RFR: 8031686: G1: assert(_hrs.max_length() == _expansion_regions) failed

[Stefan Johansson]

On 2014-01-22 22:42, Jon Masamitsu wrote:
>
> On 1/22/2014 8:41 AM, Stefan Johansson wrote:
>> Hi,
>>
>> I've got some comments offline about the test and my initial 
>> assumption that more than one GC would be tested the way the test was 
>> written might have been wrong. I've now made the test a G1 only 
>> regression test, to specifically test the issue reported in this bug.
>
> Looks good.  What would you think about changing the
> name of the test to TestG1TwoGbHeap?  Just so it's
> obviously a G1 test.
The test resides in the test/gc/g1 folder and the other tests there 
don't have G1 in their names. I had it named TestG1LargeHeap for a while 
but changed it to be more similar to the other test names.

Stefan
>
> Jon
>>
>> New webrev:
>> http://cr.openjdk.java.net/~sjohanss/8031686/webrev.02/
>>
>> Thanks,
>> Stefan
>>
>> On 2014-01-17 15:23, Stefan Johansson wrote:
>>> Hi again,
>>>
>>> I've added a regression test for the issue. See the updated review at:
>>> http://cr.openjdk.java.net/~sjohanss/8031686/webrev.01/
>>>
>>> The test is naive and just tries to run -version with a heap size of 
>>> 2g. I don't explicitly set G1 as GC but make sure I will use the 
>>> arguments given to jtreg, this will ensure testing more than one GC 
>>> and also that we do not run with G1 on unsupported platforms.
>>>
>>> Please have a look at the test and let me know if you have any 
>>> questions or suggestions.
>>>
>>> Thanks,
>>> Stefan
>>>
>>> On 2014-01-16 14:43, Stefan Johansson wrote:
>>>> Hi,
>>>>
>>>> Please review this small fix for:
>>>> https://bugs.openjdk.java.net/browse/JDK-8031686
>>>>
>>>> Webrev:
>>>> http://cr.openjdk.java.net/~sjohanss/8031686/webrev.00/
>>>>
>>>> Summary:
>>>> The failing assertion is caused by an erroneous calculation done 
>>>> when setting up the regions array in HeapRegionSeq. When using a 
>>>> large heap a pointer calculation will overflow (since ptrdiff_t is 
>>>> signed) and this will in turn lead to the regions array being 
>>>> created longer that it should be. For non debug builds this will 
>>>> lead to increased memory usage and jmap will report a faulty value 
>>>> for the number of regions. The fix is to use the 
>>>> pointer_delta-function to do the calculation.
>>>>
>>>> Note: The bug is currently targeted for 8 but is defer-requested 
>>>> and should be moved to 9 shortly. I plan to fix it in 9 and 
>>>> backport it to 8u20.
>>>>
>>>> Testing:
>>>> * Manually verified that we are able to start 32-bit VMs with more 
>>>> than 2g of heap using a debug build.
>>>>
>>>> Thanks,
>>>> Stefan
>>>
>>
>