Currently, undertow-servlet triggers Session.requestDone(...) after the response is flushed, but also after all listeners associated with exchange have already been invoked.
That we invoke Session.requestDone(...) after the response is flushed is a good thing, since the cost of session persistence/replication does not affect response times.
However, since this happens outside the context of any listenable events, e.g. ExchangeCompletionListener, HttpRequestListener, handler chain wrappers, etc.), it becomes nearly impossible (without extreme workarounds) for containers like WildFly to detect when it is safe to suspend the server.