drm/amdgpu: generalize the scheduler fence

Make it two events, one for the job being scheduled and one when it is finished.
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Christian König <christian.koenig@amd.com>
Acked-by: Chunming Zhou <david1.zhou@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drivers/gpu/drm/amd/amdgpu/amdgpu_job.c

@@ -85,7 +85,7 @@ static void amdgpu_job_free_resources(struct amdgpu_job *job)
 	unsigned i;
 
 	/* use sched fence if available */
-	f = job->base.s_fence ? &job->base.s_fence->base : job->fence;
+	f = job->base.s_fence ? &job->base.s_fence->finished : job->fence;
 
 	for (i = 0; i < job->num_ibs; ++i)
 		amdgpu_ib_free(job->adev, &job->ibs[i], f);
@@ -143,7 +143,7 @@ static struct fence *amdgpu_job_dependency(struct amd_sched_job *sched_job)
 		int r;
 
 		r = amdgpu_vm_grab_id(vm, ring, &job->sync,
-				      &job->base.s_fence->base,
+				      &job->base.s_fence->finished,
 				      &job->vm_id, &job->vm_pd_addr);
 		if (r)
 			DRM_ERROR("Error getting VM ID (%d)\n", r);

drivers/gpu/drm/amd/amdgpu/amdgpu_trace.h

@@ -102,7 +102,7 @@ TRACE_EVENT(amdgpu_cs_ioctl,
 			   __entry->adev = job->adev;
 			   __entry->sched_job = &job->base;
 			   __entry->ib = job->ibs;
-			   __entry->fence = &job->base.s_fence->base;
+			   __entry->fence = &job->base.s_fence->finished;
 			   __entry->ring_name = job->ring->name;
 			   __entry->num_ibs = job->num_ibs;
 			   ),
@@ -127,7 +127,7 @@ TRACE_EVENT(amdgpu_sched_run_job,
 			   __entry->adev = job->adev;
 			   __entry->sched_job = &job->base;
 			   __entry->ib = job->ibs;
-			   __entry->fence = &job->base.s_fence->base;
+			   __entry->fence = &job->base.s_fence->finished;
 			   __entry->ring_name = job->ring->name;
 			   __entry->num_ibs = job->num_ibs;
 			   ),

drivers/gpu/drm/amd/scheduler/gpu_sched_trace.h

@@ -26,7 +26,7 @@ TRACE_EVENT(amd_sched_job,
 	    TP_fast_assign(
 			   __entry->entity = sched_job->s_entity;
 			   __entry->sched_job = sched_job;
-			   __entry->fence = &sched_job->s_fence->base;
+			   __entry->fence = &sched_job->s_fence->finished;
 			   __entry->name = sched_job->sched->name;
 			   __entry->job_count = kfifo_len(
 				   &sched_job->s_entity->job_queue) / sizeof(sched_job);
@@ -46,7 +46,7 @@ TRACE_EVENT(amd_sched_process_job,
 		    ),
 
 	    TP_fast_assign(
-		    __entry->fence = &fence->base;
+		    __entry->fence = &fence->finished;
 		    ),
 	    TP_printk("fence=%p signaled", __entry->fence)
 );

drivers/gpu/drm/amd/scheduler/gpu_scheduler.c

@@ -140,7 +140,7 @@ int amd_sched_entity_init(struct amd_gpu_scheduler *sched,
 		return r;
 
 	atomic_set(&entity->fence_seq, 0);
-	entity->fence_context = fence_context_alloc(1);
+	entity->fence_context = fence_context_alloc(2);
 
 	return 0;
 }
@@ -251,17 +251,21 @@ static bool amd_sched_entity_add_dependency_cb(struct amd_sched_entity *entity)
 
 	s_fence = to_amd_sched_fence(fence);
 	if (s_fence && s_fence->sched == sched) {
-		/* Fence is from the same scheduler */
-		if (test_bit(AMD_SCHED_FENCE_SCHEDULED_BIT, &fence->flags)) {
-			/* Ignore it when it is already scheduled */
-			fence_put(entity->dependency);
-			return false;
-		}
 
-		/* Wait for fence to be scheduled */
-		entity->cb.func = amd_sched_entity_clear_dep;
-		list_add_tail(&entity->cb.node, &s_fence->scheduled_cb);
-		return true;
+		/*
+		 * Fence is from the same scheduler, only need to wait for
+		 * it to be scheduled
+		 */
+		fence = fence_get(&s_fence->scheduled);
+		fence_put(entity->dependency);
+		entity->dependency = fence;
+		if (!fence_add_callback(fence, &entity->cb,
+					amd_sched_entity_clear_dep))
+			return true;
+
+		/* Ignore it when it is already scheduled */
+		fence_put(fence);
+		return false;
 	}
 
 	if (!fence_add_callback(entity->dependency, &entity->cb,
@@ -389,7 +393,7 @@ void amd_sched_entity_push_job(struct amd_sched_job *sched_job)
 	struct amd_sched_entity *entity = sched_job->s_entity;
 
 	trace_amd_sched_job(sched_job);
-	fence_add_callback(&sched_job->s_fence->base, &sched_job->finish_cb,
+	fence_add_callback(&sched_job->s_fence->finished, &sched_job->finish_cb,
 			   amd_sched_job_finish_cb);
 	wait_event(entity->sched->job_scheduled,
 		   amd_sched_entity_in(sched_job));
@@ -412,7 +416,7 @@ int amd_sched_job_init(struct amd_sched_job *job,
 	INIT_DELAYED_WORK(&job->work_tdr, amd_sched_job_timedout);
 
 	if (fence)
-		*fence = &job->s_fence->base;
+		*fence = &job->s_fence->finished;
 	return 0;
 }
 
@@ -463,10 +467,10 @@ static void amd_sched_process_job(struct fence *f, struct fence_cb *cb)
 	struct amd_gpu_scheduler *sched = s_fence->sched;
 
 	atomic_dec(&sched->hw_rq_count);
-	amd_sched_fence_signal(s_fence);
+	amd_sched_fence_finished(s_fence);
 
 	trace_amd_sched_process_job(s_fence);
-	fence_put(&s_fence->base);
+	fence_put(&s_fence->finished);
 	wake_up_interruptible(&sched->wake_up_worker);
 }
 

drivers/gpu/drm/amd/scheduler/gpu_scheduler.h

@@ -27,8 +27,6 @@
 #include <linux/kfifo.h>
 #include <linux/fence.h>
 
-#define AMD_SCHED_FENCE_SCHEDULED_BIT	FENCE_FLAG_USER_BITS
-
 struct amd_gpu_scheduler;
 struct amd_sched_rq;
 
@@ -68,9 +66,9 @@ struct amd_sched_rq {
 };
 
 struct amd_sched_fence {
-	struct fence                    base;
+	struct fence                    scheduled;
+	struct fence                    finished;
 	struct fence_cb                 cb;
-	struct list_head		scheduled_cb;
 	struct amd_gpu_scheduler	*sched;
 	spinlock_t			lock;
 	void                            *owner;
@@ -86,14 +84,15 @@ struct amd_sched_job {
 	struct delayed_work		work_tdr;
 };
 
-extern const struct fence_ops amd_sched_fence_ops;
+extern const struct fence_ops amd_sched_fence_ops_scheduled;
+extern const struct fence_ops amd_sched_fence_ops_finished;
 static inline struct amd_sched_fence *to_amd_sched_fence(struct fence *f)
 {
-	struct amd_sched_fence *__f = container_of(f, struct amd_sched_fence,
-						   base);
+	if (f->ops == &amd_sched_fence_ops_scheduled)
+		return container_of(f, struct amd_sched_fence, scheduled);
 
-	if (__f->base.ops == &amd_sched_fence_ops)
-		return __f;
+	if (f->ops == &amd_sched_fence_ops_finished)
+		return container_of(f, struct amd_sched_fence, finished);
 
 	return NULL;
 }
@@ -148,7 +147,7 @@ void amd_sched_entity_push_job(struct amd_sched_job *sched_job);
 struct amd_sched_fence *amd_sched_fence_create(
 	struct amd_sched_entity *s_entity, void *owner);
 void amd_sched_fence_scheduled(struct amd_sched_fence *fence);
-void amd_sched_fence_signal(struct amd_sched_fence *fence);
+void amd_sched_fence_finished(struct amd_sched_fence *fence);
 int amd_sched_job_init(struct amd_sched_job *job,
 		       struct amd_gpu_scheduler *sched,
 		       struct amd_sched_entity *entity,

drivers/gpu/drm/amd/scheduler/sched_fence.c

@@ -37,36 +37,37 @@ struct amd_sched_fence *amd_sched_fence_create(struct amd_sched_entity *entity,
 	if (fence == NULL)
 		return NULL;
 
-	INIT_LIST_HEAD(&fence->scheduled_cb);
 	fence->owner = owner;
 	fence->sched = entity->sched;
 	spin_lock_init(&fence->lock);
 
 	seq = atomic_inc_return(&entity->fence_seq);
-	fence_init(&fence->base, &amd_sched_fence_ops, &fence->lock,
-		   entity->fence_context, seq);
+	fence_init(&fence->scheduled, &amd_sched_fence_ops_scheduled,
+		   &fence->lock, entity->fence_context, seq);
+	fence_init(&fence->finished, &amd_sched_fence_ops_finished,
+		   &fence->lock, entity->fence_context + 1, seq);
 
 	return fence;
 }
 
-void amd_sched_fence_signal(struct amd_sched_fence *fence)
+void amd_sched_fence_scheduled(struct amd_sched_fence *fence)
 {
-	int ret = fence_signal(&fence->base);
+	int ret = fence_signal(&fence->scheduled);
+
 	if (!ret)
-		FENCE_TRACE(&fence->base, "signaled from irq context\n");
+		FENCE_TRACE(&fence->scheduled, "signaled from irq context\n");
 	else
-		FENCE_TRACE(&fence->base, "was already signaled\n");
+		FENCE_TRACE(&fence->scheduled, "was already signaled\n");
 }
 
-void amd_sched_fence_scheduled(struct amd_sched_fence *s_fence)
+void amd_sched_fence_finished(struct amd_sched_fence *fence)
 {
-	struct fence_cb *cur, *tmp;
+	int ret = fence_signal(&fence->finished);
 
-	set_bit(AMD_SCHED_FENCE_SCHEDULED_BIT, &s_fence->base.flags);
-	list_for_each_entry_safe(cur, tmp, &s_fence->scheduled_cb, node) {
-		list_del_init(&cur->node);
-		cur->func(&s_fence->base, cur);
-	}
+	if (!ret)
+		FENCE_TRACE(&fence->finished, "signaled from irq context\n");
+	else
+		FENCE_TRACE(&fence->finished, "was already signaled\n");
 }
 
 static const char *amd_sched_fence_get_driver_name(struct fence *fence)
@@ -96,6 +97,7 @@ static void amd_sched_fence_free(struct rcu_head *rcu)
 {
 	struct fence *f = container_of(rcu, struct fence, rcu);
 	struct amd_sched_fence *fence = to_amd_sched_fence(f);
+
 	kmem_cache_free(sched_fence_slab, fence);
 }
 
@@ -107,16 +109,41 @@ static void amd_sched_fence_free(struct rcu_head *rcu)
  * This function is called when the reference count becomes zero.
  * It just RCU schedules freeing up the fence.
  */
-static void amd_sched_fence_release(struct fence *f)
+static void amd_sched_fence_release_scheduled(struct fence *f)
+{
+	struct amd_sched_fence *fence = to_amd_sched_fence(f);
+
+	call_rcu(&fence->finished.rcu, amd_sched_fence_free);
+}
+
+/**
+ * amd_sched_fence_release_scheduled - drop extra reference
+ *
+ * @f: fence
+ *
+ * Drop the extra reference from the scheduled fence to the base fence.
+ */
+static void amd_sched_fence_release_finished(struct fence *f)
 {
-	call_rcu(&f->rcu, amd_sched_fence_free);
+	struct amd_sched_fence *fence = to_amd_sched_fence(f);
+
+	fence_put(&fence->scheduled);
 }
 
-const struct fence_ops amd_sched_fence_ops = {
+const struct fence_ops amd_sched_fence_ops_scheduled = {
+	.get_driver_name = amd_sched_fence_get_driver_name,
+	.get_timeline_name = amd_sched_fence_get_timeline_name,
+	.enable_signaling = amd_sched_fence_enable_signaling,
+	.signaled = NULL,
+	.wait = fence_default_wait,
+	.release = amd_sched_fence_release_scheduled,
+};
+
+const struct fence_ops amd_sched_fence_ops_finished = {
 	.get_driver_name = amd_sched_fence_get_driver_name,
 	.get_timeline_name = amd_sched_fence_get_timeline_name,
 	.enable_signaling = amd_sched_fence_enable_signaling,
 	.signaled = NULL,
 	.wait = fence_default_wait,
-	.release = amd_sched_fence_release,
+	.release = amd_sched_fence_release_finished,
 };