On 06.09.2023 09:57, Boris Brezillon wrote:
>On Tue, 5 Sep 2023 19:45:18 +0100
>Adrián Larumbe <adrian.larumbe@xxxxxxxxxxxxx> wrote:
>
>> diff --git a/drivers/gpu/drm/panfrost/panfrost_job.c b/drivers/gpu/drm/panfrost/panfrost_job.c
>> index 033f5e684707..8b1bf6ac48f8 100644
>> --- a/drivers/gpu/drm/panfrost/panfrost_job.c
>> +++ b/drivers/gpu/drm/panfrost/panfrost_job.c
>> @@ -297,6 +297,11 @@ int panfrost_job_push(struct panfrost_job *job)
>>
>> kref_get(&job->refcount); /* put by scheduler job completion */
>>
>> + if (atomic_read(&pfdev->profile_mode)) {
>> + panfrost_cycle_counter_get(pfdev);
>
>This one should go in panfrost_job_hw_submit() IMO, otherwise you're
>enabling the cycle-counter before the job has its dependencies met, and
>depending on what the job depends on, it might take some time.
I think at first I decided against this because of a previous thread about
enabling the cycle counters:
https://www.mail-archive.com/dri-devel@xxxxxxxxxxxxxxxxxxxxx/msg352508.html
But it turns out the fix suggested by Steven Price was what I was doing already
in the previous patch revision, namely tagging the job with whether it
had taken the cycle counter refcnt, so I guess it should be fine.
>> + job->is_profiled = true;
>> + }
>> +
>> drm_sched_entity_push_job(&job->base);
>>
>> mutex_unlock(&pfdev->sched_lock);
>> @@ -351,6 +356,9 @@ static void panfrost_job_free(struct drm_sched_job *sched_job)
>>
>> drm_sched_job_cleanup(sched_job);
>>
>> + if (job->is_profiled)
>> + panfrost_cycle_counter_put(job->pfdev);
>
>I think I'd move this panfrost_cycle_counter_put() to
>panfrost_job_handle_{err,done}(), to release the counter as soon as
>we're done executing the job. We also need to make sure we release
>cycle counter refs in the reset path (here [1]), to keep get/put
>balanced when jobs are resubmitted.
>
>> +
>> panfrost_job_put(job);
>> }
>
>[1]https://elixir.bootlin.com/linux/latest/source/drivers/gpu/drm/panfrost/panfrost_job.c#L666
