RFR: 8290732: JNI DestroyJavaVM can start shutdown when one non-daemon thread remains

[David Holmes]

On Tue, 9 Aug 2022 09:26:44 GMT, Thomas Stuefe <stuefe at openjdk.org> wrote:

>> Please review this fix for a problem discovered by @stuart-marks in the course of examining the VM shutdown behaviour. The VM code assumed that only unattached threads called JNI's DestroyJavaVM and so they were always attached as non-daemon threads. But it is perfectly valid to call DestroyJavaVM from an already attached thread, which could be a daemon. The fix simply checks whether the caller is a daemon or not and adjusts the expected count of active threads to see. There is also an adjustment to the thread termination logic to also notify at the right time.
>> 
>> Thanks to @stuart-marks for the reproducer in JBS - the longest part of this by many hours was converting the test over the jtreg. :)
>> 
>> Testing:
>>  - the new test on all core platforms
>>  - tiers 1-3
>> Thanks.
>
> test/hotspot/jtreg/runtime/jni/daemonDestroy/TestDaemonDestroy.java line 36:
> 
>> 34:  * @run main/native TestDaemonDestroy
>> 35:  * @run main/native TestDaemonDestroy daemon
>> 36:  */
> 
> If you do two @test sections, you can give them speaking names(e.g. @test id=demon) and they can run in parallel.

I didn't think that was warranted/necessary here.

> test/hotspot/jtreg/runtime/jni/daemonDestroy/TestDaemonDestroy.java line 59:
> 
>> 57:         System.out.println("Launcher = " + launcher +
>> 58:                            (Files.exists(launcher) ? " (exists)" : " (missing)"));
>> 59: 
> 
> Nit: better arg handling?
> - If you want to just have one char arg, pass 'D' and 'N' for demon/nondemon mode instead of "1" and "".
> - Instead of using args.length, maybe actually parse the input argument, or specify D and N there too and pass it on.

I reworked it slightly so the Java side doesn't pass "" and the native code interprets any additional arg as "use daemon".

-------------

PR: https://git.openjdk.org/jdk/pull/9803