Hi Jaroslav,
On 11/19/2013 6:23 AM, Jaroslav Bachorik wrote:
Please, review this test fix.
Issue : https://bugs.openjdk.java.net/browse/JDK-6987597
Webrev: http://cr.openjdk.java.net/~jbachorik/6987597/webrev.03
The fix moves the execution of the test routine to the
GarbageCollectorMXBean notification handler - at that moment it's
certain that GC has already happened and it is safe to assert the data
provided by the MBean.
This patch may have a potential issue when GC happens during the process
of adding the notification handlers such as the number of increments is
greater than the number of registered handlers. By the time the main
thread reaches p.arriveAndAwaitAdvance, the phaser has advanced to the
next phase with few arrived parties and will wait forever (until the
next GC happens). This should rarely happen though. Note that
p.arriveAndAwaitAdvance continue to run even if the thread is
interrupted and what happens if jtreg attempts to time out the test?
Does this issue only happen to CMS with background GC thread? The
proposed patch seems a little overkill. I wonder if the test should
skip if running in CMS background mode. Does
-XX:+ExplicitGCInvokesConcurrent flag will get System.gc() to run in
foreground mode in CMS (I think that may get the GC to complete before
System.gc() returns?)
Mandy
