drm/amdgpu: add scheduler dependency callback v2

This way the scheduler doesn't wait in it's work thread any more.

v2: fix race conditions
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Jammy Zhou <Jammy.Zhou@amd.com>

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

@@ -717,6 +717,7 @@ int amdgpu_sync_resv(struct amdgpu_device *adev,
 		     void *owner);
 int amdgpu_sync_rings(struct amdgpu_sync *sync,
 		      struct amdgpu_ring *ring);
+struct fence *amdgpu_sync_get_fence(struct amdgpu_sync *sync);
 int amdgpu_sync_wait(struct amdgpu_sync *sync);
 void amdgpu_sync_free(struct amdgpu_device *adev, struct amdgpu_sync *sync,
 		      struct fence *fence);

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

@@ -27,6 +27,12 @@
 #include <drm/drmP.h>
 #include "amdgpu.h"
 
+static struct fence *amdgpu_sched_dependency(struct amd_sched_job *job)
+{
+	struct amdgpu_job *sched_job = (struct amdgpu_job *)job;
+	return amdgpu_sync_get_fence(&sched_job->ibs->sync);
+}
+
 static struct fence *amdgpu_sched_run_job(struct amd_sched_job *job)
 {
 	struct amdgpu_job *sched_job;
@@ -75,6 +81,7 @@ static void amdgpu_sched_process_job(struct amd_sched_job *job)
 }
 
 struct amd_sched_backend_ops amdgpu_sched_ops = {
+	.dependency = amdgpu_sched_dependency,
 	.run_job = amdgpu_sched_run_job,
 	.process_job = amdgpu_sched_process_job
 };

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

@@ -202,6 +202,28 @@ int amdgpu_sync_resv(struct amdgpu_device *adev,
 	return r;
 }
 
+struct fence *amdgpu_sync_get_fence(struct amdgpu_sync *sync)
+{
+	struct amdgpu_sync_entry *e;
+	struct hlist_node *tmp;
+	struct fence *f;
+	int i;
+
+	hash_for_each_safe(sync->fences, i, tmp, e, node) {
+
+		f = e->fence;
+
+		hash_del(&e->node);
+		kfree(e);
+
+		if (!fence_is_signaled(f))
+			return f;
+
+		fence_put(f);
+	}
+	return NULL;
+}
+
 int amdgpu_sync_wait(struct amdgpu_sync *sync)
 {
 	struct amdgpu_sync_entry *e;

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

@@ -192,14 +192,36 @@ void amd_sched_entity_fini(struct amd_gpu_scheduler *sched,
 	kfifo_free(&entity->job_queue);
 }
 
+static void amd_sched_entity_wakeup(struct fence *f, struct fence_cb *cb)
+{
+	struct amd_sched_entity *entity =
+		container_of(cb, struct amd_sched_entity, cb);
+	entity->dependency = NULL;
+	fence_put(f);
+	amd_sched_wakeup(entity->scheduler);
+}
+
 static struct amd_sched_job *
 amd_sched_entity_pop_job(struct amd_sched_entity *entity)
 {
+	struct amd_gpu_scheduler *sched = entity->scheduler;
 	struct amd_sched_job *job;
 
+	if (ACCESS_ONCE(entity->dependency))
+		return NULL;
+
 	if (!kfifo_out_peek(&entity->job_queue, &job, sizeof(job)))
 		return NULL;
 
+	while ((entity->dependency = sched->ops->dependency(job))) {
+
+		if (fence_add_callback(entity->dependency, &entity->cb,
+				       amd_sched_entity_wakeup))
+			fence_put(entity->dependency);
+		else
+			return NULL;
+	}
+
 	return job;
 }
 

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

@@ -45,6 +45,8 @@ struct amd_sched_entity {
 	spinlock_t			queue_lock;
 	struct amd_gpu_scheduler	*scheduler;
 	uint64_t                        fence_context;
+	struct fence			*dependency;
+	struct fence_cb			cb;
 };
 
 /**
@@ -89,6 +91,7 @@ static inline struct amd_sched_fence *to_amd_sched_fence(struct fence *f)
  * these functions should be implemented in driver side
 */
 struct amd_sched_backend_ops {
+	struct fence *(*dependency)(struct amd_sched_job *job);
 	struct fence *(*run_job)(struct amd_sched_job *job);
 	void (*process_job)(struct amd_sched_job *job);
 };