Request for Review (s) - 8060463: ParNew: SurvivorAlignmentInBytes greater then YoungPLABSize cause assert(obj != NULL || plab->words_remaining() < word_sz) failed: Else should have been able to allocate
jon.masamitsu at oracle.com
Thu Mar 24 17:26:38 UTC 2016
On 3/24/2016 9:13 AM, sangheon wrote:
> Hi Jon,
> On 03/24/2016 09:00 AM, Jon Masamitsu wrote:
>> Thanks Sangheon, Thomas, and Michail. Latest versions:
> Thank you for fixing these.
> BTW, could you add a comma in front of 'PLAB size' as well?
> And as the test is a new one, I think just '2016' seems correct not
> '2014, 2016'.
> I don't need a new webrev for these.
>> On 3/20/2016 8:04 PM, Jon Masamitsu wrote:
>>> If SurvivorAlignmentInBytes is larger than the size of the promotion
>>> LAB an allocation aligned to SurvivorAlignmentInBytes may fail. The
>>> allocation of the PLAB depends of the size of the object being
>>> allocated and if the PLAB is successfully allocated the allocation
>>> of the object should succeeded (and is checked by an assertion).
>>> For alignments larger than PLAB, a larger PLAB is sometimes needed
>>> to allow for the alignment.
>>> The fix does not try to avoid the need for a new PLAB at the
>>> next promotion by making the PLAB even larger. The large
>>> values of SurvivorAlignmentInBytes are allowed but not
>>> expected. I chose not to make the PLAB even larger for a
>>> situation I don't expect to see in practice.
>>> Tested with the newly added test which check a range
>>> of SurvivorAlignmentInBytes values.
More information about the hotspot-gc-dev