Integrated: 8297168: Provide a bulk OopHandle release mechanism with the ServiceThread

David Holmes dholmes at openjdk.org
Tue Nov 22 22:01:05 UTC 2022


On Mon, 21 Nov 2022 01:56:34 GMT, David Holmes <dholmes at openjdk.org> wrote:

> When a `JavaThread` terminates it has to release all `OopHandles` that it uses. This can't be done by the thread itself due to timing issues, so it is handed-off to the `ServiceThread` to do it - ref [JDK-8244997](https://bugs.openjdk.org/browse/JDK-8244997).
> 
> Initially there was only one `OopHandle` to handle - that of the `threadObj`, but since Loom there are another 3 `OopHandles` to process. The existing logic does the hand-off one `OopHandle` at a time but that is a potential synchronization bottleneck because each hand-off acquires the `ServiceLock`, enqueues the `OopHandle`, issues a notify to (potentially) wakeup the `ServiceThread` and then releases the `ServiceLock`. This can lead to high contention on the `ServiceLock` and also bad scheduling interactions with the `ServiceThread`.
> 
> This PR contains two commits. The first simply changes the API so that we pass the target `JavaThread` so that all 4  `OopHandles` can be extracted in one call. The second commit looks at streamlining things further by consolidating  into a single `OopHandleList` instance.
> 
> As `ServiceThread is a subclass of `JavaThread` I wasn't concerned about it having detailed knowledge of the JavaThread implementation.
> 
> Testing:
>   - tiers 1-3
>   - checked still no memory leak (ref [JDK-8296463](https://bugs.openjdk.org/browse/JDK-8296463))
> 
> Thanks.

This pull request has now been integrated.

Changeset: a6c418ea
Author:    David Holmes <dholmes at openjdk.org>
URL:       https://git.openjdk.org/jdk/commit/a6c418eaf82965d5783ab252413d6e2583944a7d
Stats:     119 lines in 4 files changed: 74 ins; 41 del; 4 mod

8297168: Provide a bulk OopHandle release mechanism with the ServiceThread

Reviewed-by: rehn, coleenp

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

PR: https://git.openjdk.org/jdk/pull/11254


More information about the hotspot-runtime-dev mailing list