}
 
        /* Fallback to software copy. */
-       spin_lock(&bo->bdev->fence_lock);
        ret = ttm_bo_wait(bo, true, intr, no_wait_gpu);
-       spin_unlock(&bo->bdev->fence_lock);
        if (ret == 0)
                ret = ttm_bo_move_memcpy(bo, evict, no_wait_gpu, new_mem);
 
        ttm_pool_unpopulate(ttm);
 }
 
+static void
+nouveau_bo_fence_unref(void **sync_obj)
+{
+       nouveau_fence_unref((struct nouveau_fence **)sync_obj);
+}
+
 void
 nouveau_bo_fence(struct nouveau_bo *nvbo, struct nouveau_fence *fence)
 {
-       struct nouveau_fence *new_fence = nouveau_fence_ref(fence);
-       struct nouveau_fence *old_fence = NULL;
-
        lockdep_assert_held(&nvbo->bo.resv->lock.base);
 
-       spin_lock(&nvbo->bo.bdev->fence_lock);
-       old_fence = nvbo->bo.sync_obj;
-       nvbo->bo.sync_obj = new_fence;
-       spin_unlock(&nvbo->bo.bdev->fence_lock);
-
-       nouveau_fence_unref(&old_fence);
-}
-
-static void
-nouveau_bo_fence_unref(void **sync_obj)
-{
-       nouveau_fence_unref((struct nouveau_fence **)sync_obj);
+       nouveau_bo_fence_unref(&nvbo->bo.sync_obj);
+       nvbo->bo.sync_obj = nouveau_fence_ref(fence);
 }
 
 static void *
 
                goto fail_unpin;
 
        /* synchronise rendering channel with the kernel's channel */
-       spin_lock(&new_bo->bo.bdev->fence_lock);
-       fence = nouveau_fence_ref(new_bo->bo.sync_obj);
-       spin_unlock(&new_bo->bo.bdev->fence_lock);
-       ret = nouveau_fence_sync(fence, chan);
-       nouveau_fence_unref(&fence);
+       ret = nouveau_fence_sync(new_bo->bo.sync_obj, chan);
        if (ret) {
                ttm_bo_unreserve(&new_bo->bo);
                goto fail_unpin;
 
        list_del(&vma->head);
 
        if (mapped) {
-               spin_lock(&nvbo->bo.bdev->fence_lock);
                fence = nouveau_fence_ref(nvbo->bo.sync_obj);
-               spin_unlock(&nvbo->bo.bdev->fence_lock);
        }
 
        if (fence) {
 static int
 validate_sync(struct nouveau_channel *chan, struct nouveau_bo *nvbo)
 {
-       struct nouveau_fence *fence = NULL;
+       struct nouveau_fence *fence = nvbo->bo.sync_obj;
        int ret = 0;
 
-       spin_lock(&nvbo->bo.bdev->fence_lock);
-       fence = nouveau_fence_ref(nvbo->bo.sync_obj);
-       spin_unlock(&nvbo->bo.bdev->fence_lock);
-
-       if (fence) {
+       if (fence)
                ret = nouveau_fence_sync(fence, chan);
-               nouveau_fence_unref(&fence);
-       }
 
        return ret;
 }
                                data |= r->vor;
                }
 
-               spin_lock(&nvbo->bo.bdev->fence_lock);
                ret = ttm_bo_wait(&nvbo->bo, false, false, false);
-               spin_unlock(&nvbo->bo.bdev->fence_lock);
                if (ret) {
                        NV_PRINTK(error, cli, "reloc wait_idle failed: %d\n", ret);
                        break;
 
        ret = ttm_bo_reserve(&nvbo->bo, true, false, false, NULL);
        if (!ret) {
-               spin_lock(&nvbo->bo.bdev->fence_lock);
                ret = ttm_bo_wait(&nvbo->bo, true, true, true);
                if (!no_wait && ret)
                        fence = nouveau_fence_ref(nvbo->bo.sync_obj);
-               spin_unlock(&nvbo->bo.bdev->fence_lock);
 
                ttm_bo_unreserve(&nvbo->bo);
        }
 
        if (stall)
                mutex_unlock(&qdev->surf_evict_mutex);
 
-       spin_lock(&surf->tbo.bdev->fence_lock);
        ret = ttm_bo_wait(&surf->tbo, true, true, !stall);
-       spin_unlock(&surf->tbo.bdev->fence_lock);
 
        if (stall)
                mutex_lock(&qdev->surf_evict_mutex);
 
 {
        void *ret;
        int retval = 0;
-       struct qxl_bo *bo = container_of(qfence, struct qxl_bo, fence);
-
-       spin_lock(&bo->tbo.bdev->fence_lock);
 
        ret = radix_tree_delete(&qfence->tree, rel_id);
        if (ret == qfence)
                DRM_DEBUG("didn't find fence in radix tree for %d\n", rel_id);
                retval = -ENOENT;
        }
-       spin_unlock(&bo->tbo.bdev->fence_lock);
        return retval;
 }
 
 
                }
                return r;
        }
-       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.bdev->fence_lock);
        ttm_bo_unreserve(&bo->tbo);
        return r;
 }
 
        glob = bo->glob;
 
        spin_lock(&glob->lru_lock);
-       spin_lock(&bdev->fence_lock);
 
        list_for_each_entry(entry, &release->bos, head) {
                bo = entry->bo;
                __ttm_bo_unreserve(bo);
                entry->reserved = false;
        }
-       spin_unlock(&bdev->fence_lock);
        spin_unlock(&glob->lru_lock);
        ww_acquire_fini(&release->ticket);
 }
 
        obj = new_radeon_fb->obj;
        new_rbo = gem_to_radeon_bo(obj);
 
-       spin_lock(&new_rbo->tbo.bdev->fence_lock);
-       if (new_rbo->tbo.sync_obj)
-               work->fence = radeon_fence_ref(new_rbo->tbo.sync_obj);
-       spin_unlock(&new_rbo->tbo.bdev->fence_lock);
-
        /* pin the new buffer */
        DRM_DEBUG_DRIVER("flip-ioctl() cur_rbo = %p, new_rbo = %p\n",
                         work->old_rbo, new_rbo);
                DRM_ERROR("failed to pin new rbo buffer before flip\n");
                goto cleanup;
        }
+       work->fence = radeon_fence_ref(new_rbo->tbo.sync_obj);
        radeon_bo_get_tiling_flags(new_rbo, &tiling_flags, NULL);
        radeon_bo_unreserve(new_rbo);
 
        drm_gem_object_unreference_unlocked(&work->old_rbo->gem_base);
        radeon_fence_unref(&work->fence);
        kfree(work);
-
        return r;
 }
 
 
        r = ttm_bo_reserve(&bo->tbo, true, no_wait, false, NULL);
        if (unlikely(r != 0))
                return r;
-       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.bdev->fence_lock);
        ttm_bo_unreserve(&bo->tbo);
        return r;
 }
 
        spin_lock(&glob->lru_lock);
        ret = __ttm_bo_reserve(bo, false, true, false, NULL);
 
-       spin_lock(&bdev->fence_lock);
-       (void) ttm_bo_wait(bo, false, false, true);
-       if (!ret && !bo->sync_obj) {
-               spin_unlock(&bdev->fence_lock);
-               put_count = ttm_bo_del_from_lru(bo);
+       if (!ret) {
+               (void) ttm_bo_wait(bo, false, false, true);
 
-               spin_unlock(&glob->lru_lock);
-               ttm_bo_cleanup_memtype_use(bo);
+               if (!bo->sync_obj) {
+                       put_count = ttm_bo_del_from_lru(bo);
 
-               ttm_bo_list_ref_sub(bo, put_count, true);
+                       spin_unlock(&glob->lru_lock);
+                       ttm_bo_cleanup_memtype_use(bo);
 
-               return;
-       }
-       if (bo->sync_obj)
-               sync_obj = driver->sync_obj_ref(bo->sync_obj);
-       spin_unlock(&bdev->fence_lock);
+                       ttm_bo_list_ref_sub(bo, put_count, true);
 
-       if (!ret) {
+                       return;
+               }
+               sync_obj = driver->sync_obj_ref(bo->sync_obj);
 
                /*
                 * Make NO_EVICT bos immediately available to
        int put_count;
        int ret;
 
-       spin_lock(&bdev->fence_lock);
        ret = ttm_bo_wait(bo, false, false, true);
 
        if (ret && !no_wait_gpu) {
                 * no new sync objects can be attached.
                 */
                sync_obj = driver->sync_obj_ref(bo->sync_obj);
-               spin_unlock(&bdev->fence_lock);
 
                __ttm_bo_unreserve(bo);
                spin_unlock(&glob->lru_lock);
                 * remove sync_obj with ttm_bo_wait, the wait should be
                 * finished, and no new wait object should have been added.
                 */
-               spin_lock(&bdev->fence_lock);
                ret = ttm_bo_wait(bo, false, false, true);
                WARN_ON(ret);
        }
-       spin_unlock(&bdev->fence_lock);
 
        if (ret || unlikely(list_empty(&bo->ddestroy))) {
                __ttm_bo_unreserve(bo);
        struct ttm_placement placement;
        int ret = 0;
 
-       spin_lock(&bdev->fence_lock);
        ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
-       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;
 
        lockdep_assert_held(&bo->resv->lock.base);
 
         * Have the driver move function wait for idle when necessary,
         * instead of doing it here.
         */
-       spin_lock(&bdev->fence_lock);
        ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
-       spin_unlock(&bdev->fence_lock);
        if (ret)
                return ret;
        mem.num_pages = bo->num_pages;
        bdev->glob = glob;
        bdev->need_dma32 = need_dma32;
        bdev->val_seq = 0;
-       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;
        int ret = 0;
 
        if (likely(bo->sync_obj == NULL))
                return 0;
 
-       while (bo->sync_obj) {
-
+       if (bo->sync_obj) {
                if (driver->sync_obj_signaled(bo->sync_obj)) {
-                       void *tmp_obj = bo->sync_obj;
-                       bo->sync_obj = NULL;
+                       driver->sync_obj_unref(&bo->sync_obj);
                        clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
-                       spin_unlock(&bdev->fence_lock);
-                       driver->sync_obj_unref(&tmp_obj);
-                       spin_lock(&bdev->fence_lock);
-                       continue;
+                       return 0;
                }
 
                if (no_wait)
                        return -EBUSY;
 
                sync_obj = driver->sync_obj_ref(bo->sync_obj);
-               spin_unlock(&bdev->fence_lock);
                ret = driver->sync_obj_wait(sync_obj,
                                            lazy, interruptible);
-               if (unlikely(ret != 0)) {
-                       driver->sync_obj_unref(&sync_obj);
-                       spin_lock(&bdev->fence_lock);
-                       return ret;
-               }
-               spin_lock(&bdev->fence_lock);
-               if (likely(bo->sync_obj == sync_obj)) {
-                       void *tmp_obj = bo->sync_obj;
-                       bo->sync_obj = NULL;
+
+               if (likely(ret == 0)) {
                        clear_bit(TTM_BO_PRIV_FLAG_MOVING,
                                  &bo->priv_flags);
-                       spin_unlock(&bdev->fence_lock);
-                       driver->sync_obj_unref(&sync_obj);
-                       driver->sync_obj_unref(&tmp_obj);
-                       spin_lock(&bdev->fence_lock);
-               } else {
-                       spin_unlock(&bdev->fence_lock);
-                       driver->sync_obj_unref(&sync_obj);
-                       spin_lock(&bdev->fence_lock);
+                       driver->sync_obj_unref(&bo->sync_obj);
                }
+               driver->sync_obj_unref(&sync_obj);
        }
-       return 0;
+       return ret;
 }
 EXPORT_SYMBOL(ttm_bo_wait);
 
 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, NULL);
        if (unlikely(ret != 0))
                return ret;
-       spin_lock(&bdev->fence_lock);
        ret = ttm_bo_wait(bo, false, true, no_wait);
-       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->bdev->fence_lock);
        ret = ttm_bo_wait(bo, false, false, false);
-       spin_unlock(&bo->bdev->fence_lock);
 
        if (unlikely(ret != 0))
                goto out;
 
        drm_vma_node_reset(&fbo->vma_node);
        atomic_set(&fbo->cpu_writers, 0);
 
-       spin_lock(&bdev->fence_lock);
        if (bo->sync_obj)
                fbo->sync_obj = driver->sync_obj_ref(bo->sync_obj);
        else
                fbo->sync_obj = NULL;
-       spin_unlock(&bdev->fence_lock);
        kref_init(&fbo->list_kref);
        kref_init(&fbo->kref);
        fbo->destroy = &ttm_transfered_destroy;
        struct ttm_buffer_object *ghost_obj;
        void *tmp_obj = NULL;
 
-       spin_lock(&bdev->fence_lock);
        if (bo->sync_obj) {
                tmp_obj = bo->sync_obj;
                bo->sync_obj = NULL;
        bo->sync_obj = driver->sync_obj_ref(sync_obj);
        if (evict) {
                ret = ttm_bo_wait(bo, false, false, false);
-               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(&bdev->fence_lock);
                if (tmp_obj)
                        driver->sync_obj_unref(&tmp_obj);
 
 
                                struct vm_area_struct *vma,
                                struct vm_fault *vmf)
 {
-       struct ttm_bo_device *bdev = bo->bdev;
        int ret = 0;
 
-       spin_lock(&bdev->fence_lock);
        if (likely(!test_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags)))
                goto out_unlock;
 
                        VM_FAULT_NOPAGE;
 
 out_unlock:
-       spin_unlock(&bdev->fence_lock);
        return ret;
 }
 
 
        glob = bo->glob;
 
        spin_lock(&glob->lru_lock);
-       spin_lock(&bdev->fence_lock);
 
        list_for_each_entry(entry, list, head) {
                bo = entry->bo;
                __ttm_bo_unreserve(bo);
                entry->reserved = false;
        }
-       spin_unlock(&bdev->fence_lock);
        spin_unlock(&glob->lru_lock);
        if (ticket)
                ww_acquire_fini(ticket);
 
  */
 static void vmw_swap_notify(struct ttm_buffer_object *bo)
 {
-       struct ttm_bo_device *bdev = bo->bdev;
-
-       spin_lock(&bdev->fence_lock);
        ttm_bo_wait(bo, false, false, false);
-       spin_unlock(&bdev->fence_lock);
 }
 
 
 
        int ret;
 
        if (flags & drm_vmw_synccpu_allow_cs) {
-               struct ttm_bo_device *bdev = bo->bdev;
+               bool nonblock = !!(flags & drm_vmw_synccpu_dontblock);
 
-               spin_lock(&bdev->fence_lock);
-               ret = ttm_bo_wait(bo, false, true,
-                                 !!(flags & drm_vmw_synccpu_dontblock));
-               spin_unlock(&bdev->fence_lock);
+               ret = ttm_bo_reserve(bo, true, nonblock, false, NULL);
+               if (!ret) {
+                       ret = ttm_bo_wait(bo, false, true, nonblock);
+                       ttm_bo_unreserve(bo);
+               }
                return ret;
        }
 
        else
                driver->sync_obj_ref(fence);
 
-       spin_lock(&bdev->fence_lock);
 
        old_fence_obj = bo->sync_obj;
        bo->sync_obj = fence;
 
-       spin_unlock(&bdev->fence_lock);
 
        if (old_fence_obj)
                vmw_fence_obj_unreference(&old_fence_obj);
 
        if (mem->mem_type != VMW_PL_MOB) {
                struct vmw_resource *res, *n;
-               struct ttm_bo_device *bdev = bo->bdev;
                struct ttm_validate_buffer val_buf;
 
                val_buf.bo = bo;
                        list_del_init(&res->mob_head);
                }
 
-               spin_lock(&bdev->fence_lock);
                (void) ttm_bo_wait(bo, false, false, false);
-               spin_unlock(&bdev->fence_lock);
        }
 }
 
 
        struct list_head io_reserve_lru;
 
        /**
-        * 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 the mentioned lock.
+        * Members protected by a bo reservation.
         */
 
        void *sync_obj;
 
  *
  * @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.
  * @vma_manager: Address space manager
  * lru_lock: Spinlock that protects the buffer+device lru lists and
  * ddestroy lists.
        struct ttm_bo_global *glob;
        struct ttm_bo_driver *driver;
        struct ttm_mem_type_manager man[TTM_NUM_MEM_TYPES];
-       spinlock_t fence_lock;
 
        /*
         * Protected by internal locks.