drm/ttm: remove TTM_BO_PRIV_FLAG_MOVING

Instead of using the flag just remember the fence of the last move operation.

This avoids waiting for command submissions pipelined after the move, but
before accessing the BO with the CPU again.

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drivers/gpu/drm/ttm/ttm_bo.c

@@ -149,6 +149,7 @@ static void ttm_bo_release_list(struct kref *list_kref)
 
 	ttm_tt_destroy(bo->ttm);
 	atomic_dec(&bo->glob->bo_count);
+	fence_put(bo->moving);
 	if (bo->resv == &bo->ttm_resv)
 		reservation_object_fini(&bo->ttm_resv);
 	mutex_destroy(&bo->wu_mutex);
@@ -1138,7 +1139,7 @@ int ttm_bo_init(struct ttm_bo_device *bdev,
 	bo->mem.page_alignment = page_alignment;
 	bo->mem.bus.io_reserved_vm = false;
 	bo->mem.bus.io_reserved_count = 0;
-	bo->priv_flags = 0;
+	bo->moving = NULL;
 	bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED);
 	bo->persistent_swap_storage = persistent_swap_storage;
 	bo->acc_size = acc_size;
@@ -1585,7 +1586,6 @@ int ttm_bo_wait(struct ttm_buffer_object *bo,
 		return -EBUSY;
 
 	reservation_object_add_excl_fence(resv, NULL);
-	clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
 	return 0;
 }
 EXPORT_SYMBOL(ttm_bo_wait);

drivers/gpu/drm/ttm/ttm_bo_util.c

@@ -465,6 +465,7 @@ static int ttm_buffer_object_transfer(struct ttm_buffer_object *bo,
 	INIT_LIST_HEAD(&fbo->lru);
 	INIT_LIST_HEAD(&fbo->swap);
 	INIT_LIST_HEAD(&fbo->io_reserve_lru);
+	fbo->moving = NULL;
 	drm_vma_node_reset(&fbo->vma_node);
 	atomic_set(&fbo->cpu_writers, 0);
 
@@ -665,7 +666,8 @@ int ttm_bo_move_accel_cleanup(struct ttm_buffer_object *bo,
 		 * operation has completed.
 		 */
 
-		set_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
+		fence_put(bo->moving);
+		bo->moving = fence_get(fence);
 
 		ret = ttm_buffer_object_transfer(bo, &ghost_obj);
 		if (ret)

drivers/gpu/drm/ttm/ttm_bo_vm.c

@@ -48,15 +48,14 @@ static int ttm_bo_vm_fault_idle(struct ttm_buffer_object *bo,
 {
 	int ret = 0;
 
-	if (likely(!test_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags)))
+	if (likely(!bo->moving))
 		goto out_unlock;
 
 	/*
 	 * Quick non-stalling check for idle.
 	 */
-	ret = ttm_bo_wait(bo, false, true);
-	if (likely(ret == 0))
-		goto out_unlock;
+	if (fence_is_signaled(bo->moving))
+		goto out_clear;
 
 	/*
 	 * If possible, avoid waiting for GPU with mmap_sem
@@ -68,17 +67,23 @@ static int ttm_bo_vm_fault_idle(struct ttm_buffer_object *bo,
 			goto out_unlock;
 
 		up_read(&vma->vm_mm->mmap_sem);
-		(void) ttm_bo_wait(bo, true, false);
+		(void) fence_wait(bo->moving, true);
 		goto out_unlock;
 	}
 
 	/*
 	 * Ordinary wait.
 	 */
-	ret = ttm_bo_wait(bo, true, false);
-	if (unlikely(ret != 0))
+	ret = fence_wait(bo->moving, true);
+	if (unlikely(ret != 0)) {
 		ret = (ret != -ERESTARTSYS) ? VM_FAULT_SIGBUS :
 			VM_FAULT_NOPAGE;
+		goto out_unlock;
+	}
+
+out_clear:
+	fence_put(bo->moving);
+	bo->moving = NULL;
 
 out_unlock:
 	return ret;

include/drm/ttm/ttm_bo_api.h

@@ -173,7 +173,7 @@ struct ttm_tt;
  * @lru: List head for the lru list.
  * @ddestroy: List head for the delayed destroy list.
  * @swap: List head for swap LRU list.
- * @priv_flags: Flags describing buffer object internal state.
+ * @moving: Fence set when BO is moving
  * @vma_node: Address space manager node.
  * @offset: The current GPU offset, which can have different meanings
  * depending on the memory type. For SYSTEM type memory, it should be 0.
@@ -239,7 +239,7 @@ struct ttm_buffer_object {
 	 * Members protected by a bo reservation.
 	 */
 
-	unsigned long priv_flags;
+	struct fence *moving;
 
 	struct drm_vma_offset_node vma_node;
 

include/drm/ttm/ttm_bo_driver.h

@@ -503,9 +503,6 @@ struct ttm_bo_global {
 
 #define TTM_NUM_MEM_TYPES 8
 
-#define TTM_BO_PRIV_FLAG_MOVING  0	/* Buffer object is moving and needs
-					   idling before CPU mapping */
-#define TTM_BO_PRIV_FLAG_MAX 1
 /**
  * struct ttm_bo_device - Buffer object driver device-specific data.
  *