RFR (XXS): 8005857: assert in GC_locker from PSOldGen::expand with -XX:+PrintGCDetails and Verbose

Thomas Schatzl thomas.schatzl at oracle.com
Fri Mar 29 01:51:54 PDT 2013


Hi all,

  please review the following change for 8005857:

In this issue, code checks whether the gc locker is active during heap
expansion to print a message that instead of gc the heap has been expanded.

The code used the method GCLocker::is_active() to get the gc locker
state, but is_active() asserts when the VM is currently not at a
safepoint and no GC is actually needed.
This assert does not hold in all cases while expanding the heap. E.g. heap
expansion may be done outside of a safepoint when satisfying an allocation
request for a large object. Heap expansion may occur also when no GC is
pending in the gc locker.

Imho the previous use of is_active() was wrong not only because of the assert.
The message text notifies of the avoidance of a gc. However, the condition did
not check whether the gc locker indicated the need for a gc, i.e. did not
check the GCLocker::needs_gc() predicate.

This change uses GCLocker::is_active_and_needs_gc() instead of
GCLocker::is_active(). Which does not check whether the vm is at a safepoint,
and also checks whether a gc is actually pending.

Jon M. suggested this fix.

JIRA:
https://jbs.oracle.com/bugs/browse/JDK-8005857

bugs.sun:
http://bugs.sun.com/view_bug.do?bug_id=8005857

Testing:
jprt, kitchensink, runthese jck tests (which contains the test case
where this bug has been reported)

Thanks,
Thomas






More information about the hotspot-gc-dev mailing list