RFR: 8351996: Behavioral updates for ClassValue::remove [v6]

[John R Rose]

On Sun, 27 Apr 2025 04:50:37 GMT, Chen Liang <liach at openjdk.org> wrote:

>> The recent patch #23866 makes calling `ClassValue::remove()` from `ClassValue::computeValue()` end up in infinite loops while fixing the stale value risk from the method.
>> 
>> The proposed fix is to preserve the stale value risk fix, and update the remove-from-compute behavior from the original designated no-op behavior to throwing an exception, as the original behavior conflicts with the stale value fix.
>> 
>> The implementation track the owner thread in promises (accessed in locked section); as a result, we can fail-fast on recursive removals from `computeValue`. I did not choose to use `ThreadTracker` as it is designed for single tracker and multiple threads, while this case here sees often just one thread, and the threads outlive the promise objects.
>> 
>> Also updated the API specs for `remove` to more concisely describe the memory effects. Please review the associated CSR as well.
>
> Chen Liang has updated the pull request with a new target base due to a merge or a rebase. The incremental webrev excludes the unrelated changes brought in by the merge/rebase. The pull request contains 12 additional commits since the last revision:
> 
>  - Cannot test for recursion eagerly - add test case
>  - More spec, eager exception, finish with existing, thanks John
>  - Merge branch 'master' of https://github.com/openjdk/jdk into fix/classvalue-compute-remove
>  - docs
>  - Remove the throwing behavior due to shallow reentrancy
>  - Merge branch 'master' of https://github.com/openjdk/jdk into fix/classvalue-compute-remove
>  - Add more tests to avoid critical errors like the last one
>  - Major typo
>  - Use identity of thread, some optimizations for single thread case
>  - Merge branch 'master' of https://github.com/openjdk/jdk into fix/classvalue-compute-remove
>  - ... and 2 more: https://git.openjdk.org/jdk/compare/b0b85ba9...edd19537

One thing I didn't note in the javadoc, but which might be significant, is that `computeValue` calls can be restarted for unpredictable reasons.  I gave some of them in the javadoc, but not all.  I did not document the possibility of crosstalk between <CV,C> pairs.  Another pair <CV,C2> might do a remove and update the internal CV version.  The purpose of htis is to flush the CV cache.  But it also creates an interference with ongoing computation of `computeValue` for the original <CV,C> pair.  Maybe we should treat such crosstalk as a bug.

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

PR Comment: https://git.openjdk.org/jdk/pull/24043#issuecomment-2833168973