Fu Wei

@mimmus @dohnto Currently, containerd is using [cgroup memory event control](https://github.com/containerd/cgroups/blob/main/memory.go#L461) to receive event with epoll. There are two events to trigger epoll_wait to return: 1) OOM event from page_fault 2)...

duplicated with #5904

> @fuweid Can rebase now that the CI should be back in order For sure. Updated

``` // Vagrant BOX:rockylinux/8 === Failed default: === FAIL: . TestShimOOMScore (0.07s) default: container_linux_test.go:1441: failed to start shim: start failed: io.containerd.runc.v2: failed to join cgroup /containerd/oomshim: write /sys/fs/cgroup/cpu/containerd/oomshim/cgroup.procs: invalid argument:...

cri-containerd plugin mounts net-ns first and then creates pause-container with that net-ns-path. For the pause-container and CNI tear-down, there is not PID reuse issue. However, the pod containers are designed...

Another flaky test case: https://api.cirrus-ci.com/v1/task/5579724511510528/logs/cri_test.log ``` default: Begin Captured StdOut/StdErr Output >> default: Forwarding from 127.0.0.1:12002 -> 12003 default: Forwarding from [::1]:12002 -> 12003 default: Handling connection for 12002 default:...

Since the containerd is reconnecting to the existing shim, the socket will show up after reload. The 1.4.13 is EOL, please use supported releases which contains https://github.com/containerd/containerd/pull/6031.

The root cause is described in #5597. Containerd state dir is temporary and supposed to be tmpfs. No need to use /mnt/ persistent storage.

@fzyzcjy if the v1.6 release works for you, please help to close the issue~ thanks

Thanks for reporting. The default namespace option of containerd client is used to inject the namespace into grpc metadata so that the server side can read it from context. But...