RFR: 8324881: ObjectSynchronizer::inflate(Thread* current...) is invoked for non-current thread [v3]

[Axel Boldt-Christmas]

On Wed, 31 Jan 2024 09:04:09 GMT, Axel Boldt-Christmas <aboldtch at openjdk.org> wrote:

>> The `ObjectSynchronizer` has always assumed that the `current` parameters are both the current thread as well as the thread that is doing the locking. The only time that we are entering on behalf of another thread is when doing re-locking in deoptimization. This has worked because the deoptee thread is suspended. However ResourceMarks have been using the wrong thread when logging is enabled.
>> 
>> This change `ObjectSynchronizer` instruments the relevant methods with both a `JavaThread* locking_thread` as well as `[Java]Thread* current` to be able to use the correct thread for ResourceMarks. 
>> 
>> Having the `inflate` care about a `locking_thread` is a little unpleasant in my opinion. But it is required for LM_LIGHTWEIGHT. 
>> Would probably be cleaner if the inflate for LM_LIGHTWEIGHT was it's own thing, as it does not share the whole INFLATING protocol. But seems like a future RFE to refactor this code.
>> 
>> Can reproduce a crash by modifying `test/jdk/com/sun/jdi/EATests.java` and using `-XX:DiagnoseSyncOnValueBasedClasses=2` with LM_LEGACY or running `test/jdk/com/sun/jdi/EATests.java` with LM_LIGHTWEIGHT/LM_MONITOR and `-Xlog:monitorinflation=trace`. 
>> 
>> Could extend this test to capture this regression in the future (or creating a new test based on the same infrastructure). Will give it an attempt, so we have a regression test for this. But these tests get rather involved as the require a lot of jvmti setup.
>
> Axel Boldt-Christmas has updated the pull request incrementally with one additional commit since the last revision:
> 
>   More restrictive API

There is definitely an issue with `ObjectMonitor::enter`. The extended test I added is currently failing with LM_MONITOR. The issue is that the the call to `ObjectMonitor::enter` races with deflation (seems like the monitorinflation trace logging made this race much more likely to happen) . But `ObjectMonitor::enter` asserts that `assert(current->thread_state() != _thread_blocked, "invariant");` before checking for deflation . And because the thread is suspended it is obviously also blocked. (Also seeing a timeout on windows-x64-debug, have to investigate if it is the same cause, it is also with LM_MONITOR)

Given this I believe that it might be the only tenable solution that we restrict the normal `enter` to only the `current` thread and create a new `enter_for` exclusively used by the re-lock code for both ObjectSynchronizer and ObjectMonitor.

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

PR Comment: https://git.openjdk.org/jdk/pull/17626#issuecomment-1920629750