diff --git a/drivers/gpu/drm/nouveau/nouveau_bo.c b/drivers/gpu/drm/nouveau/nouveau_bo.c
index 890d50e..e87e24b 100644
--- a/drivers/gpu/drm/nouveau/nouveau_bo.c
+++ b/drivers/gpu/drm/nouveau/nouveau_bo.c
@@ -1104,7 +1104,8 @@
 	if (vma->node) {
 		if (nvbo->bo.mem.mem_type != TTM_PL_SYSTEM) {
 			spin_lock(&nvbo->bo.bdev->fence_lock);
-			ttm_bo_wait(&nvbo->bo, false, false, false);
+			ttm_bo_wait(&nvbo->bo, false, false, false,
+				    TTM_USAGE_READWRITE);
 			spin_unlock(&nvbo->bo.bdev->fence_lock);
 			nouveau_vm_unmap(vma);
 		}
diff --git a/drivers/gpu/drm/nouveau/nouveau_gem.c b/drivers/gpu/drm/nouveau/nouveau_gem.c
index 5f0bc57..322bf62 100644
--- a/drivers/gpu/drm/nouveau/nouveau_gem.c
+++ b/drivers/gpu/drm/nouveau/nouveau_gem.c
@@ -589,7 +589,8 @@
 		}
 
 		spin_lock(&nvbo->bo.bdev->fence_lock);
-		ret = ttm_bo_wait(&nvbo->bo, false, false, false);
+		ret = ttm_bo_wait(&nvbo->bo, false, false, false,
+				  TTM_USAGE_READWRITE);
 		spin_unlock(&nvbo->bo.bdev->fence_lock);
 		if (ret) {
 			NV_ERROR(dev, "reloc wait_idle failed: %d\n", ret);
@@ -825,7 +826,7 @@
 	nvbo = nouveau_gem_object(gem);
 
 	spin_lock(&nvbo->bo.bdev->fence_lock);
-	ret = ttm_bo_wait(&nvbo->bo, true, true, no_wait);
+	ret = ttm_bo_wait(&nvbo->bo, true, true, no_wait, TTM_USAGE_READWRITE);
 	spin_unlock(&nvbo->bo.bdev->fence_lock);
 	drm_gem_object_unreference_unlocked(gem);
 	return ret;
diff --git a/drivers/gpu/drm/radeon/radeon_cs.c b/drivers/gpu/drm/radeon/radeon_cs.c
index fae00c0..14e8531 100644
--- a/drivers/gpu/drm/radeon/radeon_cs.c
+++ b/drivers/gpu/drm/radeon/radeon_cs.c
@@ -80,6 +80,7 @@
 			p->relocs[i].lobj.wdomain = r->write_domain;
 			p->relocs[i].lobj.rdomain = r->read_domains;
 			p->relocs[i].lobj.tv.bo = &p->relocs[i].robj->tbo;
+			p->relocs[i].lobj.tv.usage = TTM_USAGE_READWRITE;
 			p->relocs[i].handle = r->handle;
 			p->relocs[i].flags = r->flags;
 			radeon_bo_list_add_object(&p->relocs[i].lobj,
diff --git a/drivers/gpu/drm/radeon/radeon_object.h b/drivers/gpu/drm/radeon/radeon_object.h
index ede6c13..e9dc8b2 100644
--- a/drivers/gpu/drm/radeon/radeon_object.h
+++ b/drivers/gpu/drm/radeon/radeon_object.h
@@ -130,7 +130,7 @@
 	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);
+		r = ttm_bo_wait(&bo->tbo, true, true, no_wait, TTM_USAGE_READWRITE);
 	spin_unlock(&bo->tbo.bdev->fence_lock);
 	ttm_bo_unreserve(&bo->tbo);
 	return r;
diff --git a/drivers/gpu/drm/ttm/ttm_bo.c b/drivers/gpu/drm/ttm/ttm_bo.c
index a4d38d8..b824d9b 100644
--- a/drivers/gpu/drm/ttm/ttm_bo.c
+++ b/drivers/gpu/drm/ttm/ttm_bo.c
@@ -498,7 +498,7 @@
 	int ret;
 
 	spin_lock(&bdev->fence_lock);
-	(void) ttm_bo_wait(bo, false, false, true);
+	(void) ttm_bo_wait(bo, false, false, true, TTM_USAGE_READWRITE);
 	if (!bo->sync_obj) {
 
 		spin_lock(&glob->lru_lock);
@@ -566,7 +566,8 @@
 
 retry:
 	spin_lock(&bdev->fence_lock);
-	ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
+	ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu,
+			  TTM_USAGE_READWRITE);
 	spin_unlock(&bdev->fence_lock);
 
 	if (unlikely(ret != 0))
@@ -725,7 +726,8 @@
 	int ret = 0;
 
 	spin_lock(&bdev->fence_lock);
-	ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
+	ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu,
+			  TTM_USAGE_READWRITE);
 	spin_unlock(&bdev->fence_lock);
 
 	if (unlikely(ret != 0)) {
@@ -1072,7 +1074,8 @@
 	 * instead of doing it here.
 	 */
 	spin_lock(&bdev->fence_lock);
-	ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
+	ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu,
+			  TTM_USAGE_READWRITE);
 	spin_unlock(&bdev->fence_lock);
 	if (ret)
 		return ret;
@@ -1692,34 +1695,83 @@
 	return ret;
 }
 
+static void ttm_bo_unref_sync_obj_locked(struct ttm_buffer_object *bo,
+					 void *sync_obj,
+					 void **extra_sync_obj)
+{
+	struct ttm_bo_device *bdev = bo->bdev;
+	struct ttm_bo_driver *driver = bdev->driver;
+	void *tmp_obj = NULL, *tmp_obj_read = NULL, *tmp_obj_write = NULL;
+
+	/* We must unref the sync obj wherever it's ref'd.
+	 * Note that if we unref bo->sync_obj, we can unref both the read
+	 * and write sync objs too, because they can't be newer than
+	 * bo->sync_obj, so they are no longer relevant. */
+	if (sync_obj == bo->sync_obj ||
+	    sync_obj == bo->sync_obj_read) {
+		tmp_obj_read = bo->sync_obj_read;
+		bo->sync_obj_read = NULL;
+	}
+	if (sync_obj == bo->sync_obj ||
+	    sync_obj == bo->sync_obj_write) {
+		tmp_obj_write = bo->sync_obj_write;
+		bo->sync_obj_write = NULL;
+	}
+	if (sync_obj == bo->sync_obj) {
+		tmp_obj = bo->sync_obj;
+		bo->sync_obj = NULL;
+	}
+
+	clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
+	spin_unlock(&bdev->fence_lock);
+	if (tmp_obj)
+		driver->sync_obj_unref(&tmp_obj);
+	if (tmp_obj_read)
+		driver->sync_obj_unref(&tmp_obj_read);
+	if (tmp_obj_write)
+		driver->sync_obj_unref(&tmp_obj_write);
+	if (extra_sync_obj)
+		driver->sync_obj_unref(extra_sync_obj);
+	spin_lock(&bdev->fence_lock);
+}
+
 int ttm_bo_wait(struct ttm_buffer_object *bo,
-		bool lazy, bool interruptible, bool no_wait)
+		bool lazy, bool interruptible, bool no_wait,
+		enum ttm_buffer_usage usage)
 {
 	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 **bo_sync_obj;
 
-	if (likely(bo->sync_obj == NULL))
+	switch (usage) {
+	case TTM_USAGE_READ:
+		bo_sync_obj = &bo->sync_obj_read;
+		break;
+	case TTM_USAGE_WRITE:
+		bo_sync_obj = &bo->sync_obj_write;
+		break;
+	case TTM_USAGE_READWRITE:
+	default:
+		bo_sync_obj = &bo->sync_obj;
+	}
+
+	if (likely(*bo_sync_obj == NULL))
 		return 0;
 
-	while (bo->sync_obj) {
+	while (*bo_sync_obj) {
 
-		if (driver->sync_obj_signaled(bo->sync_obj, bo->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(&bdev->fence_lock);
-			driver->sync_obj_unref(&tmp_obj);
-			spin_lock(&bdev->fence_lock);
+		if (driver->sync_obj_signaled(*bo_sync_obj, bo->sync_obj_arg)) {
+			ttm_bo_unref_sync_obj_locked(bo, *bo_sync_obj, NULL);
 			continue;
 		}
 
 		if (no_wait)
 			return -EBUSY;
 
-		sync_obj = driver->sync_obj_ref(bo->sync_obj);
+		sync_obj = driver->sync_obj_ref(*bo_sync_obj);
 		sync_obj_arg = bo->sync_obj_arg;
 		spin_unlock(&bdev->fence_lock);
 		ret = driver->sync_obj_wait(sync_obj, sync_obj_arg,
@@ -1730,16 +1782,9 @@
 			return ret;
 		}
 		spin_lock(&bdev->fence_lock);
-		if (likely(bo->sync_obj == sync_obj &&
+		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(&bdev->fence_lock);
-			driver->sync_obj_unref(&sync_obj);
-			driver->sync_obj_unref(&tmp_obj);
-			spin_lock(&bdev->fence_lock);
+			ttm_bo_unref_sync_obj_locked(bo, *bo_sync_obj, &sync_obj);
 		} else {
 			spin_unlock(&bdev->fence_lock);
 			driver->sync_obj_unref(&sync_obj);
@@ -1763,7 +1808,7 @@
 	if (unlikely(ret != 0))
 		return ret;
 	spin_lock(&bdev->fence_lock);
-	ret = ttm_bo_wait(bo, false, true, no_wait);
+	ret = ttm_bo_wait(bo, false, true, no_wait, TTM_USAGE_READWRITE);
 	spin_unlock(&bdev->fence_lock);
 	if (likely(ret == 0))
 		atomic_inc(&bo->cpu_writers);
@@ -1837,7 +1882,7 @@
 	 */
 
 	spin_lock(&bo->bdev->fence_lock);
-	ret = ttm_bo_wait(bo, false, false, false);
+	ret = ttm_bo_wait(bo, false, false, false, TTM_USAGE_READWRITE);
 	spin_unlock(&bo->bdev->fence_lock);
 
 	if (unlikely(ret != 0))
diff --git a/drivers/gpu/drm/ttm/ttm_bo_util.c b/drivers/gpu/drm/ttm/ttm_bo_util.c
index ae3c6f5..6135f58 100644
--- a/drivers/gpu/drm/ttm/ttm_bo_util.c
+++ b/drivers/gpu/drm/ttm/ttm_bo_util.c
@@ -436,6 +436,8 @@
 	atomic_set(&fbo->cpu_writers, 0);
 
 	fbo->sync_obj = driver->sync_obj_ref(bo->sync_obj);
+	fbo->sync_obj_read = driver->sync_obj_ref(bo->sync_obj_read);
+	fbo->sync_obj_write = driver->sync_obj_ref(bo->sync_obj_write);
 	kref_init(&fbo->list_kref);
 	kref_init(&fbo->kref);
 	fbo->destroy = &ttm_transfered_destroy;
@@ -618,20 +620,30 @@
 	struct ttm_mem_reg *old_mem = &bo->mem;
 	int ret;
 	struct ttm_buffer_object *ghost_obj;
-	void *tmp_obj = NULL;
+	void *tmp_obj = NULL, *tmp_obj_read = NULL, *tmp_obj_write = NULL;
 
 	spin_lock(&bdev->fence_lock);
-	if (bo->sync_obj) {
+	if (bo->sync_obj)
 		tmp_obj = bo->sync_obj;
-		bo->sync_obj = NULL;
-	}
+	if (bo->sync_obj_read)
+		tmp_obj_read = bo->sync_obj_read;
+	if (bo->sync_obj_write)
+		tmp_obj_write = bo->sync_obj_write;
+
 	bo->sync_obj = driver->sync_obj_ref(sync_obj);
+	bo->sync_obj_read = driver->sync_obj_ref(sync_obj);
+	bo->sync_obj_write = driver->sync_obj_ref(sync_obj);
 	bo->sync_obj_arg = sync_obj_arg;
 	if (evict) {
-		ret = ttm_bo_wait(bo, false, false, false);
+		ret = ttm_bo_wait(bo, false, false, false,
+				  TTM_USAGE_READWRITE);
 		spin_unlock(&bdev->fence_lock);
 		if (tmp_obj)
 			driver->sync_obj_unref(&tmp_obj);
+		if (tmp_obj_read)
+			driver->sync_obj_unref(&tmp_obj_read);
+		if (tmp_obj_write)
+			driver->sync_obj_unref(&tmp_obj_write);
 		if (ret)
 			return ret;
 
@@ -655,6 +667,10 @@
 		spin_unlock(&bdev->fence_lock);
 		if (tmp_obj)
 			driver->sync_obj_unref(&tmp_obj);
+		if (tmp_obj_read)
+			driver->sync_obj_unref(&tmp_obj_read);
+		if (tmp_obj_write)
+			driver->sync_obj_unref(&tmp_obj_write);
 
 		ret = ttm_buffer_object_transfer(bo, &ghost_obj);
 		if (ret)
diff --git a/drivers/gpu/drm/ttm/ttm_bo_vm.c b/drivers/gpu/drm/ttm/ttm_bo_vm.c
index 221b924..ff1e26f 100644
--- a/drivers/gpu/drm/ttm/ttm_bo_vm.c
+++ b/drivers/gpu/drm/ttm/ttm_bo_vm.c
@@ -122,7 +122,7 @@
 
 	spin_lock(&bdev->fence_lock);
 	if (test_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags)) {
-		ret = ttm_bo_wait(bo, false, true, false);
+		ret = ttm_bo_wait(bo, false, true, false, TTM_USAGE_READWRITE);
 		spin_unlock(&bdev->fence_lock);
 		if (unlikely(ret != 0)) {
 			retval = (ret != -ERESTARTSYS) ?
diff --git a/drivers/gpu/drm/ttm/ttm_execbuf_util.c b/drivers/gpu/drm/ttm/ttm_execbuf_util.c
index 3832fe1..36d111a 100644
--- a/drivers/gpu/drm/ttm/ttm_execbuf_util.c
+++ b/drivers/gpu/drm/ttm/ttm_execbuf_util.c
@@ -221,8 +221,18 @@
 
 	list_for_each_entry(entry, list, head) {
 		bo = entry->bo;
+		entry->old_sync_obj_read = NULL;
+		entry->old_sync_obj_write = NULL;
 		entry->old_sync_obj = bo->sync_obj;
 		bo->sync_obj = driver->sync_obj_ref(sync_obj);
+		if (entry->usage & TTM_USAGE_READ) {
+			entry->old_sync_obj_read = bo->sync_obj_read;
+			bo->sync_obj_read = driver->sync_obj_ref(sync_obj);
+		}
+		if (entry->usage & TTM_USAGE_WRITE) {
+			entry->old_sync_obj_write = bo->sync_obj_write;
+			bo->sync_obj_write = driver->sync_obj_ref(sync_obj);
+		}
 		bo->sync_obj_arg = entry->new_sync_obj_arg;
 		ttm_bo_unreserve_locked(bo);
 		entry->reserved = false;
@@ -231,8 +241,15 @@
 	spin_unlock(&bdev->fence_lock);
 
 	list_for_each_entry(entry, list, head) {
-		if (entry->old_sync_obj)
+		if (entry->old_sync_obj) {
 			driver->sync_obj_unref(&entry->old_sync_obj);
+		}
+		if (entry->old_sync_obj_read) {
+			driver->sync_obj_unref(&entry->old_sync_obj_read);
+		}
+		if (entry->old_sync_obj_write) {
+			driver->sync_obj_unref(&entry->old_sync_obj_write);
+		}
 	}
 }
 EXPORT_SYMBOL(ttm_eu_fence_buffer_objects);
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c b/drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c
index 41b95ed..8ca3ddb 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c
@@ -224,6 +224,7 @@
 	if (unlikely(cur_validate_node == sw_context->cur_val_buf)) {
 		val_buf = &sw_context->val_bufs[cur_validate_node];
 		val_buf->bo = ttm_bo_reference(bo);
+		val_buf->usage = TTM_USAGE_READWRITE;
 		val_buf->new_sync_obj_arg = (void *) dev_priv;
 		list_add_tail(&val_buf->head, &sw_context->validate_nodes);
 		++sw_context->cur_val_buf;