Prevent deadlock in EventManager.trigger()

[goodsoft]

If a new event observer is registered during event execution, a deadlock occurs due to locked observersLock.

[goodsoft]

Mmkay, seems it's failing some tests.

[johncarl81]

I'm surprised to hear about deadlock wtih the EventManager, as I have a test (albeit, nondeterministic) that focuses on just that: https://github.com/johncarl81/transfuse/blob/master/transfuse-api/src/test/java/org/androidtransfuse/event/UnregisterConcurrencyTest.java#L189

[goodsoft]

Here's a distilled test case for deadlock:

        eventManager.register(Event.class, new EventObserver<Event>() {
            public void trigger(Event event) {
                eventManager.register(Event.class, otherEventObserver); // deadlock!
            }
        });
        eventManager.trigger(new Event());

[johncarl81]

Hmm, this should be protected against by a reentrant lock...

[goodsoft]

Apparently writeLock() called by register() waits for readLock() set by trigger() to be released.

[goodsoft]

Additionally, a writer can acquire the read lock, but not vice-versa, states ReentrantReadWriteLock class documentation

[johncarl81]

We could move triggerQueue() to below the finally block. The side effect of this would be that you could unregister from an event and still receive events (see https://github.com/johncarl81/transfuse/blob/master/transfuse-api/src/test/java/org/androidtransfuse/event/UnregisterConcurrencyTest.java#L162)

[goodsoft]

Yeah, I have done exactly that, and that has failed exactly this test. Not sure how to circumvent this problem

[johncarl81]

Recalling a conversation I had with some of the Otto devs, if it isn't a big deal that we might see events after calling unregister, we can probably get rid of that lock altogether.

[johncarl81]

@goodsoft: #175