[PATCH v3 4/9] drm/scheduler: Add fence deadline support

[Daniel Vetter]

On Fri, Sep 03, 2021 at 11:47:55AM -0700, Rob Clark wrote:
> From: Rob Clark <robdclark at chromium.org>
> 
> As the finished fence is the one that is exposed to userspace, and
> therefore the one that other operations, like atomic update, would
> block on, we need to propagate the deadline from from the finished
> fence to the actual hw fence.
> 
> v2: Split into drm_sched_fence_set_parent() (ckoenig)
> 
> Signed-off-by: Rob Clark <robdclark at chromium.org>
> ---
>  drivers/gpu/drm/scheduler/sched_fence.c | 34 +++++++++++++++++++++++++
>  drivers/gpu/drm/scheduler/sched_main.c  |  2 +-
>  include/drm/gpu_scheduler.h             |  8 ++++++
>  3 files changed, 43 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/gpu/drm/scheduler/sched_fence.c b/drivers/gpu/drm/scheduler/sched_fence.c
> index bcea035cf4c6..4fc41a71d1c7 100644
> --- a/drivers/gpu/drm/scheduler/sched_fence.c
> +++ b/drivers/gpu/drm/scheduler/sched_fence.c
> @@ -128,6 +128,30 @@ static void drm_sched_fence_release_finished(struct dma_fence *f)
>  	dma_fence_put(&fence->scheduled);
>  }
>  
> +static void drm_sched_fence_set_deadline_finished(struct dma_fence *f,
> +						  ktime_t deadline)
> +{
> +	struct drm_sched_fence *fence = to_drm_sched_fence(f);
> +	unsigned long flags;
> +
> +	spin_lock_irqsave(&fence->lock, flags);
> +
> +	/* If we already have an earlier deadline, keep it: */
> +	if (test_bit(DMA_FENCE_FLAG_HAS_DEADLINE_BIT, &f->flags) &&
> +	    ktime_before(fence->deadline, deadline)) {
> +		spin_unlock_irqrestore(&fence->lock, flags);
> +		return;
> +	}
> +
> +	fence->deadline = deadline;
> +	set_bit(DMA_FENCE_FLAG_HAS_DEADLINE_BIT, &f->flags);
> +
> +	spin_unlock_irqrestore(&fence->lock, flags);
> +
> +	if (fence->parent)
> +		dma_fence_set_deadline(fence->parent, deadline);
> +}
> +
>  static const struct dma_fence_ops drm_sched_fence_ops_scheduled = {
>  	.get_driver_name = drm_sched_fence_get_driver_name,
>  	.get_timeline_name = drm_sched_fence_get_timeline_name,
> @@ -138,6 +162,7 @@ static const struct dma_fence_ops drm_sched_fence_ops_finished = {
>  	.get_driver_name = drm_sched_fence_get_driver_name,
>  	.get_timeline_name = drm_sched_fence_get_timeline_name,
>  	.release = drm_sched_fence_release_finished,
> +	.set_deadline = drm_sched_fence_set_deadline_finished,
>  };
>  
>  struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f)
> @@ -152,6 +177,15 @@ struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f)
>  }
>  EXPORT_SYMBOL(to_drm_sched_fence);
>  
> +void drm_sched_fence_set_parent(struct drm_sched_fence *s_fence,
> +				struct dma_fence *fence)
> +{
> +	s_fence->parent = dma_fence_get(fence);
> +	if (test_bit(DMA_FENCE_FLAG_HAS_DEADLINE_BIT,
> +		     &s_fence->finished.flags))

Don't you need the spinlock here too to avoid races? test_bit is very
unordered, so guarantees nothing. Spinlock would need to be both around
->parent = and the test_bit.

Entirely aside, but there's discussions going on to preallocate the hw
fence somehow. If we do that we could make the deadline forwarding
lockless here. Having a spinlock just to set the parent is a bit annoying
...

Alternative is that you do this locklessly with barriers and a _lot_ of
comments. Would be good to benchmark whether the overhead matters though
first.
-Daniel

> +		dma_fence_set_deadline(fence, s_fence->deadline);
> +}
> +
>  struct drm_sched_fence *drm_sched_fence_alloc(struct drm_sched_entity *entity,
>  					      void *owner)
>  {
> diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c
> index 595e47ff7d06..27bf0ac0625f 100644
> --- a/drivers/gpu/drm/scheduler/sched_main.c
> +++ b/drivers/gpu/drm/scheduler/sched_main.c
> @@ -978,7 +978,7 @@ static int drm_sched_main(void *param)
>  		drm_sched_fence_scheduled(s_fence);
>  
>  		if (!IS_ERR_OR_NULL(fence)) {
> -			s_fence->parent = dma_fence_get(fence);
> +			drm_sched_fence_set_parent(s_fence, fence);
>  			r = dma_fence_add_callback(fence, &sched_job->cb,
>  						   drm_sched_job_done_cb);
>  			if (r == -ENOENT)
> diff --git a/include/drm/gpu_scheduler.h b/include/drm/gpu_scheduler.h
> index 7f77a455722c..158ddd662469 100644
> --- a/include/drm/gpu_scheduler.h
> +++ b/include/drm/gpu_scheduler.h
> @@ -238,6 +238,12 @@ struct drm_sched_fence {
>           */
>  	struct dma_fence		finished;
>  
> +	/**
> +	 * @deadline: deadline set on &drm_sched_fence.finished which
> +	 * potentially needs to be propagated to &drm_sched_fence.parent
> +	 */
> +	ktime_t				deadline;
> +
>          /**
>           * @parent: the fence returned by &drm_sched_backend_ops.run_job
>           * when scheduling the job on hardware. We signal the
> @@ -505,6 +511,8 @@ void drm_sched_entity_set_priority(struct drm_sched_entity *entity,
>  				   enum drm_sched_priority priority);
>  bool drm_sched_entity_is_ready(struct drm_sched_entity *entity);
>  
> +void drm_sched_fence_set_parent(struct drm_sched_fence *s_fence,
> +				struct dma_fence *fence);
>  struct drm_sched_fence *drm_sched_fence_alloc(
>  	struct drm_sched_entity *s_entity, void *owner);
>  void drm_sched_fence_init(struct drm_sched_fence *fence,
> -- 
> 2.31.1
> 

-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch