Using fence->status to determine whether or not there are callbacks
remaining on the sync_fence is racy since fence->status may have been
decremented to 0 on another CPU before fence_check_cb_func() has
completed.  By unconditionally calling fence_remove_callback() for each
fence in the sync_fence, we guarantee that each callback has either
completed (since fence_remove_callback() grabs the fence lock) or been
removed.
Signed-off-by: Andrew Bresticker <abrestic@chromium.org>
Signed-off-by: Dmitry Torokhov <dtor@chromium.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
 static void sync_fence_free(struct kref *kref)
 {
        struct sync_fence *fence = container_of(kref, struct sync_fence, kref);
-       int i, status = atomic_read(&fence->status);
+       int i;
 
        for (i = 0; i < fence->num_fences; ++i) {
-               if (status)
-                       fence_remove_callback(fence->cbs[i].sync_pt,
-                                             &fence->cbs[i].cb);
+               fence_remove_callback(fence->cbs[i].sync_pt, &fence->cbs[i].cb);
                fence_put(fence->cbs[i].sync_pt);
        }