Thomas,
On 3/25/22 16:59, Thomas Hoffmann (Speed4Trade GmbH) wrote:
-----Ursprüngliche Nachricht-----
Von: Christopher Schultz <ch...@christopherschultz.net>
Gesendet: Freitag, 25. März 2022 14:05
An: users@tomcat.apache.org
Betreff: Re: AW: Question to possible memory leak by Threadlocal variable
Thomas,
On 3/24/22 05:49, Thomas Hoffmann (Speed4Trade GmbH) wrote:
-----Ursprüngliche Nachricht-----
Von: Mark Thomas <ma...@apache.org>
Gesendet: Donnerstag, 24. März 2022 09:32
An: users@tomcat.apache.org
Betreff: Re: Question to possible memory leak by Threadlocal variable
On 24/03/2022 07:57, Thomas Hoffmann (Speed4Trade GmbH) wrote:
<snip/>
Is it correct, that every spawned thread must call tl.remove() to
cleanup all
the references to prevent the logged warning (and not only the main
thread)?
Yes. Or the threads need to exit.
Second question is: How might it cause a memory leak?
The threads are terminated and hold a reference to this static
variable. But
on the other side, that class A is also eligible for garbage
collection after undeployment.
So both, the thread class and the class A are ready to get garbage
collected. Maybe I missed something (?)
It sounds as if the clean-up is happening too late. Tomcat expects
clean-up to be completed once contextDestroyed() has returned for all
ServLetContextListeners. If the clean-up is happening asynchronously
(e.g.
the call is made to stop the threads but doesn't wait until the
threads have
stopped) you could see this message.
In this case it sounds as if you aren't going to get a memory leak
but Tomcat can't tell that at the point it checks.
Mark
---------------------------------------------------------------------
To unsubscribe, e-mail: users-unsubscr...@tomcat.apache.org
For additional commands, e-mail: users-h...@tomcat.apache.org
Hello Mark,
thanks for the information.
The shutdown of the framework is currently placed within the destroy()
method of a servlet (with load on startup).
At least the debugger shows that servlet-->destroy() is executed before
the method checkThreadLocalMapForLeaks() runs.
I will take a look, whether the threads already exited.
Tomcat only checks its own request-processing threads for ThreadLocals, so
any threads created by the application or that library are unrelated to the
warning you are seeing.
Any library which saves ThreadLocals from request-processing threads is
going to have this problem if the objects are of types loaded by the webapp
ClassLoader.
There are a few ways to mitigate this, but they are ugly and it would be
better if the library didn't use ThreadLocal storage, or if it would use vanilla
classes from java.* and not its own types.
You say that those objects are eligible for GC after the library shuts down,
but that's not true: anything you stick in ThreadLocal storage is being held ...
by the ThreadLocal storage and won't be GC'd. If an object can't be collected,
the java.lang.Class defining it can't be collected, and therefore the
ClassLoader which loaded it (the webapp
ClassLoader) can't be free'd. We call this a "pinned ClassLoader" and it still
contains all of the java.lang.Class instances that the ClassLoader ever loaded
during its lifetime. If you reload repeatedly, you'll see un-collectable
ClassLoader instances piling up in memory which is
*definitely* a leak.
The good news for you is that Tomcat has noticed the problem and will, over
time, retire and replace each of the affected Threads in its request-
processing thread pool. As those Thread objects are garbage-collected, the
TheradLocal storage for each is also collected, etc. and *eventually* your leak
will be resolved. But it would be better not to have one in the first place.
Why not name the library? Why anonymize the object type if it's
org.apache.something?
-chris
Hello Chris,
I didn't want to blame any library 😉 But as you ask for it, I send more details.
Regarding the ThreadLocal thing:
I thought that the threadlocal variables are stored within the
Thread-class in the member variable "ThreadLocal.ThreadLocalMap
threadLocals":
https://github.com/AdoptOpenJDK/openjdk-jdk11/blob/master/src/java.base/share/classes/java/lang/Thread.java
So I thought, when the thread dies, these variables will also be
released and automatically removed from the ThreadLocal variable /
instance (?)
This is correct, but if the ThreadLocal is being stored in the
request-processing thread, then when your web application is redeployed,
the request processing threads outlive that event. Maybe you thought
your application gets a private set of threads for its own use, but
that's not the case: Tomcat pools threads across all applications
deployed on the server. You can play some games to segregate some
applications from others, but it's a lot of work for not much gain IMO.
Since the threads outlive the application, you can see the problem, now.
I considered the ThreadLocal class as just the manager of the
thread's member variable "threadLocals".
Basically, yes.
Regarding the library:
The full log-message is:
12-Mar-2022 15:01:16.302 SCHWERWIEGEND [Thread-15]
org.apache.catalina.loader.WebappClassLoaderBase.checkThreadLocalMapForLeaks
The web application [ROOT] created a ThreadLocal with key of type
[java.lang.ThreadLocal.SuppliedThreadLocal] (value
[java.lang.ThreadLocal$SuppliedThreadLocal@2121cbad]) and a value of type
[org.apache.camel.impl.DefaultCamelContext.OptionHolder] (value
[org.apache.camel.impl.DefaultCamelContext$OptionHolder@338d0413]) but failed
to remove it when the web application was stopped. Threads are going to be
renewed over time to try and avoid a probable memory leak.
>
The blamed class is this version:
https://github.com/apache/camel/blob/camel-3.14.x/core/camel-core-engine/src/main/java/org/apache/camel/impl/DefaultCamelContext.java
Interesting that Camel is storing a ThreadLocal. Maybe there is a better
way to use Camel in the context of a web application?
Within our app we have a startup servlet with:
Servlet-> init: context = new DefaultCamelContext();
Servlet -> destroy: context.stop();
The stop-method will call the doStop() method of this class (via the
class hierarchy DefaultCamelContext --> SimpleCamelContext -->
AbstractCamelContext). >
After the destroy-method is executed, all spawned threads of Camel
are stopped / vanished. There is also no log entry, that some
orphaned threads exist when undeploying the app.
So I don’t know, what's the mistake within this class. What would be
the right way to clean up the ThreadLocal variable? Just stopping the
threads didn’t seem to clean it up properly.
The saved ThreadLocal was done from within one of the request-processing
threads that Tomcat owns. This wasn't a thread spawned by the library,
which is likely to already be cleaned-up when stop() is completed, as
expected.
It looks like Camel may be capturing some values and storing them in
ThreadLocal when it doesn't make sense (in a web application) to do so.
Are you able to instrument your application to see when those
ThreadLocals are set?
-chris
---------------------------------------------------------------------
To unsubscribe, e-mail: users-unsubscr...@tomcat.apache.org
For additional commands, e-mail: users-h...@tomcat.apache.org
