RFR(xs): 8134749: SoftReferences declared dead too early
per.liden at oracle.com
Thu Mar 10 14:55:22 UTC 2016
On 2015-09-01 12:57, Per Liden wrote:
> Hi Jon,
> On 2015-08-31 20:20, Jon Masamitsu wrote:
>> Change looks good.
>> Could you please run some tests with your fix and
>> -XX:+UseDynamicNumberOfGCThreads and
>> -XX:+UseParallelGC threads? UseDynamicNumberOfGCThreads sometimes moves
>> work among the queues to get them into queues that will be processed by
>> an active
>> GC worker. Thanks.
> Thanks for looking at the patch. I'll do some test runs with those options.
Apologies for the delay, I thought I had replied, but is seems this bug
completely fell of my radar. Anyway, I just re-ran all tonga tests
related to reference processing, with the following flag combinations:
1) -XX:+UseG1GC -XX:ParallelGCThreads=8 -XX:+UseDynamicNumberOfGCThreads
2) -XX:+UseG1GC -XX:ParallelGCThreads=8 -XX:-UseDynamicNumberOfGCThreads
3) -XX:+UseParallelGC -XX:ParallelGCThreads=8
4) -XX:+UseParallelGC -XX:ParallelGCThreads=8
They all passed.
I also revisited the code and I can't see anything problematic
interactions with UseDynamicNumberOfGCThreads.
>> On 08/31/2015 07:15 AM, Per Liden wrote:
>>> While preparing a different patch I looked through the
>>> ReferenceProcessor and noticed what looks like a silent mistreatment
>>> of SoftReferences.
>>> Summary: RefProcPhase1Task uses WorkerThread::id() to select the list
>>> to process. After the recent changes to how WorkGang works, this no
>>> longer guarantees that all reference lists are processed. This in turn
>>> means we might kill SoftReferences even when the policy says they
>>> should be kept alive.
>>> The fix here would be to use the worker_id (named "i" in this code)
>>> passed into the Task's work() function instead of using
>>>  There's no longer any guarantee that all GangWorker threads will
>>> execute the Task's work() function, since a quick-to-wakeup worker
>>> thread might steal the work intended for a slow-to-wakeup worker thread.
>>> Webrev: http://cr.openjdk.java.net/~pliden/8134749/webrev.0/
>>> Bug: https://bugs.openjdk.java.net/browse/JDK-8134749
>>> Testing: jprt
>>> Btw, I'll to send out a follow-up patch (JDK-8134744) which makes sure
>>> WorkerThread::id() and the worker_id always match to avoid this type
>>> of bugs in the future.
More information about the hotspot-gc-dev