On Wed, 7 Jun 2023 23:17:18 GMT, Chris Plummer <cjplum...@openjdk.org> wrote:
> The test waits for a ThreadDeathEvent for "main". Once that arrives, it then
> waits for the next ThreadStartEvent (for any thread). Once it arrives, the
> test tries to create and enable a StepRequest on the "main" thread. Since
> "main" is supposedly dead, the expectation is an IllegalThreadStateException.
> However, it turns out that sometimes the enabling can in fact succeed.
>
> Just because a ThreadDeathEvent has been received for a thread does not mean
> you can no longer do things with the thread like suspend it or enable a
> StepEvent. There is a short delay in the debug agent after sending the
> ThreadDeathEvent before it stops tracking the thread. The thread can still be
> acted upon until then. The JDWP and JDI specs seem to support doing this:
>
>> Notification of a completed thread in the target VM. The notification is
>> generated by the dying thread before it terminates. Because of this timing,
>> it is possible for {@link VirtualMachine#allThreads} to return this thread
>> after this event is received.
>>
>> Note that this event gives no information about the lifetime of the thread
>> object. It may or may not be collected soon depending on what references
>> exist in the target VM.
>
> What this means is that when the test receives some arbitrary
> ThreadStartEvent after the "main" ThreadDeathEvent has been received, the
> test may in fact still be able to enable a StepRequest on the "main" thread,
> causing the test to fail. What seems to trigger the failure is receiving an
> unexpected spurious ThreadStartEvent such as from the Common-Clean thread or
> a carrier thread, although even then it only fails some of the time. In fact
> if I modify the test to enable the StepRequest when it receives the
> ThreadDeathEvent for "main", it still almost always passes, but will fail
> more frequently than it normally does.
>
> It seems if the test always waits for the ThreadStartEvent for
> "DestroyJavaVM", then the "main" thread is truly gone by then and the test
> always passes, so this is how I've chosen to fix the issue.
>
> Tested with tier1, tier2 svc tests, and tier5 svc tests.
I think it makes sense!
-------------
Marked as reviewed by kevinw (Committer).
PR Review: https://git.openjdk.org/jdk/pull/14372#pullrequestreview-1470630717
