Hi Ivan,
On 03/27/2014 08:26 AM, Ivan Gerasimov wrote:
David, Mandy, thank you for comments!
Here's what we want to achieve in the test:
Two EarlyTimeout child threads should both be blocked in
remove(TIMEOUT) at the moment the weakReference gets enqueued.
This is the situation, when the bug 6853696 manifests itself.
If we made sure the weakReference is enqueued before remove() is
called by the child threads, then they both would wait for specified
amount of time even without the fix.
That's why we make the main thread wait for the children to start
before calling System.gc().
I understand it may seem strange at the first glance.
But we want the child thread to be inside the remove() function when
the main thread is unblocked, that's why it is done this way.
Unfortunately, we cannot easily make sure that remove() was called by
the child threads at the moment the GC decides to enqueue the
weakReference.
There could be a little delay (say half the timeout: 500ms) specified
after main thread returns from the startedSignal.await(); and before
setting referent = null; and doing System.gc(). This would decrease the
chance that the reference is enqueued before EarlyTimeout threads enter
queue.remove(1000), thus making the test more reliable in failing with
unpatched code.
Now even if the referent is released and System.gc() is called, that
does not guarantee that a WeakReference is going to be enqueued before
the EarlyTimeout threads timeout and the result could as well be 0
collected references. To increase the chance that the reference is
enqueued in a timely manner, main thread could, immediately after
System.gc(), call:
SharedSecrets.getJavaLangRefAccess().tryHandlePendingReference();
(since SharedSecrets is in sun.misc protected package, JavaLangRefAccess
instance would have to be obtained using reflection unfortunately).
Therefore, the test is probabilistic by its nature, and we cannot be
absolutely sure that at least one child thread will get non-null
reference returned from remove().
I'm withdrawing the second webrev: While moving startedSignal.await()
after System.gc() reduces the number of cases when nonNullRefCount ==
0, it also greatly decreases the chances of occurring the situation
the test was designed for.
I believe the correct fix here is to ignore the (quite rare) situation
when nonNullRefCount upon the completion.
I suggest to return to the very first trivial fix:
http://cr.openjdk.java.net/~igerasim/8038333/1/webrev/
But this webrev *is moving startedSignal.await() after System.gc()* ...
Regards, Peter
Sincerely yours,
Ivan
On 27.03.2014 4:30, Mandy Chung wrote:
On 3/26/2014 3:57 PM, David Holmes wrote:
Ivan,
I think the problem is that the EarlyTimeout threads can complete
their remove(TIMEOUT) before the main thread has started them all,
cleared the reference and called System.gc().
Depending on exactly what is being tested, the EarlyTimeout threads
may need to wait on another latch that is signalled by the main
thread after the gc() call returns. Still no guarantee that the gc
will do its work before the timeout elapses.
This is similar to what I have been thinking. I believe the
EarlyTimeout threads don't need the startedSignal.countDown; instead
it can have a signal latch with 1 count. The EarlyTimeout threads
awaits on the signal latch. The main thread will call
signal.countDown after System.gc() and we can add a check
weakReference.isEnqueued to make sure before we awake the threads to
proceed the queue.remove call.
Ivan - for the error message, perhaps you can simply do this:
if (nonNullRefCount != 1) {
throw new RuntimeException(nonNullRefCount + " references
were removed from queue");
}
Mandy
David
On 27/03/2014 6:18 AM, Ivan Gerasimov wrote:
Thank you Mandy!
Are you able to reproduce the test failure?
Yes, I could easily reproduce the failure when I reduced the
timeout to
10 ms.
With the timeout reduced, the test fails every third time on my
machine.
I think the test verifies that only one thread gets the reference
is a
good test.
But if none of the threads could get the reference, it should not
cause
the failure of the test.
It only means that during this particular run we could not have tested
what we needed.
We could retry, but I'm not sure it's worth complicating the test.
It's easier to ignore the failure, taking into account that it happens
rarely.
I think the race is due to the threads get to call queue.remove as
soon as both threads decrement the count of the latch that can be
well
before the reference is enqueued.
The problem is that we have no way to block the main thread until
there
is a reference in the queue.
I improved the situation a bit, having moved the await() after the
call
of gc().
It'd be good to add additional information in the test to help
diagnosing test failure.
I added reporting to stderr about being unable to remove a reference
from the queue.
I believe we shouldn't treat it as an error, and should simply
ignore it.
Would you please have a look at the updated webrev:
http://cr.openjdk.java.net/~igerasim/8038333/1/webrev/
Sincerely yours,
Ivan
