drm/amdgpu: fix last_vm_update fence is not effetive for sched fence

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

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

@@ -705,7 +705,7 @@ struct amdgpu_sync {
 	struct amdgpu_semaphore *semaphores[AMDGPU_NUM_SYNCS];
 	struct amdgpu_fence	*sync_to[AMDGPU_MAX_RINGS];
 	DECLARE_HASHTABLE(fences, 4);
-	struct amdgpu_fence	*last_vm_update;
+	struct fence	        *last_vm_update;
 };
 
 void amdgpu_sync_create(struct amdgpu_sync *sync);
@@ -963,7 +963,7 @@ struct amdgpu_vm_id {
 	unsigned		id;
 	uint64_t		pd_gpu_addr;
 	/* last flushed PD/PT update */
-	struct amdgpu_fence	*flushed_updates;
+	struct fence	        *flushed_updates;
 	/* last use of vmid */
 	struct amdgpu_fence	*last_id_use;
 };
@@ -2349,7 +2349,7 @@ int amdgpu_vm_grab_id(struct amdgpu_vm *vm, struct amdgpu_ring *ring,
 		      struct amdgpu_sync *sync);
 void amdgpu_vm_flush(struct amdgpu_ring *ring,
 		     struct amdgpu_vm *vm,
-		     struct amdgpu_fence *updates);
+		     struct fence *updates);
 void amdgpu_vm_fence(struct amdgpu_device *adev,
 		     struct amdgpu_vm *vm,
 		     struct amdgpu_fence *fence);

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

@@ -119,5 +119,6 @@ int amdgpu_sched_ib_submit_kernel_helper(struct amdgpu_device *adev,
 			return r;
 		*f = fence_get(&ibs[num_ibs - 1].fence->base);
 	}
+
 	return 0;
 }

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

@@ -58,6 +58,29 @@ void amdgpu_sync_create(struct amdgpu_sync *sync)
 	sync->last_vm_update = NULL;
 }
 
+static bool amdgpu_sync_same_dev(struct amdgpu_device *adev, struct fence *f)
+{
+	struct amdgpu_fence *a_fence = to_amdgpu_fence(f);
+	struct amd_sched_fence *s_fence = to_amd_sched_fence(f);
+
+	if (a_fence)
+		return a_fence->ring->adev == adev;
+	if (s_fence)
+		return (struct amdgpu_device *)s_fence->scheduler->priv == adev;
+	return false;
+}
+
+static bool amdgpu_sync_test_owner(struct fence *f, void *owner)
+{
+	struct amdgpu_fence *a_fence = to_amdgpu_fence(f);
+	struct amd_sched_fence *s_fence = to_amd_sched_fence(f);
+	if (s_fence)
+		return s_fence->owner == owner;
+	if (a_fence)
+		return a_fence->owner == owner;
+	return false;
+}
+
 /**
  * amdgpu_sync_fence - remember to sync to this fence
  *
@@ -71,10 +94,23 @@ int amdgpu_sync_fence(struct amdgpu_device *adev, struct amdgpu_sync *sync,
 	struct amdgpu_sync_entry *e;
 	struct amdgpu_fence *fence;
 	struct amdgpu_fence *other;
+	struct fence *tmp, *later;
 
 	if (!f)
 		return 0;
 
+	if (amdgpu_sync_same_dev(adev, f) &&
+	    amdgpu_sync_test_owner(f, AMDGPU_FENCE_OWNER_VM)) {
+		if (sync->last_vm_update) {
+			tmp = sync->last_vm_update;
+			BUG_ON(f->context != tmp->context);
+			later = (f->seqno - tmp->seqno <= INT_MAX) ? f : tmp;
+			sync->last_vm_update = fence_get(later);
+			fence_put(tmp);
+		} else
+			sync->last_vm_update = fence_get(f);
+	}
+
 	fence = to_amdgpu_fence(f);
 	if (!fence || fence->ring->adev != adev) {
 		hash_for_each_possible(sync->fences, e, node, f->context) {
@@ -103,13 +139,6 @@ int amdgpu_sync_fence(struct amdgpu_device *adev, struct amdgpu_sync *sync,
 		amdgpu_fence_later(fence, other));
 	amdgpu_fence_unref(&other);
 
-	if (fence->owner == AMDGPU_FENCE_OWNER_VM) {
-		other = sync->last_vm_update;
-		sync->last_vm_update = amdgpu_fence_ref(
-			amdgpu_fence_later(fence, other));
-		amdgpu_fence_unref(&other);
-	}
-
 	return 0;
 }
 
@@ -296,5 +325,5 @@ void amdgpu_sync_free(struct amdgpu_device *adev,
 	for (i = 0; i < AMDGPU_MAX_RINGS; ++i)
 		amdgpu_fence_unref(&sync->sync_to[i]);
 
-	amdgpu_fence_unref(&sync->last_vm_update);
+	fence_put(sync->last_vm_update);
 }

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

@@ -200,19 +200,29 @@ int amdgpu_vm_grab_id(struct amdgpu_vm *vm, struct amdgpu_ring *ring,
  */
 void amdgpu_vm_flush(struct amdgpu_ring *ring,
 		     struct amdgpu_vm *vm,
-		     struct amdgpu_fence *updates)
+		     struct fence *updates)
 {
 	uint64_t pd_addr = amdgpu_bo_gpu_offset(vm->page_directory);
 	struct amdgpu_vm_id *vm_id = &vm->ids[ring->idx];
-	struct amdgpu_fence *flushed_updates = vm_id->flushed_updates;
+	struct fence *flushed_updates = vm_id->flushed_updates;
+	bool is_earlier = false;
+
+	if (flushed_updates && updates) {
+		BUG_ON(flushed_updates->context != updates->context);
+		is_earlier = (updates->seqno - flushed_updates->seqno <=
+			      INT_MAX) ? true : false;
+	}
 
 	if (pd_addr != vm_id->pd_gpu_addr || !flushed_updates ||
-	    (updates && amdgpu_fence_is_earlier(flushed_updates, updates))) {
+	    is_earlier) {
 
 		trace_amdgpu_vm_flush(pd_addr, ring->idx, vm_id->id);
-		vm_id->flushed_updates = amdgpu_fence_ref(
-			amdgpu_fence_later(flushed_updates, updates));
-		amdgpu_fence_unref(&flushed_updates);
+		if (is_earlier) {
+			vm_id->flushed_updates = fence_get(updates);
+			fence_put(flushed_updates);
+		}
+		if (!flushed_updates)
+			vm_id->flushed_updates = fence_get(updates);
 		vm_id->pd_gpu_addr = pd_addr;
 		amdgpu_ring_emit_vm_flush(ring, vm_id->id, vm_id->pd_gpu_addr);
 	}
@@ -1347,7 +1357,7 @@ void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
 	fence_put(vm->page_directory_fence);
 
 	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
-		amdgpu_fence_unref(&vm->ids[i].flushed_updates);
+		fence_put(vm->ids[i].flushed_updates);
 		amdgpu_fence_unref(&vm->ids[i].last_id_use);
 	}