RFR: 8253877: gc/g1/TestGCLogMessages.java fails - missing "Evacuation failure" message

[Thomas Schatzl]

On Tue, 13 Oct 2020 12:44:04 GMT, Stefan Johansson <sjohanss at openjdk.org> wrote:

>> Hi all,
>> 
>>   can I have reviews for this change to gc/g1/TestGCLogMessages to hopefully make it more reliable?
>> 
>> One subtest tries to force evacuation failure, but apparently this does not always succeed. Particularly with upcoming
>> young gen sizing changes, there seems to be a larger than before chance that the current mechanism does not work.
>> This mechanism assumes that before that test the young gen has been sized that much that the allocation of a half-heap
>> humongous object causes an evacuation failure next time - of course that's a somewhat dodgy assumption that apparently
>> recently started failing more.  The change employs the G1 internal evacuation failure debugging mechanism to force this
>> kind of situation. The advantage is that this is guaranteed to work, the disadvantage is that it relies on that
>> feature.  Additionally I did some slight renaming to the helper that used the term "ToSpaceExhaustion" instead of
>> "EvacuationFailure" which is the name of the message that is actually checked for.
>> Testing: 2k+ tests of previous version without reproduction (meaning that at least for me without context of other
>> tests, the issue has not been reproducable, at least not with the mentioned 1% failure rate), 3k tests with new version
>> without reproduction  Thanks,
>>   Thomas
>
> I think it is good to make use of the debug feature to ensure stable testing and I'm approving this change.
> 
> I guess an approach if we don't want to use it is to just have an allocation loop that keeps everything live, it should
> trigger evacuation failures before throwing OOME.

Exactly the catch-OOME approach seemed too error-prone. We may not recover from the OOME handler.

-------------

PR: https://git.openjdk.java.net/jdk/pull/624