Hi Daniel,
On Friday 11 November 2016 10:12 PM, Daniel Fuchs wrote:
Hi Harsha,
// We reached here because the executor was shutdown.
// If executor was supplied by client, then it was shutdown
// abruptly or JMXConnector was shutdown along with executor
// while this thread was suspended at L564.
if (!(executor instanceof LinearExecutor)) {
// Spawn new executor that will do cleanup if JMXConnector is closed
// or keep notif system running otherwise
executor = new LinearExecutor();
....
I find it a bit strange to 'keep notif system running' if
the executor was shutdown.
It may happen that user will not be aware of shutdown sequence
(client->server->executor) and may shutdown executor first. In that
case, we may need to keep notif system running.
If the user that supplied the executor shuts it down,
then does it make sense to continue running? Also
if the old executor has already been shutdown - and you
create a new one - then do we have any guarantee that
the new executor will be shut down properly?
Old executor will be shutdown by the user. We are only swapping a user
supplied executor with new linear executor. It will follow the same
shutdown sequence as original one.
I wonder whether you should instead perform the cleaning
action directly in the catch clause - that is - copy
the lines 553-562 over there, or maybe simply
call this::doRun() - without submitting 'this'
to an executor... But then we would have to have
some guarantee that doRun() will terminate and
not call doRun() again (hence my suggestion to
copy lines 553-562 instead)...
Well, that will not cover the case where user shuts down executor but
keeps the client/server alive. So we will need a executor that can keep
notif system running as well as do clean-up if client/server is closed.
This code is very hairy - with states and multiple
critical sections and wait() and notifyAll(), so it
is difficult to reason about, and I can't guarantee
that it would be the right thing to do...
But maybe it would be worth prototyping it and see if
it also passes all the non-reg tests?
Yes Daniel. It is a precarious section of code loosely held together by
states. However, I think the changes are not too intrusive since it does
not rely on states. It only looks for RejectedExecutionException. I have
written test-case that can validate my changes. I haven't JPRT yet but
would be happy to share test results if required.
best regards,
-- daniel
On 11/11/16 16:02, Harsha Wardhana B wrote:
Hello,
Please review the fix for
issue : https://bugs.openjdk.java.net/browse/JDK-8141591
webrev : http://cr.openjdk.java.net/~hb/8141591/webrev.00/
Fix details:
The root-cause for the issue is that NotifFetcher thread was suspended
at L562 after which main thread ran to completion shutting down both
client and server and stopping the executor. Fixing as
synchronized (ClientNotifForwarder.this) {
if (state == STARTED) {
executor.execute(new NotifFetcher());
}
}
will not solve the problem since NotifFetcher can get suspended after
checking state==STARTED and still hit the same problem.
One way to solve would be to catch RejectedExecutionException and then
fallback on linearExecutor. This will take care of necessary cleanup
if client/server are shutdown or keeps the system running if executor
is abrupty shutdown by the client without stopping the client/server.
The comments section of the issue also has discussion about various ways
of fixing the issue and merits/demerits of different approaches. It also
helps in getting a perspective on issue and the fix.
Thanks
Harsha
Thanks
Harsha
