RFR 8242263: Diagnose synchronization on primitive wrappers

[David Holmes]

Hi Patricio,

On 11/08/2020 6:03 am, Patricio Chilano wrote:
> Hi Erik,
> 
> Thanks for looking into this and for the offline discussions.

If there are still issues with this regarding JFR then I would suggest 
we simply drop the use of JFR for reporting this.

Meanwhile looking at v5 incremental that all seems fine to me.

Thanks,
David
-----

> On 8/10/20 10:19 AM, Erik Gahlin wrote:
>>
>> Hi Patricio,
>>
>> I have tried to review the JFR changes, but I need more information on 
>> how the feature is going to be used.
>>
>> Is this something temporary that will be used gain insight for the 
>> Valhalla project, or do you think this will be needed longer term? If 
>> it is longer term, the bar is higher on what can be accepted.
>>
>> If this is a temporary feature, we could mark the event as 
>> experimental in metadata.xml, similar to what we did with GC events 
>> during the development of ZGC. Events that are experimental do not 
>> show up by default in visualization tools such as JMC and can be 
>> removed when they are not needed, or we have a better solution.
>>
> Right, it's temporary so we can mark it as experimental. I also changed 
> the category name to be "Java Virtual Machine, Diagnostics".
> 
>> For events to be enabled in default.jfc, they should not cause more 
>> than 1% overhead, not even in pathological cases. To me, it seems this 
>> could happen if you make a loop where it is triggered all the time.
>>
>> For events in profile.jfc, the overhead should still be low (1-2%), 
>> but the target is the typical application. For example, allocation 
>> profiling is only enabled in profile.jfc. Some application that are 
>> allocation intensive could cause a higher overhead than 1%, but that 
>> is OK because that configuration is only to be used for a short period 
>> of time. That said, there is still a budget on how much space in a 
>> recording an event can take up, so it also depends on how important 
>> the information is for the user. To me, it seems this event will not 
>> be that important for the average user, which makes me think it should 
>> be disabled by default.
>>
> After our off-list discussion I left it as enabled. I also removed the 
> threshold setting and added the startTime=false in metadata.xml.
> 
>> To complicate things, I noticed that aa command line flag is also used 
>> to enable the event. This is something we have worked hard to avoid 
>> when it comes to JFR events. All configuration should happen using 
>> configuration files. We had problem with this in the past where users 
>> gets confused why their events are not enabled. When we ported JFR to 
>> Hotspot, we got rid of those of those flags/events and it is a much 
>> better situation. Again, if this a temporary event to discover usage 
>> pattern for a release or two, it might be OK, but if we believe this 
>> event to stick around, I think we should look into alternatives, or 
>> not use JFR at all for this.
>>
> Yes, I see your point. Since we want to diagnose synchronization on 
> these classes but we don't want to affect performance the flag has to be 
> specified at startup otherwise the diagnostic instrumentation is 
> disabled, so I don't see a clear way to enable the events from JFR 
> alone. In any case, this will be a temporary event.
> 
> Here is v5. Let me know if you are okay with the changes.
> 
> Inc: http://cr.openjdk.java.net/~pchilanomate/8242263/v5/inc/webrev/
> Full: http://cr.openjdk.java.net/~pchilanomate/8242263/v5/webrev/
> 
> Thanks Erik!
> 
> Patricio
>>
>> Thanks
>>
>> Erik
>>
>>> On 28 Jul 2020, at 21:16, Patricio Chilano 
>>> <patricio.chilano.mateo at oracle.com 
>>> <mailto:patricio.chilano.mateo at oracle.com>> wrote:
>>>
>>> Hi all,
>>>
>>> Please review the following change that adds diagnostic capabilities 
>>> when synchronizing on primitive wrapper classes.
>>>
>>> Bug: https://bugs.openjdk.java.net/browse/JDK-8242263
>>> Webrev: http://cr.openjdk.java.net/~pchilanomate/8242263/v1/webrev/
>>>
>>> The new flag allows to identify synchronization on these classes and 
>>> to take one of the following actions: exit the VM with fatal error, 
>>> log a warning message, or issue a JFR event. The implementation uses 
>>> a simple approach where a check is added at the beginning of the 
>>> monitorenter generated code when the flag is enabled to check whether 
>>> the object is of a primitive wrapper class. If it is, we jump to the 
>>> slow path, otherwise we just continue as always. The extra 
>>> instructions will be: load the klass of the object, load the 
>>> _access_flags field for that klass, AND with a constant, and branch 
>>> based on the result. The code will only be generated whenever the new 
>>> opt-in diagnostic flag is enabled so performance won't be affected 
>>> when off.
>>>
>>> In addition to the purpose described in the description of the bug, 
>>> this flag will also be useful when trying to diagnose possible 
>>> synchronization issues if these classes ever become inline types as 
>>> part of the Valhalla project.
>>>
>>> I added test SyncOnPrimitiveWrapperTest.java that tests for the exit 
>>> and logging cases. I added test TestSyncOnPrimitiveWrapperEvent.java 
>>> to test for the JFR event case.
>>> I tested the patch running tiers1-6 in mach5 with the flag set to 
>>> DiagnoseSyncOnPrimitiveWrappers=2.
>>> I checked it builds with arm32 and ppc but can't run any tests on 
>>> those platforms, so it would be good if somebody can run the new test 
>>> included in the patch.
>>>
>>> Let me know if you think I should run or add any more tests.
>>>
>>> Thanks!
>>>
>>> Patricio
>>
>