RFR: 8293540: [Metrics] Incorrectly detected resource limits with additional cgroup fs mounts

[Severin Gehwolf]

On Tue, 13 Sep 2022 13:06:10 GMT, Severin Gehwolf <sgehwolf at openjdk.org> wrote:

> Similar issue to the hotspot change discussed in https://bugs.openjdk.org/browse/JDK-8293472. The Java metrics implementation may get the resource limits wrong if there are additional cgroup fs mounts. Apparently that's more common than one might think. I've reproduced this with these existing tests on cg v2:
> 
> 
> test/hotspot/jtreg/containers/docker/TestMemoryAwareness.java
> test/jdk/jdk/internal/platform/docker/TestDockerCpuMetrics.java
> test/jdk/jdk/internal/platform/docker/TestDockerMemoryMetrics.java 
> 
> 
> I've also added `test/jdk/jdk/internal/platform/docker/TestDockerBasic.java` and amended `test/jdk/jdk/internal/platform/cgroup/TestCgroupSubsystemFactory.java` which unconditionally fails (irrespective of cgroup version in use). The fix is fairly straight forward and is an extension which we already do for the `cpuset` controller: Allow duplicates, and if there are any prefer those mounted at `/sys/fs/cgroup`.
> 
> Testing:
> - [x] fastdebug build on cgroups v2 and cgroups v1 (before and after the product fix)
> - [x] added tests fail before, pass after the product fix.
> - [x] Some manual testing using `cgcreate` and `cgexec` on cg1 and cg2. Still pass. 
> - [ ] GHA still running.
> 
> Please review! Many thanks in advance.

Thanks! I've updated the PR title to match the bug synopsis.

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

PR: https://git.openjdk.org/jdk/pull/10248