In a previous post about nproc limit, I wrote that I had to investigate the nproc limit with the number of threads because my Oracle 12c EM agent was having thousands of threads. This post is a short feedback about this issue and the way I have found the root cause. It concerns the enterprise manager agent 12c on Grid Infrasctructure >= 126.96.36.199
The issue was:
ps -o nlwp,pid,lwp,args -u oracle | sort -n NLWP PID LWP COMMAND 1 8444 8444 oracleOPRODP3 (LOCAL=NO) 1 9397 9397 oracleOPRODP3 (LOCAL=NO) 1 9542 9542 oracleOPRODP3 (LOCAL=NO) 1 9803 9803 /u00/app/oracle/product/agent12c/core/188.8.131.52.0/perl/bin/perl /u00/app/oracle/product/agent12c/core/184.108.40.206.0/bin/emwd.pl agent /u00/app/oracle/product/agent12c/agent_inst/sysman/log/emagent.nohup 19 11966 11966 /u00/app/11.2.0/grid/bin/oraagent.bin 1114 9963 9963 /u00/app/oracle/product/agent12c/core/220.127.116.11.0/jdk/bin/java ... emagentSDK.jar oracle.sysman.gcagent.tmmain.TMMain
By default ps has only one entry per process, but each processes can have several threads – implemented on linux as light-weight process (LWP). Here, the NLWP column shows that I have 1114 threads for my EM 12c agent – and it was increasing every day until it reached the limit and the node failed (‘Resource temporarily unavailable’).
The first thing to do is to know what those threads are. The ps entries do not have a lot of information, but I discovered jstack which every java developer should know, I presume. You probably know that java has very verbose (lengthy) stack traces. Jstack was able to show me thousands of them in only one command:
$ jstack 9963 2014-06-03 13:29:04 Full thread dump Java HotSpot(TM) 64-Bit Server VM (20.14-b01 mixed mode): "Attach Listener" daemon prio=10 tid=0x00007f3368002000 nid=0x4c9b waiting on condition [0x0000000000000000] java.lang.Thread.State: RUNNABLE "CRSeOns" prio=10 tid=0x00007f32c80b6800 nid=0x3863 in Object.wait() [0x00007f31fe11f000] java.lang.Thread.State: TIMED_WAITING (on object monitor) at java.lang.Object.wait(Native Method) at oracle.eons.impl.NotificationQueue.internalDequeue(NotificationQueue.java:278) - locked (a java.lang.Object) at oracle.eons.impl.NotificationQueue.dequeue(NotificationQueue.java:255) at oracle.eons.proxy.impl.client.base.SubscriberImpl.receive(SubscriberImpl.java:98) at oracle.eons.proxy.impl.client.base.SubscriberImpl.receive(SubscriberImpl.java:79) at oracle.eons.proxy.impl.client.ProxySubscriber.receive(ProxySubscriber.java:29) at oracle.sysman.db.receivelet.eons.EonsMetric.beginSubscription(EonsMetric.java:872) at oracle.sysman.db.receivelet.eons.EonsMetricWlm.run(EonsMetricWlm.java:139) at oracle.sysman.gcagent.target.interaction.execution.ReceiveletInteractionMgr$3$1.run(ReceiveletInteractionMgr.java:1401) at java.util.concurrent.ThreadPoolExecutor$Worker.runTask(ThreadPoolExecutor.java:895) at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:918) at oracle.sysman.gcagent.util.system.GCAThread$RunnableWrapper.run(GCAThread.java:184) at java.lang.Thread.run(Thread.java:662) ...
I don’t paste all of them here. We have the ‘main’, we have a few GCs and ‘Gang workers’ which are present in all JVMs and we have a few enterprise manager threads. And what was interesting was that I had thousands of “CRSeOns” that seemed to be increasing.
Some guesses: I’m on RAC, and I have a ‘ons’ resource and the EM agent tries to subscribe to it. Goggle search returned nothing, and that’s the reason I put that in a blog post now. Then I searched MOS, and bingo, there is a note: Doc ID 1486626.1. It has nothing to do with my issue, but has an interesting comment in it:
In cluster version 18.104.22.168 and higher, the ora.eons resource functionality has been moved to EVM. Because of this the ora.eons resource no longer exists or is controlled by crsctl.
It also explains how to disable EM agent subscription:
emctl setproperty agent -name disableEonsRcvlet -value true
I’m in 22.214.171.124 and I have thousands of threads related to a functionality that doesn’t exist anymore. And that leads to some failures in my 4 nodes cluster.
The solution was simple: disable it.
For a long time I have seen a lot of memory leaks or CPU usage leaks related to the enterprise manager agent. With this new issue, I discovered a thread leak and I also faced a SR leak when trying to get support for the ‘Resource temporarily unavailable’ error, going back and forth between OS, Database, Cluster and EM support teams…