RFR: 8308614: Enabling JVMTI ClassLoad event slows down vthread creation by factor 10 [v5]

[Serguei Spitsyn]

On Thu, 30 Nov 2023 21:27:27 GMT, Serguei Spitsyn <sspitsyn at openjdk.org> wrote:

>> This is a fix of a performance/scalability related issue. The `JvmtiThreadState` objects for virtual thread filtered events enabled globally are created eagerly because it is needed when the `interp_only_mode` is enabled. Otherwise, some events which are generated in `interp_only_mode` from the debugging version of interpreter chunks can be missed.
>> However, it has to be okay to avoid eager creation of these object if no `interp_only_mode` has ever been requested.
>> It seems to be an extremely important optimization to create JvmtiThreadState objects lazily in such cases.
>> It is done by introducing the flag `JvmtiThreadState::_seen_interp_only_mode` which indicates when the `JvmtiThreadState` objects have to be created eagerly.
>> 
>> Additionally, the fix includes the following related changes:
>>  - Use condition double checking idiom for `MutexLocker mu(JvmtiThreadState_lock)` in the function `JvmtiVTMSTransitionDisabler::VTMS_mount_end` which is on a performance-critical path and looks like this:
>> 
>>   JvmtiThreadState* state = thread->jvmti_thread_state();
>>   if (state != nullptr && state->is_pending_interp_only_mode()) {
>>     MutexLocker mu(JvmtiThreadState_lock);
>>     state = thread->jvmti_thread_state();
>>     if (state != nullptr && state->is_pending_interp_only_mode()) {
>>       JvmtiEventController::enter_interp_only_mode();
>>     }
>>   }
>> 
>> 
>>  - Add extra check of `JvmtiExport::can_support_virtual_threads()` when virtual thread mount and unmount are posted.
>>  - Minor: Added a `ThreadsListHandle` to the `JvmtiEventControllerPrivate::enter_interp_only_mode`. It is needed because of the dynamic creation of compensating carrier threads which is racy for JVMTI `SetEventNotificationMode` implementation.
>> 
>> Performance mesurements:
>> - Without this fix the test provided by the bug submitter gives execution numbers:
>>   - no ClassLoad events enabled:     3251 ms
>>   - ClassLoad events are enabled: 40534 ms
>>   
>> - With the fix:
>>   - no ClassLoad events enabled:    3270 ms
>>   - ClassLoad events are enabled:   3385 ms
>>   
>> Testing:
>>  - Ran mach5 tiers 1-6, no regressions are noticed
>
> Serguei Spitsyn has updated the pull request incrementally with two additional commits since the last revision:
> 
>  - review: one more minor correction of a comment
>  - review: minor correction of a comment

Chris and Alex, thank you for review!

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

PR Comment: https://git.openjdk.org/jdk/pull/16686#issuecomment-1836761160