RFR: 8241354: ZGC still crashes in containers with NUMA due to get_mempolicy is disabled by default
Jie Fu
jiefu at openjdk.java.net
Wed Apr 28 08:20:17 UTC 2021
On Mon, 26 Apr 2021 23:19:46 GMT, Jie Fu <jiefu at openjdk.org> wrote:
> Hi all,
>
> ZGC still crashes in containers hosting on NUMA machines due to an fatal error [1], which really limits the usage of zgc.
>
> The reason is that get_mempolicy is blocked by default in dockers [2].
> With the default docker configuration, ZGC will always crash with NUMA.
>
> Other collectors are also somehow affected by this restriction.
> But unlike zgc, they won't crash and can still work in dockers.
> So we'd better fix this bug as soon as possible since more and more people would like to use zgc.
>
> The patch was prepared under the guidance of StefanK [3] and many thanks to him.
> And we'd like to provide a user friendly -Xlog:gc+init log in a separate RFE after this issue.
>
> Testing:
> - tier1 ~ tier3 on Linux/x64
>
> Thanks.
> Best regards,
> Jie
>
> [1] https://github.com/openjdk/jdk/blob/master/src/hotspot/os/linux/gc/z/zNUMA_linux.cpp#L64
> [2] https://docs.docker.com/engine/security/seccomp/
> [3] https://mail.openjdk.java.net/pipermail/hotspot-gc-dev/2020-March/028927.html
> The approach looks good. However, I'd like to suggest the following adjustments [pliden at 9cff734](https://github.com/pliden/jdk/commit/9cff734de0f03b4990cab55c7052e3e7f3b6a8f1) on top of this PR, which breaks out the syscall into a separate function and renames a few things.
Thanks @pliden for your review and help.
Updated.
-------------
PR: https://git.openjdk.java.net/jdk/pull/3704
More information about the hotspot-gc-dev
mailing list