Request for Review (s) - 8157620: Guarantee in run_task(task, num_workers) fails

[Derek White]

Hi Jon,

The new code looks good!

While reviewing this I looked over WorkerManager again, and saw a 
problem in the error handling:

  - Line 60 checks for errors creating the WorkerThread or OS thread.
  - Line 66-67 unconditionally update created_workers and tries to start 
a (possibly non-existent) thread.
  - And I don't like the spacing of line 61, but that's a nit.

I'm not sure if this should get fixed as a separate change or as part of 
this one, but it looks bad.

- Derek

On 6/7/16 2:46 PM, Jon Masamitsu wrote:
> New webrev and delta, respectively, with changes indicated below.
>
> http://cr.openjdk.java.net/~jmasa/8157620/webrev.01/
>
> http://cr.openjdk.java.net/~jmasa/8157620/webrev_delta.00_01/
>
> Jon
>
> On 6/7/2016 1:36 AM, Thomas Schatzl wrote:
>> Hi Jon,
>>
>> On Fri, 2016-06-03 at 14:04 -0700, Jon Masamitsu wrote:
>>> https://bugs.openjdk.java.net/browse/JDK-8157620
>>>
>>> A phase of a G1 collection chooses its own number of GC workers
>>> but doesn't add more workers if more are needed than have been
>>> created.   Add additional workers as needed.  Also changed the
>>> guarantee to check the number of workers against the total number
>>> of workers since the number of GC workers used to execute the
>>> task may be different than active_workers (as in this case).
>>>
>>> Added an execution of TestGCOld with parameters that evoked the
>>> failure.
>>>
>>> http://cr.openjdk.java.net/~jmasa/8157620/webrev.00/
>>>
>>> Tested fix with TestGCOld.
>>> Stability tested with gc_test_suite
>>>
>>    looks good, with two comments:
>>
>> - I would _prefer_ if run_task(workgang, uint) always called
>> add_workers() itself. Otherwise all of these invocations need to call
>> add_workers() first themselves (if they do not limit themselves to
>> active_workers).
>>
>> The run_task(workgang, uint) method has kind of been implemented as a
>> short/quick way to run a workgang with a given number of threads, so
>> this requirement adds additional boiler plate code again...
>>
>> - in the test, maybe put the -Xmx/-Xms right after the @run
>> main/othervm like in the other tests just for consistency.
>>
>> ---------------------
>>
>> One comment related to the code changed, but unrelated to this CR: the
>> code in G1ConcurrentMark::calc_parallel_marking_threads() looks
>> (almost?) the same as AdaptiveSizePolicy::calc_active_conc_workers().
>> Any arguments about cleaning this up at some point?
>>
>> When looking at the calculation of the
>> AdaptiveSizePolicy::calc_default_active_workers() method, wouldn't it
>> be better if instead of using Universe::heap()->capacity(), use
>> something like Universe::heap()->used() (and actually e.g. for g1 one
>> could subtract young gen and areas covered by primitive arrays too).
>>
>> Actually, this calculation should be different for young/old gen gc
>> imho, as for both completely different areas are covered by the
>> collection... (just some random thoughts when looking at this).
>>
>> Thanks,
>>    Thomas
>>
>