[PATCH 1/2] drm/sched: Refactor ring mirror list handling.

[PATCH 1/2] drm/sched: Refactor ring mirror list handling.

[Andrey Grodzovsky]

Decauple sched threads stop and start and ring mirror
list handling from the policy of what to do about the
guilty jobs.
When stoppping the sched thread and detaching sched fences
from non signaled HW fenes wait for all signaled HW fences
to complete before rerunning the jobs.

Suggested-by: Christian Koenig <Christian.Koenig@amd.com>
Signed-off-by: Andrey Grodzovsky <andrey.grodzovsky@amd.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 17 +++---
 drivers/gpu/drm/etnaviv/etnaviv_sched.c    |  8 +--
 drivers/gpu/drm/scheduler/sched_main.c     | 86 +++++++++++++++++++-----------
 drivers/gpu/drm/v3d/v3d_sched.c            | 11 ++--
 include/drm/gpu_scheduler.h                | 10 ++--
 5 files changed, 83 insertions(+), 49 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
index ef36cc5..42111d5 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
@@ -3292,17 +3292,16 @@ static int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev,
 	/* block all schedulers and reset given job's ring */
 	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
 		struct amdgpu_ring *ring = adev->rings[i];
+		bool park_only = job && job->base.sched != &ring->sched;
 
 		if (!ring || !ring->sched.thread)
 			continue;
 
-		kthread_park(ring->sched.thread);
+		drm_sched_stop(&ring->sched, job ? &job->base : NULL, park_only);
 
-		if (job && job->base.sched != &ring->sched)
+		if (park_only)
 			continue;
 
-		drm_sched_hw_job_reset(&ring->sched, job ? &job->base : NULL);
-
 		/* after all hw jobs are reset, hw fence is meaningless, so force_completion */
 		amdgpu_fence_driver_force_completion(ring);
 	}
@@ -3445,6 +3444,7 @@ static void amdgpu_device_post_asic_reset(struct amdgpu_device *adev,
 					  struct amdgpu_job *job)
 {
 	int i;
+	bool unpark_only;
 
 	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
 		struct amdgpu_ring *ring = adev->rings[i];
@@ -3456,10 +3456,13 @@ static void amdgpu_device_post_asic_reset(struct amdgpu_device *adev,
 		 * or all rings (in the case @job is NULL)
 		 * after above amdgpu_reset accomplished
 		 */
-		if ((!job || job->base.sched == &ring->sched) && !adev->asic_reset_res)
-			drm_sched_job_recovery(&ring->sched);
+		unpark_only = (job && job->base.sched != &ring->sched) ||
+			       adev->asic_reset_res;
+
+		if (!unpark_only)
+			drm_sched_resubmit_jobs(&ring->sched);
 
-		kthread_unpark(ring->sched.thread);
+		drm_sched_start(&ring->sched, unpark_only);
 	}
 
 	if (!amdgpu_device_has_dc_support(adev)) {
diff --git a/drivers/gpu/drm/etnaviv/etnaviv_sched.c b/drivers/gpu/drm/etnaviv/etnaviv_sched.c
index 49a6763..fab3b51 100644
--- a/drivers/gpu/drm/etnaviv/etnaviv_sched.c
+++ b/drivers/gpu/drm/etnaviv/etnaviv_sched.c
@@ -109,16 +109,16 @@ static void etnaviv_sched_timedout_job(struct drm_sched_job *sched_job)
 	}
 
 	/* block scheduler */
-	kthread_park(gpu->sched.thread);
-	drm_sched_hw_job_reset(&gpu->sched, sched_job);
+	drm_sched_stop(&gpu->sched, sched_job, false);
 
 	/* get the GPU back into the init state */
 	etnaviv_core_dump(gpu);
 	etnaviv_gpu_recover_hang(gpu);
 
+	drm_sched_resubmit_jobs(&gpu->sched);
+
 	/* restart scheduler after GPU is usable again */
-	drm_sched_job_recovery(&gpu->sched);
-	kthread_unpark(gpu->sched.thread);
+	drm_sched_start(&gpu->sched);
 }
 
 static void etnaviv_sched_free_job(struct drm_sched_job *sched_job)
diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c
index dbb6906..8fb7f86 100644
--- a/drivers/gpu/drm/scheduler/sched_main.c
+++ b/drivers/gpu/drm/scheduler/sched_main.c
@@ -60,8 +60,6 @@
 
 static void drm_sched_process_job(struct dma_fence *f, struct dma_fence_cb *cb);
 
-static void drm_sched_expel_job_unlocked(struct drm_sched_job *s_job);
-
 /**
  * drm_sched_rq_init - initialize a given run queue struct
  *
@@ -342,13 +340,21 @@ static void drm_sched_job_timedout(struct work_struct *work)
  * @bad: bad scheduler job
  *
  */
-void drm_sched_hw_job_reset(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad)
+void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad,
+		    bool park_only)
 {
 	struct drm_sched_job *s_job;
 	struct drm_sched_entity *entity, *tmp;
 	unsigned long flags;
+	struct list_head wait_list;
 	int i;
 
+	kthread_park(sched->thread);
+	if (park_only)
+		return;
+
+	INIT_LIST_HEAD(&wait_list);
+
 	spin_lock_irqsave(&sched->job_list_lock, flags);
 	list_for_each_entry_reverse(s_job, &sched->ring_mirror_list, node) {
 		if (s_job->s_fence->parent &&
@@ -358,9 +364,24 @@ void drm_sched_hw_job_reset(struct drm_gpu_scheduler *sched, struct drm_sched_jo
 			s_job->s_fence->parent = NULL;
 			atomic_dec(&sched->hw_rq_count);
 		}
+		else {
+			/* TODO Is it get/put neccessey here ? */
+			dma_fence_get(&s_job->s_fence->finished);
+			list_add(&s_job->finish_node, &wait_list);
+		}
 	}
 	spin_unlock_irqrestore(&sched->job_list_lock, flags);
 
+	/*
+	 * Verify all the signaled jobs in mirror list are removed from the ring
+	 * We rely on the fact that any finish_work in progress will wait for this
+	 * handler to complete before releasing all of the jobs we iterate.
+	 */
+	list_for_each_entry(s_job, &wait_list, finish_node) {
+		dma_fence_wait(&s_job->s_fence->finished, false);
+		dma_fence_put(&s_job->s_fence->finished);
+	}
+
 	if (bad && bad->s_priority != DRM_SCHED_PRIORITY_KERNEL) {
 		atomic_inc(&bad->karma);
 		/* don't increase @bad's karma if it's from KERNEL RQ,
@@ -385,7 +406,7 @@ void drm_sched_hw_job_reset(struct drm_gpu_scheduler *sched, struct drm_sched_jo
 		}
 	}
 }
-EXPORT_SYMBOL(drm_sched_hw_job_reset);
+EXPORT_SYMBOL(drm_sched_stop);
 
 /**
  * drm_sched_job_recovery - recover jobs after a reset
@@ -393,14 +414,17 @@ EXPORT_SYMBOL(drm_sched_hw_job_reset);
  * @sched: scheduler instance
  *
  */
-void drm_sched_job_recovery(struct drm_gpu_scheduler *sched)
+void drm_sched_start(struct drm_gpu_scheduler *sched, bool unpark_only)
 {
 	struct drm_sched_job *s_job, *tmp;
 	bool found_guilty = false;
 	unsigned long flags;
 	int r;
 
-	spin_lock_irqsave(&sched->job_list_lock, flags);
+	if (unpark_only)
+		goto unpark;
+
+	spin_lock_irqsave(&sched->job_list_lock, flags);	
 	list_for_each_entry_safe(s_job, tmp, &sched->ring_mirror_list, node) {
 		struct drm_sched_fence *s_fence = s_job->s_fence;
 		struct dma_fence *fence;
@@ -414,12 +438,9 @@ void drm_sched_job_recovery(struct drm_gpu_scheduler *sched)
 		if (found_guilty && s_job->s_fence->scheduled.context == guilty_context)
 			dma_fence_set_error(&s_fence->finished, -ECANCELED);
 
-		spin_unlock_irqrestore(&sched->job_list_lock, flags);
-		fence = sched->ops->run_job(s_job);
-		atomic_inc(&sched->hw_rq_count);
 
+		fence = s_job->s_fence->parent;
 		if (fence) {
-			s_fence->parent = dma_fence_get(fence);
 			r = dma_fence_add_callback(fence, &s_fence->cb,
 						   drm_sched_process_job);
 			if (r == -ENOENT)
@@ -427,18 +448,35 @@ void drm_sched_job_recovery(struct drm_gpu_scheduler *sched)
 			else if (r)
 				DRM_ERROR("fence add callback failed (%d)\n",
 					  r);
-			dma_fence_put(fence);
-		} else {
-			if (s_fence->finished.error < 0)
-				drm_sched_expel_job_unlocked(s_job);
+		} else
 			drm_sched_process_job(NULL, &s_fence->cb);
-		}
-		spin_lock_irqsave(&sched->job_list_lock, flags);
 	}
+
 	drm_sched_start_timeout(sched);
 	spin_unlock_irqrestore(&sched->job_list_lock, flags);
+
+unpark:
+	kthread_unpark(sched->thread);
 }
-EXPORT_SYMBOL(drm_sched_job_recovery);
+EXPORT_SYMBOL(drm_sched_start);
+
+/**
+ * drm_sched_resubmit_jobs - helper to relunch job from mirror ring list
+ *
+ * @sched: scheduler instance
+ *
+ */
+void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched)
+{
+	struct drm_sched_job *s_job, *tmp;
+
+	/*TODO DO we need spinlock here ? */
+	list_for_each_entry_safe(s_job, tmp, &sched->ring_mirror_list, node) {
+		s_job->s_fence->parent = sched->ops->run_job(s_job);
+		atomic_inc(&sched->hw_rq_count);
+	}
+}
+EXPORT_SYMBOL(drm_sched_resubmit_jobs);
 
 /**
  * drm_sched_job_init - init a scheduler job
@@ -634,26 +672,14 @@ static int drm_sched_main(void *param)
 				DRM_ERROR("fence add callback failed (%d)\n",
 					  r);
 			dma_fence_put(fence);
-		} else {
-			if (s_fence->finished.error < 0)
-				drm_sched_expel_job_unlocked(sched_job);
+		} else
 			drm_sched_process_job(NULL, &s_fence->cb);
-		}
 
 		wake_up(&sched->job_scheduled);
 	}
 	return 0;
 }
 
-static void drm_sched_expel_job_unlocked(struct drm_sched_job *s_job)
-{
-	struct drm_gpu_scheduler *sched = s_job->sched;
-
-	spin_lock(&sched->job_list_lock);
-	list_del_init(&s_job->node);
-	spin_unlock(&sched->job_list_lock);
-}
-
 /**
  * drm_sched_init - Init a gpu scheduler instance
  *
diff --git a/drivers/gpu/drm/v3d/v3d_sched.c b/drivers/gpu/drm/v3d/v3d_sched.c
index 445b2ef..f99346a 100644
--- a/drivers/gpu/drm/v3d/v3d_sched.c
+++ b/drivers/gpu/drm/v3d/v3d_sched.c
@@ -178,18 +178,19 @@ v3d_job_timedout(struct drm_sched_job *sched_job)
 	for (q = 0; q < V3D_MAX_QUEUES; q++) {
 		struct drm_gpu_scheduler *sched = &v3d->queue[q].sched;
 
-		kthread_park(sched->thread);
-		drm_sched_hw_job_reset(sched, (sched_job->sched == sched ?
-					       sched_job : NULL));
+		drm_sched_stop(sched, (sched_job->sched == sched ?
+					       sched_job : NULL), false);
 	}
 
 	/* get the GPU back into the init state */
 	v3d_reset(v3d);
 
+	for (q = 0; q < V3D_MAX_QUEUES; q++)
+		drm_sched_resubmit_jobs(sched_job->sched);
+
 	/* Unblock schedulers and restart their jobs. */
 	for (q = 0; q < V3D_MAX_QUEUES; q++) {
-		drm_sched_job_recovery(&v3d->queue[q].sched);
-		kthread_unpark(v3d->queue[q].sched.thread);
+		drm_sched_start(&v3d->queue[q].sched, false);
 	}
 
 	mutex_unlock(&v3d->reset_lock);
diff --git a/include/drm/gpu_scheduler.h b/include/drm/gpu_scheduler.h
index 47e1979..c94b592 100644
--- a/include/drm/gpu_scheduler.h
+++ b/include/drm/gpu_scheduler.h
@@ -175,6 +175,7 @@ struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f);
  *               finished to remove the job from the
  *               @drm_gpu_scheduler.ring_mirror_list.
  * @node: used to append this struct to the @drm_gpu_scheduler.ring_mirror_list.
+ * @finish_node: used in a list to wait on before resetting the scheduler
  * @id: a unique id assigned to each job scheduled on the scheduler.
  * @karma: increment on every hang caused by this job. If this exceeds the hang
  *         limit of the scheduler then the job is marked guilty and will not
@@ -193,6 +194,7 @@ struct drm_sched_job {
 	struct dma_fence_cb		finish_cb;
 	struct work_struct		finish_work;
 	struct list_head		node;
+	struct list_head		finish_node;
 	uint64_t			id;
 	atomic_t			karma;
 	enum drm_sched_priority		s_priority;
@@ -298,9 +300,11 @@ int drm_sched_job_init(struct drm_sched_job *job,
 		       void *owner);
 void drm_sched_job_cleanup(struct drm_sched_job *job);
 void drm_sched_wakeup(struct drm_gpu_scheduler *sched);
-void drm_sched_hw_job_reset(struct drm_gpu_scheduler *sched,
-			    struct drm_sched_job *job);
-void drm_sched_job_recovery(struct drm_gpu_scheduler *sched);
+void drm_sched_stop(struct drm_gpu_scheduler *sched,
+		    struct drm_sched_job *job,
+		    bool park_only);
+void drm_sched_start(struct drm_gpu_scheduler *sched, bool unpark_only);
+void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched);
 bool drm_sched_dependency_optimized(struct dma_fence* fence,
 				    struct drm_sched_entity *entity);
 void drm_sched_fault(struct drm_gpu_scheduler *sched);
-- 
2.7.4

_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel

[PATCH 2/2] drm/sched: Rework HW fence processing.

[Andrey Grodzovsky]

Expedite job deletion from ring mirror list to the HW fence signal
callback instead from finish_work, together with waiting for all
such fences to signal in drm_sched_stop we garantee that
already signaled job will not be processed twice.
Remove the sched finish fence callback and just submit finish_work
directly from the HW fence callback.

Suggested-by: Christian Koenig <Christian.Koenig@amd.com>
Signed-off-by: Andrey Grodzovsky <andrey.grodzovsky@amd.com>
---
 drivers/gpu/drm/scheduler/sched_fence.c |  4 +++-
 drivers/gpu/drm/scheduler/sched_main.c  | 39 ++++++++++++++++-----------------
 include/drm/gpu_scheduler.h             | 10 +++++++--
 3 files changed, 30 insertions(+), 23 deletions(-)

diff --git a/drivers/gpu/drm/scheduler/sched_fence.c b/drivers/gpu/drm/scheduler/sched_fence.c
index d8d2dff..e62c239 100644
--- a/drivers/gpu/drm/scheduler/sched_fence.c
+++ b/drivers/gpu/drm/scheduler/sched_fence.c
@@ -151,7 +151,8 @@ struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f)
 EXPORT_SYMBOL(to_drm_sched_fence);
 
 struct drm_sched_fence *drm_sched_fence_create(struct drm_sched_entity *entity,
-					       void *owner)
+					       void *owner,
+					       struct drm_sched_job *s_job)
 {
 	struct drm_sched_fence *fence = NULL;
 	unsigned seq;
@@ -163,6 +164,7 @@ struct drm_sched_fence *drm_sched_fence_create(struct drm_sched_entity *entity,
 	fence->owner = owner;
 	fence->sched = entity->rq->sched;
 	spin_lock_init(&fence->lock);
+	fence->s_job = s_job;
 
 	seq = atomic_inc_return(&entity->fence_seq);
 	dma_fence_init(&fence->scheduled, &drm_sched_fence_ops_scheduled,
diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c
index 8fb7f86..2860037 100644
--- a/drivers/gpu/drm/scheduler/sched_main.c
+++ b/drivers/gpu/drm/scheduler/sched_main.c
@@ -284,31 +284,17 @@ static void drm_sched_job_finish(struct work_struct *work)
 	cancel_delayed_work_sync(&sched->work_tdr);
 
 	spin_lock_irqsave(&sched->job_list_lock, flags);
-	/* remove job from ring_mirror_list */
-	list_del_init(&s_job->node);
-	/* queue TDR for next job */
 	drm_sched_start_timeout(sched);
 	spin_unlock_irqrestore(&sched->job_list_lock, flags);
 
 	sched->ops->free_job(s_job);
 }
 
-static void drm_sched_job_finish_cb(struct dma_fence *f,
-				    struct dma_fence_cb *cb)
-{
-	struct drm_sched_job *job = container_of(cb, struct drm_sched_job,
-						 finish_cb);
-	schedule_work(&job->finish_work);
-}
-
 static void drm_sched_job_begin(struct drm_sched_job *s_job)
 {
 	struct drm_gpu_scheduler *sched = s_job->sched;
 	unsigned long flags;
 
-	dma_fence_add_callback(&s_job->s_fence->finished, &s_job->finish_cb,
-			       drm_sched_job_finish_cb);
-
 	spin_lock_irqsave(&sched->job_list_lock, flags);
 	list_add_tail(&s_job->node, &sched->ring_mirror_list);
 	drm_sched_start_timeout(sched);
@@ -418,13 +404,17 @@ void drm_sched_start(struct drm_gpu_scheduler *sched, bool unpark_only)
 {
 	struct drm_sched_job *s_job, *tmp;
 	bool found_guilty = false;
-	unsigned long flags;
 	int r;
 
 	if (unpark_only)
 		goto unpark;
 
-	spin_lock_irqsave(&sched->job_list_lock, flags);	
+	/*
+	 * Locking the list is not required here as the sched thread is parked
+	 * so no new jobs are being pushed in to HW and in drm_sched_stop we
+	 * flushed any in flight jobs who didn't signal yet. Also concurrent
+	 * GPU recovers can't run in parallel.
+	 */
 	list_for_each_entry_safe(s_job, tmp, &sched->ring_mirror_list, node) {
 		struct drm_sched_fence *s_fence = s_job->s_fence;
 		struct dma_fence *fence;
@@ -453,7 +443,6 @@ void drm_sched_start(struct drm_gpu_scheduler *sched, bool unpark_only)
 	}
 
 	drm_sched_start_timeout(sched);
-	spin_unlock_irqrestore(&sched->job_list_lock, flags);
 
 unpark:
 	kthread_unpark(sched->thread);
@@ -505,7 +494,7 @@ int drm_sched_job_init(struct drm_sched_job *job,
 	job->sched = sched;
 	job->entity = entity;
 	job->s_priority = entity->rq - sched->sched_rq;
-	job->s_fence = drm_sched_fence_create(entity, owner);
+	job->s_fence = drm_sched_fence_create(entity, owner, job);
 	if (!job->s_fence)
 		return -ENOMEM;
 	job->id = atomic64_inc_return(&sched->job_id_count);
@@ -593,15 +582,25 @@ static void drm_sched_process_job(struct dma_fence *f, struct dma_fence_cb *cb)
 	struct drm_sched_fence *s_fence =
 		container_of(cb, struct drm_sched_fence, cb);
 	struct drm_gpu_scheduler *sched = s_fence->sched;
+	struct drm_sched_job *s_job = s_fence->s_job;
+	unsigned long flags;
+
+	cancel_delayed_work(&sched->work_tdr);
 
-	dma_fence_get(&s_fence->finished);
 	atomic_dec(&sched->hw_rq_count);
 	atomic_dec(&sched->num_jobs);
+
+	spin_lock_irqsave(&sched->job_list_lock, flags);
+	/* remove job from ring_mirror_list */
+	list_del_init(&s_job->node);
+	spin_unlock_irqrestore(&sched->job_list_lock, flags);
+
 	drm_sched_fence_finished(s_fence);
 
 	trace_drm_sched_process_job(s_fence);
-	dma_fence_put(&s_fence->finished);
 	wake_up_interruptible(&sched->wake_up_worker);
+
+	schedule_work(&s_job->finish_work);
 }
 
 /**
diff --git a/include/drm/gpu_scheduler.h b/include/drm/gpu_scheduler.h
index c94b592..23855c6 100644
--- a/include/drm/gpu_scheduler.h
+++ b/include/drm/gpu_scheduler.h
@@ -115,6 +115,8 @@ struct drm_sched_rq {
 	struct drm_sched_entity		*current_entity;
 };
 
+struct drm_sched_job;
+
 /**
  * struct drm_sched_fence - fences corresponding to the scheduling of a job.
  */
@@ -160,6 +162,9 @@ struct drm_sched_fence {
          * @owner: job owner for debugging
          */
 	void				*owner;
+
+	/* Back pointer to owning job */
+	struct drm_sched_job 		*s_job;
 };
 
 struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f);
@@ -330,8 +335,9 @@ 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);
 
-struct drm_sched_fence *drm_sched_fence_create(
-	struct drm_sched_entity *s_entity, void *owner);
+struct drm_sched_fence *drm_sched_fence_create(struct drm_sched_entity *s_entity,
+					       void *owner,
+					       struct drm_sched_job *s_job);
 void drm_sched_fence_scheduled(struct drm_sched_fence *fence);
 void drm_sched_fence_finished(struct drm_sched_fence *fence);
 
-- 
2.7.4

_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Re: [PATCH 2/2] drm/sched: Rework HW fence processing.

[Grodzovsky, Andrey]

On 12/06/2018 12:41 PM, Andrey Grodzovsky wrote:
> Expedite job deletion from ring mirror list to the HW fence signal
> callback instead from finish_work, together with waiting for all
> such fences to signal in drm_sched_stop we garantee that
> already signaled job will not be processed twice.
> Remove the sched finish fence callback and just submit finish_work
> directly from the HW fence callback.
>
> Suggested-by: Christian Koenig <Christian.Koenig@amd.com>
> Signed-off-by: Andrey Grodzovsky <andrey.grodzovsky@amd.com>
> ---
>   drivers/gpu/drm/scheduler/sched_fence.c |  4 +++-
>   drivers/gpu/drm/scheduler/sched_main.c  | 39 ++++++++++++++++-----------------
>   include/drm/gpu_scheduler.h             | 10 +++++++--
>   3 files changed, 30 insertions(+), 23 deletions(-)
>
> diff --git a/drivers/gpu/drm/scheduler/sched_fence.c b/drivers/gpu/drm/scheduler/sched_fence.c
> index d8d2dff..e62c239 100644
> --- a/drivers/gpu/drm/scheduler/sched_fence.c
> +++ b/drivers/gpu/drm/scheduler/sched_fence.c
> @@ -151,7 +151,8 @@ struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f)
>   EXPORT_SYMBOL(to_drm_sched_fence);
>   
>   struct drm_sched_fence *drm_sched_fence_create(struct drm_sched_entity *entity,
> -					       void *owner)
> +					       void *owner,
> +					       struct drm_sched_job *s_job)
>   {
>   	struct drm_sched_fence *fence = NULL;
>   	unsigned seq;
> @@ -163,6 +164,7 @@ struct drm_sched_fence *drm_sched_fence_create(struct drm_sched_entity *entity,
>   	fence->owner = owner;
>   	fence->sched = entity->rq->sched;
>   	spin_lock_init(&fence->lock);
> +	fence->s_job = s_job;
>   
>   	seq = atomic_inc_return(&entity->fence_seq);
>   	dma_fence_init(&fence->scheduled, &drm_sched_fence_ops_scheduled,
> diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c
> index 8fb7f86..2860037 100644
> --- a/drivers/gpu/drm/scheduler/sched_main.c
> +++ b/drivers/gpu/drm/scheduler/sched_main.c
> @@ -284,31 +284,17 @@ static void drm_sched_job_finish(struct work_struct *work)
>   	cancel_delayed_work_sync(&sched->work_tdr);
>   
>   	spin_lock_irqsave(&sched->job_list_lock, flags);
> -	/* remove job from ring_mirror_list */
> -	list_del_init(&s_job->node);
> -	/* queue TDR for next job */
>   	drm_sched_start_timeout(sched);
>   	spin_unlock_irqrestore(&sched->job_list_lock, flags);
>   
>   	sched->ops->free_job(s_job);
>   }
>   
> -static void drm_sched_job_finish_cb(struct dma_fence *f,
> -				    struct dma_fence_cb *cb)
> -{
> -	struct drm_sched_job *job = container_of(cb, struct drm_sched_job,
> -						 finish_cb);
> -	schedule_work(&job->finish_work);
> -}
> -
>   static void drm_sched_job_begin(struct drm_sched_job *s_job)
>   {
>   	struct drm_gpu_scheduler *sched = s_job->sched;
>   	unsigned long flags;
>   
> -	dma_fence_add_callback(&s_job->s_fence->finished, &s_job->finish_cb,
> -			       drm_sched_job_finish_cb);
> -
>   	spin_lock_irqsave(&sched->job_list_lock, flags);
>   	list_add_tail(&s_job->node, &sched->ring_mirror_list);
>   	drm_sched_start_timeout(sched);
> @@ -418,13 +404,17 @@ void drm_sched_start(struct drm_gpu_scheduler *sched, bool unpark_only)
>   {
>   	struct drm_sched_job *s_job, *tmp;
>   	bool found_guilty = false;
> -	unsigned long flags;
>   	int r;
>   
>   	if (unpark_only)
>   		goto unpark;
>   
> -	spin_lock_irqsave(&sched->job_list_lock, flags);	
> +	/*
> +	 * Locking the list is not required here as the sched thread is parked
> +	 * so no new jobs are being pushed in to HW and in drm_sched_stop we
> +	 * flushed any in flight jobs who didn't signal yet.

The comment is inaccurate here - it's supposed to be ' any in flight 
jobs who already have their
sched finished signaled and they are removed from the mirror ring list 
at that point already anyway'
I will fix this text later with other comments received on the patches.

Andrey
> Also concurrent
> +	 * GPU recovers can't run in parallel.
> +	 */
>   	list_for_each_entry_safe(s_job, tmp, &sched->ring_mirror_list, node) {
>   		struct drm_sched_fence *s_fence = s_job->s_fence;
>   		struct dma_fence *fence;
> @@ -453,7 +443,6 @@ void drm_sched_start(struct drm_gpu_scheduler *sched, bool unpark_only)
>   	}
>   
>   	drm_sched_start_timeout(sched);
> -	spin_unlock_irqrestore(&sched->job_list_lock, flags);
>   
>   unpark:
>   	kthread_unpark(sched->thread);
> @@ -505,7 +494,7 @@ int drm_sched_job_init(struct drm_sched_job *job,
>   	job->sched = sched;
>   	job->entity = entity;
>   	job->s_priority = entity->rq - sched->sched_rq;
> -	job->s_fence = drm_sched_fence_create(entity, owner);
> +	job->s_fence = drm_sched_fence_create(entity, owner, job);
>   	if (!job->s_fence)
>   		return -ENOMEM;
>   	job->id = atomic64_inc_return(&sched->job_id_count);
> @@ -593,15 +582,25 @@ static void drm_sched_process_job(struct dma_fence *f, struct dma_fence_cb *cb)
>   	struct drm_sched_fence *s_fence =
>   		container_of(cb, struct drm_sched_fence, cb);
>   	struct drm_gpu_scheduler *sched = s_fence->sched;
> +	struct drm_sched_job *s_job = s_fence->s_job;
> +	unsigned long flags;
> +
> +	cancel_delayed_work(&sched->work_tdr);
>   
> -	dma_fence_get(&s_fence->finished);
>   	atomic_dec(&sched->hw_rq_count);
>   	atomic_dec(&sched->num_jobs);
> +
> +	spin_lock_irqsave(&sched->job_list_lock, flags);
> +	/* remove job from ring_mirror_list */
> +	list_del_init(&s_job->node);
> +	spin_unlock_irqrestore(&sched->job_list_lock, flags);
> +
>   	drm_sched_fence_finished(s_fence);
>   
>   	trace_drm_sched_process_job(s_fence);
> -	dma_fence_put(&s_fence->finished);
>   	wake_up_interruptible(&sched->wake_up_worker);
> +
> +	schedule_work(&s_job->finish_work);
>   }
>   
>   /**
> diff --git a/include/drm/gpu_scheduler.h b/include/drm/gpu_scheduler.h
> index c94b592..23855c6 100644
> --- a/include/drm/gpu_scheduler.h
> +++ b/include/drm/gpu_scheduler.h
> @@ -115,6 +115,8 @@ struct drm_sched_rq {
>   	struct drm_sched_entity		*current_entity;
>   };
>   
> +struct drm_sched_job;
> +
>   /**
>    * struct drm_sched_fence - fences corresponding to the scheduling of a job.
>    */
> @@ -160,6 +162,9 @@ struct drm_sched_fence {
>            * @owner: job owner for debugging
>            */
>   	void				*owner;
> +
> +	/* Back pointer to owning job */
> +	struct drm_sched_job 		*s_job;
>   };
>   
>   struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f);
> @@ -330,8 +335,9 @@ 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);
>   
> -struct drm_sched_fence *drm_sched_fence_create(
> -	struct drm_sched_entity *s_entity, void *owner);
> +struct drm_sched_fence *drm_sched_fence_create(struct drm_sched_entity *s_entity,
> +					       void *owner,
> +					       struct drm_sched_job *s_job);
>   void drm_sched_fence_scheduled(struct drm_sched_fence *fence);
>   void drm_sched_fence_finished(struct drm_sched_fence *fence);
>   

_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Re: [PATCH 1/2] drm/sched: Refactor ring mirror list handling.

[Christian König]

Am 06.12.18 um 18:41 schrieb Andrey Grodzovsky:
> Decauple sched threads stop and start and ring mirror
> list handling from the policy of what to do about the
> guilty jobs.
> When stoppping the sched thread and detaching sched fences
> from non signaled HW fenes wait for all signaled HW fences
> to complete before rerunning the jobs.
>
> Suggested-by: Christian Koenig <Christian.Koenig@amd.com>
> Signed-off-by: Andrey Grodzovsky <andrey.grodzovsky@amd.com>

Just briefly skimmed over this, but it looks exactly like what I had in 
mind.

Need to give that a more detailed thought tomorrow,
Christian.

> ---
>   drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 17 +++---
>   drivers/gpu/drm/etnaviv/etnaviv_sched.c    |  8 +--
>   drivers/gpu/drm/scheduler/sched_main.c     | 86 +++++++++++++++++++-----------
>   drivers/gpu/drm/v3d/v3d_sched.c            | 11 ++--
>   include/drm/gpu_scheduler.h                | 10 ++--
>   5 files changed, 83 insertions(+), 49 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> index ef36cc5..42111d5 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> @@ -3292,17 +3292,16 @@ static int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev,
>   	/* block all schedulers and reset given job's ring */
>   	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
>   		struct amdgpu_ring *ring = adev->rings[i];
> +		bool park_only = job && job->base.sched != &ring->sched;
>   
>   		if (!ring || !ring->sched.thread)
>   			continue;
>   
> -		kthread_park(ring->sched.thread);
> +		drm_sched_stop(&ring->sched, job ? &job->base : NULL, park_only);
>   
> -		if (job && job->base.sched != &ring->sched)
> +		if (park_only)
>   			continue;
>   
> -		drm_sched_hw_job_reset(&ring->sched, job ? &job->base : NULL);
> -
>   		/* after all hw jobs are reset, hw fence is meaningless, so force_completion */
>   		amdgpu_fence_driver_force_completion(ring);
>   	}
> @@ -3445,6 +3444,7 @@ static void amdgpu_device_post_asic_reset(struct amdgpu_device *adev,
>   					  struct amdgpu_job *job)
>   {
>   	int i;
> +	bool unpark_only;
>   
>   	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
>   		struct amdgpu_ring *ring = adev->rings[i];
> @@ -3456,10 +3456,13 @@ static void amdgpu_device_post_asic_reset(struct amdgpu_device *adev,
>   		 * or all rings (in the case @job is NULL)
>   		 * after above amdgpu_reset accomplished
>   		 */
> -		if ((!job || job->base.sched == &ring->sched) && !adev->asic_reset_res)
> -			drm_sched_job_recovery(&ring->sched);
> +		unpark_only = (job && job->base.sched != &ring->sched) ||
> +			       adev->asic_reset_res;
> +
> +		if (!unpark_only)
> +			drm_sched_resubmit_jobs(&ring->sched);
>   
> -		kthread_unpark(ring->sched.thread);
> +		drm_sched_start(&ring->sched, unpark_only);
>   	}
>   
>   	if (!amdgpu_device_has_dc_support(adev)) {
> diff --git a/drivers/gpu/drm/etnaviv/etnaviv_sched.c b/drivers/gpu/drm/etnaviv/etnaviv_sched.c
> index 49a6763..fab3b51 100644
> --- a/drivers/gpu/drm/etnaviv/etnaviv_sched.c
> +++ b/drivers/gpu/drm/etnaviv/etnaviv_sched.c
> @@ -109,16 +109,16 @@ static void etnaviv_sched_timedout_job(struct drm_sched_job *sched_job)
>   	}
>   
>   	/* block scheduler */
> -	kthread_park(gpu->sched.thread);
> -	drm_sched_hw_job_reset(&gpu->sched, sched_job);
> +	drm_sched_stop(&gpu->sched, sched_job, false);
>   
>   	/* get the GPU back into the init state */
>   	etnaviv_core_dump(gpu);
>   	etnaviv_gpu_recover_hang(gpu);
>   
> +	drm_sched_resubmit_jobs(&gpu->sched);
> +
>   	/* restart scheduler after GPU is usable again */
> -	drm_sched_job_recovery(&gpu->sched);
> -	kthread_unpark(gpu->sched.thread);
> +	drm_sched_start(&gpu->sched);
>   }
>   
>   static void etnaviv_sched_free_job(struct drm_sched_job *sched_job)
> diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c
> index dbb6906..8fb7f86 100644
> --- a/drivers/gpu/drm/scheduler/sched_main.c
> +++ b/drivers/gpu/drm/scheduler/sched_main.c
> @@ -60,8 +60,6 @@
>   
>   static void drm_sched_process_job(struct dma_fence *f, struct dma_fence_cb *cb);
>   
> -static void drm_sched_expel_job_unlocked(struct drm_sched_job *s_job);
> -
>   /**
>    * drm_sched_rq_init - initialize a given run queue struct
>    *
> @@ -342,13 +340,21 @@ static void drm_sched_job_timedout(struct work_struct *work)
>    * @bad: bad scheduler job
>    *
>    */
> -void drm_sched_hw_job_reset(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad)
> +void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad,
> +		    bool park_only)
>   {
>   	struct drm_sched_job *s_job;
>   	struct drm_sched_entity *entity, *tmp;
>   	unsigned long flags;
> +	struct list_head wait_list;
>   	int i;
>   
> +	kthread_park(sched->thread);
> +	if (park_only)
> +		return;
> +
> +	INIT_LIST_HEAD(&wait_list);
> +
>   	spin_lock_irqsave(&sched->job_list_lock, flags);
>   	list_for_each_entry_reverse(s_job, &sched->ring_mirror_list, node) {
>   		if (s_job->s_fence->parent &&
> @@ -358,9 +364,24 @@ void drm_sched_hw_job_reset(struct drm_gpu_scheduler *sched, struct drm_sched_jo
>   			s_job->s_fence->parent = NULL;
>   			atomic_dec(&sched->hw_rq_count);
>   		}
> +		else {
> +			/* TODO Is it get/put neccessey here ? */
> +			dma_fence_get(&s_job->s_fence->finished);
> +			list_add(&s_job->finish_node, &wait_list);
> +		}
>   	}
>   	spin_unlock_irqrestore(&sched->job_list_lock, flags);
>   
> +	/*
> +	 * Verify all the signaled jobs in mirror list are removed from the ring
> +	 * We rely on the fact that any finish_work in progress will wait for this
> +	 * handler to complete before releasing all of the jobs we iterate.
> +	 */
> +	list_for_each_entry(s_job, &wait_list, finish_node) {
> +		dma_fence_wait(&s_job->s_fence->finished, false);
> +		dma_fence_put(&s_job->s_fence->finished);
> +	}
> +
>   	if (bad && bad->s_priority != DRM_SCHED_PRIORITY_KERNEL) {
>   		atomic_inc(&bad->karma);
>   		/* don't increase @bad's karma if it's from KERNEL RQ,
> @@ -385,7 +406,7 @@ void drm_sched_hw_job_reset(struct drm_gpu_scheduler *sched, struct drm_sched_jo
>   		}
>   	}
>   }
> -EXPORT_SYMBOL(drm_sched_hw_job_reset);
> +EXPORT_SYMBOL(drm_sched_stop);
>   
>   /**
>    * drm_sched_job_recovery - recover jobs after a reset
> @@ -393,14 +414,17 @@ EXPORT_SYMBOL(drm_sched_hw_job_reset);
>    * @sched: scheduler instance
>    *
>    */
> -void drm_sched_job_recovery(struct drm_gpu_scheduler *sched)
> +void drm_sched_start(struct drm_gpu_scheduler *sched, bool unpark_only)
>   {
>   	struct drm_sched_job *s_job, *tmp;
>   	bool found_guilty = false;
>   	unsigned long flags;
>   	int r;
>   
> -	spin_lock_irqsave(&sched->job_list_lock, flags);
> +	if (unpark_only)
> +		goto unpark;
> +
> +	spin_lock_irqsave(&sched->job_list_lock, flags);	
>   	list_for_each_entry_safe(s_job, tmp, &sched->ring_mirror_list, node) {
>   		struct drm_sched_fence *s_fence = s_job->s_fence;
>   		struct dma_fence *fence;
> @@ -414,12 +438,9 @@ void drm_sched_job_recovery(struct drm_gpu_scheduler *sched)
>   		if (found_guilty && s_job->s_fence->scheduled.context == guilty_context)
>   			dma_fence_set_error(&s_fence->finished, -ECANCELED);
>   
> -		spin_unlock_irqrestore(&sched->job_list_lock, flags);
> -		fence = sched->ops->run_job(s_job);
> -		atomic_inc(&sched->hw_rq_count);
>   
> +		fence = s_job->s_fence->parent;
>   		if (fence) {
> -			s_fence->parent = dma_fence_get(fence);
>   			r = dma_fence_add_callback(fence, &s_fence->cb,
>   						   drm_sched_process_job);
>   			if (r == -ENOENT)
> @@ -427,18 +448,35 @@ void drm_sched_job_recovery(struct drm_gpu_scheduler *sched)
>   			else if (r)
>   				DRM_ERROR("fence add callback failed (%d)\n",
>   					  r);
> -			dma_fence_put(fence);
> -		} else {
> -			if (s_fence->finished.error < 0)
> -				drm_sched_expel_job_unlocked(s_job);
> +		} else
>   			drm_sched_process_job(NULL, &s_fence->cb);
> -		}
> -		spin_lock_irqsave(&sched->job_list_lock, flags);
>   	}
> +
>   	drm_sched_start_timeout(sched);
>   	spin_unlock_irqrestore(&sched->job_list_lock, flags);
> +
> +unpark:
> +	kthread_unpark(sched->thread);
>   }
> -EXPORT_SYMBOL(drm_sched_job_recovery);
> +EXPORT_SYMBOL(drm_sched_start);
> +
> +/**
> + * drm_sched_resubmit_jobs - helper to relunch job from mirror ring list
> + *
> + * @sched: scheduler instance
> + *
> + */
> +void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched)
> +{
> +	struct drm_sched_job *s_job, *tmp;
> +
> +	/*TODO DO we need spinlock here ? */
> +	list_for_each_entry_safe(s_job, tmp, &sched->ring_mirror_list, node) {
> +		s_job->s_fence->parent = sched->ops->run_job(s_job);
> +		atomic_inc(&sched->hw_rq_count);
> +	}
> +}
> +EXPORT_SYMBOL(drm_sched_resubmit_jobs);
>   
>   /**
>    * drm_sched_job_init - init a scheduler job
> @@ -634,26 +672,14 @@ static int drm_sched_main(void *param)
>   				DRM_ERROR("fence add callback failed (%d)\n",
>   					  r);
>   			dma_fence_put(fence);
> -		} else {
> -			if (s_fence->finished.error < 0)
> -				drm_sched_expel_job_unlocked(sched_job);
> +		} else
>   			drm_sched_process_job(NULL, &s_fence->cb);
> -		}
>   
>   		wake_up(&sched->job_scheduled);
>   	}
>   	return 0;
>   }
>   
> -static void drm_sched_expel_job_unlocked(struct drm_sched_job *s_job)
> -{
> -	struct drm_gpu_scheduler *sched = s_job->sched;
> -
> -	spin_lock(&sched->job_list_lock);
> -	list_del_init(&s_job->node);
> -	spin_unlock(&sched->job_list_lock);
> -}
> -
>   /**
>    * drm_sched_init - Init a gpu scheduler instance
>    *
> diff --git a/drivers/gpu/drm/v3d/v3d_sched.c b/drivers/gpu/drm/v3d/v3d_sched.c
> index 445b2ef..f99346a 100644
> --- a/drivers/gpu/drm/v3d/v3d_sched.c
> +++ b/drivers/gpu/drm/v3d/v3d_sched.c
> @@ -178,18 +178,19 @@ v3d_job_timedout(struct drm_sched_job *sched_job)
>   	for (q = 0; q < V3D_MAX_QUEUES; q++) {
>   		struct drm_gpu_scheduler *sched = &v3d->queue[q].sched;
>   
> -		kthread_park(sched->thread);
> -		drm_sched_hw_job_reset(sched, (sched_job->sched == sched ?
> -					       sched_job : NULL));
> +		drm_sched_stop(sched, (sched_job->sched == sched ?
> +					       sched_job : NULL), false);
>   	}
>   
>   	/* get the GPU back into the init state */
>   	v3d_reset(v3d);
>   
> +	for (q = 0; q < V3D_MAX_QUEUES; q++)
> +		drm_sched_resubmit_jobs(sched_job->sched);
> +
>   	/* Unblock schedulers and restart their jobs. */
>   	for (q = 0; q < V3D_MAX_QUEUES; q++) {
> -		drm_sched_job_recovery(&v3d->queue[q].sched);
> -		kthread_unpark(v3d->queue[q].sched.thread);
> +		drm_sched_start(&v3d->queue[q].sched, false);
>   	}
>   
>   	mutex_unlock(&v3d->reset_lock);
> diff --git a/include/drm/gpu_scheduler.h b/include/drm/gpu_scheduler.h
> index 47e1979..c94b592 100644
> --- a/include/drm/gpu_scheduler.h
> +++ b/include/drm/gpu_scheduler.h
> @@ -175,6 +175,7 @@ struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f);
>    *               finished to remove the job from the
>    *               @drm_gpu_scheduler.ring_mirror_list.
>    * @node: used to append this struct to the @drm_gpu_scheduler.ring_mirror_list.
> + * @finish_node: used in a list to wait on before resetting the scheduler
>    * @id: a unique id assigned to each job scheduled on the scheduler.
>    * @karma: increment on every hang caused by this job. If this exceeds the hang
>    *         limit of the scheduler then the job is marked guilty and will not
> @@ -193,6 +194,7 @@ struct drm_sched_job {
>   	struct dma_fence_cb		finish_cb;
>   	struct work_struct		finish_work;
>   	struct list_head		node;
> +	struct list_head		finish_node;
>   	uint64_t			id;
>   	atomic_t			karma;
>   	enum drm_sched_priority		s_priority;
> @@ -298,9 +300,11 @@ int drm_sched_job_init(struct drm_sched_job *job,
>   		       void *owner);
>   void drm_sched_job_cleanup(struct drm_sched_job *job);
>   void drm_sched_wakeup(struct drm_gpu_scheduler *sched);
> -void drm_sched_hw_job_reset(struct drm_gpu_scheduler *sched,
> -			    struct drm_sched_job *job);
> -void drm_sched_job_recovery(struct drm_gpu_scheduler *sched);
> +void drm_sched_stop(struct drm_gpu_scheduler *sched,
> +		    struct drm_sched_job *job,
> +		    bool park_only);
> +void drm_sched_start(struct drm_gpu_scheduler *sched, bool unpark_only);
> +void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched);
>   bool drm_sched_dependency_optimized(struct dma_fence* fence,
>   				    struct drm_sched_entity *entity);
>   void drm_sched_fault(struct drm_gpu_scheduler *sched);

_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

Re: [PATCH 1/2] drm/sched: Refactor ring mirror list handling.

[Grodzovsky, Andrey]

On 12/06/2018 01:33 PM, Christian König wrote:
> Am 06.12.18 um 18:41 schrieb Andrey Grodzovsky:
>> Decauple sched threads stop and start and ring mirror
>> list handling from the policy of what to do about the
>> guilty jobs.
>> When stoppping the sched thread and detaching sched fences
>> from non signaled HW fenes wait for all signaled HW fences
>> to complete before rerunning the jobs.
>>
>> Suggested-by: Christian Koenig <Christian.Koenig@amd.com>
>> Signed-off-by: Andrey Grodzovsky <andrey.grodzovsky@amd.com>
>
> Just briefly skimmed over this, but it looks exactly like what I had 
> in mind.
>
> Need to give that a more detailed thought tomorrow,
> Christian.

Please note I've already resent V2 after finding refactoring error.

Andrey

>
>> ---
>>   drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 17 +++---
>>   drivers/gpu/drm/etnaviv/etnaviv_sched.c    |  8 +--
>>   drivers/gpu/drm/scheduler/sched_main.c     | 86 
>> +++++++++++++++++++-----------
>>   drivers/gpu/drm/v3d/v3d_sched.c            | 11 ++--
>>   include/drm/gpu_scheduler.h                | 10 ++--
>>   5 files changed, 83 insertions(+), 49 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c 
>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>> index ef36cc5..42111d5 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>> @@ -3292,17 +3292,16 @@ static int 
>> amdgpu_device_pre_asic_reset(struct amdgpu_device *adev,
>>       /* block all schedulers and reset given job's ring */
>>       for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
>>           struct amdgpu_ring *ring = adev->rings[i];
>> +        bool park_only = job && job->base.sched != &ring->sched;
>>             if (!ring || !ring->sched.thread)
>>               continue;
>>   -        kthread_park(ring->sched.thread);
>> +        drm_sched_stop(&ring->sched, job ? &job->base : NULL, 
>> park_only);
>>   -        if (job && job->base.sched != &ring->sched)
>> +        if (park_only)
>>               continue;
>>   -        drm_sched_hw_job_reset(&ring->sched, job ? &job->base : 
>> NULL);
>> -
>>           /* after all hw jobs are reset, hw fence is meaningless, so 
>> force_completion */
>>           amdgpu_fence_driver_force_completion(ring);
>>       }
>> @@ -3445,6 +3444,7 @@ static void 
>> amdgpu_device_post_asic_reset(struct amdgpu_device *adev,
>>                         struct amdgpu_job *job)
>>   {
>>       int i;
>> +    bool unpark_only;
>>         for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
>>           struct amdgpu_ring *ring = adev->rings[i];
>> @@ -3456,10 +3456,13 @@ static void 
>> amdgpu_device_post_asic_reset(struct amdgpu_device *adev,
>>            * or all rings (in the case @job is NULL)
>>            * after above amdgpu_reset accomplished
>>            */
>> -        if ((!job || job->base.sched == &ring->sched) && 
>> !adev->asic_reset_res)
>> -            drm_sched_job_recovery(&ring->sched);
>> +        unpark_only = (job && job->base.sched != &ring->sched) ||
>> +                   adev->asic_reset_res;
>> +
>> +        if (!unpark_only)
>> +            drm_sched_resubmit_jobs(&ring->sched);
>>   -        kthread_unpark(ring->sched.thread);
>> +        drm_sched_start(&ring->sched, unpark_only);
>>       }
>>         if (!amdgpu_device_has_dc_support(adev)) {
>> diff --git a/drivers/gpu/drm/etnaviv/etnaviv_sched.c 
>> b/drivers/gpu/drm/etnaviv/etnaviv_sched.c
>> index 49a6763..fab3b51 100644
>> --- a/drivers/gpu/drm/etnaviv/etnaviv_sched.c
>> +++ b/drivers/gpu/drm/etnaviv/etnaviv_sched.c
>> @@ -109,16 +109,16 @@ static void etnaviv_sched_timedout_job(struct 
>> drm_sched_job *sched_job)
>>       }
>>         /* block scheduler */
>> -    kthread_park(gpu->sched.thread);
>> -    drm_sched_hw_job_reset(&gpu->sched, sched_job);
>> +    drm_sched_stop(&gpu->sched, sched_job, false);
>>         /* get the GPU back into the init state */
>>       etnaviv_core_dump(gpu);
>>       etnaviv_gpu_recover_hang(gpu);
>>   +    drm_sched_resubmit_jobs(&gpu->sched);
>> +
>>       /* restart scheduler after GPU is usable again */
>> -    drm_sched_job_recovery(&gpu->sched);
>> -    kthread_unpark(gpu->sched.thread);
>> +    drm_sched_start(&gpu->sched);
>>   }
>>     static void etnaviv_sched_free_job(struct drm_sched_job *sched_job)
>> diff --git a/drivers/gpu/drm/scheduler/sched_main.c 
>> b/drivers/gpu/drm/scheduler/sched_main.c
>> index dbb6906..8fb7f86 100644
>> --- a/drivers/gpu/drm/scheduler/sched_main.c
>> +++ b/drivers/gpu/drm/scheduler/sched_main.c
>> @@ -60,8 +60,6 @@
>>     static void drm_sched_process_job(struct dma_fence *f, struct 
>> dma_fence_cb *cb);
>>   -static void drm_sched_expel_job_unlocked(struct drm_sched_job 
>> *s_job);
>> -
>>   /**
>>    * drm_sched_rq_init - initialize a given run queue struct
>>    *
>> @@ -342,13 +340,21 @@ static void drm_sched_job_timedout(struct 
>> work_struct *work)
>>    * @bad: bad scheduler job
>>    *
>>    */
>> -void drm_sched_hw_job_reset(struct drm_gpu_scheduler *sched, struct 
>> drm_sched_job *bad)
>> +void drm_sched_stop(struct drm_gpu_scheduler *sched, struct 
>> drm_sched_job *bad,
>> +            bool park_only)
>>   {
>>       struct drm_sched_job *s_job;
>>       struct drm_sched_entity *entity, *tmp;
>>       unsigned long flags;
>> +    struct list_head wait_list;
>>       int i;
>>   +    kthread_park(sched->thread);
>> +    if (park_only)
>> +        return;
>> +
>> +    INIT_LIST_HEAD(&wait_list);
>> +
>>       spin_lock_irqsave(&sched->job_list_lock, flags);
>>       list_for_each_entry_reverse(s_job, &sched->ring_mirror_list, 
>> node) {
>>           if (s_job->s_fence->parent &&
>> @@ -358,9 +364,24 @@ void drm_sched_hw_job_reset(struct 
>> drm_gpu_scheduler *sched, struct drm_sched_jo
>>               s_job->s_fence->parent = NULL;
>>               atomic_dec(&sched->hw_rq_count);
>>           }
>> +        else {
>> +            /* TODO Is it get/put neccessey here ? */
>> +            dma_fence_get(&s_job->s_fence->finished);
>> +            list_add(&s_job->finish_node, &wait_list);
>> +        }
>>       }
>>       spin_unlock_irqrestore(&sched->job_list_lock, flags);
>>   +    /*
>> +     * Verify all the signaled jobs in mirror list are removed from 
>> the ring
>> +     * We rely on the fact that any finish_work in progress will 
>> wait for this
>> +     * handler to complete before releasing all of the jobs we iterate.
>> +     */
>> +    list_for_each_entry(s_job, &wait_list, finish_node) {
>> +        dma_fence_wait(&s_job->s_fence->finished, false);
>> +        dma_fence_put(&s_job->s_fence->finished);
>> +    }
>> +
>>       if (bad && bad->s_priority != DRM_SCHED_PRIORITY_KERNEL) {
>>           atomic_inc(&bad->karma);
>>           /* don't increase @bad's karma if it's from KERNEL RQ,
>> @@ -385,7 +406,7 @@ void drm_sched_hw_job_reset(struct 
>> drm_gpu_scheduler *sched, struct drm_sched_jo
>>           }
>>       }
>>   }
>> -EXPORT_SYMBOL(drm_sched_hw_job_reset);
>> +EXPORT_SYMBOL(drm_sched_stop);
>>     /**
>>    * drm_sched_job_recovery - recover jobs after a reset
>> @@ -393,14 +414,17 @@ EXPORT_SYMBOL(drm_sched_hw_job_reset);
>>    * @sched: scheduler instance
>>    *
>>    */
>> -void drm_sched_job_recovery(struct drm_gpu_scheduler *sched)
>> +void drm_sched_start(struct drm_gpu_scheduler *sched, bool unpark_only)
>>   {
>>       struct drm_sched_job *s_job, *tmp;
>>       bool found_guilty = false;
>>       unsigned long flags;
>>       int r;
>>   -    spin_lock_irqsave(&sched->job_list_lock, flags);
>> +    if (unpark_only)
>> +        goto unpark;
>> +
>> +    spin_lock_irqsave(&sched->job_list_lock, flags);
>>       list_for_each_entry_safe(s_job, tmp, &sched->ring_mirror_list, 
>> node) {
>>           struct drm_sched_fence *s_fence = s_job->s_fence;
>>           struct dma_fence *fence;
>> @@ -414,12 +438,9 @@ void drm_sched_job_recovery(struct 
>> drm_gpu_scheduler *sched)
>>           if (found_guilty && s_job->s_fence->scheduled.context == 
>> guilty_context)
>>               dma_fence_set_error(&s_fence->finished, -ECANCELED);
>>   -        spin_unlock_irqrestore(&sched->job_list_lock, flags);
>> -        fence = sched->ops->run_job(s_job);
>> -        atomic_inc(&sched->hw_rq_count);
>>   +        fence = s_job->s_fence->parent;
>>           if (fence) {
>> -            s_fence->parent = dma_fence_get(fence);
>>               r = dma_fence_add_callback(fence, &s_fence->cb,
>>                              drm_sched_process_job);
>>               if (r == -ENOENT)
>> @@ -427,18 +448,35 @@ void drm_sched_job_recovery(struct 
>> drm_gpu_scheduler *sched)
>>               else if (r)
>>                   DRM_ERROR("fence add callback failed (%d)\n",
>>                         r);
>> -            dma_fence_put(fence);
>> -        } else {
>> -            if (s_fence->finished.error < 0)
>> -                drm_sched_expel_job_unlocked(s_job);
>> +        } else
>>               drm_sched_process_job(NULL, &s_fence->cb);
>> -        }
>> -        spin_lock_irqsave(&sched->job_list_lock, flags);
>>       }
>> +
>>       drm_sched_start_timeout(sched);
>>       spin_unlock_irqrestore(&sched->job_list_lock, flags);
>> +
>> +unpark:
>> +    kthread_unpark(sched->thread);
>>   }
>> -EXPORT_SYMBOL(drm_sched_job_recovery);
>> +EXPORT_SYMBOL(drm_sched_start);
>> +
>> +/**
>> + * drm_sched_resubmit_jobs - helper to relunch job from mirror ring 
>> list
>> + *
>> + * @sched: scheduler instance
>> + *
>> + */
>> +void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched)
>> +{
>> +    struct drm_sched_job *s_job, *tmp;
>> +
>> +    /*TODO DO we need spinlock here ? */
>> +    list_for_each_entry_safe(s_job, tmp, &sched->ring_mirror_list, 
>> node) {
>> +        s_job->s_fence->parent = sched->ops->run_job(s_job);
>> +        atomic_inc(&sched->hw_rq_count);
>> +    }
>> +}
>> +EXPORT_SYMBOL(drm_sched_resubmit_jobs);
>>     /**
>>    * drm_sched_job_init - init a scheduler job
>> @@ -634,26 +672,14 @@ static int drm_sched_main(void *param)
>>                   DRM_ERROR("fence add callback failed (%d)\n",
>>                         r);
>>               dma_fence_put(fence);
>> -        } else {
>> -            if (s_fence->finished.error < 0)
>> -                drm_sched_expel_job_unlocked(sched_job);
>> +        } else
>>               drm_sched_process_job(NULL, &s_fence->cb);
>> -        }
>>             wake_up(&sched->job_scheduled);
>>       }
>>       return 0;
>>   }
>>   -static void drm_sched_expel_job_unlocked(struct drm_sched_job *s_job)
>> -{
>> -    struct drm_gpu_scheduler *sched = s_job->sched;
>> -
>> -    spin_lock(&sched->job_list_lock);
>> -    list_del_init(&s_job->node);
>> -    spin_unlock(&sched->job_list_lock);
>> -}
>> -
>>   /**
>>    * drm_sched_init - Init a gpu scheduler instance
>>    *
>> diff --git a/drivers/gpu/drm/v3d/v3d_sched.c 
>> b/drivers/gpu/drm/v3d/v3d_sched.c
>> index 445b2ef..f99346a 100644
>> --- a/drivers/gpu/drm/v3d/v3d_sched.c
>> +++ b/drivers/gpu/drm/v3d/v3d_sched.c
>> @@ -178,18 +178,19 @@ v3d_job_timedout(struct drm_sched_job *sched_job)
>>       for (q = 0; q < V3D_MAX_QUEUES; q++) {
>>           struct drm_gpu_scheduler *sched = &v3d->queue[q].sched;
>>   -        kthread_park(sched->thread);
>> -        drm_sched_hw_job_reset(sched, (sched_job->sched == sched ?
>> -                           sched_job : NULL));
>> +        drm_sched_stop(sched, (sched_job->sched == sched ?
>> +                           sched_job : NULL), false);
>>       }
>>         /* get the GPU back into the init state */
>>       v3d_reset(v3d);
>>   +    for (q = 0; q < V3D_MAX_QUEUES; q++)
>> +        drm_sched_resubmit_jobs(sched_job->sched);
>> +
>>       /* Unblock schedulers and restart their jobs. */
>>       for (q = 0; q < V3D_MAX_QUEUES; q++) {
>> -        drm_sched_job_recovery(&v3d->queue[q].sched);
>> -        kthread_unpark(v3d->queue[q].sched.thread);
>> +        drm_sched_start(&v3d->queue[q].sched, false);
>>       }
>>         mutex_unlock(&v3d->reset_lock);
>> diff --git a/include/drm/gpu_scheduler.h b/include/drm/gpu_scheduler.h
>> index 47e1979..c94b592 100644
>> --- a/include/drm/gpu_scheduler.h
>> +++ b/include/drm/gpu_scheduler.h
>> @@ -175,6 +175,7 @@ struct drm_sched_fence *to_drm_sched_fence(struct 
>> dma_fence *f);
>>    *               finished to remove the job from the
>>    *               @drm_gpu_scheduler.ring_mirror_list.
>>    * @node: used to append this struct to the 
>> @drm_gpu_scheduler.ring_mirror_list.
>> + * @finish_node: used in a list to wait on before resetting the 
>> scheduler
>>    * @id: a unique id assigned to each job scheduled on the scheduler.
>>    * @karma: increment on every hang caused by this job. If this 
>> exceeds the hang
>>    *         limit of the scheduler then the job is marked guilty and 
>> will not
>> @@ -193,6 +194,7 @@ struct drm_sched_job {
>>       struct dma_fence_cb        finish_cb;
>>       struct work_struct        finish_work;
>>       struct list_head        node;
>> +    struct list_head        finish_node;
>>       uint64_t            id;
>>       atomic_t            karma;
>>       enum drm_sched_priority        s_priority;
>> @@ -298,9 +300,11 @@ int drm_sched_job_init(struct drm_sched_job *job,
>>                  void *owner);
>>   void drm_sched_job_cleanup(struct drm_sched_job *job);
>>   void drm_sched_wakeup(struct drm_gpu_scheduler *sched);
>> -void drm_sched_hw_job_reset(struct drm_gpu_scheduler *sched,
>> -                struct drm_sched_job *job);
>> -void drm_sched_job_recovery(struct drm_gpu_scheduler *sched);
>> +void drm_sched_stop(struct drm_gpu_scheduler *sched,
>> +            struct drm_sched_job *job,
>> +            bool park_only);
>> +void drm_sched_start(struct drm_gpu_scheduler *sched, bool 
>> unpark_only);
>> +void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched);
>>   bool drm_sched_dependency_optimized(struct dma_fence* fence,
>>                       struct drm_sched_entity *entity);
>>   void drm_sched_fault(struct drm_gpu_scheduler *sched);
>

_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

RE: [PATCH 2/2] drm/sched: Rework HW fence processing.

[Zhou, David(ChunMing)]

> -----Original Message-----
> From: dri-devel <dri-devel-bounces@lists.freedesktop.org> On Behalf Of
> Andrey Grodzovsky
> Sent: Friday, December 07, 2018 1:41 AM
> To: dri-devel@lists.freedesktop.org; amd-gfx@lists.freedesktop.org;
> ckoenig.leichtzumerken@gmail.com; eric@anholt.net;
> etnaviv@lists.freedesktop.org
> Cc: Liu, Monk <Monk.Liu@amd.com>
> Subject: [PATCH 2/2] drm/sched: Rework HW fence processing.
> 
> Expedite job deletion from ring mirror list to the HW fence signal callback
> instead from finish_work, together with waiting for all such fences to signal in
> drm_sched_stop we garantee that already signaled job will not be processed
> twice.
> Remove the sched finish fence callback and just submit finish_work directly
> from the HW fence callback.
> 
> Suggested-by: Christian Koenig <Christian.Koenig@amd.com>
> Signed-off-by: Andrey Grodzovsky <andrey.grodzovsky@amd.com>
> ---
>  drivers/gpu/drm/scheduler/sched_fence.c |  4 +++-
> drivers/gpu/drm/scheduler/sched_main.c  | 39 ++++++++++++++++----------
> -------
>  include/drm/gpu_scheduler.h             | 10 +++++++--
>  3 files changed, 30 insertions(+), 23 deletions(-)
> 
> diff --git a/drivers/gpu/drm/scheduler/sched_fence.c
> b/drivers/gpu/drm/scheduler/sched_fence.c
> index d8d2dff..e62c239 100644
> --- a/drivers/gpu/drm/scheduler/sched_fence.c
> +++ b/drivers/gpu/drm/scheduler/sched_fence.c
> @@ -151,7 +151,8 @@ struct drm_sched_fence
> *to_drm_sched_fence(struct dma_fence *f)
> EXPORT_SYMBOL(to_drm_sched_fence);
> 
>  struct drm_sched_fence *drm_sched_fence_create(struct
> drm_sched_entity *entity,
> -					       void *owner)
> +					       void *owner,
> +					       struct drm_sched_job *s_job)
>  {
>  	struct drm_sched_fence *fence = NULL;
>  	unsigned seq;
> @@ -163,6 +164,7 @@ struct drm_sched_fence
> *drm_sched_fence_create(struct drm_sched_entity *entity,
>  	fence->owner = owner;
>  	fence->sched = entity->rq->sched;
>  	spin_lock_init(&fence->lock);
> +	fence->s_job = s_job;
> 
>  	seq = atomic_inc_return(&entity->fence_seq);
>  	dma_fence_init(&fence->scheduled,
> &drm_sched_fence_ops_scheduled, diff --git
> a/drivers/gpu/drm/scheduler/sched_main.c
> b/drivers/gpu/drm/scheduler/sched_main.c
> index 8fb7f86..2860037 100644
> --- a/drivers/gpu/drm/scheduler/sched_main.c
> +++ b/drivers/gpu/drm/scheduler/sched_main.c
> @@ -284,31 +284,17 @@ static void drm_sched_job_finish(struct
> work_struct *work)
>  	cancel_delayed_work_sync(&sched->work_tdr);
> 
>  	spin_lock_irqsave(&sched->job_list_lock, flags);
> -	/* remove job from ring_mirror_list */
> -	list_del_init(&s_job->node);
> -	/* queue TDR for next job */
>  	drm_sched_start_timeout(sched);
>  	spin_unlock_irqrestore(&sched->job_list_lock, flags);
> 
>  	sched->ops->free_job(s_job);
>  }
> 
> -static void drm_sched_job_finish_cb(struct dma_fence *f,
> -				    struct dma_fence_cb *cb)
> -{
> -	struct drm_sched_job *job = container_of(cb, struct drm_sched_job,
> -						 finish_cb);
> -	schedule_work(&job->finish_work);
> -}
> -
>  static void drm_sched_job_begin(struct drm_sched_job *s_job)  {
>  	struct drm_gpu_scheduler *sched = s_job->sched;
>  	unsigned long flags;
> 
> -	dma_fence_add_callback(&s_job->s_fence->finished, &s_job-
> >finish_cb,
> -			       drm_sched_job_finish_cb);
> -
>  	spin_lock_irqsave(&sched->job_list_lock, flags);
>  	list_add_tail(&s_job->node, &sched->ring_mirror_list);
>  	drm_sched_start_timeout(sched);
> @@ -418,13 +404,17 @@ void drm_sched_start(struct drm_gpu_scheduler
> *sched, bool unpark_only)  {
>  	struct drm_sched_job *s_job, *tmp;
>  	bool found_guilty = false;
> -	unsigned long flags;
>  	int r;
> 
>  	if (unpark_only)
>  		goto unpark;
> 
> -	spin_lock_irqsave(&sched->job_list_lock, flags);
> +	/*
> +	 * Locking the list is not required here as the sched thread is parked
> +	 * so no new jobs are being pushed in to HW and in drm_sched_stop
> we
> +	 * flushed any in flight jobs who didn't signal yet. Also concurrent
> +	 * GPU recovers can't run in parallel.
> +	 */
>  	list_for_each_entry_safe(s_job, tmp, &sched->ring_mirror_list,
> node) {
>  		struct drm_sched_fence *s_fence = s_job->s_fence;
>  		struct dma_fence *fence;
> @@ -453,7 +443,6 @@ void drm_sched_start(struct drm_gpu_scheduler
> *sched, bool unpark_only)
>  	}
> 
>  	drm_sched_start_timeout(sched);
> -	spin_unlock_irqrestore(&sched->job_list_lock, flags);
> 
>  unpark:
>  	kthread_unpark(sched->thread);
> @@ -505,7 +494,7 @@ int drm_sched_job_init(struct drm_sched_job *job,
>  	job->sched = sched;
>  	job->entity = entity;
>  	job->s_priority = entity->rq - sched->sched_rq;
> -	job->s_fence = drm_sched_fence_create(entity, owner);
> +	job->s_fence = drm_sched_fence_create(entity, owner, job);
>  	if (!job->s_fence)
>  		return -ENOMEM;
>  	job->id = atomic64_inc_return(&sched->job_id_count);
> @@ -593,15 +582,25 @@ static void drm_sched_process_job(struct
> dma_fence *f, struct dma_fence_cb *cb)
>  	struct drm_sched_fence *s_fence =
>  		container_of(cb, struct drm_sched_fence, cb);
>  	struct drm_gpu_scheduler *sched = s_fence->sched;
> +	struct drm_sched_job *s_job = s_fence->s_job;


Seems we are back to original, Christian argued s_fence and s_job are with different lifetime , Do their lifetime become same now?

-David

> +	unsigned long flags;
> +
> +	cancel_delayed_work(&sched->work_tdr);
> 
> -	dma_fence_get(&s_fence->finished);
>  	atomic_dec(&sched->hw_rq_count);
>  	atomic_dec(&sched->num_jobs);
> +
> +	spin_lock_irqsave(&sched->job_list_lock, flags);
> +	/* remove job from ring_mirror_list */
> +	list_del_init(&s_job->node);
> +	spin_unlock_irqrestore(&sched->job_list_lock, flags);
> +
>  	drm_sched_fence_finished(s_fence);
> 
>  	trace_drm_sched_process_job(s_fence);
> -	dma_fence_put(&s_fence->finished);
>  	wake_up_interruptible(&sched->wake_up_worker);
> +
> +	schedule_work(&s_job->finish_work);
>  }
> 
>  /**
> diff --git a/include/drm/gpu_scheduler.h b/include/drm/gpu_scheduler.h
> index c94b592..23855c6 100644
> --- a/include/drm/gpu_scheduler.h
> +++ b/include/drm/gpu_scheduler.h
> @@ -115,6 +115,8 @@ struct drm_sched_rq {
>  	struct drm_sched_entity		*current_entity;
>  };
> 
> +struct drm_sched_job;
> +
>  /**
>   * struct drm_sched_fence - fences corresponding to the scheduling of a job.
>   */
> @@ -160,6 +162,9 @@ struct drm_sched_fence {
>           * @owner: job owner for debugging
>           */
>  	void				*owner;
> +
> +	/* Back pointer to owning job */
> +	struct drm_sched_job 		*s_job;
>  };
> 
>  struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f); @@
> -330,8 +335,9 @@ 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);
> 
> -struct drm_sched_fence *drm_sched_fence_create(
> -	struct drm_sched_entity *s_entity, void *owner);
> +struct drm_sched_fence *drm_sched_fence_create(struct
> drm_sched_entity *s_entity,
> +					       void *owner,
> +					       struct drm_sched_job *s_job);
>  void drm_sched_fence_scheduled(struct drm_sched_fence *fence);  void
> drm_sched_fence_finished(struct drm_sched_fence *fence);
> 
> --
> 2.7.4
> 
> _______________________________________________
> dri-devel mailing list
> dri-devel@lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/dri-devel
_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Re: [PATCH 2/2] drm/sched: Rework HW fence processing.

[Christian König]

Am 07.12.18 um 04:18 schrieb Zhou, David(ChunMing):
>
>> -----Original Message-----
>> From: dri-devel <dri-devel-bounces@lists.freedesktop.org> On Behalf Of
>> Andrey Grodzovsky
>> Sent: Friday, December 07, 2018 1:41 AM
>> To: dri-devel@lists.freedesktop.org; amd-gfx@lists.freedesktop.org;
>> ckoenig.leichtzumerken@gmail.com; eric@anholt.net;
>> etnaviv@lists.freedesktop.org
>> Cc: Liu, Monk <Monk.Liu@amd.com>
>> Subject: [PATCH 2/2] drm/sched: Rework HW fence processing.
>>
>> Expedite job deletion from ring mirror list to the HW fence signal callback
>> instead from finish_work, together with waiting for all such fences to signal in
>> drm_sched_stop we garantee that already signaled job will not be processed
>> twice.
>> Remove the sched finish fence callback and just submit finish_work directly
>> from the HW fence callback.
>>
>> Suggested-by: Christian Koenig <Christian.Koenig@amd.com>
>> Signed-off-by: Andrey Grodzovsky <andrey.grodzovsky@amd.com>
>> ---
>>   drivers/gpu/drm/scheduler/sched_fence.c |  4 +++-
>> drivers/gpu/drm/scheduler/sched_main.c  | 39 ++++++++++++++++----------
>> -------
>>   include/drm/gpu_scheduler.h             | 10 +++++++--
>>   3 files changed, 30 insertions(+), 23 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/scheduler/sched_fence.c
>> b/drivers/gpu/drm/scheduler/sched_fence.c
>> index d8d2dff..e62c239 100644
>> --- a/drivers/gpu/drm/scheduler/sched_fence.c
>> +++ b/drivers/gpu/drm/scheduler/sched_fence.c
>> @@ -151,7 +151,8 @@ struct drm_sched_fence
>> *to_drm_sched_fence(struct dma_fence *f)
>> EXPORT_SYMBOL(to_drm_sched_fence);
>>
>>   struct drm_sched_fence *drm_sched_fence_create(struct
>> drm_sched_entity *entity,
>> -					       void *owner)
>> +					       void *owner,
>> +					       struct drm_sched_job *s_job)
>>   {
>>   	struct drm_sched_fence *fence = NULL;
>>   	unsigned seq;
>> @@ -163,6 +164,7 @@ struct drm_sched_fence
>> *drm_sched_fence_create(struct drm_sched_entity *entity,
>>   	fence->owner = owner;
>>   	fence->sched = entity->rq->sched;
>>   	spin_lock_init(&fence->lock);
>> +	fence->s_job = s_job;
>>
>>   	seq = atomic_inc_return(&entity->fence_seq);
>>   	dma_fence_init(&fence->scheduled,
>> &drm_sched_fence_ops_scheduled, diff --git
>> a/drivers/gpu/drm/scheduler/sched_main.c
>> b/drivers/gpu/drm/scheduler/sched_main.c
>> index 8fb7f86..2860037 100644
>> --- a/drivers/gpu/drm/scheduler/sched_main.c
>> +++ b/drivers/gpu/drm/scheduler/sched_main.c
>> @@ -284,31 +284,17 @@ static void drm_sched_job_finish(struct
>> work_struct *work)
>>   	cancel_delayed_work_sync(&sched->work_tdr);
>>
>>   	spin_lock_irqsave(&sched->job_list_lock, flags);
>> -	/* remove job from ring_mirror_list */
>> -	list_del_init(&s_job->node);
>> -	/* queue TDR for next job */
>>   	drm_sched_start_timeout(sched);
>>   	spin_unlock_irqrestore(&sched->job_list_lock, flags);
>>
>>   	sched->ops->free_job(s_job);
>>   }
>>
>> -static void drm_sched_job_finish_cb(struct dma_fence *f,
>> -				    struct dma_fence_cb *cb)
>> -{
>> -	struct drm_sched_job *job = container_of(cb, struct drm_sched_job,
>> -						 finish_cb);
>> -	schedule_work(&job->finish_work);
>> -}
>> -
>>   static void drm_sched_job_begin(struct drm_sched_job *s_job)  {
>>   	struct drm_gpu_scheduler *sched = s_job->sched;
>>   	unsigned long flags;
>>
>> -	dma_fence_add_callback(&s_job->s_fence->finished, &s_job-
>>> finish_cb,
>> -			       drm_sched_job_finish_cb);
>> -
>>   	spin_lock_irqsave(&sched->job_list_lock, flags);
>>   	list_add_tail(&s_job->node, &sched->ring_mirror_list);
>>   	drm_sched_start_timeout(sched);
>> @@ -418,13 +404,17 @@ void drm_sched_start(struct drm_gpu_scheduler
>> *sched, bool unpark_only)  {
>>   	struct drm_sched_job *s_job, *tmp;
>>   	bool found_guilty = false;
>> -	unsigned long flags;
>>   	int r;
>>
>>   	if (unpark_only)
>>   		goto unpark;
>>
>> -	spin_lock_irqsave(&sched->job_list_lock, flags);
>> +	/*
>> +	 * Locking the list is not required here as the sched thread is parked
>> +	 * so no new jobs are being pushed in to HW and in drm_sched_stop
>> we
>> +	 * flushed any in flight jobs who didn't signal yet. Also concurrent
>> +	 * GPU recovers can't run in parallel.
>> +	 */
>>   	list_for_each_entry_safe(s_job, tmp, &sched->ring_mirror_list,
>> node) {
>>   		struct drm_sched_fence *s_fence = s_job->s_fence;
>>   		struct dma_fence *fence;
>> @@ -453,7 +443,6 @@ void drm_sched_start(struct drm_gpu_scheduler
>> *sched, bool unpark_only)
>>   	}
>>
>>   	drm_sched_start_timeout(sched);
>> -	spin_unlock_irqrestore(&sched->job_list_lock, flags);
>>
>>   unpark:
>>   	kthread_unpark(sched->thread);
>> @@ -505,7 +494,7 @@ int drm_sched_job_init(struct drm_sched_job *job,
>>   	job->sched = sched;
>>   	job->entity = entity;
>>   	job->s_priority = entity->rq - sched->sched_rq;
>> -	job->s_fence = drm_sched_fence_create(entity, owner);
>> +	job->s_fence = drm_sched_fence_create(entity, owner, job);
>>   	if (!job->s_fence)
>>   		return -ENOMEM;
>>   	job->id = atomic64_inc_return(&sched->job_id_count);
>> @@ -593,15 +582,25 @@ static void drm_sched_process_job(struct
>> dma_fence *f, struct dma_fence_cb *cb)
>>   	struct drm_sched_fence *s_fence =
>>   		container_of(cb, struct drm_sched_fence, cb);
>>   	struct drm_gpu_scheduler *sched = s_fence->sched;
>> +	struct drm_sched_job *s_job = s_fence->s_job;
>
> Seems we are back to original, Christian argued s_fence and s_job are with different lifetime , Do their lifetime become same now?

No, that still doesn't work. The scheduler fence lives much longer than 
the job, so we would have a dangling pointer here.

The real question is why we actually need that? I mean we just need to 
move the callback structure into the job, don't we?

Christian.

>
> -David
>
>> +	unsigned long flags;
>> +
>> +	cancel_delayed_work(&sched->work_tdr);
>>
>> -	dma_fence_get(&s_fence->finished);
>>   	atomic_dec(&sched->hw_rq_count);
>>   	atomic_dec(&sched->num_jobs);
>> +
>> +	spin_lock_irqsave(&sched->job_list_lock, flags);
>> +	/* remove job from ring_mirror_list */
>> +	list_del_init(&s_job->node);
>> +	spin_unlock_irqrestore(&sched->job_list_lock, flags);
>> +
>>   	drm_sched_fence_finished(s_fence);
>>
>>   	trace_drm_sched_process_job(s_fence);
>> -	dma_fence_put(&s_fence->finished);
>>   	wake_up_interruptible(&sched->wake_up_worker);
>> +
>> +	schedule_work(&s_job->finish_work);
>>   }
>>
>>   /**
>> diff --git a/include/drm/gpu_scheduler.h b/include/drm/gpu_scheduler.h
>> index c94b592..23855c6 100644
>> --- a/include/drm/gpu_scheduler.h
>> +++ b/include/drm/gpu_scheduler.h
>> @@ -115,6 +115,8 @@ struct drm_sched_rq {
>>   	struct drm_sched_entity		*current_entity;
>>   };
>>
>> +struct drm_sched_job;
>> +
>>   /**
>>    * struct drm_sched_fence - fences corresponding to the scheduling of a job.
>>    */
>> @@ -160,6 +162,9 @@ struct drm_sched_fence {
>>            * @owner: job owner for debugging
>>            */
>>   	void				*owner;
>> +
>> +	/* Back pointer to owning job */
>> +	struct drm_sched_job 		*s_job;
>>   };
>>
>>   struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f); @@
>> -330,8 +335,9 @@ 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);
>>
>> -struct drm_sched_fence *drm_sched_fence_create(
>> -	struct drm_sched_entity *s_entity, void *owner);
>> +struct drm_sched_fence *drm_sched_fence_create(struct
>> drm_sched_entity *s_entity,
>> +					       void *owner,
>> +					       struct drm_sched_job *s_job);
>>   void drm_sched_fence_scheduled(struct drm_sched_fence *fence);  void
>> drm_sched_fence_finished(struct drm_sched_fence *fence);
>>
>> --
>> 2.7.4
>>
>> _______________________________________________
>> dri-devel mailing list
>> dri-devel@lists.freedesktop.org
>> https://lists.freedesktop.org/mailman/listinfo/dri-devel

_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Re: [PATCH 2/2] drm/sched: Rework HW fence processing.

[Grodzovsky, Andrey]

On 12/07/2018 03:19 AM, Christian König wrote:
> Am 07.12.18 um 04:18 schrieb Zhou, David(ChunMing):
>>
>>> -----Original Message-----
>>> From: dri-devel <dri-devel-bounces@lists.freedesktop.org> On Behalf Of
>>> Andrey Grodzovsky
>>> Sent: Friday, December 07, 2018 1:41 AM
>>> To: dri-devel@lists.freedesktop.org; amd-gfx@lists.freedesktop.org;
>>> ckoenig.leichtzumerken@gmail.com; eric@anholt.net;
>>> etnaviv@lists.freedesktop.org
>>> Cc: Liu, Monk <Monk.Liu@amd.com>
>>> Subject: [PATCH 2/2] drm/sched: Rework HW fence processing.
>>>
>>> Expedite job deletion from ring mirror list to the HW fence signal 
>>> callback
>>> instead from finish_work, together with waiting for all such fences 
>>> to signal in
>>> drm_sched_stop we garantee that already signaled job will not be 
>>> processed
>>> twice.
>>> Remove the sched finish fence callback and just submit finish_work 
>>> directly
>>> from the HW fence callback.
>>>
>>> Suggested-by: Christian Koenig <Christian.Koenig@amd.com>
>>> Signed-off-by: Andrey Grodzovsky <andrey.grodzovsky@amd.com>
>>> ---
>>>   drivers/gpu/drm/scheduler/sched_fence.c |  4 +++-
>>> drivers/gpu/drm/scheduler/sched_main.c  | 39 ++++++++++++++++----------
>>> -------
>>>   include/drm/gpu_scheduler.h             | 10 +++++++--
>>>   3 files changed, 30 insertions(+), 23 deletions(-)
>>>
>>> diff --git a/drivers/gpu/drm/scheduler/sched_fence.c
>>> b/drivers/gpu/drm/scheduler/sched_fence.c
>>> index d8d2dff..e62c239 100644
>>> --- a/drivers/gpu/drm/scheduler/sched_fence.c
>>> +++ b/drivers/gpu/drm/scheduler/sched_fence.c
>>> @@ -151,7 +151,8 @@ struct drm_sched_fence
>>> *to_drm_sched_fence(struct dma_fence *f)
>>> EXPORT_SYMBOL(to_drm_sched_fence);
>>>
>>>   struct drm_sched_fence *drm_sched_fence_create(struct
>>> drm_sched_entity *entity,
>>> -                           void *owner)
>>> +                           void *owner,
>>> +                           struct drm_sched_job *s_job)
>>>   {
>>>       struct drm_sched_fence *fence = NULL;
>>>       unsigned seq;
>>> @@ -163,6 +164,7 @@ struct drm_sched_fence
>>> *drm_sched_fence_create(struct drm_sched_entity *entity,
>>>       fence->owner = owner;
>>>       fence->sched = entity->rq->sched;
>>>       spin_lock_init(&fence->lock);
>>> +    fence->s_job = s_job;
>>>
>>>       seq = atomic_inc_return(&entity->fence_seq);
>>>       dma_fence_init(&fence->scheduled,
>>> &drm_sched_fence_ops_scheduled, diff --git
>>> a/drivers/gpu/drm/scheduler/sched_main.c
>>> b/drivers/gpu/drm/scheduler/sched_main.c
>>> index 8fb7f86..2860037 100644
>>> --- a/drivers/gpu/drm/scheduler/sched_main.c
>>> +++ b/drivers/gpu/drm/scheduler/sched_main.c
>>> @@ -284,31 +284,17 @@ static void drm_sched_job_finish(struct
>>> work_struct *work)
>>>       cancel_delayed_work_sync(&sched->work_tdr);
>>>
>>>       spin_lock_irqsave(&sched->job_list_lock, flags);
>>> -    /* remove job from ring_mirror_list */
>>> -    list_del_init(&s_job->node);
>>> -    /* queue TDR for next job */
>>>       drm_sched_start_timeout(sched);
>>>       spin_unlock_irqrestore(&sched->job_list_lock, flags);
>>>
>>>       sched->ops->free_job(s_job);
>>>   }
>>>
>>> -static void drm_sched_job_finish_cb(struct dma_fence *f,
>>> -                    struct dma_fence_cb *cb)
>>> -{
>>> -    struct drm_sched_job *job = container_of(cb, struct drm_sched_job,
>>> -                         finish_cb);
>>> -    schedule_work(&job->finish_work);
>>> -}
>>> -
>>>   static void drm_sched_job_begin(struct drm_sched_job *s_job)  {
>>>       struct drm_gpu_scheduler *sched = s_job->sched;
>>>       unsigned long flags;
>>>
>>> - dma_fence_add_callback(&s_job->s_fence->finished, &s_job-
>>>> finish_cb,
>>> -                   drm_sched_job_finish_cb);
>>> -
>>>       spin_lock_irqsave(&sched->job_list_lock, flags);
>>>       list_add_tail(&s_job->node, &sched->ring_mirror_list);
>>>       drm_sched_start_timeout(sched);
>>> @@ -418,13 +404,17 @@ void drm_sched_start(struct drm_gpu_scheduler
>>> *sched, bool unpark_only)  {
>>>       struct drm_sched_job *s_job, *tmp;
>>>       bool found_guilty = false;
>>> -    unsigned long flags;
>>>       int r;
>>>
>>>       if (unpark_only)
>>>           goto unpark;
>>>
>>> -    spin_lock_irqsave(&sched->job_list_lock, flags);
>>> +    /*
>>> +     * Locking the list is not required here as the sched thread is 
>>> parked
>>> +     * so no new jobs are being pushed in to HW and in drm_sched_stop
>>> we
>>> +     * flushed any in flight jobs who didn't signal yet. Also 
>>> concurrent
>>> +     * GPU recovers can't run in parallel.
>>> +     */
>>>       list_for_each_entry_safe(s_job, tmp, &sched->ring_mirror_list,
>>> node) {
>>>           struct drm_sched_fence *s_fence = s_job->s_fence;
>>>           struct dma_fence *fence;
>>> @@ -453,7 +443,6 @@ void drm_sched_start(struct drm_gpu_scheduler
>>> *sched, bool unpark_only)
>>>       }
>>>
>>>       drm_sched_start_timeout(sched);
>>> -    spin_unlock_irqrestore(&sched->job_list_lock, flags);
>>>
>>>   unpark:
>>>       kthread_unpark(sched->thread);
>>> @@ -505,7 +494,7 @@ int drm_sched_job_init(struct drm_sched_job *job,
>>>       job->sched = sched;
>>>       job->entity = entity;
>>>       job->s_priority = entity->rq - sched->sched_rq;
>>> -    job->s_fence = drm_sched_fence_create(entity, owner);
>>> +    job->s_fence = drm_sched_fence_create(entity, owner, job);
>>>       if (!job->s_fence)
>>>           return -ENOMEM;
>>>       job->id = atomic64_inc_return(&sched->job_id_count);
>>> @@ -593,15 +582,25 @@ static void drm_sched_process_job(struct
>>> dma_fence *f, struct dma_fence_cb *cb)
>>>       struct drm_sched_fence *s_fence =
>>>           container_of(cb, struct drm_sched_fence, cb);
>>>       struct drm_gpu_scheduler *sched = s_fence->sched;
>>> +    struct drm_sched_job *s_job = s_fence->s_job;
>>
>> Seems we are back to original, Christian argued s_fence and s_job are 
>> with different lifetime , Do their lifetime become same now?
>
> No, that still doesn't work. The scheduler fence lives much longer 
> than the job, so we would have a dangling pointer here.
>
> The real question is why we actually need that? I mean we just need to 
> move the callback structure into the job, don't we?
>
> Christian.

So let's say I move the .cb from drm_sched_fence to drm_sched_job, and 
there it's ok to reference
job->s_fence because s_fence life as at least as the job, right ?

Andrey

>
>>
>> -David
>>
>>> +    unsigned long flags;
>>> +
>>> +    cancel_delayed_work(&sched->work_tdr);
>>>
>>> -    dma_fence_get(&s_fence->finished);
>>>       atomic_dec(&sched->hw_rq_count);
>>>       atomic_dec(&sched->num_jobs);
>>> +
>>> +    spin_lock_irqsave(&sched->job_list_lock, flags);
>>> +    /* remove job from ring_mirror_list */
>>> +    list_del_init(&s_job->node);
>>> +    spin_unlock_irqrestore(&sched->job_list_lock, flags);
>>> +
>>>       drm_sched_fence_finished(s_fence);
>>>
>>>       trace_drm_sched_process_job(s_fence);
>>> -    dma_fence_put(&s_fence->finished);
>>>       wake_up_interruptible(&sched->wake_up_worker);
>>> +
>>> +    schedule_work(&s_job->finish_work);
>>>   }
>>>
>>>   /**
>>> diff --git a/include/drm/gpu_scheduler.h b/include/drm/gpu_scheduler.h
>>> index c94b592..23855c6 100644
>>> --- a/include/drm/gpu_scheduler.h
>>> +++ b/include/drm/gpu_scheduler.h
>>> @@ -115,6 +115,8 @@ struct drm_sched_rq {
>>>       struct drm_sched_entity        *current_entity;
>>>   };
>>>
>>> +struct drm_sched_job;
>>> +
>>>   /**
>>>    * struct drm_sched_fence - fences corresponding to the scheduling 
>>> of a job.
>>>    */
>>> @@ -160,6 +162,9 @@ struct drm_sched_fence {
>>>            * @owner: job owner for debugging
>>>            */
>>>       void                *owner;
>>> +
>>> +    /* Back pointer to owning job */
>>> +    struct drm_sched_job         *s_job;
>>>   };
>>>
>>>   struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f); @@
>>> -330,8 +335,9 @@ 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);
>>>
>>> -struct drm_sched_fence *drm_sched_fence_create(
>>> -    struct drm_sched_entity *s_entity, void *owner);
>>> +struct drm_sched_fence *drm_sched_fence_create(struct
>>> drm_sched_entity *s_entity,
>>> +                           void *owner,
>>> +                           struct drm_sched_job *s_job);
>>>   void drm_sched_fence_scheduled(struct drm_sched_fence *fence);  void
>>> drm_sched_fence_finished(struct drm_sched_fence *fence);
>>>
>>> -- 
>>> 2.7.4
>>>
>>> _______________________________________________
>>> dri-devel mailing list
>>> dri-devel@lists.freedesktop.org
>>> https://lists.freedesktop.org/mailman/listinfo/dri-devel
>

_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Re: [PATCH 2/2] drm/sched: Rework HW fence processing.

[Koenig, Christian]

Am 07.12.18 um 16:22 schrieb Grodzovsky, Andrey:
>
> On 12/07/2018 03:19 AM, Christian König wrote:
>> Am 07.12.18 um 04:18 schrieb Zhou, David(ChunMing):
>>>> -----Original Message-----
>>>> From: dri-devel <dri-devel-bounces@lists.freedesktop.org> On Behalf Of
>>>> Andrey Grodzovsky
>>>> Sent: Friday, December 07, 2018 1:41 AM
>>>> To: dri-devel@lists.freedesktop.org; amd-gfx@lists.freedesktop.org;
>>>> ckoenig.leichtzumerken@gmail.com; eric@anholt.net;
>>>> etnaviv@lists.freedesktop.org
>>>> Cc: Liu, Monk <Monk.Liu@amd.com>
>>>> Subject: [PATCH 2/2] drm/sched: Rework HW fence processing.
>>>>
>>>> Expedite job deletion from ring mirror list to the HW fence signal
>>>> callback
>>>> instead from finish_work, together with waiting for all such fences
>>>> to signal in
>>>> drm_sched_stop we garantee that already signaled job will not be
>>>> processed
>>>> twice.
>>>> Remove the sched finish fence callback and just submit finish_work
>>>> directly
>>>> from the HW fence callback.
>>>>
>>>> Suggested-by: Christian Koenig <Christian.Koenig@amd.com>
>>>> Signed-off-by: Andrey Grodzovsky <andrey.grodzovsky@amd.com>
>>>> ---
>>>>    drivers/gpu/drm/scheduler/sched_fence.c |  4 +++-
>>>> drivers/gpu/drm/scheduler/sched_main.c  | 39 ++++++++++++++++----------
>>>> -------
>>>>    include/drm/gpu_scheduler.h             | 10 +++++++--
>>>>    3 files changed, 30 insertions(+), 23 deletions(-)
>>>>
>>>> diff --git a/drivers/gpu/drm/scheduler/sched_fence.c
>>>> b/drivers/gpu/drm/scheduler/sched_fence.c
>>>> index d8d2dff..e62c239 100644
>>>> --- a/drivers/gpu/drm/scheduler/sched_fence.c
>>>> +++ b/drivers/gpu/drm/scheduler/sched_fence.c
>>>> @@ -151,7 +151,8 @@ struct drm_sched_fence
>>>> *to_drm_sched_fence(struct dma_fence *f)
>>>> EXPORT_SYMBOL(to_drm_sched_fence);
>>>>
>>>>    struct drm_sched_fence *drm_sched_fence_create(struct
>>>> drm_sched_entity *entity,
>>>> -                           void *owner)
>>>> +                           void *owner,
>>>> +                           struct drm_sched_job *s_job)
>>>>    {
>>>>        struct drm_sched_fence *fence = NULL;
>>>>        unsigned seq;
>>>> @@ -163,6 +164,7 @@ struct drm_sched_fence
>>>> *drm_sched_fence_create(struct drm_sched_entity *entity,
>>>>        fence->owner = owner;
>>>>        fence->sched = entity->rq->sched;
>>>>        spin_lock_init(&fence->lock);
>>>> +    fence->s_job = s_job;
>>>>
>>>>        seq = atomic_inc_return(&entity->fence_seq);
>>>>        dma_fence_init(&fence->scheduled,
>>>> &drm_sched_fence_ops_scheduled, diff --git
>>>> a/drivers/gpu/drm/scheduler/sched_main.c
>>>> b/drivers/gpu/drm/scheduler/sched_main.c
>>>> index 8fb7f86..2860037 100644
>>>> --- a/drivers/gpu/drm/scheduler/sched_main.c
>>>> +++ b/drivers/gpu/drm/scheduler/sched_main.c
>>>> @@ -284,31 +284,17 @@ static void drm_sched_job_finish(struct
>>>> work_struct *work)
>>>>        cancel_delayed_work_sync(&sched->work_tdr);
>>>>
>>>>        spin_lock_irqsave(&sched->job_list_lock, flags);
>>>> -    /* remove job from ring_mirror_list */
>>>> -    list_del_init(&s_job->node);
>>>> -    /* queue TDR for next job */
>>>>        drm_sched_start_timeout(sched);
>>>>        spin_unlock_irqrestore(&sched->job_list_lock, flags);
>>>>
>>>>        sched->ops->free_job(s_job);
>>>>    }
>>>>
>>>> -static void drm_sched_job_finish_cb(struct dma_fence *f,
>>>> -                    struct dma_fence_cb *cb)
>>>> -{
>>>> -    struct drm_sched_job *job = container_of(cb, struct drm_sched_job,
>>>> -                         finish_cb);
>>>> -    schedule_work(&job->finish_work);
>>>> -}
>>>> -
>>>>    static void drm_sched_job_begin(struct drm_sched_job *s_job)  {
>>>>        struct drm_gpu_scheduler *sched = s_job->sched;
>>>>        unsigned long flags;
>>>>
>>>> - dma_fence_add_callback(&s_job->s_fence->finished, &s_job-
>>>>> finish_cb,
>>>> -                   drm_sched_job_finish_cb);
>>>> -
>>>>        spin_lock_irqsave(&sched->job_list_lock, flags);
>>>>        list_add_tail(&s_job->node, &sched->ring_mirror_list);
>>>>        drm_sched_start_timeout(sched);
>>>> @@ -418,13 +404,17 @@ void drm_sched_start(struct drm_gpu_scheduler
>>>> *sched, bool unpark_only)  {
>>>>        struct drm_sched_job *s_job, *tmp;
>>>>        bool found_guilty = false;
>>>> -    unsigned long flags;
>>>>        int r;
>>>>
>>>>        if (unpark_only)
>>>>            goto unpark;
>>>>
>>>> -    spin_lock_irqsave(&sched->job_list_lock, flags);
>>>> +    /*
>>>> +     * Locking the list is not required here as the sched thread is
>>>> parked
>>>> +     * so no new jobs are being pushed in to HW and in drm_sched_stop
>>>> we
>>>> +     * flushed any in flight jobs who didn't signal yet. Also
>>>> concurrent
>>>> +     * GPU recovers can't run in parallel.
>>>> +     */
>>>>        list_for_each_entry_safe(s_job, tmp, &sched->ring_mirror_list,
>>>> node) {
>>>>            struct drm_sched_fence *s_fence = s_job->s_fence;
>>>>            struct dma_fence *fence;
>>>> @@ -453,7 +443,6 @@ void drm_sched_start(struct drm_gpu_scheduler
>>>> *sched, bool unpark_only)
>>>>        }
>>>>
>>>>        drm_sched_start_timeout(sched);
>>>> -    spin_unlock_irqrestore(&sched->job_list_lock, flags);
>>>>
>>>>    unpark:
>>>>        kthread_unpark(sched->thread);
>>>> @@ -505,7 +494,7 @@ int drm_sched_job_init(struct drm_sched_job *job,
>>>>        job->sched = sched;
>>>>        job->entity = entity;
>>>>        job->s_priority = entity->rq - sched->sched_rq;
>>>> -    job->s_fence = drm_sched_fence_create(entity, owner);
>>>> +    job->s_fence = drm_sched_fence_create(entity, owner, job);
>>>>        if (!job->s_fence)
>>>>            return -ENOMEM;
>>>>        job->id = atomic64_inc_return(&sched->job_id_count);
>>>> @@ -593,15 +582,25 @@ static void drm_sched_process_job(struct
>>>> dma_fence *f, struct dma_fence_cb *cb)
>>>>        struct drm_sched_fence *s_fence =
>>>>            container_of(cb, struct drm_sched_fence, cb);
>>>>        struct drm_gpu_scheduler *sched = s_fence->sched;
>>>> +    struct drm_sched_job *s_job = s_fence->s_job;
>>> Seems we are back to original, Christian argued s_fence and s_job are
>>> with different lifetime , Do their lifetime become same now?
>> No, that still doesn't work. The scheduler fence lives much longer
>> than the job, so we would have a dangling pointer here.
>>
>> The real question is why we actually need that? I mean we just need to
>> move the callback structure into the job, don't we?
>>
>> Christian.
> So let's say I move the .cb from drm_sched_fence to drm_sched_job, and
> there it's ok to reference
> job->s_fence because s_fence life as at least as the job, right ?

Exactly. Well, we could actually make the live of s_fence a bit shorter 
and drop it as soon as it is signaled.

Christian.

>
> Andrey
>
>>> -David
>>>
>>>> +    unsigned long flags;
>>>> +
>>>> +    cancel_delayed_work(&sched->work_tdr);
>>>>
>>>> -    dma_fence_get(&s_fence->finished);
>>>>        atomic_dec(&sched->hw_rq_count);
>>>>        atomic_dec(&sched->num_jobs);
>>>> +
>>>> +    spin_lock_irqsave(&sched->job_list_lock, flags);
>>>> +    /* remove job from ring_mirror_list */
>>>> +    list_del_init(&s_job->node);
>>>> +    spin_unlock_irqrestore(&sched->job_list_lock, flags);
>>>> +
>>>>        drm_sched_fence_finished(s_fence);
>>>>
>>>>        trace_drm_sched_process_job(s_fence);
>>>> -    dma_fence_put(&s_fence->finished);
>>>>        wake_up_interruptible(&sched->wake_up_worker);
>>>> +
>>>> +    schedule_work(&s_job->finish_work);
>>>>    }
>>>>
>>>>    /**
>>>> diff --git a/include/drm/gpu_scheduler.h b/include/drm/gpu_scheduler.h
>>>> index c94b592..23855c6 100644
>>>> --- a/include/drm/gpu_scheduler.h
>>>> +++ b/include/drm/gpu_scheduler.h
>>>> @@ -115,6 +115,8 @@ struct drm_sched_rq {
>>>>        struct drm_sched_entity        *current_entity;
>>>>    };
>>>>
>>>> +struct drm_sched_job;
>>>> +
>>>>    /**
>>>>     * struct drm_sched_fence - fences corresponding to the scheduling
>>>> of a job.
>>>>     */
>>>> @@ -160,6 +162,9 @@ struct drm_sched_fence {
>>>>             * @owner: job owner for debugging
>>>>             */
>>>>        void                *owner;
>>>> +
>>>> +    /* Back pointer to owning job */
>>>> +    struct drm_sched_job         *s_job;
>>>>    };
>>>>
>>>>    struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f); @@
>>>> -330,8 +335,9 @@ 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);
>>>>
>>>> -struct drm_sched_fence *drm_sched_fence_create(
>>>> -    struct drm_sched_entity *s_entity, void *owner);
>>>> +struct drm_sched_fence *drm_sched_fence_create(struct
>>>> drm_sched_entity *s_entity,
>>>> +                           void *owner,
>>>> +                           struct drm_sched_job *s_job);
>>>>    void drm_sched_fence_scheduled(struct drm_sched_fence *fence);  void
>>>> drm_sched_fence_finished(struct drm_sched_fence *fence);
>>>>
>>>> -- 
>>>> 2.7.4
>>>>
>>>> _______________________________________________
>>>> dri-devel mailing list
>>>> dri-devel@lists.freedesktop.org
>>>> https://lists.freedesktop.org/mailman/listinfo/dri-devel

_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel