diff --git a/src/gallium/drivers/nouveau/nouveau_buffer.c b/src/gallium/drivers/nouveau/nouveau_buffer.c
index 3e04049..95905a8 100644
--- a/src/gallium/drivers/nouveau/nouveau_buffer.c
+++ b/src/gallium/drivers/nouveau/nouveau_buffer.c
@@ -205,6 +205,9 @@ nouveau_transfer_write(struct nouveau_context *nv, struct nouveau_transfer *tx,
                   base, size / 4, (const uint32_t *)data);
    else
       nv->push_data(nv, buf->bo, buf->offset + base, buf->domain, size, data);
+
+   nouveau_fence_ref(nv->screen->fence.current, &buf->fence);
+   nouveau_fence_ref(nv->screen->fence.current, &buf->fence_wr);
 }
 
 
diff --git a/src/gallium/drivers/nouveau/nouveau_fence.c b/src/gallium/drivers/nouveau/nouveau_fence.c
index dea146c..c686710 100644
--- a/src/gallium/drivers/nouveau/nouveau_fence.c
+++ b/src/gallium/drivers/nouveau/nouveau_fence.c
@@ -189,16 +189,15 @@ nouveau_fence_wait(struct nouveau_fence *fence)
    /* wtf, someone is waiting on a fence in flush_notify handler? */
    assert(fence->state != NOUVEAU_FENCE_STATE_EMITTING);
 
-   if (fence->state < NOUVEAU_FENCE_STATE_EMITTED) {
+   if (fence->state < NOUVEAU_FENCE_STATE_EMITTED)
       nouveau_fence_emit(fence);
 
-      if (fence == screen->fence.current)
-         nouveau_fence_new(screen, &screen->fence.current, FALSE);
-   }
-   if (fence->state < NOUVEAU_FENCE_STATE_FLUSHED) {
+   if (fence->state < NOUVEAU_FENCE_STATE_FLUSHED)
       if (nouveau_pushbuf_kick(screen->pushbuf, screen->pushbuf->channel))
          return FALSE;
-   }
+
+   if (fence == screen->fence.current)
+      nouveau_fence_next(screen);
 
    do {
       nouveau_fence_update(screen, FALSE);
diff --git a/src/gallium/drivers/nouveau/nv50/nv50_vbo.c b/src/gallium/drivers/nouveau/nv50/nv50_vbo.c
index c6162b5..947c67d 100644
--- a/src/gallium/drivers/nouveau/nv50/nv50_vbo.c
+++ b/src/gallium/drivers/nouveau/nv50/nv50_vbo.c
@@ -597,6 +597,15 @@ nv50_draw_elements(struct nv50_context *nv50, boolean shorten,
 
       assert(nouveau_resource_mapped_by_gpu(nv50->idxbuf.buffer));
 
+      /* This shouldn't have to be here. The going theory is that the buffer
+       * is being filled in by PGRAPH, and it's not done yet by the time it
+       * gets submitted to PFIFO, which in turn starts immediately prefetching
+       * the not-yet-written data. Ideally this wait would only happen on
+       * pushbuf submit, but it's probably not a big performance difference.
+       */
+      if (buf->fence_wr && !nouveau_fence_signalled(buf->fence_wr))
+         nouveau_fence_wait(buf->fence_wr);
+
       while (instance_count--) {
          BEGIN_NV04(push, NV50_3D(VERTEX_BEGIN_GL), 1);
          PUSH_DATA (push, prim);