RFR: JDK-8061259: ParNew promotion failed is serialized on a lock

Jungwoo Ha jwha at google.com
Mon Nov 3 17:16:01 UTC 2014


Nice catch!
PTAL
http://cr.openjdk.java.net/~rasbold/8061259/webrev.03/

On Thu Oct 30 2014 at 1:33:04 PM Bengt Rutisson <bengt.rutisson at oracle.com>
wrote:

>
> Hi Jungwoo,
>
>
> On 10/30/14 6:24 PM, Jungwoo Ha wrote:
>
> PTAL
> http://cr.openjdk.java.net/~rasbold/8061259/webrev.02/
>
>
> Thanks! Looks good except for one detail.
>
> 1125   bool set_promotion_failed()   { _has_promotion_failed = 1; }
> 1126   bool reset_promotion_failed() { _has_promotion_failed = 0; }
>
> Since _has_promotion_failed is now a bool I don't think we should be
> assigning 1 and 0 to it. We should be using true and false.
>
> Other than that it looks good to me.
>
> Thanks!
>
> Bengt
>
>
>
>
> On Thu Oct 30 2014 at 12:28:19 AM Bengt Rutisson <
> bengt.rutisson at oracle.com> wrote:
>
>>
>> Hi again,
>>
>> One more minor thing.
>>
>> The methods has_promotion_failed(), set_promotion_failed() and
>> reset_promotion_failed() are protected but they could be made private
>> instead.
>>
>> Thanks,
>> Bengt
>>
>>
>> On 2014-10-30 08:09, Bengt Rutisson wrote:
>>
>>
>>
>> Hi Jungwoo,
>>
>> On 2014-10-29 23:51, Jungwoo Ha wrote:
>>
>>  PTAL
>> http://cr.openjdk.java.net/~rasbold/8061259/webrev.01/
>>
>>  I've looked a bit at the webrev. A couple of comments:
>>>
>>> Why do you use OrderAccess methods for writing and reading the
>>> _has_promo_failed flag in has_promo_failed() and set_promot_failed() ?
>>>
>>
>>   I think that has no effect on x86, but I assumed that processors with
>> weak memory model may want ordering of set/reset/has call.
>>
>>
>> You don't need the OrderAccess methods for the weak memory models here
>> either. You just race on reading the variable and if you see the "wrong"
>> value you eventually take a lock (which will order all memory accesses) to
>> read the variable properly.
>>
>> By removing the use of OrderAccess you can make
>> ConcurrentMarkSweepGeneration::_has_promotion_failed a bool instead of a
>> juint which simplifies the code a bit.
>>
>>
>>  Can we write out the full word "promotion" instead of just "promo" in
>>> the variables and methods?
>>>
>>
>>   Done.
>>
>>   Can we change the name of the flag from UseCMSFastPromotionFailure to
>>> CMSFastPromotionFailure? Most CMS flags start with CMS and I don't think we
>>> need the "Use" prefix.
>>>
>>
>>   Done.
>>
>>  What do you think about making the flag true by default? At least for
>>> JDK 9. If we decide to backport to JDK 8 or 7 it might be a good idea to
>>> keep the default value as false.
>>>
>>
>>   Done.
>> Let me know if there is anything for me to do to backport to JDK8 and 7.
>>
>>
>> I think this fix would be worth backporting to JDK 8. I don't think there
>> is much action required on your side. I created a backport bug for JDK 8
>> just to make sure that we don't forget it. It will be a little while before
>> the 8 update repos are in a state to accept enhancements again. So, it is
>> nice to have the backport bug to keep track of this.
>>
>> Backporting to JDK 7 requires some more work. Unless you have good
>> arguments for why it is important to backport this to JDK 7 I don't think
>> it is worth doing.
>>
>>
>>  Did you find the information provided by _fast_promo_failure_hitcount
>>> useful for debugging? If it not too useful I would consider removing it
>>> since it is cluttering up the code a bit.
>>>
>>
>>   I removed it.
>> It was useful to for development, but I think it is no longer needed.
>>
>>
>> Great. Thanks.
>>
>> One more comment. This code comment appears in two places just after we
>> have taken the lock.
>>
>> 3365   if (CMSFastPromotionFailure && has_promotion_failed()) {
>> 3366     // Caller must have checked already without synchronization.
>> 3367     // Check again here while holding the lock.
>> 3368     return NULL;
>> 3369   }
>>
>> There is actually really no requirement that the caller must have checked
>> has_promotion_failed() before calling the method. That's just an
>> optimization. I think the first comment can be skipped and we just leave
>> the second comment "// Check again here while holding the lock.". I would
>> also suggest moving that comment up to the line just before the if
>> statement.
>>
>> Thanks,
>> Bengt
>>
>>
>>
>>
>>
>>
>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.openjdk.java.net/pipermail/hotspot-gc-dev/attachments/20141103/bfcb2af6/attachment.html>


More information about the hotspot-gc-dev mailing list