logging-log4j-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Leon Finker (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (LOG4J2-1457) Class loader deadlock when using async logging
Date Wed, 06 Jul 2016 20:52:10 GMT

    [ https://issues.apache.org/jira/browse/LOG4J2-1457?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15365057#comment-15365057
] 

Leon Finker commented on LOG4J2-1457:
-------------------------------------

I was able to reproduce it consistently in debugger. I've set a conditional breakpoint in:
	at org.apache.logging.log4j.core.util.Loader.initializeClass(Loader.java:241)
The condition being that the className.contains the class name that has the static initializer.

Another breakpoint was in the static initializer of the class in question. I also added a
for loop to make sure to fill up the disruptor buffer from initializer (-DAsyncLogger.RingBufferSize=512):
 static {
        Exception e = new Exception();
        for(int i =0; i<1024; ++i) {
            logger.error(e);
        }
    }

Once the breakpoint in class static initializer was hit, I let it log the exception once and
kept it suspended. This caused the breakpoint in initializeClass to be hit. I kept it suspended.
I then switched back to the thread with static initializer and let it go through the loop.
This filled up the buffer and blocked the class initializer. I then resumed the thread with
initializeClass. They have both deadlocked.

Then I have made local log4j2 changes to load class without initializer. The deadlock doesn't
happen any longer. So this does solve the issue. Even if the other thread is still in breakpoint
in class initializer, the call to Class.forName(className, false, loader); still succeeds
and doesn't block anymore.

I now realize what caused the bug to show up. Apparently there were two exceptions logged
from the static initializer and the second exception log happened just at the time when the
ring buffer was full. This caused the deadlock.



> Class loader deadlock when using async logging
> ----------------------------------------------
>
>                 Key: LOG4J2-1457
>                 URL: https://issues.apache.org/jira/browse/LOG4J2-1457
>             Project: Log4j 2
>          Issue Type: Bug
>    Affects Versions: 2.6.1
>         Environment: On CentOS 6.7 and Java 1.8.0_60.
>            Reporter: Leon Finker
>            Priority: Critical
>         Attachments: threaddump.txt
>
>
> We've encountered a class loading deadlock. Please review attached thread dump. Is it
possible to have an option of pre-initializing the exception's thread stack on the caller's
thread? It's hard to predict what libraries are doing in their classes' static initializers
and may eventually end up logging and causing deadlock.
> In the attached thread dump here are the threads of interest:
> "Log4j2-AsyncLogger[AsyncContext@18b4aac2]1" #16 daemon prio=5 os_prio=0 tid=0x00007ff870c7b000
nid=0x79f3 in Object.wait() [0x00007ff839142000]
>    java.lang.Thread.State: RUNNABLE
>         at java.lang.Class.forName0(Native Method)
> ...
> and
> "1A03340:Company:japan" #568 prio=5 os_prio=0 tid=0x00007ff871677000 nid=0x725 runnable
[0x00007ff74bd27000]
> ...<clinit>...



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

---------------------------------------------------------------------
To unsubscribe, e-mail: log4j-dev-unsubscribe@logging.apache.org
For additional commands, e-mail: log4j-dev-help@logging.apache.org


Mime
View raw message