Acceptor is dead with lot's of Broken pipe exceptions in error log

Hi,

could you possibly try 6.1.26, as we made some changes to the selector manager that could improve handling of closing connections (JETTY-1296, JETTY-748, JETTY-1264). While none of these are directly related to a Broken Pipe, they may improve the selectors ability to recover from such and error.

How frequently does this happen?

Interestingly, Glassfish/Grizzly has had a similar issue reported: http://java.net/jira/browse/GRIZZLY-813, so we may have a look at their fix to see if this is an NIO issue.

The Grizzly fix was simply to catch an additional IOException and the cancel the key or close the exception.

6.1.26 has already improved the exception handling around that area, so this may be fixed.
If you can reproduce with 6.1.26, then we should similarly be able to cancel the key and re-register.

I'm running 6.1.26 currently. It is reproduceable. BUT it seems that bug is reproduceable only if I set acceptors to 2, can't reproduce bug with acceptors set to 4.

Today bug was reproduced again with acceptors set to 4...

System Debian-5
Kernel: 2.6.26-2-amd64
Java: java version "1.6.0_24"
Java(TM) SE Runtime Environment (build 1.6.0_24-b07)
Java HotSpot(TM) 64-Bit Server VM (build 19.1-b02, mixed mode)

Additional java opts: -Xmx6144m
Load is (all info is in seconds):
avg: 0,0009, min: 0,0001, max: 3,4729, count: 449395, perSec: 748,9917

Also some config options:

<New class="org.mortbay.thread.QueuedThreadPool">
<Set name="minThreads">10</Set>
<Set name="maxThreads">300</Set>
<Set name="lowThreads">20</Set>
<Set name="SpawnOrShrinkAt">10</Set>
</New>

<Call name="addConnector">
<Arg>
<New class="org.mortbay.jetty.nio.SelectChannelConnector">
<Set name="host"><SystemProperty name="jetty.host" /></Set>
<Set name="port"><SystemProperty name="jetty.port" default="8080"/></Set>
<Set name="maxIdleTime">60000</Set>
<Set name="Acceptors">4</Set>
<Set name="acceptQueueSize">12000</Set>
<Set name="statsOn">false</Set>
<Set name="confidentialPort">8443</Set>
<Set name="lowResourcesConnections">10000</Set>
<Set name="lowResourcesMaxIdleTime">7000</Set>
</New>
</Arg>
</Call>

Jetty works behind nginx.
Nginx configured to accept 20 000 connections.

I've had another report about this issue. I will try to add some explicit handling into the server for it. As it appears the selector is dead or dying, I think the solution will have to remake a new selector (as we've done for other JVM bugs). Question is, will we lose existing connections?

Also I wonder why we've not had this reported on jetty-7?

I'm looking and better handling IOException in our wakeup code, but the problem is that IOException is not actually thrown by selector.wakeup. Can you tell from the log if the exception was caught and logged by jetty, or was it printed out by the JVM library itself?

It was caught and logged by jetty.
Also we have migrated to jetty-7 already and it seems that there is no such problem with it... So currently will be not able to reproduce/test this bug... Jetty-7 is running for four days without problems already.

We're seeing this under high SOAP load with Jetty 6.1.22 (embedded in our server).

583675662 2011-03-17 16:17:43,413 [30949593@qtp-10365435-9 - Acceptor0 SelectChannelConnector@10.64.24.157:8019] WARN  org.mortbay.log  - EXCEPTION 
java.io.IOException: Broken pipe (errno:32)
                at sun.nio.ch.PollArrayWrapper.interrupt(Native Method)
                at sun.nio.ch.PollArrayWrapper.interrupt(PollArrayWrapper.java:175)
                at sun.nio.ch.PollSelectorImpl.wakeup(PollSelectorImpl.java:84)
                at org.mortbay.io.nio.SelectorManager$SelectSet.wakeup(SelectorManager.java:681)
                at org.mortbay.io.nio.SelectorManager$SelectSet.doSelect(SelectorManager.java:563)
                at org.mortbay.io.nio.SelectorManager.doSelect(SelectorManager.java:189)
                at org.mortbay.jetty.nio.SelectChannelConnector.accept(SelectChannelConnector.java:124)
                at org.mortbay.jetty.AbstractConnector$Acceptor.run(AbstractConnector.java:706)
                at org.mortbay.thread.QueuedThreadPool$PoolThread.run(QueuedThreadPool.java:520)

This stack trace certainly looks similar to the GRIZZLY-813 issue.

This continually occurs for an hour and then these errors start to show up as we run out of file descriptors.

587316759 2011-03-17 17:18:24,510 [30949593@qtp-10365435-9 - Acceptor0 SelectChannelConnector@10.64.24.157:8019] WARN  org.mortbay.log  - EXCEPTION 
java.io.IOException: Too many open files (errno:24)
                at sun.nio.ch.ServerSocketChannelImpl.accept0(Native Method)
                at sun.nio.ch.ServerSocketChannelImpl.accept(ServerSocketChannelImpl.java:145)
                at org.mortbay.jetty.nio.SelectChannelConnector$1.acceptChannel(SelectChannelConnector.java:75)
                at org.mortbay.io.nio.SelectorManager$SelectSet.doSelect(SelectorManager.java:540)
                at org.mortbay.io.nio.SelectorManager.doSelect(SelectorManager.java:189)
                at org.mortbay.jetty.nio.SelectChannelConnector.accept(SelectChannelConnector.java:124)
                at org.mortbay.jetty.AbstractConnector$Acceptor.run(AbstractConnector.java:706)
                at org.mortbay.thread.QueuedThreadPool$PoolThread.run(QueuedThreadPool.java:520)

Java(TM) 2 Runtime Environment, Standard Edition (build 1.5.0.14-_25_jul_2008_05_19)
Java HotSpot(TM) Server VM (build 1.5.0.14 jinteg:07.25.08-11:41 IA64, mixed mode)

One little note: we have migrated to jetty 7 and we have no more such issue...

Does anyone know if there is a fix for this that could be put into Jetty 6?

I don't know for sure if it is a fix, but I've committed some defensive code to jetty-6 trunk. If anybody that suffers from this bug could try it out and confirm that it works, then I would do a maintenance release of jetty-6

As no-one has commented since May, I'm assuming that Greg's fixes in 6.1 trunk are OK. If this is not the case, then please reopen this issue, providing stack traces etc and we will take another look at it, hopefully before 6.1.27 goes out the door.

Jan