(original) (raw)

David Holmes <david.holmes@oracle.com> wrote:

Hi Jaroslav,

I still see some suspect uses of p.arrive() instead of 
p.arriveAndAwaitAdvance().

David

On 21/01/2014 11:14 PM, Jaroslav Bachorik wrote:
Based on the experience while fixing
https://bugs.openjdk.java.net/browse/JDK-8032377 I've updated the patch
not to require an exact number for the blocked count - it will pass
whenever the reported blocked count is at least large as the the
expected number.

Webrev: http://cr.openjdk.java.net/\~jbachorik/6309226/webrev.06

Thanks,

-JB-

On 20.12.2013 09:24, Jaroslav Bachorik wrote:
On 20.12.2013 03:27, David Holmes wrote:
Sorry for the delay - still catching up after vacation (only 435 openjdk
emails left :( ).

NP


On 9/12/2013 9:46 PM, Jaroslav Bachorik wrote:
On 19.11.2013 19:55, David Holmes wrote:
On 20/11/2013 12:43 AM, Jaroslav Bachorik wrote:
Hi David,

On 18.11.2013 22:03, David Holmes wrote:
Hi Jaroslav,

I think your phaser usage is incorrect:

88 public void run() {
89 p.arriveAndAwaitAdvance(); // phase\[1\]
90 synchronized(lock1) {
91 System.out.println("\[LockerThread obtained
Lock1\]");
92 p.arrive(); // phase\[2\]
93 }
94 p.arriveAndAwaitAdvance(); // phase\[3\]
95 }

Here the current thread can release itself at line 94\. You have
assumed
that the phase\[2\] arrive will be a trigger to release the main
thread
but it may not have reached its arriveAndAwaitAdvance phase\[2\]
statement
yet, so the current thread arrives then does arrive-and-wait but the
number of arrivals is 2 so it doesn't wait.

A CyclicBarrier per phase may be clearer.

Yes, thanks for pointing this out. But, wouldn't
p.arriveAndAwaitAdvance() instead of p.arrive() suffice? I've
tried to
replace Phaser with CyclicBarrier but while it seems to work as
expected
it pollutes code with the necessity to catch various exceptions
(InterruptedException, BarrierException etc.). So, if the simple fix
with Phaser would do the trick I would better use that.

In this case yes arriveAndAwaitAdvance would fix it. I don't know if
other tests have the same issue - the concern would be ensuring
there's
no possibility of deadlocks in general.

I've updated the webrev with the one using p.arriveAndAwaitAdvance() at
line 92 - http://cr.openjdk.java.net/\~jbachorik/6309226/webrev.04

Are you ok with this change then?

I think you have the same problem anywhere that arrive() is used eg:

129 p.arrive(); // phase\[2\]
130 p.arriveAndAwaitAdvance(); // phase\[3\]

and

185 p.arrive(); // phase\[2\]
186 }
187 p.arriveAndAwaitAdvance(); // phase\[3\]

Very probable :( Also, when analysing the recurrence of JDK-8029890 I've
found out that Phaser.arriveAndAwaitAdvance() actually blocked the
thread in a way that it increased the number of contentions reported :(
CyclicBarrier seems to wait instead - I will have to use CyclicBarrier
when testing the number of reported contentions and Phaser when testing
the number of reported waits :/

-JB-


David
------

Thanks,

-JB-



Thanks,
David

-JB-


David

On 18/11/2013 7:59 PM, Jaroslav Bachorik wrote:
Hi,

after discussing this with Mandy I've rewritten the test to use the
j.u.concurrent for synchronization - this also makes it much
easier to
follow the test logic.

The waited time, the blocked time and the waited counts are only
checked
for sanity (increasing values) since it is not possible to do the
reliable checks against hard numbers.

I ran the test in a tight loop for 1500 times using -Xcomp and
-Xint
and
the test seems to pass constantly.

New webrev: http://cr.openjdk.java.net/\~jbachorik/6309226/webrev.03

Thanks,

-JB-


On 21.10.2013 13:55, Jaroslav Bachorik wrote:
Please, review this small patch for a test failing due to the
updated
implementation in JDK6.

Issue: https://bugs.openjdk.java.net/browse/JDK-6309226
Webrev: http://cr.openjdk.java.net/\~jbachorik/6309226/webrev.00/

The test fails due to the change in mustang where
ThreadMXBean.getThreadInfo().getWaitedTime() and
ThreadMXBean.getThreadInfo().getWaitedCount() include
Thread.sleep()
too. Unfortunately, Thread.sleep() is used throughout the test for
synchronization purposes and this breaks the test.

In the patch I propose to replace Thread.sleep() with busy wait
and
hinting the scheduler by Thread.yield(). While not very elegant it
successfully works around inclusion of unknown number of
Thread.sleep()s
(they are called in loop).

Thanks,

-JB-