[jdk21u-dev] RFR: 8315373: Change VirtualThread to unmount after freezing, re-mount before thawing

[Aleksey Shipilev]

On Thu, 8 Feb 2024 11:17:02 GMT, Aleksey Shipilev <shade at openjdk.org> wrote:

> This is a series of Loom backports that brings in the fixes from JDK 22 and mainline into JDK 21, where Loom is production-ready. This batch fixes a few thread state and JVMTI problems.
> 
> Unfortunately, these backports are tied together, as they change the same code a few times. It is very hard to make them into separate backports without breaking stuff along the way, so instead I clobbered them together in this PR. They follow the application order from mainline. All patches together touch only the Virtual Thread parts, so the risk for the rest of the code should be minimal.
> 
> Brief explanation of fixes:
> 
> * [JDK-8315373](https://bugs.openjdk.org/browse/JDK-8315373): In mainline since Sep 2023, there is no bug tail. Changes how VTs interact with JVMTI. We want this fix to provide the ground for the rest of backports. Notably, it makes [JDK-8312777](https://bugs.openjdk.org/browse/JDK-8312777) backport clean. 
> * [JDK-8312498](https://bugs.openjdk.org/browse/JDK-8312498): In mainline since Sep 2023. There is a bug tail, [JDK-8322818](https://bugs.openjdk.org/browse/JDK-8322818). Fixes what thread state VT reports when timed-parked. This fixes the actual Loom bug.
> * [JDK-8312777](https://bugs.openjdk.org/browse/JDK-8312777): In mainline since Oct 2023, there is no bug tail. Fixes the JVMTI race with mount/unmount events. Makes [JDK-8322818](https://bugs.openjdk.org/browse/JDK-8322818) backport clean. This fixes the actual Loom bug.
> * [JDK-8321270](https://bugs.openjdk.org/browse/JDK-8321270): In mainline since Dec 2023, there is no bug tail. Makes [JDK-8322818](https://bugs.openjdk.org/browse/JDK-8322818) a clean backport. This fixes the actual Loom performance problem.
> * [JDK-8322818](https://bugs.openjdk.org/browse/JDK-8322818): In mainline since Jan 2023. Fixes the bug introduced by [JDK-8312498](https://bugs.openjdk.org/browse/JDK-8312498). Has two additional follow-ups: [JDK-8323002](https://bugs.openjdk.org/browse/JDK-8323002) and [JDK-8323296](https://bugs.openjdk.org/browse/JDK-8323296).
> * [JDK-8323002](https://bugs.openjdk.org/browse/JDK-8323002): In mainline since Jan 2023. Follow-up test fix for [JDK-8322818](https://bugs.openjdk.org/browse/JDK-8322818). Test-only change.
> * [JDK-8323296](https://bugs.openjdk.org/browse/JDK-8323296): In mainline since Jan 2023. Follow-up test fix for [JDK-8322818](https://bugs.openjdk.org/browse/JDK-8322818). Test-only change.
> * [JDK-8316924](https://bugs.openjdk.org/browse/JDK-8316924): In mainline since Sep 2023. Fo...

After discussing this with other engineers, and seeing more of the soft push back here, I am now leaning to keep this PR out of 21.0.3 and push it early into 21.0.4, some time next week. This would be less stressful for everyone involved.

The reason to have these patches is to make 21u conveniently usable for Loom adopters. However, there are other fixes (e.g. JDK-8320275) that would not make it to 21.0.3. Plus, there are dependent fixes (e.g. JDK-8322846) that we also want to get in after this lands. So backporting the performance fixes when there are not yet backported functional fixes does not gain us much ground for Loom stabilization for 21.0.3.

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

PR Comment: https://git.openjdk.org/jdk21u-dev/pull/245#issuecomment-1956334815