[
https://issues.apache.org/jira/browse/AMQ-6937?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16440868#comment-16440868
]
Tomas Pavelka commented on AMQ-6937:
------------------------------------
I have discovered that the CPU spin loop can happen even on Linux: whenever the
process runs out of file descriptors the accept enters a loop that spins the
CPU at 100%. I have attached a patch that slows down exception handling in such
case.
> Recycling TCP/IP stack on z/OS causes an infinite error loop in transport
> server
> --------------------------------------------------------------------------------
>
> Key: AMQ-6937
> URL: https://issues.apache.org/jira/browse/AMQ-6937
> Project: ActiveMQ
> Issue Type: Bug
> Components: Broker
> Affects Versions: 5.15.3
> Reporter: Tomas Pavelka
> Priority: Major
> Attachments: AMQ-6937-cpu-spin-prevention.patch
>
>
> The ActiveMQ transport servers (e.g. TcpTransportServer) run the socket
> accept (java.net.ServerSocket#accept) in an infinite loop. The accept call
> can repeatedly fail with an exception spinning the CPU at full speed an
> filling up logs quickly.
> Here is an example of an exception that gets repeated indefinitely:
> java.net.SocketException: EDC5122I Input/output error. (Accept failed)
> at java.net.ServerSocket.implAccept(ServerSocket.java:623)
> at java.net.ServerSocket.accept(ServerSocket.java:582)
> at
> org.apache.activemq.transport.tcp.TcpTransportServer.run(TcpTransportServer.java:351)
> at java.lang.Thread.run(Thread.java:785)
> This is a common problem on z/OS because the pattern of running accept in a
> loop is used in many open source projects. For example, here is the same
> issue in Derby:
> https://issues.apache.org/jira/browse/DERBY-5347
> And here in Jetty:
> [https://github.com/eclipse/jetty.project/issues/283]
> Whenever the problem appears the socket becomes unusable. Would it be
> possible for ActiveMQ to allow to insert a custom
> org.apache.activemq.transport.TransportAcceptListener that would detect the
> problem and do a re-bind on the socket?
>
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
