diff --git a/drivers/gpu/drm/nouveau/nouveau_bo.c b/drivers/gpu/drm/nouveau/nouveau_bo.c
index 84f85183d041..f6f44779d82f 100644
--- a/drivers/gpu/drm/nouveau/nouveau_bo.c
+++ b/drivers/gpu/drm/nouveau/nouveau_bo.c
@@ -36,6 +36,21 @@
 #include <linux/log2.h>
 #include <linux/slab.h>
 
+int
+nouveau_bo_sync_gpu(struct nouveau_bo *nvbo, struct nouveau_channel *chan)
+{
+	struct nouveau_fence *prev_fence = nvbo->bo.sync_obj;
+	int ret;
+
+	if (!prev_fence || nouveau_fence_channel(prev_fence) == chan)
+		return 0;
+
+	spin_lock(&nvbo->bo.lock);
+	ret = ttm_bo_wait(&nvbo->bo, false, false, false);
+	spin_unlock(&nvbo->bo.lock);
+	return ret;
+}
+
 static void
 nouveau_bo_del_ttm(struct ttm_buffer_object *bo)
 {
diff --git a/drivers/gpu/drm/nouveau/nouveau_drv.h b/drivers/gpu/drm/nouveau/nouveau_drv.h
index e424bf74d706..1e093a069b7b 100644
--- a/drivers/gpu/drm/nouveau/nouveau_drv.h
+++ b/drivers/gpu/drm/nouveau/nouveau_drv.h
@@ -1165,6 +1165,7 @@ extern u16 nouveau_bo_rd16(struct nouveau_bo *nvbo, unsigned index);
 extern void nouveau_bo_wr16(struct nouveau_bo *nvbo, unsigned index, u16 val);
 extern u32 nouveau_bo_rd32(struct nouveau_bo *nvbo, unsigned index);
 extern void nouveau_bo_wr32(struct nouveau_bo *nvbo, unsigned index, u32 val);
+extern int nouveau_bo_sync_gpu(struct nouveau_bo *, struct nouveau_channel *);
 
 /* nouveau_fence.c */
 struct nouveau_fence;
diff --git a/drivers/gpu/drm/nouveau/nouveau_gem.c b/drivers/gpu/drm/nouveau/nouveau_gem.c
index 547f2c24c1e7..a915dcdd9a49 100644
--- a/drivers/gpu/drm/nouveau/nouveau_gem.c
+++ b/drivers/gpu/drm/nouveau/nouveau_gem.c
@@ -361,16 +361,11 @@ validate_list(struct nouveau_channel *chan, struct list_head *list,
 
 	list_for_each_entry(nvbo, list, entry) {
 		struct drm_nouveau_gem_pushbuf_bo *b = &pbbo[nvbo->pbbo_index];
-		struct nouveau_fence *prev_fence = nvbo->bo.sync_obj;
 
-		if (prev_fence && nouveau_fence_channel(prev_fence) != chan) {
-			spin_lock(&nvbo->bo.lock);
-			ret = ttm_bo_wait(&nvbo->bo, false, false, false);
-			spin_unlock(&nvbo->bo.lock);
-			if (unlikely(ret)) {
-				NV_ERROR(dev, "fail wait other chan\n");
-				return ret;
-			}
+		ret = nouveau_bo_sync_gpu(nvbo, chan);
+		if (unlikely(ret)) {
+			NV_ERROR(dev, "fail pre-validate sync\n");
+			return ret;
 		}
 
 		ret = nouveau_gem_set_domain(nvbo->gem, b->read_domains,
@@ -381,7 +376,7 @@ validate_list(struct nouveau_channel *chan, struct list_head *list,
 			return ret;
 		}
 
-		nvbo->channel = chan;
+		nvbo->channel = (b->read_domains & (1 << 31)) ? NULL : chan;
 		ret = ttm_bo_validate(&nvbo->bo, &nvbo->placement,
 				      false, false, false);
 		nvbo->channel = NULL;
@@ -390,6 +385,12 @@ validate_list(struct nouveau_channel *chan, struct list_head *list,
 			return ret;
 		}
 
+		ret = nouveau_bo_sync_gpu(nvbo, chan);
+		if (unlikely(ret)) {
+			NV_ERROR(dev, "fail post-validate sync\n");
+			return ret;
+		}
+
 		if (nvbo->bo.offset == b->presumed.offset &&
 		    ((nvbo->bo.mem.mem_type == TTM_PL_VRAM &&
 		      b->presumed.domain & NOUVEAU_GEM_DOMAIN_VRAM) ||
@@ -615,6 +616,21 @@ nouveau_gem_ioctl_pushbuf(struct drm_device *dev, void *data,
 
 	mutex_lock(&dev->struct_mutex);
 
+	/* Mark push buffers as being used on PFIFO, the validation code
+	 * will then make sure that if the pushbuf bo moves, that they
+	 * happen on the kernel channel, which will in turn cause a sync
+	 * to happen before we try and submit the push buffer.
+	 */
+	for (i = 0; i < req->nr_push; i++) {
+		if (push[i].bo_index >= req->nr_buffers) {
+			NV_ERROR(dev, "push %d buffer not in list\n", i);
+			ret = -EINVAL;
+			goto out;
+		}
+
+		bo[push[i].bo_index].read_domains |= (1 << 31);
+	}
+
 	/* Validate buffer list */
 	ret = nouveau_gem_pushbuf_validate(chan, file_priv, bo, req->buffers,
 					   req->nr_buffers, &op, &do_reloc);