HASingletons can be stopped on all cluster members.

org.jboss.ha.singleton.HASingletonSupport#_stopOldMaster is called locally on each node when a new view is processed and it is not the master.
It is also called via an asynchronous cluster RPC from the new master.

The RPC call is not synchronized, and can cause a race condition when the master node changes quickly. The _stopOldMaster call from the previous master can arrive after the new master has taken over, causing it to stop again.