On Mon, 6 Nov 2023 20:04:44 GMT, Serguei Spitsyn <sspit...@openjdk.org> wrote:
>> Is the issue here that agent thread started by Debuggee.checkStatus is
>> racing with the test? The producer/consumer thread do 1000 put/take ops and
>> it looks like it can complete and VM commence shutdown while the agent
>> thread is observing, do I read this correctly?
>
> Thank you for the comments.
> Alex, you are right.
> I've misread this code at line 201:
>
> 199 err = jvmti->SuspendThread(vthread);
> 200 if (err == JVMTI_ERROR_THREAD_NOT_ALIVE) {
> 201 continue;
> 202 }
> 203 check_jvmti_status(jni, err, "Error in SuspendThread");
> 204 // LOG("Agent: suspended vt: %s ct: %s\n", vname, cname);
> 205
> 206 check_vthread_consistency_suspended(jvmti, jni, vthread);
>
> The function `check_vthread_consistency_suspended()` is not called when if
> the `SuspendThread` returned `JVMTI_ERROR_THREAD_NOT_ALIVE` error code.
>
> It does not look as a test bug then.
David and Alan,
Yes, there is always a race between the app and the JVMTI agent execution. This
was always a general issue with the JVMTI. The agent code executing JVMTI
functions and events is not protected from the `JVMTI_ERROR_WRONG_PHASE`
errors. This error code can be returned by JVMTI functions even in the middle
of the event callback execution. We had this problem in the JDWP agent and
fixed it by adding proper synchronization into the agent itself.
-------------
PR Review Comment: https://git.openjdk.org/jdk/pull/16488#discussion_r1383923778
