RFR: Fix TestUnifiedLoggingSwitchStress test failure

[Aleksey Shipilev]

On 06/01/2018 06:24 PM, Zhengyu Gu wrote:
> 
> 
> On 06/01/2018 12:02 PM, Aleksey Shipilev wrote:
>> On 06/01/2018 05:20 PM, Zhengyu Gu wrote:
>>> The test failed due to degenerated cycle, which takes over task queue from earlier cycle. Resetting
>>> task queue statistics under these circumstances, results assertion failures during
>>> print_statistics().
>>>
>>> Webrev: http://cr.openjdk.java.net/~zgu/shenandoah/degen_tq_stats/webrev.00/
>>
>> What is the assert? Maybe we should do what G1 is doing, namely:
> 
> 428   TASKQUEUE_STATS_ONLY(print_taskqueue_stats());
> 
> will invoke TaskQueueStats::verify() and fail following assert:
> 
>  assert(get(push) == get(pop) + get(steal),
>          "push=" SIZE_FORMAT " pop=" SIZE_FORMAT " steal=" SIZE_FORMAT,
>          get(push), get(pop), get(steal));

So this is because taskqueues are not quiescent yet, right?


>> If we still need to do what you suggest, I think we are better off using the multiline-version:
>>
>> #if TASKQUEUE_STATS
>>    // Do not reset task queue statistics if we are carrying over work
>>    // in the task queues
>>    if (!sh->is_degenerated_gc_in_progress()) {
>>      reset_taskqueue_stats();
>>    }
>> #endif
> 
> okay: http://cr.openjdk.java.net/~zgu/shenandoah/degen_tq_stats/webrev.01/

Looks good.

But I still wonder why can't we simplify the whole thing? Remove the TQ stats block from
SCM::mark_from_roots() completely, move reset past the print in SCM::finish_mark_from_roots()? Let
the queues finish, and print+reset.

For example:

http://cr.openjdk.java.net/~shade/shenandoah/tq-why/webrev.01/src/hotspot/share/gc/shenandoah/shenandoahConcurrentMark.cpp.sdiff.html

-Aleksey