[jboss-jira] [JBoss JIRA] (WFLY-3652) Network connection leak

[Stuart Douglas (JIRA)]

    [ https://issues.jboss.org/browse/WFLY-3652?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13006276#comment-13006276 ] 

Stuart Douglas commented on WFLY-3652:
--------------------------------------

Which side of the network stack is half closed in this case? I assume the server has closed its side, and the client is keeping it open? 

By default Undertow does not close the request stream when the response stream is closed, as it is perfectly valid to continue to process request data after the response has already been written (although probably not that common).

I could change the behaviour to force close (i.e. RST) the request when the response is done, but this has some negative consequences. Force closing when async servlets are in use could potentially be a solution, although really the application should just be calling AsyncContext.complete() when the request is done.

> Network connection leak
> -----------------------
>
>                 Key: WFLY-3652
>                 URL: https://issues.jboss.org/browse/WFLY-3652
>             Project: WildFly
>          Issue Type: Bug
>          Components: Web (Undertow)
>    Affects Versions: 8.1.0.Final
>         Environment: Linux 2.6.38-16-server
> Java(TM) SE Runtime Environment (build 1.7.0_45-b18)
>            Reporter: Jan Vanhercke
>            Assignee: Stuart Douglas
>             Fix For: 9.0.0.Alpha1
>
>
> When using Asynchronous servlets and AsyncListeners for long polling we observe a connection leak in the undertow subsystem.
> Heap dumps show a large number of org.xnio.io.NioSocketConduit, io.undertow.server.protocol.http.HttpServerConnection and related objects.
> However, since the effective number of connections is far less, nearly all AsyncContext instances we find are in a complete state and lsof output returns a large number of sockets with 'can't identify protocol' entries indicating that sockets are kept open by the JVM but are in fact half closed by the network stack.
> Not all connections appear to be leaking, but over time, depending on the load, the server instance fills up. 



--
This message was sent by Atlassian JIRA
(v6.3.1#6329)