[jdk21u-dev] RFR: 8348402: PerfDataManager stalls shutdown for 1ms

[Aleksey Shipilev]

On Thu, 7 Aug 2025 08:43:45 GMT, Goetz Lindenmaier <goetz at openjdk.org> wrote:

> OK, I can set it aside for now and pick it up again if there is an additional trigger to backport it.

I am not blocking this, just giving a perspective :)

(+): The positive impact on small applications is noticeable. We have argued about the negative impact back when it was done in mainline, and reasoned it is minor at worst. The synchronization is on `FutuleWakeups` path, where we already are in the world of pain.
(+): Doing the proper CS synchronization between perf data counters and their manager allows more reliable shutdown. I can see how this might mitigate some JVM shutdown crashes in the wild.
(-): We see lots of performance-sensitive applications run with `-XX:-UsePerfData`, where neither of these issues manifest. So the positive impact is smaller.
(-): Since this CS-taking code path is summarily removed in JDK 25. So if we backport, only JDK 21 would be taking it. This also tells us waiting for real-world JDK 25 testing is futile, since there is nothing to test.

So on balance, we can take it into JDK 21. Let's not do any prior releases, though.

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

PR Comment: https://git.openjdk.org/jdk21u-dev/pull/2058#issuecomment-3163422985