[jdk17u-dev] RFR: 8351500: G1: NUMA migrations cause crashes in region allocation

[Severin Gehwolf]

On Thu, 22 May 2025 13:16:06 GMT, Thomas Stuefe <stuefe at openjdk.org> wrote:

> This is not a clean backport. The effected G1Allocator and G1Collector methods have changed since JDK17.
> 
> So this backports reimplements the patch in a minimally invasive way while retaining as much similarity as possible with the original patch.
> 
> The gist of the patch is clear: instead of finding out the NUMA node index at every instance of G1Allocator::allocate_xxx, and then be subject to NUMA node migrations, we fix the NUMA node index once and use that one.
> 
> I tested this patch with my "FakeNUMA" addition (I plan to upstream that one at some point). This FakeNUMA mode mimics a lot of NUMA node migrations. I can verify that without this patch the JVM crashes quickly, with the patch it does not crash.

I'm not really an expert in GC code. Somebody else needs to review this. @adinn might be able to help, perhaps?

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

PR Comment: https://git.openjdk.org/jdk17u-dev/pull/3607#issuecomment-3044373096