RFR: 8305670: Performance regression in LockSupport.unpark with lots of idle threads

[Andrei Pangin]

On Tue, 18 Apr 2023 21:22:26 GMT, Daniel D. Daugherty <dcubed at openjdk.org> wrote:

>> Address the performance regression in `LockSupport.unpark()` with the following changes:
>> - Add a `quick_mode` parameter to `ThreadsListHandle::cv_internal_thread_to_JavaThread()` that tells the conversion function to skip searching the ThreadsList for the target `JavaThread*`.
>> - Update `Unsafe_Unpark()` to verify that `java_lang_Thread::thread(thread_oop) != nullptr` both before and after the creation of the ThreadsListHandle; this verification that a `JavaThread*` is saved in the `java.lang.Thread` object allows us to know that the target `JavaThread*` is protected by the ThreadsListHandle without searching it.
>> - Move the storing of the `JavaThread*` in the `java.lang.Thread` object until after the `JavaThread*` is added to the system ThreadsList.
>> - This change in the order permits the check of `java_lang_Thread::thread(thread_oop) != nullptr` before the creation of the ThreadsListHandle to have the meaning that we need for a `JavaThread*` that is in the early stages of running.
>> - The check `java_lang_Thread::thread(thread_oop) != nullptr` that happens after the creation of the ThreadsListHandle permits the meaning that we need for a `JavaThread*` that is in the end stages of running.
>> - So the combination of the two `java_lang_Thread::thread(thread_oop) != nullptr` checks allows us to safely skip the ThreadsList search in `ThreadsListHandle::cv_internal_thread_to_JavaThread()`.
>> 
>> Additional changes:
>> - Remove the `EnableThreadSMRExtraValidityChecks` option since that option implies that the ThreadsList search is optional in all situations which is not the case. Also remove comments that mention the option.
>> - Update the comments in `ThreadsListHandle::cv_internal_thread_to_JavaThread()` to clarify what `java_lang_Thread::thread(thread_oop) != nullptr` means in the JavaThread's lifecycle and that function's attempts to convert a `jthread` into a `JavaThread*` that's known to be protected by the ThreadsListHandle.
>
> @dholmes-ora and @robehn - The merge of my (mostly comments) changes with @robehn's
> "quick_mode" changes are tested and ready for review.

@dcubed-ojdk Thank you for working on this. The changes make total sense to me (I'm not a reviewer though).
I tested the patch with the `UnparkRegression` benchmark from the bug report and it now produces 151K calls per second comparing to 70K without the patch. That is a quite good improvement.

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

PR Comment: https://git.openjdk.org/jdk/pull/13519#issuecomment-1513903333