Yuriy Sidelnikov created AMQ-5020:
-------------------------------------
Summary: Memory leak due to race condition in
VMTransportServer/VMTransport classes for spoke broker when hub is down.
Key: AMQ-5020
URL: https://issues.apache.org/jira/browse/AMQ-5020
Project: ActiveMQ
Issue Type: Bug
Components: Broker
Affects Versions: 5.9.0
Reporter: Yuriy Sidelnikov
When spoke is not connected for some reason to a hub(for example, the hub is
down) it is trying to reconnect to the hub with increasing timeout between
attempts. In attempt to reconnect VMTransportServer class creates linked pair
of VMTransport objects(client and server) and started them asynchronously to be
used in bridge between local and remote brokers. If hub is still down TCP
transport raise exception and then clean up process is trying to destroy this
pair of VMtransport objects.
However if due to race condition VMtransport objects have not been fully
initialized yet it is impossible to destroy them properly.
This excerpt describe how VMtransport tries to shutdown a peer:
try {
peer.transportListener.onCommand(new ShutdownInfo());
} catch (Exception ignore) {
}
// let any requests pending a response see an exception
try {
peer.transportListener.onException(new
TransportDisposedIOException("peer (" + this + ") stopped."));
} catch (Exception ignore) {
}
In some cases(high load, slow servers, etc.) due to race condition
peer.transportListener is not initialized yet on a moment of stopping so it
leads to NPE (which will be just ignored) and corresponded VMTransport is not
destroyed.
The issue can be fixed by attached patch.
--
This message was sent by Atlassian JIRA
(v6.1.5#6160)
