See https://issues.apache.org/jira/browse/PROTON-2112

"If an epoll proactor is dealing with file descriptor resource starvation at the time of shutdown, it can try to initiate new IO in conflict with other shutting down activity. Specifically, when freeing a connection's socket, it tries to resume any pause listeners.

I suspect there is a missing check somewhere to see if proactor->shutting_down is set before attempting the resume. But it may require a more thorough dismantling of the overflow listener list before commencing to tear down straggling connections."