drm/amdgpu: explicitely sync to VM updates v2

Allows us to reduce the overhead while syncing to fences a bit.

v2: also drop adev parameter from the functions

Signed-off-by: Christian König <[e-mail]>
Reviewed-by: Felix Kuehling <[e-mail]>
Signed-off-by: Alex Deucher <[e-mail]>

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

···
 	if (ret)
 		return ret;
 
-	return amdgpu_sync_fence(NULL, sync, vm->last_update, false);
+	return amdgpu_sync_fence(sync, vm->last_update, false);
 }
 
 static uint64_t get_pte_flags(struct amdgpu_device *adev, struct kgd_mem *mem)
···
 
 	amdgpu_vm_clear_freed(adev, vm, &bo_va->last_pt_update);
 
-	amdgpu_sync_fence(NULL, sync, bo_va->last_pt_update, false);
+	amdgpu_sync_fence(sync, bo_va->last_pt_update, false);
 
 	return 0;
 }
···
 		return ret;
 	}
 
-	return amdgpu_sync_fence(NULL, sync, bo_va->last_pt_update, false);
+	return amdgpu_sync_fence(sync, bo_va->last_pt_update, false);
 }
 
 static int map_bo_to_gpuvm(struct amdgpu_device *adev,
···
 			pr_debug("Memory eviction: Validate BOs failed. Try again\n");
 			goto validate_map_fail;
 		}
-		ret = amdgpu_sync_fence(NULL, &sync_obj, bo->tbo.moving, false);
+		ret = amdgpu_sync_fence(&sync_obj, bo->tbo.moving, false);
 		if (ret) {
 			pr_debug("Memory eviction: Sync BO fence failed. Try again\n");
 			goto validate_map_fail;

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

···
 	if (r)
 		return r;
 
-	r = amdgpu_sync_fence(adev, &p->job->sync,
-			      fpriv->prt_va->last_pt_update, false);
+	r = amdgpu_sync_vm_fence(&p->job->sync, fpriv->prt_va->last_pt_update);
 	if (r)
 		return r;
 
 	if (amdgpu_mcbp || amdgpu_sriov_vf(adev)) {
-		struct dma_fence *f;
-
 		bo_va = fpriv->csa_va;
 		BUG_ON(!bo_va);
 		r = amdgpu_vm_bo_update(adev, bo_va, false);
 		if (r)
 			return r;
 
-		f = bo_va->last_pt_update;
-		r = amdgpu_sync_fence(adev, &p->job->sync, f, false);
+		r = amdgpu_sync_vm_fence(&p->job->sync, bo_va->last_pt_update);
 		if (r)
 			return r;
 	}
 
 	amdgpu_bo_list_for_each_entry(e, p->bo_list) {
-		struct dma_fence *f;
-
 		/* ignore duplicates */
 		bo = ttm_to_amdgpu_bo(e->tv.bo);
 		if (!bo)
···
 		if (r)
 			return r;
 
-		f = bo_va->last_pt_update;
-		r = amdgpu_sync_fence(adev, &p->job->sync, f, false);
+		r = amdgpu_sync_vm_fence(&p->job->sync, bo_va->last_pt_update);
 		if (r)
 			return r;
 	}
···
 	if (r)
 		return r;
 
-	r = amdgpu_sync_fence(adev, &p->job->sync, vm->last_update, false);
+	r = amdgpu_sync_vm_fence(&p->job->sync, vm->last_update);
 	if (r)
 		return r;
 
···
 			dma_fence_put(old);
 		}
 
-		r = amdgpu_sync_fence(p->adev, &p->job->sync, fence, true);
+		r = amdgpu_sync_fence(&p->job->sync, fence, true);
 		dma_fence_put(fence);
 		if (r)
 			return r;
···
 		return r;
 	}
 
-	r = amdgpu_sync_fence(p->adev, &p->job->sync, fence, true);
+	r = amdgpu_sync_fence(&p->job->sync, fence, true);
 	dma_fence_put(fence);
 
 	return r;

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

···
 	int r;
 
 	if (ring->vmid_wait && !dma_fence_is_signaled(ring->vmid_wait))
-		return amdgpu_sync_fence(adev, sync, ring->vmid_wait, false);
+		return amdgpu_sync_fence(sync, ring->vmid_wait, false);
 
 	fences = kmalloc_array(sizeof(void *), id_mgr->num_ids, GFP_KERNEL);
 	if (!fences)
···
 			return -ENOMEM;
 		}
 
-		r = amdgpu_sync_fence(adev, sync, &array->base, false);
+		r = amdgpu_sync_fence(sync, &array->base, false);
 		dma_fence_put(ring->vmid_wait);
 		ring->vmid_wait = &array->base;
 		return r;
···
 		tmp = amdgpu_sync_peek_fence(&(*id)->active, ring);
 		if (tmp) {
 			*id = NULL;
-			r = amdgpu_sync_fence(adev, sync, tmp, false);
+			r = amdgpu_sync_fence(sync, tmp, false);
 			return r;
 		}
 		needs_flush = true;
···
 	/* Good we can use this VMID. Remember this submission as
 	* user of the VMID.
 	*/
-	r = amdgpu_sync_fence(ring->adev, &(*id)->active, fence, false);
+	r = amdgpu_sync_fence(&(*id)->active, fence, false);
 	if (r)
 		return r;
 
···
 		/* Good, we can use this VMID. Remember this submission as
 		 * user of the VMID.
 		 */
-		r = amdgpu_sync_fence(ring->adev, &(*id)->active, fence, false);
+		r = amdgpu_sync_fence(&(*id)->active, fence, false);
 		if (r)
 			return r;
 
···
 			id = idle;
 
 			/* Remember this submission as user of the VMID */
-			r = amdgpu_sync_fence(ring->adev, &id->active,
-					      fence, false);
+			r = amdgpu_sync_fence(&id->active, fence, false);
 			if (r)
 				goto error;
 

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

···
 	fence = amdgpu_sync_get_fence(&job->sync, &explicit);
 	if (fence && explicit) {
 		if (drm_sched_dependency_optimized(fence, s_entity)) {
-			r = amdgpu_sync_fence(ring->adev, &job->sched_sync,
-					      fence, false);
+			r = amdgpu_sync_fence(&job->sched_sync, fence, false);
 			if (r)
 				DRM_ERROR("Error adding fence (%d)\n", r);
 		}

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

···
  * Tries to add the fence to an existing hash entry. Returns true when an entry
  * was found, false otherwise.
  */
-static bool amdgpu_sync_add_later(struct amdgpu_sync *sync, struct dma_fence *f, bool explicit)
+static bool amdgpu_sync_add_later(struct amdgpu_sync *sync, struct dma_fence *f,
+				  bool explicit)
 {
 	struct amdgpu_sync_entry *e;
 
···
  * amdgpu_sync_fence - remember to sync to this fence
  *
  * @sync: sync object to add fence to
- * @fence: fence to sync to
+ * @f: fence to sync to
+ * @explicit: if this is an explicit dependency
  *
+ * Add the fence to the sync object.
  */
-int amdgpu_sync_fence(struct amdgpu_device *adev, struct amdgpu_sync *sync,
-		      struct dma_fence *f, bool explicit)
+int amdgpu_sync_fence(struct amdgpu_sync *sync, struct dma_fence *f,
+		      bool explicit)
 {
 	struct amdgpu_sync_entry *e;
 
 	if (!f)
 		return 0;
-	if (amdgpu_sync_same_dev(adev, f) &&
-	    amdgpu_sync_get_owner(f) == AMDGPU_FENCE_OWNER_VM)
-		amdgpu_sync_keep_later(&sync->last_vm_update, f);
 
 	if (amdgpu_sync_add_later(sync, f, explicit))
 		return 0;
···
 	hash_add(sync->fences, &e->node, f->context);
 	e->fence = dma_fence_get(f);
 	return 0;
+}
+
+/**
+ * amdgpu_sync_vm_fence - remember to sync to this VM fence
+ *
+ * @adev: amdgpu device
+ * @sync: sync object to add fence to
+ * @fence: the VM fence to add
+ *
+ * Add the fence to the sync object and remember it as VM update.
+ */
+int amdgpu_sync_vm_fence(struct amdgpu_sync *sync, struct dma_fence *fence)
+{
+	if (!fence)
+		return 0;
+
+	amdgpu_sync_keep_later(&sync->last_vm_update, fence);
+	return amdgpu_sync_fence(sync, fence, false);
 }
 
 /**
···
 
 	/* always sync to the exclusive fence */
 	f = dma_resv_get_excl(resv);
-	r = amdgpu_sync_fence(adev, sync, f, false);
+	r = amdgpu_sync_fence(sync, f, false);
 
 	flist = dma_resv_get_list(resv);
 	if (!flist || r)
···
 				continue;
 		}
 
-		r = amdgpu_sync_fence(adev, sync, f, false);
+		r = amdgpu_sync_fence(sync, f, false);
 		if (r)
 			break;
 	}
···
 	hash_for_each_safe(source->fences, i, tmp, e, node) {
 		f = e->fence;
 		if (!dma_fence_is_signaled(f)) {
-			r = amdgpu_sync_fence(NULL, clone, f, e->explicit);
+			r = amdgpu_sync_fence(clone, f, e->explicit);
 			if (r)
 				return r;
 		} else {

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

···
 };
 
 void amdgpu_sync_create(struct amdgpu_sync *sync);
-int amdgpu_sync_fence(struct amdgpu_device *adev, struct amdgpu_sync *sync,
-		      struct dma_fence *f, bool explicit);
+int amdgpu_sync_fence(struct amdgpu_sync *sync, struct dma_fence *f,
+		      bool explicit);
+int amdgpu_sync_vm_fence(struct amdgpu_sync *sync, struct dma_fence *fence);
 int amdgpu_sync_resv(struct amdgpu_device *adev,
 		     struct amdgpu_sync *sync,
 		     struct dma_resv *resv,
···
 		     bool explicit_sync);
 struct dma_fence *amdgpu_sync_peek_fence(struct amdgpu_sync *sync,
 				     struct amdgpu_ring *ring);
-struct dma_fence *amdgpu_sync_get_fence(struct amdgpu_sync *sync, bool *explicit);
+struct dma_fence *amdgpu_sync_get_fence(struct amdgpu_sync *sync,
+					bool *explicit);
 int amdgpu_sync_clone(struct amdgpu_sync *source, struct amdgpu_sync *clone);
 int amdgpu_sync_wait(struct amdgpu_sync *sync, bool intr);
 void amdgpu_sync_free(struct amdgpu_sync *sync);

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

···
 	p->num_dw_left = ndw;
 
 	/* Wait for moves to be completed */
-	r = amdgpu_sync_fence(p->adev, &p->job->sync, exclusive, false);
+	r = amdgpu_sync_fence(&p->job->sync, exclusive, false);
 	if (r)
 		return r;