[jira] Commented: (HARMONY-3647) [drlvm][thread] Issues with native threads creation and termination

["Li-Gang Wang (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/HARMONY-3647?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#action_12489669 ] 

Li-Gang Wang commented on HARMONY-3647:
---------------------------------------

Gregory, thanks for your attention on this issue. I think the problem still exists, but I need more time to test. Currently my solution is to use the patches in the given URL links as fixes (or workrounds since I am not sure if they are the true fixes).

Actually there is one more things with native thread that has associated java.lang.Thread object, i.e., be treated as Java thread when it runs Java methods. The issue is, in current implementation when I attach a java thread object with a native thread, the j.l.Thread object constructor sets its contextClassLoader to be NULL, and this thread by default is in the same ThreadGroup with main thread. 

The problem arises with test in kernel.test "java.lang.ThreadTest", which enumerates the threads in the group and accesses their contextClassloaders. Then the NULL contectClassLoader in the native threads causes failures. 

This issue is addressed in JIRA H3671 by assigning a loader to the object forcefully, which I guess is only a workaround. Please have a look at it  https://issues.apache.org/jira/browse/HARMONY-3671.  

Thanks.



> [drlvm][thread] Issues with native threads creation and termination
> -------------------------------------------------------------------
>
>                 Key: HARMONY-3647
>                 URL: https://issues.apache.org/jira/browse/HARMONY-3647
>             Project: Harmony
>          Issue Type: Bug
>          Components: DRLVM
>         Environment: Win/Linux
>            Reporter: Li-Gang Wang
>
> While I tested the native finref (finalizer & ref enqueue) thread implementation, I met two issues about native thread creation and termination. I have only temporary workarounds for them currently. The basic issue is how to create/terminate a native thread that has associated java.lang.Thread object.
> 1.        On Win & Linux, jvmti.test failed occasionally
> CAUSE: We couldn't attach finref threads as Java threads atomically with the according interface, attach_current_thread(). The attaching action may contend some lock with the main thread. When we ran jvmti.test, if the main thread held the lock firstly, finref threads had to run Thread class's constructor before they were associated to a Java thread structure. This would incur failure.
> The related issue is on Jira is https://issues.apache.org/jira/browse/HARMONY-3578.
> In this patch I make the main thread wait until all finref threads finish their attaching actions.
>  
> 2.    On Win, gc.FinAlloc test hung after it printed "PASSED" once per tens of times.
> CAUSE: In vm_destroy(), finref threads are signaled to terminate and then the main thread terminates itself. But the termination of native threads is not atomic, so sometimes finref threads were not terminated properly and held a shared lock after termination. In this situation when the main thread was calling Thread.detach() method to detach itself, it could not hold the lock and was blocked.
> The related issue is on Jira is https://issues.apache.org/jira/browse/HARMONY-3588.
> In this patch, when VM signals finref threads to terminate, it will wait until the termination action finishes.
> With the exploration, I suspect that current design in DRLVM might require this kind of native thread (that has associated j.l.thread object) to be treated as real Java thread, and created in Java code. That is, it doesn't allow the VM to simply create a native thread to be associated with a j.l.thread object. We need this kind of thread because we need them to run Java methods (the finalizers). I guess  this should be a bug in current design, since the main thread must be created from native thread but able to run main() method of the entry class.

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.