|Lars Engholm Johansen||Jun 16, 2014 2:40 am|
|André Warnier||Jun 16, 2014 8:56 am|
|Lars Engholm Johansen||Jun 19, 2014 6:10 am|
|Christopher Schultz||Jun 19, 2014 6:31 am|
|Konstantin Kolinko||Jun 19, 2014 6:36 am|
|André Warnier||Jun 19, 2014 2:44 pm|
|Filip Hanik||Jun 19, 2014 5:49 pm|
|Lars Engholm Johansen||Jun 26, 2014 6:56 am|
|André Warnier||Jun 26, 2014 8:09 am|
|Christopher Schultz||Jun 26, 2014 3:59 pm|
|Christopher Schultz||Jun 26, 2014 4:01 pm|
|Filip Hanik||Jun 27, 2014 8:36 am|
|Christopher Schultz||Jun 27, 2014 12:02 pm|
|Lars Engholm Johansen||Sep 18, 2014 1:30 am|
|Filip Hanik||Sep 18, 2014 10:03 am|
|Lars Engholm Johansen||Oct 6, 2014 2:11 am|
|Mark Thomas||Oct 6, 2014 5:14 am|
|Lars Engholm Johansen||Nov 10, 2014 1:57 am|
|Mark Thomas||Nov 12, 2014 3:15 am|
|Lars Engholm Johansen||Nov 18, 2014 5:18 am|
|Subject:||Re: Connection count explosion due to thread http-nio-80-ClientPoller-x death|
|From:||Lars Engholm Johansen (lar...@gmail.com)|
|Date:||Nov 18, 2014 5:18:35 am|
Thanks for looking further into this, Mark,
We are running:
java version "1.7.0_65" OpenJDK Runtime Environment (IcedTea 2.5.1) (7u65-2.5.1-4ubuntu1~0.12.04.2) OpenJDK 64-Bit Server VM (build 24.65-b04, mixed mode)
Linux 3.11.0-15-generic #25~precise1-Ubuntu SMP Thu Jan 30 17:39:31 UTC 2014 x86_64 x86_64 x86_64 GNU/Linux Intel(R) Xeon(R) CPU E5-1650 v2 @ 3.50GHz
The bug happened very seldom compared to the 100M text WebSocket messages a day, about once a week, so I guess it could be an external problem like the JVM.
I already added a temporary fix to our Tomcat installation as mentioned earlier, but other users could also suffer from this. How do we proceed?
Best regards, Lars Engholm Johansen
On Wed, Nov 12, 2014 at 12:15 PM, Mark Thomas <mar...@apache.org> wrote:
On 10/11/2014 09:57, Lars Engholm Johansen wrote:
I looked into our javax.websocket.Endpoint implementation and found the following suspicious code:
When we need to close the WebSocket session already in .onOpen() method (rejecting a connection), we are calling session.close() asynchronously after 1 second via a java.util.Timer task. This was due to bug https://issues.apache.org/bugzilla/show_bug.cgi?id=54716, that I can see should be fixed long time ago (thanks).
Can this cause the selector's keyset to be accessed by more than one
I don't see how.
I've just double checked the NIO Poller code and the only places the keyset is used is in Poller.run() and Poller.timeout() - both of which are only ever accessed from the Poller thread.
I've also looked over the run() and timeout() methods and haven't yet found anything that could trigger this.
There are multiple Pollers but each Poller has a distinct set of sockets to manage.
I'm beginning to wonder if there is a JVM bug here. Which JVM are you using?
Best regards, Lars Engholm Johansen
On Mon, Oct 6, 2014 at 2:14 PM, Mark Thomas <mar...@apache.org> wrote:
On 06/10/2014 10:11, Lars Engholm Johansen wrote:
I have good news as I have identified the reason for the devastating NioEndpoint.Poller thread death:
In rare circumstances a ConcurrentModification can occur in the Poller's connection timeout handling called from OUTSIDE the
java.util.ConcurrentModificationException at java.util.HashMap$HashIterator.nextEntry(HashMap.java:922) at java.util.HashMap$KeyIterator.next(HashMap.java:956) at
Somehow the Poller's Selector object gets modified from another thread.
Any idea how? I've been looking through that code for some time now (this stack trace appears to be from 7.0.55 for those that want to look at this themselves) and I can't see anywhere where the selector's keyset is accessed by more than one thread.
As a remedy until fixed properly by the Tomcat team, I have added a try-catch(ConcurrentModificationException) surrounding the for loop in Poller.timeout(). That way, in case of the rare problem, a full iteration of the Selector will be retried in the next call to Poller.timeout().
That seems like a reasonable work-around but before we start making changes to the Tomcat code I'd really like to understand the root cause(s) of the issue else we might not be fixing the actual issue and could make it worse for some folks.
I am really happy now as all our production servers have been rock stable for two weeks now.
Best regards to all, Lars Engholm Johansen
On Thu, Sep 18, 2014 at 7:03 PM, Filip Hanik <fil...@hanik.com> wrote:
Thanks Lars, if you are indeed experiencing a non caught error, let us know what it is.
On Thu, Sep 18, 2014 at 2:30 AM, Lars Engholm Johansen < lar...@gmail.com> wrote:
Thanks guys for all the feedback.
I have tried the following suggested tasks:
- Upgrading Tomcat to the newest 7.0.55 on all our servers -> Problem still persists - Force a System.gc() when connection count is on the loose -> Connection count is not dropping - Lowering the log level of NioEndpoint class that contains the Poller code -> No info about why the poller thread exits in any tomcat logs - Reverting the JVM stack size per thread to the default is discussed previously -> Problem still persists
I have now checked out the NioEndpoint source code and recompiled it with a logging try-catch surrounding the whole of the Poller.run() implementation as I noticed that the outer try-catch here only catches OOME. I will report back with my findings as soon as the problem arises again.
On Fri, Jun 27, 2014 at 9:02 PM, Christopher Schultz < chr...@christopherschultz.net> wrote:
On 6/27/14, 11:36 AM, Filip Hanik wrote:
Are there any log entries that would indicate that the poller thread has died? This/these thread/s start when Tomcat starts. and a stack over flow on a processing thread should never affect the poller thread.
OP reported in the initial post that the thread had disappeared:
On 6/16/14, 5:40 AM, Lars Engholm Johansen wrote:
We have no output in tomcat or our logs at the time when this event occurs. The only sign is when comparing full java thread dump with a dump from a newly launched Tomcat:
One of http-nio-80-ClientPoller-0 or http-nio-80-ClientPoller-1 is missing/has died.