RFR S: 7127792 Add the ability to change an existing PeriodicTask's execution interval

[David Holmes]

On 9/10/2012 9:42 PM, Rickard Bäckman wrote:
> David,
> see inline.
>
> On Oct 9, 2012, at 1:18 PM, David Holmes wrote:
>
>> On 9/10/2012 7:36 PM, Rickard Bäckman wrote:
>>> David,
>>>
>>> thanks for your reply!
>>> I've changed the code according to the suggestions, I've also changed the types in PeriodicTask from being a size_t to being a
>>> jint (see updated webrev for details).
>>
>> But now the type changes have been pushed out to the task creators. Most task creations pass an int already but BiasedLocking uses a size_t. Just shows how messed up the typing was to begin with.
>
> Agreed. Two ways of solving it, 1) change the callers to use an int. 2) Do the cast in the constructor (should be safe since we check the possible interval).

int -> size_t shouldn't cause a warning so callers currently passing int 
are ok. So keeping it as size_t in constructor arg and casting to int 
before storing internally seems okay.

>>
>> Minor nit: should be an int rather than jint as these are not Java types.
>
> Will fix.
>
>>
>>> To prevent the waiting for very long time (which could overflow, etc) when we don't have any active task, I added an extra if
>>> so that if we are waiting while no tasks are available, we reset the time_before_wait and consider time_slept to be zero after sleeping.
>>> That means the first task added will always sleep for the period requested.
>>
>> Those semantics seem reasonable.
>>
>> The only thing that concerns me here is the affect of calling real_time_tick(0). I can't quite tell what the profiling code does.
>
> I could avoid calling that if time_slept = 0

Ok. You could skip real_time_tick altogether on zero.

I'm generally okay with the approach being taken here, but the changes 
are disruptive enough that I can't see for sure that existing tasks will 
be unaffected. I guess time will tell. And let's see what others might spot.

Thanks,
David (signing off for the night)

> Thanks
> /R
>
>>
>> David
>> -----
>>
>>> Updated webrev:
>>> http://cr.openjdk.java.net/~rbackman/7127792.1/
>>>
>>> Thanks
>>> /R
>>>
>>> On Oct 9, 2012, at 9:46 AM, David Holmes wrote:
>>>
>>>> Sorry Rickard, missed the original RFR :)
>>>>
>>>> So to be clear here the synopsis concerns changing the period, while the mechanism implemented is for a more general dynamic disenroll / enroll. So changing a period is effected by removing a task and then adding it with the new period.
>>>>
>>>> And for anyone not reading the fine-print when you dynamically enroll a task its first firing is somewhat arbitrary - somewhere between the time of enrollment and that time plus its period.
>>>>
>>>> src/share/vm/runtime/thread.hpp
>>>>
>>>> Can you add a comment:
>>>>
>>>>     static void stop();
>>>> +   // Only allow start once the VM is sufficiently initialized
>>>> +   // Otherwise the first task to enroll will trigger the start
>>>> +   static void make_startable();
>>>>
>>>> ---
>>>>
>>>> src/share/vm/runtime/thread.cpp
>>>>
>>>> We have a bit of type mixing here:
>>>>
>>>> - size_t time_to_wait
>>>> - jlong time_slept
>>>> - int remaining = time_to_wait
>>>>
>>>> I don't understand why Task is using size_t for time intervals. If you make "remaining" a size_t then it will cause issues when you pass it to wait. But I would expect the initialization of "remaining" to cause an unsigned-to-signed conversion warning, so perhaps an explicit cast to silence that.
>>>>
>>>> 1203     bool status = PeriodicTask_lock->wait(Mutex::_no_safepoint_check_flag, remaining);
>>>> 1204     if (status || _should_terminate) {
>>>> 1205       break;
>>>> 1206     }
>>>>
>>>> Can you rename status to timedout to make the logic more obvious. Also note that you will potentially return with time_slept still at zero, even though you may have slept for an arbitrary amount of time. That seems wrong as zero will then be passed to "tick".
>>>>
>>>> 1208     // spurious wakeup of some kind
>>>>
>>>> This comment is no longer accurate as you may have been woken up due to a change in the task list, I suggest:
>>>>
>>>>      // Change to task list or spurious wakeup of some kind
>>>>
>>>> 1213     remaining = PeriodicTask::time_to_wait();
>>>> 1214     if (remaining == 0) {
>>>> 1215         continue;
>>>> 1216     }
>>>>
>>>> Can you insert a comment before continue:
>>>>
>>>>   // Last task was just disenrolled so loop around and wait until
>>>>   // another task gets enrolled
>>>>   continue;
>>>>
>>>>
>>>> 1218     remaining -= time_slept;
>>>>
>>>> Again type mixing: subtracting a long from an int. Again a potential warning to get rid of.
>>>>
>>>> Also in a long running VM perhaps there have been no periodic tasks for many days and then one turns up. The subtraction could wrap and cause remaining to remain positive.  I know you've now documented the uncertainty in the first fire time but this seems somewhat too random. Let's see what others think. :)
>>>>
>>>>
>>>> 1329   PeriodicTask_lock->notify_all();
>>>>
>>>> There is only one thread that waits so notify() will suffice.
>>>>
>>>> Thanks,
>>>> David
>>>>
>>>> On 9/10/2012 4:34 PM, Rickard Bäckman wrote:
>>>>> Trying again,
>>>>>
>>>>> can I have a couple of reviews, please?
>>>>>
>>>>> /R
>>>>>
>>>>> On Oct 4, 2012, at 3:01 PM, Rickard Bäckman wrote:
>>>>>
>>>>>> Hi all,
>>>>>>
>>>>>> can I please have a couple of reviews on the following change:
>>>>>>
>>>>>> http://cr.openjdk.java.net/~rbackman/7127792/
>>>>>> http://bugs.sun.com/bugdatabase/view_bug.do?bug_id=7127792
>>>>>>
>>>>>> In short the purpose is to enable tasks to change the interval they are executed in. We
>>>>>> would also like to be able to add (and remove) tasks after the WatcherThread has started.
>>>>>>
>>>>>> Thanks
>>>>>> /R
>>>>>
>>>
>