In both NAKACK and UNICAST, we have atomic booleans, which are consulted before trying to acquire the lock. So if a lock is released eagerly in a ProtocolStack.down() call, although the lock is released, the atomic var isn't, and so subsequent threads will still not be able to process different messages.
A simple solution would be to release the lock and the atomic var in down().
OTOH, we should see whether we can use either locks or atomic vars, but not both.

This will need to be redesigned anyway with scopes in place (https://jira.jboss.org/jira/browse/JGRP-822).