[WIP] Use format of the parent cgroup to determine whether to use cgroupfs or systemd cgroup driver #1071
Conversation
|
Hi @filbranden. Thanks for your PR. I'm waiting for a containerd member to verify that this patch is reasonable to test. If it is, they should reply with Once the patch is verified, the new status will be reflected by the I understand the commands that are listed here. Instructions for interacting with me using PR comments are available here. If you have questions or suggestions related to my behavior, please file an issue against the kubernetes/test-infra repository. |
The parent cgroup will always be a systemd "slice" which can be detected by looking at a ".slice" suffix in the cgroup name. Use that information to drive whether to tell `runc` to use systemd or cgroupfs. The advantage of this approach is that the full control of which cgroup driver to use is now in the CRI client. `containerd-cri` will simply match what the CRI client requested, and ask `runc` to do the same. This way it is possible to switch cgroup drivers by simply reconfiguring the CRI client (Kubelet.) Note: There's a corner case for the root slice, which in theory would be "/" for both cases and auto-detection would not work. But the correct way to specify the root slice for the systemd case is "-.slice" (or "/-.slice", since we only consider the last component of the path), so that corner case actually needs to be handled in the CRI client and not here. Tested: TODO! Signed-off-by: Filipe Brandenburger <[email protected]>
Systemd slice names are always hierarchical, so the nested slices will always have the name of their parent as a prefix. Fix the example and test case to use valid slice names. Signed-off-by: Filipe Brandenburger <[email protected]>
This field was deprecated in favor of the same one in Runtime.Options, but we're no longer looking at that field either, so document it as doubly deprecated now. Deprecating the one in Runtime.Options is a bit more tricky, since it's in a protobuf that's part of an official API. Need to figure out what to do about that one. Signed-off-by: Filipe Brandenburger <[email protected]>
The field is doubly deprecated and doesn't really do anything anymore, so let's just stop documenting it altogether. Signed-off-by: Filipe Brandenburger <[email protected]>
|
/ok-to-test |
|
I've actually included this as part of #1070. :p |
|
This has been covered by #1224. Thanks for suggesting this @filbranden Close this one. Feel free to reopen if you want to add more things. :) |
The parent cgroup will always be a systemd "slice" which can be detected by looking at a ".slice" suffix in the cgroup name. Use that information to drive whether to tell
runcto use systemd or cgroupfs.The advantage of this approach is that the full control of which cgroup driver to use is now in the CRI client.
containerd-criwill simply match what the CRI client requested, and askruncto do the same. This way it is possible to switch cgroup drivers by simply reconfiguring the CRI client (Kubelet.)Note: There's a corner case for the root slice, which in theory would be "/" for both cases and auto-detection would not work. But the correct way to specify the root slice for the systemd case is "-.slice" (or "/-.slice", since we only consider the last component of the path), so that corner case actually needs to be handled in the CRI client and not here.
This still needs testing! But sending it along to get some early feedback on the idea itself.
There are a few follow up commits updating comments, documentation and other minor fixes.
It's unclear to me what can be done about the
systemd_cgroupfield of Runtime.Options, since that seems to be part of an external API, so not sure how easy it can be deprecated... Opinions?/cc @Random-Liu @yujuhong