Paul McKenney convinced me that there's no *guarantee* that all archs
will refuse the speculate the atomic ops above the state test, eg:

CPU0 (stopref_set_state)		CPU1 (stopref)

 atomic_set(&ack, 0);			if (state == XXX)
 wmb();						atomic_inc(&ack);
 state = XXX;

Certainly Alpha needs a rmb() inside stopref to guarantee it sees
the same ordering.