<AWT Dev> OpenJdk11-28-EA JDialog hanging

[Martin Balao]

Hi all,

I was not sure if my runs of SequencedEventTest test were triggering the
problematic execution flow, because this issue should manifest even after
JDK-8152974 [1] fix. After a few tries I attached a debugger. Attaching a
debugger obviously interferred with robot steps on the test but, anyways,
handled it and reproduced the problem.

I had a quick look and this is my hypothesis for JDK-8204142 [2] issue:

Let's say we have EDT0 (EventQueue0, AppContext0) and EDT1 (EventQueue1,
AppContext1). Both threads are dispatching events from their respective
EventQueue queues. In particular, dispatching SequencedEvent events. We
also have SequencedEvent.list list, in which SequencedEvent events from
both AppContext0 and AppContext1 are interleaved. These SequencedEvent
events will be dispatched by both EDT0 and EDT1 (EDT0 dispatching
AppContext0 events from EventQueue0 and EDT1 dispatching AppContext1 events
from EventQueue1).

Now, we may have the following sequence of events in the
SequencedEvent.list queue: EV0-AppContext0, EV0-AppContext1,
EV1-AppContext1, EV1-AppContext0. Events, though, may come to EventQueue
queues out of order (I.e: EventQueue1 may have EV1-AppContext1 first and
EV0-AppContext1 then, while EventQueue0 may have EV0-AppContext0 first and
EV1-AppContext0 then).

In the previous scenario, EDT1 will try to dispatch EV1-AppContext1 first.
However, it notices that it is not the first event in the
SequencedEvent.list queue. After JDK-8152974 [1] fix, EDT1 will try to
dispatch other SequencedEvents from his EventQueue queue. So it goes
straight to EV0-AppContext1, which is the next one in EventQueue1. It
notices that this is not the first event in SequencedEvent.list queue and
tries to do the same. However, no more SequencedEvents are available for
him to dispatch (no more SequencedEvents in EventQueue1). It will then
block indefinitely waiting for SequencedEvents to come to EventQueue1. Now
it's time for EDT0. EDT0 will successfully dispatch EV0-AppContext0 and
then try to dispatch EV1-AppContext0. When trying to dispatch the latter,
it will notice that it is not the first one in SequencedEvent.list queue
and blocks the same way EDT1 did before. Now we have both EDT0 and EDT1
blocked waiting for SequencedEvent events to come to their respective
EventQueue queues. EDT1 should have woken up and proceeded dispatching
EV0-AppContext1.

The underlying problem is that SequencedEvent.dispatch function is not
prepared for SequencedEvent.list list to have events from different
EventQueue queues, dispatched by different EDTs. The assumption that if the
event is not the first one on SequencedEvent.list list we will successfully
be able to dispatch an event from the EventQueue and unblock the sequence
is wrong. We may wait for this event indefinitely and, at some point, we
will be the ones blocking the whole thing because our event now is the
first one. Please note here that the wake-up scheme in
SequencedEvent.dispose function fails because it's sending SentEvent events
and the waiting thread (after JDK-8152974 [1]) is filtering SequencedEvent
events.

Kind regards,
Martin.-

--
[1] - https://bugs.openjdk.java.net/browse/JDK-8152974
[2] - https://bugs.openjdk.java.net/browse/JDK-8204142
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.openjdk.java.net/pipermail/awt-dev/attachments/20181010/5291c01c/attachment.html>