RFR: 8295303: cleanup debug agent's confusing use of EI_GC_FINISH [v3]

[Chris Plummer]

On Tue, 8 Nov 2022 02:30:19 GMT, Chris Plummer <cjplummer at openjdk.org> wrote:

>> This PR renamed EI_GC_FINISH to EI_CLASS_UNLOAD and does some other minor cleanups related to EI_GC_FINISH.
>> 
>> The debug agent deals with 3 types of events: JVMTI (per the spec), JDWP (per the spec), and an event indexing that the debug agent refers to as EI (Event Index) events. We have the following EI_GC_FINISH event. The indexing is into the array of jdwp event handlers which are created when the debugger sends an EventRequest command.
>> 
>> Note there is no EI_CLASS_UNLOAD event that maps to the JDWP CLASS_UNLOAD event, but instead we have:
>> 
>> `        EI_GC_FINISH = 8,`
>> 
>> And these are the mappings between EI_GC_FINISH and JDWP and JVMTI events
>> 
>> 
>>   case JVMTI_EVENT_GARBAGE_COLLECTION_FINISH:
>>       return EI_GC_FINISH;
>> 
>>   case JDWP_EVENT(CLASS_UNLOAD):
>>       return EI_GC_FINISH;
>> 
>>   index2jvmti[EI_GC_FINISH - EI_min] = JVMTI_EVENT_GARBAGE_COLLECTION_FINISH;
>> 
>>   index2jdwp[EI_GC_FINISH - EI_min] = JDWP_EVENT(CLASS_UNLOAD);
>> 
>> 
>> So there is this odd relationship between EI_GC_FINISH and the JDWP CLASS_UNLOAD event. Note that JVMTI does not have a CLASS_UNLOAD (except for unused support in the extension mechanism), and JDWP has no GC_FINISH event.
>> 
>> This relationship between EI_GC_FINISH and the JDWP CLASS_UNLOAD events stems from the fact that at one point JVMTI_EVENT_GARBAGE_COLLECTION_FINISH was used to trigger the synthesizing all of JDWP CLASS_UNLOAD events for classes that unloaded during the last GC. That is no longer the case, and instead each time a JVMTI OBJECT_FREE event is triggered for a Class instance, the JDWP CLASS_UNLOAD is generated.
>> 
>> Since JDWP CLASS_UNLOAD maps to EI_GC_FINISH, we have the following:
>> 
>> `    node = getHandlerChain(EI_GC_FINISH)->first;`
>> 
>> By looking at this line of code, you would never guess that this is how you fetch the event handler chain for JDWP CLASS_UNLOAD EventRequests, but it is.
>> 
>> To clean this up I renamed EI_GC_FINISH to EI_CLASS_UNLOAD. However, that still leaves the EI_GC_FINISH to JVMTI_EVENT_GARBAGE_COLLECTION_FINISH mapping to deal with. It's not needed anymore. When we get a JVMTI_EVENT_GARBAGE_COLLECTION_FINISH, this is all we ever do with it:
>> 
>> 
>> static void JNICALL
>> cbGarbageCollectionFinish(jvmtiEnv *jvmti_env)
>> {
>>     LOG_CB(("cbGarbageCollectionFinish"));
>>     ++garbageCollected;
>>     LOG_MISC(("END cbGarbageCollectionFinish"));
>> }
>> 
>> 
>> So the event is not passed around at all, and doesn't trigger other events or mapping to a JDWP event. It is never used as an "event index". This means jvmti2EventIndex should assert if it is ever passed in, since following mapping should never be needed:
>> 
>> 
>>         case JVMTI_EVENT_GARBAGE_COLLECTION_FINISH:
>>             return EI_GC_FINISH; 
>> 
>> 
>> This is accomplished by simply deleting the above code, and failing through to the default error handling code.
>> 
>> Also no longer needed is the following entry:
>> 
>>     index2jvmti[EI_GC_FINISH          -EI_min] = JVMTI_EVENT_GARBAGE_COLLECTION_FINISH
>> 
>> For this we just assign the entry to 0, which will result in an error if the entry is ever referenced.
>
> Chris Plummer has updated the pull request with a new target base due to a merge or a rebase. The pull request now contains four commits:
> 
>  - merge
>  - Fixed ei range check logic errors.
>  - fix extra whitespace
>  - Rename EI_GC_FINISH to EI_CLASS_UNLOAD and other cleanups related to EI_GC_FINISH

Thanks Alex and Serguei!

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

PR: https://git.openjdk.org/jdk/pull/10887