RFR: 8041626: [Event Request] Shutdown reason

David Holmes david.holmes at oracle.com
Fri Feb 9 04:22:01 UTC 2018

Hi Robin,

On 9/02/2018 1:50 AM, Robin Westberg wrote:
> Hi David,
>> On 8 Feb 2018, at 04:28, David Holmes <david.holmes at oracle.com> wrote:
>> Hi Robin,
>> Adding in hotspot-runtime-dev as all the hotspot changes belong to runtime.
> Thanks, sorry about that..
>> I had an initial look through this.
>> To be honest I don't like it. We seem to have to record little bits of information all over the place just so they can be reported by the shutdown event. It seems untidy. :(
>> Can you rename _starting_thread to _main_thread please. The use of "starting" in thread.hpp/cpp is really a naming anomaly. The main thread is the thread that loads the JVM. And that would be consistent with set_exception_in_main_thread.
>> Though why do we care if the main thread exited with an unhandled exception? And if we do care, why do we only care when the shutdown reason is ""No remaining non-daemon Java threads"?
>> I don't like the need to add os::get_abort_exit_code. Do we really need it? What useful information does it convey?
> Right, almost all the runtime changes are made in order to try to figure out what the process exit code from the launcher will eventually be. For example, the launcher returns 1 if the main thread exited with an unhandled exception, but 0 otherwise. But I actually agree that this information is probably only of marginal use (it could always be captured from wherever Java is launched if someone really wants it), so it is perhaps not worth the trouble.

Yes and that particular example of launcher behaviour is an archaic 
throwback to C programming - where end of "main" means end of the 
program - IMHO. I don't think it is of interest - and certainly not 
worth jumping through so many hoops to make the info available.

> Discussed it a bit with Erik Gahlin, and perhaps the best option here is to simply remove the status code field from the event, that would simplify things and make the code you mention above go away.

That sounds good to me. :)

>> It is unfortunate that you need to add beforeHalt to deal with the event mechanism itself being turned off (?) by a shutdown hook. That would seem to potentially lose a lot of event information given hooks can run in arbitrary order and execute arbitrary Java code. And essentially you end up recording the initial reason shutdown started, though potentially it could end up terminating for a different reason.
> In this case I think it actually conveys useful information if you are trying to diagnose an unexpected shutdown. It should be useful to find the initial request of an orderly shutdown, even if something else happens during the shutdown sequence like a finalizer calling exit (in which case you could possibly end up with two shutdown events, but both may contain interesting information).

Yes that is useful. But I find the need to code things the way they are, 
unfortunate. But given current constraints, so be it.


> Best regards,
> Robin
>> Let's see what others think ...
>> Thanks,
>> David
>> On 8/02/2018 1:18 AM, Robin Westberg wrote:
>>> Hi all,
>>> Please review the following change that adds an event-based tracing event that is generated when the VM shuts down. The intent is to capture shutdowns that occur after the VM has been properly initialized (as initialization problems would most likely mean that the tracing framework hasn’t been properly started either).
>>> Issue: https://bugs.openjdk.java.net/browse/JDK-8041626
>>> Webrev: http://cr.openjdk.java.net/~rwestberg/8041626/webrev.00/
>>> Testing: hs-tier1,hs-tier2,jdk-tier1,jdk-tier2
>>> Best regards,
>>> Robin

More information about the core-libs-dev mailing list