On Thu, 1 Oct 2020 15:49:39 +0100 Steven Price steven.price@arm.com wrote:

On 01/10/2020 15:01, Boris Brezillon wrote:

...

If more than two or more jobs end up timeout-ing concurrently, only one of them (the one attached to the scheduler acquiring the lock) is fully handled. The other one remains in a dangling state where it's no longer part of the scheduling queue, but still blocks something in scheduler thus leading to repetitive timeouts when new jobs are queued.

Let's make sure all bad jobs are properly handled by the thread acquiring the lock.

Signed-off-by: Boris Brezillon boris.brezillon@collabora.com Fixes: f3ba91228e8e ("drm/panfrost: Add initial panfrost driver") Cc: stable@vger.kernel.org

---

drivers/gpu/drm/panfrost/panfrost_job.c | 18 ++++++++++++++---- 1 file changed, 14 insertions(+), 4 deletions(-)

diff --git a/drivers/gpu/drm/panfrost/panfrost_job.c b/drivers/gpu/drm/panfrost/panfrost_job.c index 30e7b7196dab..e87edca51d84 100644 --- a/drivers/gpu/drm/panfrost/panfrost_job.c +++ b/drivers/gpu/drm/panfrost/panfrost_job.c @@ -25,7 +25,7 @@

struct panfrost_queue_state { struct drm_gpu_scheduler sched;

• struct drm_sched_job *bad; u64 fence_context; u64 emit_seqno; };

@@ -392,19 +392,29 @@ static void panfrost_job_timedout(struct drm_sched_job *sched_job) job_read(pfdev, JS_TAIL_LO(js)), sched_job);

• /*

• * Collect the bad job here so it can be processed by the thread
  

• * acquiring the reset lock.
  

• */
  

• pfdev->js->queue[js].bad = sched_job;

• if (!mutex_trylock(&pfdev->reset_lock)) return;

  for (i = 0; i < NUM_JOB_SLOTS; i++) { struct drm_gpu_scheduler *sched = &pfdev->js->queue[i].sched;

• drm_sched_stop(sched, sched_job);
  

  if (js != i) /* Ensure any timeouts on other slots have finished */ cancel_delayed_work_sync(&sched->work_tdr);

• }

• drm_sched_increase_karma(sched_job);

• drm_sched_stop(sched, pfdev->js->queue[i].bad);  
  

So I can see that the call to drm_sched_stop() needs to move below the cancel_delayed_work_sync() to ensure that the update to queue->bad is synchronised.

Exactly, and it probably deserves a comment :-).

What I'm not so sure about is whether it's possible for the scheduler to make progress between the 'cancel' and the 'stop' - there is a reason I wrote it the other way round...

Hm, we should probably stop the queue on which the timeout happened before acquiring the lock, this way we guarantee that nothing executes on queues that reported a timeout until the reset actually happens.

That leaves queues on which no timeout was reported, and I wonder why it'd be bad to let those continue executing jobs. Do you see any case where that'd be a problem?

The hole for things to go round is clearly much smaller with this change, but I'm not sure it's completely plugged. Am I missing something?

...

• if (pfdev->js->queue[i].bad)
  

• 	drm_sched_increase_karma(pfdev->js->queue[i].bad);
  

• pfdev->js->queue[i].bad = NULL;
  

• }

  spin_lock_irqsave(&pfdev->js->job_lock, flags); for (i = 0; i < NUM_JOB_SLOTS; i++) {

While we're on potential holes... some more context:

...

if (pfdev->jobs[i]) {
	pm_runtime_put_noidle(pfdev->dev);
	panfrost_devfreq_record_idle(pfdev);
	pfdev->jobs[i] = NULL;
}

} spin_unlock_irqrestore(&pfdev->js->job_lock, flags);

panfrost_device_reset(pfdev);

for (i = 0; i < NUM_JOB_SLOTS; i++) drm_sched_resubmit_jobs(&pfdev->js->queue[i].sched);

/* restart scheduler after GPU is usable again */ for (i = 0; i < NUM_JOB_SLOTS; i++) drm_sched_start(&pfdev->js->queue[i].sched, true);

mutex_unlock(&pfdev->reset_lock);

I'm wondering whether the mutex_unlock() should actually happen before the drm_sched_start() - in the (admittedly very unlikely) case where a timeout occurs before all the drm_sched_start() calls have completed it's possible for the timeout to be completely missed because the mutex is still held.

Oh, good catch.

On 01/10/2020 16:49, Boris Brezillon wrote:

On Thu, 1 Oct 2020 15:49:39 +0100 Steven Price steven.price@arm.com wrote:

...

On 01/10/2020 15:01, Boris Brezillon wrote:

...

If more than two or more jobs end up timeout-ing concurrently, only one of them (the one attached to the scheduler acquiring the lock) is fully handled. The other one remains in a dangling state where it's no longer part of the scheduling queue, but still blocks something in scheduler thus leading to repetitive timeouts when new jobs are queued.

Let's make sure all bad jobs are properly handled by the thread acquiring the lock.

Signed-off-by: Boris Brezillon boris.brezillon@collabora.com Fixes: f3ba91228e8e ("drm/panfrost: Add initial panfrost driver") Cc: stable@vger.kernel.org

---

drivers/gpu/drm/panfrost/panfrost_job.c | 18 ++++++++++++++---- 1 file changed, 14 insertions(+), 4 deletions(-)

diff --git a/drivers/gpu/drm/panfrost/panfrost_job.c b/drivers/gpu/drm/panfrost/panfrost_job.c index 30e7b7196dab..e87edca51d84 100644 --- a/drivers/gpu/drm/panfrost/panfrost_job.c +++ b/drivers/gpu/drm/panfrost/panfrost_job.c @@ -25,7 +25,7 @@

struct panfrost_queue_state { struct drm_gpu_scheduler sched;

• struct drm_sched_job *bad; u64 fence_context; u64 emit_seqno; };

@@ -392,19 +392,29 @@ static void panfrost_job_timedout(struct drm_sched_job *sched_job) job_read(pfdev, JS_TAIL_LO(js)), sched_job);

• /*

• * Collect the bad job here so it can be processed by the thread
  

• * acquiring the reset lock.
  

• */
  

• pfdev->js->queue[js].bad = sched_job;

• if (!mutex_trylock(&pfdev->reset_lock)) return;

  for (i = 0; i < NUM_JOB_SLOTS; i++) { struct drm_gpu_scheduler *sched = &pfdev->js->queue[i].sched;

• drm_sched_stop(sched, sched_job);
  if (js != i)
  	/* Ensure any timeouts on other slots have finished */
  	cancel_delayed_work_sync(&sched->work_tdr);
  

• }

• drm_sched_increase_karma(sched_job);

• drm_sched_stop(sched, pfdev->js->queue[i].bad);
  

So I can see that the call to drm_sched_stop() needs to move below the cancel_delayed_work_sync() to ensure that the update to queue->bad is synchronised. What I'm not so sure about is whether it's possible for the scheduler to make progress between the 'cancel' and the 'stop' - there is a reason I wrote it the other way round...

Ok, I think see what you mean now. So, there's indeed a race between the cancel_delayed_work_sync() and drm_sched_stop() calls, and a timeout might go undetected because of that. This being said, the only problem I see is that we would not increase karma on that job. In any case the job will be re-queued, and unless we keep having timeouts on the other queues it should be detected at some point. I can also try to retrieve the deadline before canceling the delayed work and check it before stopping the scheduler, but I'm not sure it's worth it.

Yes - well explained, I was trying to write an explanation but you beat me to it ;)

However I think it's worse than just the karma - the second timeout of a job could take a while to execute and actually execute *after* the first timeout has completed. I'm wondering whether we just need something like:

cancel_delayed_work_sync() drm_sched_stop() cancel_delayed_work_sync()

that way we ensure that any other timeouts will definitely have completed before performing the reset.

There's still a minor race regarding the karma - but that shouldn't matter as you mention.

Steve

...

The hole for things to go round is clearly much smaller with this change, but I'm not sure it's completely plugged. Am I missing something?

...

• if (pfdev->js->queue[i].bad)
  

• 	drm_sched_increase_karma(pfdev->js->queue[i].bad);
  

• pfdev->js->queue[i].bad = NULL;
  

• }

  spin_lock_irqsave(&pfdev->js->job_lock, flags); for (i = 0; i < NUM_JOB_SLOTS; i++) {

While we're on potential holes... some more context:

...

if (pfdev->jobs[i]) {
	pm_runtime_put_noidle(pfdev->dev);
	panfrost_devfreq_record_idle(pfdev);
	pfdev->jobs[i] = NULL;
}

} spin_unlock_irqrestore(&pfdev->js->job_lock, flags);

panfrost_device_reset(pfdev);

for (i = 0; i < NUM_JOB_SLOTS; i++) drm_sched_resubmit_jobs(&pfdev->js->queue[i].sched);

/* restart scheduler after GPU is usable again */ for (i = 0; i < NUM_JOB_SLOTS; i++) drm_sched_start(&pfdev->js->queue[i].sched, true);

mutex_unlock(&pfdev->reset_lock);

I'm wondering whether the mutex_unlock() should actually happen before the drm_sched_start() - in the (admittedly very unlikely) case where a timeout occurs before all the drm_sched_start() calls have completed it's possible for the timeout to be completely missed because the mutex is still held.

Thanks,

Steve