drm/ttm/radeon/nouveau: Kill the bo lock in favour of a bo device fence_lock

The bo lock used only to protect the bo sync object members, and since it
is a per bo lock, fencing a buffer list will see a lot of locks and unlocks.
Replace it with a per-device lock that protects the sync object members on
*all* bos. Reading and setting these members will always be very quick, so
the risc of heavy lock contention is microscopic. Note that waiting for
sync objects will always take place outside of this lock.

The bo device fence lock will eventually be replaced with a seqlock /
rcu mechanism so we can determine that a bo is idle under a
rcu / read seqlock.

However this change will allow us to batch fencing and unreserving of
buffers with a minimal amount of locking.

Signed-off-by: Thomas Hellstrom <[e-mail]>
Reviewed-by: Jerome Glisse <[e-mail]>
Signed-off-by: Dave Airlie <[e-mail]>

drivers/gpu/drm/nouveau/nouveau_gem.c

···
 		if (likely(fence)) {
 			struct nouveau_fence *prev_fence;
 
-			spin_lock(&nvbo->bo.lock);
+			spin_lock(&nvbo->bo.bdev->fence_lock);
 			prev_fence = nvbo->bo.sync_obj;
 			nvbo->bo.sync_obj = nouveau_fence_ref(fence);
-			spin_unlock(&nvbo->bo.lock);
+			spin_unlock(&nvbo->bo.bdev->fence_lock);
 			nouveau_fence_unref((void *)&prev_fence);
 		}
 
···
 				data |= r->vor;
 		}
 
-		spin_lock(&nvbo->bo.lock);
+		spin_lock(&nvbo->bo.bdev->fence_lock);
 		ret = ttm_bo_wait(&nvbo->bo, false, false, false);
-		spin_unlock(&nvbo->bo.lock);
+		spin_unlock(&nvbo->bo.bdev->fence_lock);
 		if (ret) {
 			NV_ERROR(dev, "reloc wait_idle failed: %d\n", ret);
 			break;
···
 	}
 
 	if (req->flags & NOUVEAU_GEM_CPU_PREP_NOBLOCK) {
-		spin_lock(&nvbo->bo.lock);
+		spin_lock(&nvbo->bo.bdev->fence_lock);
 		ret = ttm_bo_wait(&nvbo->bo, false, false, no_wait);
-		spin_unlock(&nvbo->bo.lock);
+		spin_unlock(&nvbo->bo.bdev->fence_lock);
 	} else {
 		ret = ttm_bo_synccpu_write_grab(&nvbo->bo, no_wait);
 		if (ret == 0)

drivers/gpu/drm/radeon/radeon_object.c

···
 
 	list_for_each_entry(lobj, head, list) {
 		bo = lobj->bo;
-		spin_lock(&bo->tbo.lock);
+		spin_lock(&bo->tbo.bdev->fence_lock);
 		old_fence = (struct radeon_fence *)bo->tbo.sync_obj;
 		bo->tbo.sync_obj = radeon_fence_ref(fence);
 		bo->tbo.sync_obj_arg = NULL;
-		spin_unlock(&bo->tbo.lock);
+		spin_unlock(&bo->tbo.bdev->fence_lock);
 		if (old_fence) {
 			radeon_fence_unref(&old_fence);
 		}

drivers/gpu/drm/radeon/radeon_object.h

···
 	r = ttm_bo_reserve(&bo->tbo, true, no_wait, false, 0);
 	if (unlikely(r != 0))
 		return r;
-	spin_lock(&bo->tbo.lock);
+	spin_lock(&bo->tbo.bdev->fence_lock);
 	if (mem_type)
 		*mem_type = bo->tbo.mem.mem_type;
 	if (bo->tbo.sync_obj)
 		r = ttm_bo_wait(&bo->tbo, true, true, no_wait);
-	spin_unlock(&bo->tbo.lock);
+	spin_unlock(&bo->tbo.bdev->fence_lock);
 	ttm_bo_unreserve(&bo->tbo);
 	return r;
 }

drivers/gpu/drm/ttm/ttm_bo.c

···
 	}
 
 	if (bo->mem.mm_node) {
-		spin_lock(&bo->lock);
 		bo->offset = (bo->mem.start << PAGE_SHIFT) +
 		    bdev->man[bo->mem.mem_type].gpu_offset;
 		bo->cur_placement = bo->mem.placement;
-		spin_unlock(&bo->lock);
 	} else
 		bo->offset = 0;
 
···
 	int put_count;
 	int ret;
 
-	spin_lock(&bo->lock);
+	spin_lock(&bdev->fence_lock);
 	(void) ttm_bo_wait(bo, false, false, true);
 	if (!bo->sync_obj) {
 
 		spin_lock(&glob->lru_lock);
 
 		/**
-		 * Lock inversion between bo::reserve and bo::lock here,
+		 * Lock inversion between bo:reserve and bdev::fence_lock here,
 		 * but that's OK, since we're only trylocking.
 		 */
 
···
 		if (unlikely(ret == -EBUSY))
 			goto queue;
 
-		spin_unlock(&bo->lock);
+		spin_unlock(&bdev->fence_lock);
 		put_count = ttm_bo_del_from_lru(bo);
 
 		spin_unlock(&glob->lru_lock);
···
 	kref_get(&bo->list_kref);
 	list_add_tail(&bo->ddestroy, &bdev->ddestroy);
 	spin_unlock(&glob->lru_lock);
-	spin_unlock(&bo->lock);
+	spin_unlock(&bdev->fence_lock);
 
 	if (sync_obj) {
 		driver->sync_obj_flush(sync_obj, sync_obj_arg);
···
 			       bool no_wait_reserve,
 			       bool no_wait_gpu)
 {
+	struct ttm_bo_device *bdev = bo->bdev;
 	struct ttm_bo_global *glob = bo->glob;
 	int put_count;
 	int ret = 0;
 
 retry:
-	spin_lock(&bo->lock);
+	spin_lock(&bdev->fence_lock);
 	ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
-	spin_unlock(&bo->lock);
+	spin_unlock(&bdev->fence_lock);
 
 	if (unlikely(ret != 0))
 		return ret;
···
 	struct ttm_placement placement;
 	int ret = 0;
 
-	spin_lock(&bo->lock);
+	spin_lock(&bdev->fence_lock);
 	ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
-	spin_unlock(&bo->lock);
+	spin_unlock(&bdev->fence_lock);
 
 	if (unlikely(ret != 0)) {
 		if (ret != -ERESTARTSYS) {
···
 {
 	int ret = 0;
 	struct ttm_mem_reg mem;
+	struct ttm_bo_device *bdev = bo->bdev;
 
 	BUG_ON(!atomic_read(&bo->reserved));
 
···
 	 * Have the driver move function wait for idle when necessary,
 	 * instead of doing it here.
 	 */
-	spin_lock(&bo->lock);
+	spin_lock(&bdev->fence_lock);
 	ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
-	spin_unlock(&bo->lock);
+	spin_unlock(&bdev->fence_lock);
 	if (ret)
 		return ret;
 	mem.num_pages = bo->num_pages;
···
 	}
 	bo->destroy = destroy;
 
-	spin_lock_init(&bo->lock);
 	kref_init(&bo->kref);
 	kref_init(&bo->list_kref);
 	atomic_set(&bo->cpu_writers, 0);
···
 	bdev->dev_mapping = NULL;
 	bdev->glob = glob;
 	bdev->need_dma32 = need_dma32;
-
+	spin_lock_init(&bdev->fence_lock);
 	mutex_lock(&glob->device_list_mutex);
 	list_add_tail(&bdev->device_list, &glob->device_list);
 	mutex_unlock(&glob->device_list_mutex);
···
 		bool lazy, bool interruptible, bool no_wait)
 {
 	struct ttm_bo_driver *driver = bo->bdev->driver;
+	struct ttm_bo_device *bdev = bo->bdev;
 	void *sync_obj;
 	void *sync_obj_arg;
 	int ret = 0;
···
 			void *tmp_obj = bo->sync_obj;
 			bo->sync_obj = NULL;
 			clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
-			spin_unlock(&bo->lock);
+			spin_unlock(&bdev->fence_lock);
 			driver->sync_obj_unref(&tmp_obj);
-			spin_lock(&bo->lock);
+			spin_lock(&bdev->fence_lock);
 			continue;
 		}
 
···
 
 		sync_obj = driver->sync_obj_ref(bo->sync_obj);
 		sync_obj_arg = bo->sync_obj_arg;
-		spin_unlock(&bo->lock);
+		spin_unlock(&bdev->fence_lock);
 		ret = driver->sync_obj_wait(sync_obj, sync_obj_arg,
 					    lazy, interruptible);
 		if (unlikely(ret != 0)) {
 			driver->sync_obj_unref(&sync_obj);
-			spin_lock(&bo->lock);
+			spin_lock(&bdev->fence_lock);
 			return ret;
 		}
-		spin_lock(&bo->lock);
+		spin_lock(&bdev->fence_lock);
 		if (likely(bo->sync_obj == sync_obj &&
 			   bo->sync_obj_arg == sync_obj_arg)) {
 			void *tmp_obj = bo->sync_obj;
 			bo->sync_obj = NULL;
 			clear_bit(TTM_BO_PRIV_FLAG_MOVING,
 				  &bo->priv_flags);
-			spin_unlock(&bo->lock);
+			spin_unlock(&bdev->fence_lock);
 			driver->sync_obj_unref(&sync_obj);
 			driver->sync_obj_unref(&tmp_obj);
-			spin_lock(&bo->lock);
+			spin_lock(&bdev->fence_lock);
 		} else {
-			spin_unlock(&bo->lock);
+			spin_unlock(&bdev->fence_lock);
 			driver->sync_obj_unref(&sync_obj);
-			spin_lock(&bo->lock);
+			spin_lock(&bdev->fence_lock);
 		}
 	}
 	return 0;
···
 
 int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait)
 {
+	struct ttm_bo_device *bdev = bo->bdev;
 	int ret = 0;
 
 	/*
···
 	ret = ttm_bo_reserve(bo, true, no_wait, false, 0);
 	if (unlikely(ret != 0))
 		return ret;
-	spin_lock(&bo->lock);
+	spin_lock(&bdev->fence_lock);
 	ret = ttm_bo_wait(bo, false, true, no_wait);
-	spin_unlock(&bo->lock);
+	spin_unlock(&bdev->fence_lock);
 	if (likely(ret == 0))
 		atomic_inc(&bo->cpu_writers);
 	ttm_bo_unreserve(bo);
···
 	 * Wait for GPU, then move to system cached.
 	 */
 
-	spin_lock(&bo->lock);
+	spin_lock(&bo->bdev->fence_lock);
 	ret = ttm_bo_wait(bo, false, false, false);
-	spin_unlock(&bo->lock);
+	spin_unlock(&bo->bdev->fence_lock);
 
 	if (unlikely(ret != 0))
 		goto out;

drivers/gpu/drm/ttm/ttm_bo_util.c

···
 	 * TODO: Explicit member copy would probably be better here.
 	 */
 
-	spin_lock_init(&fbo->lock);
 	init_waitqueue_head(&fbo->event_queue);
 	INIT_LIST_HEAD(&fbo->ddestroy);
 	INIT_LIST_HEAD(&fbo->lru);
···
 	struct ttm_buffer_object *ghost_obj;
 	void *tmp_obj = NULL;
 
-	spin_lock(&bo->lock);
+	spin_lock(&bdev->fence_lock);
 	if (bo->sync_obj) {
 		tmp_obj = bo->sync_obj;
 		bo->sync_obj = NULL;
···
 	bo->sync_obj_arg = sync_obj_arg;
 	if (evict) {
 		ret = ttm_bo_wait(bo, false, false, false);
-		spin_unlock(&bo->lock);
+		spin_unlock(&bdev->fence_lock);
 		if (tmp_obj)
 			driver->sync_obj_unref(&tmp_obj);
 		if (ret)
···
 		 */
 
 		set_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
-		spin_unlock(&bo->lock);
+		spin_unlock(&bdev->fence_lock);
 		if (tmp_obj)
 			driver->sync_obj_unref(&tmp_obj);
 

drivers/gpu/drm/ttm/ttm_bo_vm.c

···
 	 * move.
 	 */
 
-	spin_lock(&bo->lock);
+	spin_lock(&bdev->fence_lock);
 	if (test_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags)) {
 		ret = ttm_bo_wait(bo, false, true, false);
-		spin_unlock(&bo->lock);
+		spin_unlock(&bdev->fence_lock);
 		if (unlikely(ret != 0)) {
 			retval = (ret != -ERESTARTSYS) ?
 			    VM_FAULT_SIGBUS : VM_FAULT_NOPAGE;
 			goto out_unlock;
 		}
 	} else
-		spin_unlock(&bo->lock);
+		spin_unlock(&bdev->fence_lock);
 
 
 	ret = ttm_mem_io_reserve(bdev, &bo->mem);

drivers/gpu/drm/ttm/ttm_execbuf_util.c

···
 
 	list_for_each_entry(entry, list, head) {
 		struct ttm_buffer_object *bo = entry->bo;
-		struct ttm_bo_driver *driver = bo->bdev->driver;
+		struct ttm_bo_device *bdev = bo->bdev;
+		struct ttm_bo_driver *driver = bdev->driver;
 		void *old_sync_obj;
 
-		spin_lock(&bo->lock);
+		spin_lock(&bdev->fence_lock);
 		old_sync_obj = bo->sync_obj;
 		bo->sync_obj = driver->sync_obj_ref(sync_obj);
 		bo->sync_obj_arg = entry->new_sync_obj_arg;
-		spin_unlock(&bo->lock);
+		spin_unlock(&bdev->fence_lock);
 		ttm_bo_unreserve(bo);
 		entry->reserved = false;
 		if (old_sync_obj)

include/drm/ttm/ttm_bo_api.h

···
  * keeps one refcount. When this refcount reaches zero,
  * the object is destroyed.
  * @event_queue: Queue for processes waiting on buffer object status change.
- * @lock: spinlock protecting mostly synchronization members.
  * @mem: structure describing current placement.
  * @persistant_swap_storage: Usually the swap storage is deleted for buffers
  * pinned in physical memory. If this behaviour is not desired, this member
···
 	struct kref kref;
 	struct kref list_kref;
 	wait_queue_head_t event_queue;
-	spinlock_t lock;
 
 	/**
 	 * Members protected by the bo::reserved lock.
···
 	atomic_t reserved;
 
 	/**
-	 * Members protected by the bo::lock
+	 * Members protected by struct buffer_object_device::fence_lock
 	 * In addition, setting sync_obj to anything else
 	 * than NULL requires bo::reserved to be held. This allows for
-	 * checking NULL while reserved but not holding bo::lock.
+	 * checking NULL while reserved but not holding the mentioned lock.
 	 */
 
 	void *sync_obj_arg;

include/drm/ttm/ttm_bo_driver.h

···
  *
  * @driver: Pointer to a struct ttm_bo_driver struct setup by the driver.
  * @man: An array of mem_type_managers.
+ * @fence_lock: Protects the synchronizing members on *all* bos belonging
+ * to this device.
  * @addr_space_mm: Range manager for the device address space.
  * lru_lock: Spinlock that protects the buffer+device lru lists and
  * ddestroy lists.
···
 	struct ttm_bo_driver *driver;
 	rwlock_t vm_lock;
 	struct ttm_mem_type_manager man[TTM_NUM_MEM_TYPES];
+	spinlock_t fence_lock;
 	/*
 	 * Protected by the vm lock.
 	 */