8351500: G1: NUMA migrations cause crashes in region allocation #3607
Conversation
|
👋 Welcome back stuefe! A progress list of the required criteria for merging this PR into |
|
❗ This change is not yet ready to be integrated. |
openjdk
bot
changed the title
|
This backport pull request has now been updated with issue from the original commit. |
|
/approval request This is a solution for a quite hairy problem that hits customers on NUMA machines. It is rare, intermittent, and quite difficult to pinpoint. I therefore would appreciate it if I got approval for backporting. Risk: lowish. The mechanism is quite clear, the patch, albeit unclean, simple in essence. I also took care to test its functioning well (see JBS description for details.) |
|
Ping @sjohanss - could I get a review for this? |
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.
Progress
Issue
Reviewing
Using
gitCheckout this PR locally:
$ git fetch https://git.openjdk.org/jdk17u-dev.git pull/3607/head:pull/3607$ git checkout pull/3607Update a local copy of the PR:
$ git checkout pull/3607$ git pull https://git.openjdk.org/jdk17u-dev.git pull/3607/headUsing Skara CLI tools
Checkout this PR locally:
$ git pr checkout 3607View PR using the GUI difftool:
$ git pr show -t 3607Using diff file
Download this PR as a diff file:
https://git.openjdk.org/jdk17u-dev/pull/3607.diff
Using Webrev
Link to Webrev Comment