AM, Neil Armstrong wrote:
>>>>> Hi Andrey,
>>>>>
>>>>> Le 02/10/2019 à 16:40, Grodzovsky, Andrey a écrit :
>>>>>> On 9/30/19 10:52 AM, Hillf Danton wrote:
>>>>>>> On Mon, 30 Sep 2019 11:17:45 +0200 Neil Armstrong wrote:
>>>>>>>> Did a new run from 5.3:
>>>>>>>>
>>>>>>>> [ 35.971972] Call trace:
>>>>>>>> [ 35.974391] drm_sched_increase_karma+0x5c/0xf0
>>>>>>>> ffff000010667f38 FFFF000010667F94
>>>>>>>> drivers/gpu/drm/scheduler/sched_main.c:335
>>>>>>>>
>>>>>>>> The crashing line is :
>>>>>>>> if (bad->s_fence->scheduled.context ==
>>>>>>>> entity->fence_context) {
>>>>>>>>
>>>>>>>> Doesn't seem related to guilty job.
>>>>>>> Bail out if s_fence is no longer fresh.
>>>>>>>
>>>>>>> --- a/drivers/gpu/drm/scheduler/sched_main.c
>>>>>>> +++ b/drivers/gpu/drm/scheduler/sched_main.c
>>>>>>> @@ -333,6 +333,10 @@ void drm_sched_increase_karma(struct drm
>>>>>>>
>>>>>>> spin_lock(&rq->lock);
>>>>>>> list_for_each_entry_safe(entity, tmp, &rq->entities, list) {
>>>>>>> + if (!smp_load_acquire(&bad->s_fence)) {
>>>>>>> + spin_unlock(&rq->lock);
>>>>>>> + return;
>>>>>>> + }
>>>>>>> if (bad->s_fence->scheduled.context ==
>>>>>>> entity->fence_context) {
>>>>>>> if (atomic_read(&bad->karma) >
>>>>>>> @@ -543,7 +547,7 @@ EXPORT_SYMBOL(drm_sched_job_init);
>>>>>>> void drm_sched_job_cleanup(struct drm_sched_job *job)
>>>>>>> {
>>>>>>> dma_fence_put(&job->s_fence->finished);
>>>>>>> - job->s_fence = NULL;
>>>>>>> + smp_store_release(&job->s_fence, 0);
>>>>>>> }
>>>>>>> EXPORT_SYMBOL(drm_sched_job_cleanup);
>>>>> This fixed the problem on the 10 CI runs.
>>>>>
>>>>> Neil
>>>>
>>>> These are good news but I still fail to see how this fixes the problem -
>>>> Hillf, do you mind explaining how you came up with this particular fix -
>>>> what was the bug you saw ?
>>> As Steven explained, seems the same job was submitted on both HW slots,
>>> and then when timeout occurs each thread calls panfrost_job_timedout
>>> which leads to drm_sched_stop() on the first call and on the
>>> second call the job was already freed.
>>>
>>> Steven proposed a working fix, and this one does the same but on
>>> the drm_sched side. This one looks cleaner, but panfrost should
>>> not call drm_sched_stop() twice for the same job.
>> I'm not sure that Hillf's fix is sufficient. In particular in
>> drm_sched_increase_karma() I don't understand how the smp_load_acquire()
>> call prevents bad->s_fence becoming NULL immediately afterwards (but
>> admittedly the window is much smaller). But really this is just a
>> Panfrost bug (calling drm_sched_stop() twice on the same job).
>>
>> The part of my change that I'd welcome feedback on is changing
>> cancel_delayed_work() to cancel_delayed_work_sync() in drm_sched_stop()
>> when called on different scheduler to the bad job. It's not clear to me
>> exactly what the semantics of the function should be, and I haven't
>> tested the effect of the change on drivers other than Panfrost.
>
> Yeah, at least of hand that change doesn't seem to make sense to me.

We need to ensure that any other timeouts that might have started
processing are complete before actually resetting the hardware.
Otherwise after the reset another thread could come along and attempt to
reset the hardware again (and cause a double free of a job). My change
to use the _sync() variant prevents this happening.

> Multiple timeout workers can perfectly run in parallel, Panfrost needs
> to make sure that they don't affect each other.
>
> The simplest way of doing this is to have a mutex you trylock when
> starting the reset.
>
> The first one grabbing it wins, all other just silently return.

Ah that would simplify my change removing the need for the new variable.
I hadn't thought to use trylock. I'll give that a spin and post a new
simpler version.

Thanks,

Steve

> Regards,
> Christian.
>
>>
>> Steve
>>
>>> Neil
>>>
>>>> Andrey
>>>>
>>>>>> Does this change help the problem ? Note that drm_sched_job_cleanup is
>>>>>> called from scheduler thread which is stopped at all times when work_tdr
>>>>>> thread is running and anyway the 'bad' job is still in the
>>>>>> ring_mirror_list while it's being accessed from
>>>>>> drm_sched_increase_karma so I don't think drm_sched_job_cleanup can be
>>>>>> called for it BEFORE or while drm_sched_increase_karma is executed.
>>>>>>
>>>>>> Andrey
>>>>>>
>>>>>>
>>>>>>>
>>>>>>> --
>>>>>>>
>>> _______________________________________________
>>> dri-devel mailing list
>>> dri-devel@lists.freedesktop.org
>>> https://lists.freedesktop.org/mailman/listinfo/dri-devel
>>>
>
> _______________________________________________
> dri-devel mailing list
> dri-devel@lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/dri-devel
>

\
 
 \ /
  Last update: 2019-10-04 18:03    [W:0.064 / U:0.876 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site