drm/amdgpu: move explicit sync check into the CS

This moves the memory allocation out of the critical code path.
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Luben Tuikov <luben.tuikov@amd.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20221014084641.128280-8-christian.koenig@amd.com

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

@@ -449,8 +449,19 @@ static int amdgpu_syncobj_lookup_and_add(struct amdgpu_cs_parser *p,
 	}
 
 	r = amdgpu_sync_fence(&p->gang_leader->sync, fence);
-	dma_fence_put(fence);
+	if (r)
+		goto error;
+
+	/*
+	 * When we have an explicit dependency it might be necessary to insert a
+	 * pipeline sync to make sure that all caches etc are flushed and the
+	 * next job actually sees the results from the previous one.
+	 */
+	if (fence->context == p->gang_leader->base.entity->fence_context)
+		r = amdgpu_sync_fence(&p->gang_leader->explicit_sync, fence);
 
+error:
+	dma_fence_put(fence);
 	return r;
 }
 

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

@@ -182,7 +182,7 @@ int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs,
 
 	need_ctx_switch = ring->current_ctx != fence_ctx;
 	if (ring->funcs->emit_pipeline_sync && job &&
-	    ((tmp = amdgpu_sync_get_fence(&job->sched_sync)) ||
+	    ((tmp = amdgpu_sync_get_fence(&job->explicit_sync)) ||
 	     (amdgpu_sriov_vf(adev) && need_ctx_switch) ||
 	     amdgpu_vm_need_pipeline_sync(ring, job))) {
 		need_pipe_sync = true;

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

@@ -108,7 +108,7 @@ int amdgpu_job_alloc(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 	(*job)->vm = vm;
 
 	amdgpu_sync_create(&(*job)->sync);
-	amdgpu_sync_create(&(*job)->sched_sync);
+	amdgpu_sync_create(&(*job)->explicit_sync);
 	(*job)->vram_lost_counter = atomic_read(&adev->vram_lost_counter);
 	(*job)->vm_pd_addr = AMDGPU_BO_INVALID_OFFSET;
 
@@ -176,7 +176,7 @@ static void amdgpu_job_free_cb(struct drm_sched_job *s_job)
 	drm_sched_job_cleanup(s_job);
 
 	amdgpu_sync_free(&job->sync);
-	amdgpu_sync_free(&job->sched_sync);
+	amdgpu_sync_free(&job->explicit_sync);
 
 	dma_fence_put(&job->hw_fence);
 }
@@ -204,7 +204,7 @@ void amdgpu_job_free(struct amdgpu_job *job)
 
 	amdgpu_job_free_resources(job);
 	amdgpu_sync_free(&job->sync);
-	amdgpu_sync_free(&job->sched_sync);
+	amdgpu_sync_free(&job->explicit_sync);
 	if (job->gang_submit != &job->base.s_fence->scheduled)
 		dma_fence_put(job->gang_submit);
 
@@ -251,12 +251,6 @@ amdgpu_job_dependency(struct drm_sched_job *sched_job,
 	int r;
 
 	fence = amdgpu_sync_get_fence(&job->sync);
-	if (fence && drm_sched_dependency_optimized(fence, s_entity)) {
-		r = amdgpu_sync_fence(&job->sched_sync, fence);
-		if (r)
-			DRM_ERROR("Error adding fence (%d)\n", r);
-	}
-
 	while (!fence && job->vm && !job->vmid) {
 		r = amdgpu_vmid_grab(job->vm, ring, job, &fence);
 		if (r)

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

@@ -48,7 +48,7 @@ struct amdgpu_job {
 	struct drm_sched_job    base;
 	struct amdgpu_vm	*vm;
 	struct amdgpu_sync	sync;
-	struct amdgpu_sync	sched_sync;
+	struct amdgpu_sync	explicit_sync;
 	struct dma_fence	hw_fence;
 	struct dma_fence	*gang_submit;
 	uint32_t		preamble_status;