Open mythi opened 4 years ago
Can you do the initial local "touch" testing? Rebuild systemd with the configure
changes, make repoadd
, then use our tooling to create an image and boot it locally to assure that at least it boots normally?
@ahkok this issue is another cgroups related finding that I noticed but independent from the cgroup-v2 topic.
LGTM
containerd is trickier. I believe this is needed
--- vendor/github.com/containerd/cri/pkg/config/config.go.old 2019-09-26 16:02:33.147075178 +0300
+++ vendor/github.com/containerd/cri/pkg/config/config.go 2019-09-26 16:01:24.411290489 +0300
@@ -197,7 +197,7 @@
},
SandboxImage: "k8s.gcr.io/pause:3.1",
StatsCollectPeriod: 10,
- SystemdCgroup: false,
+ SystemdCgroup: true,
MaxContainerLogLineSize: 16 * 1024,
Registry: Registry{
Mirrors: map[string]Mirror{
containerd 1.3.0 config.toml
needs
[plugins."io.containerd.grpc.v1.cri".containerd.runtimes.runc.options]
SystemdCgroup = true
containerd still defaults to cgroupfs
.
AFAIU, containerd should auto-detect in 1.3.0.
@mythi I think Kata has issues with systemd based cgroups. Are there any issues with keeping cgroupfs for the short term while we work through https://github.com/kata-containers/runtime/issues/2182 ?
@egernst kubeadm.yaml
and crio.conf
in Clear Linux are already set to use systemd (cloud-native-setup defaults to crio). Only containerd ships without config.
I've not realized containerd can auto-detect. In pre 1.3.0, the setting was global and then got deprecated (and broken) in 1.3.0 in favor of runtime options specific setting.
Per k8s setup docs it's encouraged to have kubelet/CRIs to use systemd as the cgroupd driver when systemd is used.
/cc @kad @mcastelino @CraigSterrett @dklyle