[ https://issues.apache.org/activemq/browse/AMQ-932?page=comments#action_37000 ] John Heitmann commented on AMQ-932: -----------------------------------
Those are the exact symptoms that occur as a result of the problem I saw. For every bad connection there was a bunch of individual object leakage (one each of the major transport filters, a couple different mbeans, a connection object etc), but by far the most serious leakage were the empty pre-allocated marshaler cache arrays. This looks like a common symptom of anything that would leak transports/connections, so it might not be the same root cause. If you can reproduce it and want help I'd be happy to take a look. If you can't find the problem one thing that could really help mitigate the severity of these kinds of leaks is to make the initialization of the marshaling caches conditional on the wireFormat.cacheEnabled flag. This would give a lot more leak headroom. > Quickly broken client connections lead to memory leaks > ------------------------------------------------------ > > Key: AMQ-932 > URL: https://issues.apache.org/activemq/browse/AMQ-932 > Project: ActiveMQ > Issue Type: Bug > Components: Broker > Affects Versions: 4.0.2 > Reporter: John Heitmann > Attachments: 73.diff > > > Connections to the openwire port that are pathologically broken (for example > any http request) or that die in some other way extremely quickly will lead > to memory losses of aout 64Kb each time. This happens because many services > are stop()ed directly in the middle of start(), and then never stopped for > real, or stopped again but on an object tree with an inconsistent state. This > is usually also accompanied by the JMX message: > WARN ManagedTransportConnection - Failed to unregister mbean: > org.apache.activemq:BrokerName=localhost,Type=Connection,ConnectorName=default,Connection=25 > But that is a cosmetic symptom and not critical (and this has otherwise > nothing to do with JMX). > My patch is a band-aid that is functional but I'm not very happy with it. The > patch changes some service logic so that if stop is called in the middle of > start, the stop is instead queued and called at the end of start. There will > still be multiple stops, and you'll still see the cosmetic JMX error from the > second ineffectual stop, but the first stop cleans up correctly so there are > no leaks. > I think there's probably a better solution, but it was tough to see what. I'd > appreciate better ideas. Possibly something involving moving the dangerous > operations (wire format negotiation etc) out of start? > I am working on a unit test, but I can't promise I will have something to > submit. I'm having to play JVM games to detect the problem in a unit test and > that might not fly for general purpose use. -- This message is automatically generated by JIRA. - If you think it was sent incorrectly contact one of the administrators: https://issues.apache.org/activemq/secure/Administrators.jspa - For more information on JIRA, see: http://www.atlassian.com/software/jira
