Re: [PATCH v5] drm/sched: Make sure we wait for all dependencies in kill_jobs_cb()

[Boris Brezillon <boris.brezillon@collabora.com>]

On Fri, 16 Jun 2023 19:12:48 +0200
Boris Brezillon <boris.brezillon@collabora.com> wrote:

> drm_sched_entity_kill_jobs_cb() logic is omitting the last fence popped
> from the dependency array that was waited upon before
> drm_sched_entity_kill() was called (drm_sched_entity::dependency field),
> so we're basically waiting for all dependencies except one.
> 
> In theory, this wait shouldn't be needed because resources should have
> their users registered to the dma_resv object, thus guaranteeing that
> future jobs wanting to access these resources wait on all the previous
> users (depending on the access type, of course). But we want to keep
> these explicit waits in the kill entity path just in case.
> 
> Let's make sure we keep all dependencies in the array in
> drm_sched_job_dependency(), so we can iterate over the array and wait
> in drm_sched_entity_kill_jobs_cb().
> 
> We make sure we wait on drm_sched_fence::finished if we were
> originally asked to wait on drm_sched_fence::scheduled. In that case,
> we assume the intent was to delegate the wait to the firmware/GPU or
> rely on the pipelining done at the entity/scheduler level, but when
> killing jobs, we really want to wait for completion not just scheduling.
> 
> v5:
> - Flag deps on which we should only wait for the scheduled event
>   at insertion time.
> 
> v4:
> - Fix commit message
> - Fix a use-after-free bug
> 
> v3:
> - Always wait for drm_sched_fence::finished fences in
>   drm_sched_entity_kill_jobs_cb() when we see a sched_fence
> 
> v2:
> - Don't evict deps in drm_sched_job_dependency()
> 
> Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
> Suggested-by: "Christian König" <christian.koenig@amd.com>
> Cc: Frank Binns <frank.binns@imgtec.com>
> Cc: Sarah Walker <sarah.walker@imgtec.com>
> Cc: Donald Robson <donald.robson@imgtec.com>
> Cc: Luben Tuikov <luben.tuikov@amd.com>
> Cc: David Airlie <airlied@gmail.com>
> Cc: Daniel Vetter <daniel@ffwll.ch>
> Cc: Sumit Semwal <sumit.semwal@linaro.org>
> Cc: "Christian König" <christian.koenig@amd.com>
> ---
> Hello Christian,
> 
> It turns out the approach you suggested just moves the complexity to
> the insertion path, and it actually makes the 'should we replace or
> drop fence' test a bit more tedious. We might end up with less
> duplicates if some drivers start mixing scheduled/finished fences, but
> I doubt this will be the case in practice, and I'm not sure it's worth
> the extra complexity.
> 
> Let me know what you think.
> 
> Regards,
> 
> Boris
> ---
>  drivers/gpu/drm/scheduler/sched_entity.c | 29 ++++++++++++--
>  drivers/gpu/drm/scheduler/sched_main.c   | 50 +++++++++++++++++++++++-
>  include/drm/gpu_scheduler.h              |  7 ++++
>  3 files changed, 80 insertions(+), 6 deletions(-)
> 
> diff --git a/drivers/gpu/drm/scheduler/sched_entity.c b/drivers/gpu/drm/scheduler/sched_entity.c
> index 68e807ae136a..ffdee2544f42 100644
> --- a/drivers/gpu/drm/scheduler/sched_entity.c
> +++ b/drivers/gpu/drm/scheduler/sched_entity.c
> @@ -176,13 +176,14 @@ static void drm_sched_entity_kill_jobs_cb(struct dma_fence *f,
>  {
>  	struct drm_sched_job *job = container_of(cb, struct drm_sched_job,
>  						 finish_cb);
> +	unsigned long index;
>  	int r;
>  
>  	dma_fence_put(f);
>  
>  	/* Wait for all dependencies to avoid data corruptions */
> -	while (!xa_empty(&job->dependencies)) {
> -		f = xa_erase(&job->dependencies, job->last_dependency++);
> +	xa_for_each(&job->dependencies, index, f) {
> +		xa_erase(&job->dependencies, index);
>  		r = dma_fence_add_callback(f, &job->finish_cb,
>  					   drm_sched_entity_kill_jobs_cb);
>  		if (!r)
> @@ -415,8 +416,28 @@ static struct dma_fence *
>  drm_sched_job_dependency(struct drm_sched_job *job,
>  			 struct drm_sched_entity *entity)
>  {
> -	if (!xa_empty(&job->dependencies))
> -		return xa_erase(&job->dependencies, job->last_dependency++);
> +	struct dma_fence *f;
> +
> +	/* We keep the fence around, so we can iterate over all dependencies
> +	 * in drm_sched_entity_kill_jobs_cb() to ensure all deps are signaled
> +	 * before killing the job.
> +	 */
> +	f = xa_load(&job->dependencies, job->last_dependency);
> +	if (f) {
> +		/* If the entry is flagged as 'wait-for-scheduled', we are
> +		 * dealing with a drm_sched_fence and we want to wait for
> +		 * the scheduled event only.
> +		 */
> +		if (xa_get_mark(&job->dependencies, job->last_dependency,
> +				DRM_SCHED_DEP_WAIT_FOR_SCHEDULED)) {
> +			struct drm_sched_fence *s_fence = to_drm_sched_fence(f);
> +
> +			f = &s_fence->scheduled;
> +		}
> +
> +		job->last_dependency++;
> +		return dma_fence_get(f);
> +	}
>  
>  	if (job->sched->ops->prepare_job)
>  		return job->sched->ops->prepare_job(job, entity);
> diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c
> index 394010a60821..8ac207cbd713 100644
> --- a/drivers/gpu/drm/scheduler/sched_main.c
> +++ b/drivers/gpu/drm/scheduler/sched_main.c
> @@ -691,6 +691,8 @@ EXPORT_SYMBOL(drm_sched_job_arm);
>  int drm_sched_job_add_dependency(struct drm_sched_job *job,
>  				 struct dma_fence *fence)
>  {
> +	bool fence_wait_for_scheduled = false;
> +	struct drm_sched_fence *s_fence;
>  	struct dma_fence *entry;
>  	unsigned long index;
>  	u32 id = 0;
> @@ -699,20 +701,64 @@ int drm_sched_job_add_dependency(struct drm_sched_job *job,
>  	if (!fence)
>  		return 0;
>  
> +	s_fence = to_drm_sched_fence(fence);
> +	if (s_fence && fence == &s_fence->scheduled) {
> +		/* Make sure the finished fence hasn't been destroyed. */
> +		fence = dma_fence_get_rcu(&s_fence->finished);
> +		if (WARN_ON(!fence))
> +			return -EINVAL;
> +
> +		/* Release the scheduled fence ref, now that we have a
> +		 * ref on the finished fence.
> +		 */
> +		dma_fence_put(&s_fence->scheduled);
> +
> +		/* Waiting for scheduled event only. */
> +		fence_wait_for_scheduled = true;
> +	}
> +
>  	/* Deduplicate if we already depend on a fence from the same context.
>  	 * This lets the size of the array of deps scale with the number of
>  	 * engines involved, rather than the number of BOs.
>  	 */
>  	xa_for_each(&job->dependencies, index, entry) {
> +		bool entry_wait_for_scheduled, fence_is_later;
> +
>  		if (entry->context != fence->context)
>  			continue;
>  
> -		if (dma_fence_is_later(fence, entry)) {
> +		entry_wait_for_scheduled = xa_get_mark(&job->dependencies,
> +						       index,
> +						       DRM_SCHED_DEP_WAIT_FOR_SCHEDULED);
> +		fence_is_later = dma_fence_is_later(fence, entry);
> +
> +		/*
> +		 * We can replace entry by fence when:
> +		 * - fence is after entry and both are waiting on scheduled
> +		 * - fence is after entry and both are waiting on finished
> +		 * - fence is after entry, entry is waiting on scheduled and fence on finished
> +		 *
> +		 * We can keep entry and drop fence when:
> +		 * - fence is before entry and both are waiting on scheduled
> +		 * - fence is before entry and both are waiting on finished
> +		 * - fence is before entry, fence is waiting on scheduled and entry on finished
> +		 *
> +		 * In all other situations, we can't guarantee the order and have to duplicate.
> +		 */
> +		if (fence_is_later && entry_wait_for_scheduled >= fence_wait_for_scheduled) {
>  			dma_fence_put(entry);
>  			xa_store(&job->dependencies, index, fence, GFP_KERNEL);
> -		} else {
> +			if (fence_wait_for_scheduled) {
> +				xa_set_mark(&job->dependencies, index,
> +					    DRM_SCHED_DEP_WAIT_FOR_SCHEDULED);
> +			}
> +		} else if (!fence_is_later &&
> +			   entry_wait_for_scheduled <= fence_wait_for_scheduled) {
>  			dma_fence_put(fence);
> +		} else {
> +			continue;
>  		}
> +
>  		return 0;
>  	}

Forgot to flag wait-for-scheduled in the insertion path:

        ret = xa_alloc(&job->dependencies, &id, fence, xa_limit_32b, GFP_KERNEL);
	if (ret != 0)
		dma_fence_put(fence);
	else if (fence_wait_for_scheduled)
		xa_set_mark(&job->dependencies, index, DRM_SCHED_DEP_WAIT_FOR_SCHEDULED);

	return ret;

>  
> diff --git a/include/drm/gpu_scheduler.h b/include/drm/gpu_scheduler.h
> index e95b4837e5a3..fbdb01242997 100644
> --- a/include/drm/gpu_scheduler.h
> +++ b/include/drm/gpu_scheduler.h
> @@ -262,6 +262,13 @@ struct drm_sched_rq {
>  	struct rb_root_cached		rb_tree_root;
>  };
>  
> +/*
> + * Used to flag dependency entries that are backed by drm_sched_fence object
> + * and on which we should only wait for the scheduled events.
> + * For internal use only.
> + */
> +#define DRM_SCHED_DEP_WAIT_FOR_SCHEDULED	XA_MARK_1
> +
>  /**
>   * struct drm_sched_fence - fences corresponding to the scheduling of a job.
>   */