[PATCH 3/3] drm/scheduler: Clean up jobs when the scheduler is torn down.

[Asahi Lina]

On 15/07/2023 16.14, Luben Tuikov wrote:
> On 2023-07-14 04:21, Asahi Lina wrote:
>> drm_sched_fini() currently leaves any pending jobs dangling, which
>> causes segfaults and other badness when job completion fences are
>> signaled after the scheduler is torn down.
> 
> If there are pending jobs, ideally we want to call into the driver,
> so that it can release resources it may be holding for those.
> The idea behind "pending" is that they are pending in the hardware
> and we don't know their state until signalled/the callback called.
> (Or unless the device is reset and we get a notification of that fact.)

That's what the job->free_job() callback does, then the driver is free 
to do whatever it wants with those jobs. A driver could opt to 
synchronously kill those jobs (if it can) or account for them 
separately/asynchronously.

What this patch basically says is that if you destroy a scheduler with 
pending jobs, it immediately considers them terminated with an error, 
and returns ownership back to the driver for freeing. Then the driver 
can decide how to handle the rest and whatever the underlying hardware 
state is.

>> Explicitly detach all jobs from their completion callbacks and free
>> them. This makes it possible to write a sensible safe abstraction for
>> drm_sched, without having to externally duplicate the tracking of
>> in-flight jobs.
>>
>> This shouldn't regress any existing drivers, since calling
>> drm_sched_fini() with any pending jobs is broken and this change should
>> be a no-op if there are no pending jobs.
> 
> While this statement is true on its own, it kind of contradicts
> the premise of the first paragraph.

I mean right *now* it's broken, before this patch. I'm trying to make it 
safe, but it shouldn't regress any exiting drivers since if they trigger 
this code path they are broken today.

> 
>> Signed-off-by: Asahi Lina <lina at asahilina.net>
>> ---
>>   drivers/gpu/drm/scheduler/sched_main.c | 32 ++++++++++++++++++++++++++++++--
>>   1 file changed, 30 insertions(+), 2 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c
>> index 1f3bc3606239..a4da4aac0efd 100644
>> --- a/drivers/gpu/drm/scheduler/sched_main.c
>> +++ b/drivers/gpu/drm/scheduler/sched_main.c
>> @@ -1186,10 +1186,38 @@ EXPORT_SYMBOL(drm_sched_init);
>>   void drm_sched_fini(struct drm_gpu_scheduler *sched)
>>   {
>>   	struct drm_sched_entity *s_entity;
>> +	struct drm_sched_job *s_job, *tmp;
>>   	int i;
>>   
>> -	if (sched->thread)
>> -		kthread_stop(sched->thread);
>> +	if (!sched->thread)
>> +		return;
>> +
>> +	/*
>> +	 * Stop the scheduler, detaching all jobs from their hardware callbacks
>> +	 * and cleaning up complete jobs.
>> +	 */
>> +	drm_sched_stop(sched, NULL);
>> +
>> +	/*
>> +	 * Iterate through the pending job list and free all jobs.
>> +	 * This assumes the driver has either guaranteed jobs are already stopped, or that
>> +	 * otherwise it is responsible for keeping any necessary data structures for
>> +	 * in-progress jobs alive even when the free_job() callback is called early (e.g. by
>> +	 * putting them in its own queue or doing its own refcounting).
>> +	 */
>> +	list_for_each_entry_safe(s_job, tmp, &sched->pending_list, list) {
>> +		spin_lock(&sched->job_list_lock);
>> +		list_del_init(&s_job->list);
>> +		spin_unlock(&sched->job_list_lock);
>> +
>> +		dma_fence_set_error(&s_job->s_fence->finished, -ESRCH);
>> +		drm_sched_fence_finished(s_job->s_fence);
> 
> I'd imagine it's better to rebase this on top of drm-misc-next where
> drm_sched_fence_finished() takes one more parameter--the error.

Ah, sure! I can do that.

> 
>> +
>> +		WARN_ON(s_job->s_fence->parent);
>> +		sched->ops->free_job(s_job);
>> +	}
>> +
>> +	kthread_stop(sched->thread);
>>   
>>   	for (i = DRM_SCHED_PRIORITY_COUNT - 1; i >= DRM_SCHED_PRIORITY_MIN; i--) {
>>   		struct drm_sched_rq *rq = &sched->sched_rq[i];
>>
> 
> Conceptually I don't mind this patch--I see what it is trying to achieve,
> but technically, we want the driver to detect GPU removal and return shared
> resources back, such as "jobs", which DRM is also aware of.

I think you missed the context of why I'm doing this, so in short: my 
use case (like Xe's) involves using a separate drm_sched instance *per 
file* since these queues are scheduled directly by the firmware. So this 
isn't about GPU removal, but rather about a GPU context going away while 
jobs are in flight (e.g. the process got killed). We want that to 
quickly kill the "DRM view" of the world, including signaling all the 
fences with an error and freeing resources like the scheduler itself.

In the case of this particular GPU, there is no known way to actively 
and synchronously abort GPU jobs, so we need to let them run to 
completion (or failure), but we don't want that to block process cleanup 
and freeing a bunch of high-level resources. The driver is architected 
roughly along the lines of a firmware abstraction layer that maps to the 
firmware shared memory structures, and then a layer on top that 
implements the DRM view. When a process gets killed, the DRM side (which 
includes the scheduler, etc.) gets torn down immediately, and it makes 
sense to handle this cleanup inside drm_sched since it already has a 
view into what jobs are in flight. Otherwise, I would have to duplicate 
job tracking in the driver (actually worse: in the Rust abstraction for 
safety), which doesn't make much sense.

But what I *do* have in the driver is tracking of the firmware 
structures. So when the drm_sched gets torn down and all the jobs 
killed, the underlying firmware jobs do run to completion, and the 
resources they use are all cleaned up after that (it's all reference 
counted). The primitive involved here is that in-flight firmware jobs 
are assigned an event completion slot, and that keeps a reference to 
them from a global array until the events fire and all the jobs are 
known to have completed. This keeps things memory-safe, since we 
absolutely cannot free/destroy firmware structures while they are in use 
(otherwise the firmware crashes, which is fatal on these GPUs - requires 
a full system reboot to recover).

In practice, with the VM map model that we use, what ends up happening 
when a process gets killed is that all the user objects for in-flight 
jobs get unmapped, which usually causes the GPU hardware (not firmware) 
to fault. This then triggers early termination of jobs anyway via the 
firmware fault recovery flow. But even that takes some short amount of 
time, and by then all the drm_sched stuff is long gone and we're just 
dealing with the in-flight firmware stuff.

> In the case where we're initiating the tear, we should notify the driver that
> we're about to forget jobs (resources), so that it knows to return them back
> or that it shouldn't notify us for them (since we've notified we're forgetting them.)

That contradicts Christian's comment. I tried to document that (after 
this patch) the scheduler no longer cares about hw fences and whether 
they are signaled or not after it's destroyed, and I got a strongly 
worded NAK for it. Sooo... which is it? Is it okay for drivers not to 
signal the hw fence after a scheduler teardown, or not?

But really, I don't see a use case for an explicit "about to forget job" 
callback. The job free callback already serves the purpose of telling 
the driver to clean up resources associated with a job. If it wants to 
synchronously abort things there, it could easily take over its own 
fence signaling and do something with the underlying stuff if the fence 
is not signaled yet.

In my case, since the driver is written in Rust and free_job() just maps 
to the destructor (Drop impl), that just ends up freeing a bunch of 
memory and other objects, and I don't particularly care about the state 
of the firmware side any more after that. The flow is the same whether 
it was a successful job completion, a failure, or an early destruction 
due to the drm_sched getting torn down.

> (Note also that in this latter case, traditionally, the device would be reset,
> so that we can guarantee that it has forgotten all shared resources which
> we are to tear up. This is somewhat more complicated with GPUs, thus the method
> pointed out above.)

Yeah, in the firmware scheduling case we can't do this at all unless the 
firmware has an explicit teardown/forget op (which I'm not aware of) and 
a full GPU reset isn't something we can do either. Hence we just let the 
underlying jobs complete. In practice they tend to die pretty quickly 
anyway once all the buffers are unmapped.

~~ Lina