Unfortunately this patch is not safe: theoretically ->detach can call
bdrv_unapply_subtree_drain, and if it polls, will can call a bh that
for example reads the graph, finding it in an inconsistent state, since
it is between the two writes QLIST_REMOVE(child, next_parent); and
QLIST_REMOVE(child, n
Doing the opposite can make ->detach() (more precisely
bdrv_unapply_subtree_drain() in bdrv_child_cb_detach) undo the subtree_drain
just performed to protect the removal of the child from the graph,
thus making the fully-enabled assert_bdrv_graph_writable fail.
Note that assert_bdrv_graph_writable