drm/amdgpu: make sure the fence is emitted before ring to get it.

Signed-off-by: Chunming Zhou <david1.zhou@amd.com>
Acked-by: Christian K?nig <christian.koenig@amd.com>
Reviewed-by: Jammy Zhou <Jammy.Zhou@amd.com>

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

@@ -81,6 +81,7 @@ extern int amdgpu_vm_size;
 extern int amdgpu_vm_block_size;
 extern int amdgpu_enable_scheduler;
 
+#define AMDGPU_WAIT_IDLE_TIMEOUT_IN_MS	        3000
 #define AMDGPU_MAX_USEC_TIMEOUT			100000	/* 100 ms */
 #define AMDGPU_FENCE_JIFFIES_TIMEOUT		(HZ / 2)
 /* AMDGPU_IB_POOL_SIZE must be a power of 2 */
@@ -1239,6 +1240,7 @@ struct amdgpu_cs_parser {
 	/* user fence */
 	struct amdgpu_user_fence uf;
 
+	struct amdgpu_ring *ring;
 	struct mutex job_lock;
 	struct work_struct job_work;
 	int (*prepare_job)(struct amdgpu_cs_parser *sched_job);

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

@@ -915,7 +915,7 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
 				goto out;
 		} else
 			parser->prepare_job = amdgpu_cs_parser_prepare_job;
-
+		parser->ring = ring;
 		parser->run_job = amdgpu_cs_parser_run_job;
 		parser->free_job = amdgpu_cs_parser_free_job;
 		amd_sched_push_job(ring->scheduler,
@@ -965,24 +965,16 @@ int amdgpu_cs_wait_ioctl(struct drm_device *dev, void *data,
 	ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id);
 	if (ctx == NULL)
 		return -EINVAL;
-	if (amdgpu_enable_scheduler) {
-		r = amd_sched_wait_ts(&ctx->rings[ring->idx].c_entity,
-				      wait->in.handle, true, timeout);
-		if (r)
-			return r;
-		r = 1;
-	} else {
-		fence = amdgpu_ctx_get_fence(ctx, ring, wait->in.handle);
-		if (IS_ERR(fence))
-			r = PTR_ERR(fence);
 
-		else if (fence) {
-			r = fence_wait_timeout(fence, true, timeout);
-			fence_put(fence);
+	fence = amdgpu_ctx_get_fence(ctx, ring, wait->in.handle);
+	if (IS_ERR(fence))
+		r = PTR_ERR(fence);
+	else if (fence) {
+		r = fence_wait_timeout(fence, true, timeout);
+		fence_put(fence);
+	} else
+		r = 1;
 
-		} else
-			r = 1;
-	}
 	amdgpu_ctx_put(ctx);
 	if (r < 0)
 		return r;

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

@@ -261,6 +261,16 @@ struct fence *amdgpu_ctx_get_fence(struct amdgpu_ctx *ctx,
 	struct amdgpu_ctx_ring *cring = & ctx->rings[ring->idx];
 	struct fence *fence;
 	uint64_t queued_seq;
+	int r;
+
+	if (amdgpu_enable_scheduler) {
+		r = amd_sched_wait_emit(&cring->c_entity,
+					seq,
+					true,
+					AMDGPU_WAIT_IDLE_TIMEOUT_IN_MS);
+		if (r)
+			return NULL;
+	}
 
 	spin_lock(&ctx->ring_lock);
 	if (amdgpu_enable_scheduler)

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

@@ -56,12 +56,15 @@ static void amdgpu_sched_run_job(struct amd_gpu_scheduler *sched,
 			       sched_job->filp);
 	if (r)
 		goto err;
-
 	if (sched_job->run_job) {
 		r = sched_job->run_job(sched_job);
 		if (r)
 			goto err;
 	}
+	atomic64_set(&c_entity->last_emitted_v_seq,
+		     sched_job->uf.sequence);
+	wake_up_all(&c_entity->wait_emit);
+
 	mutex_unlock(&sched_job->job_lock);
 	return;
 err: