READY notification sometimes not accepted by systemd

[indyjo]

Sometimes systemd doesn't accept READY notifications generated by runc, causing starting containers to remain in activating state and be killed after timeout.

This applies to configurations in which an OCI container is started as a systemd service with Type=notify.

In those cases, systemd produces a log message like Cannot find unit for notify message of PID 1234, ignoring.

The reason seems to be that runc terminates immediately after sending the READY=1 signal. systemd needs to find out the cgroup of the message sender which only works if that process is still around.

This is a well-known issue: systemd/systemd#2739
In systemd-notify it was fixed by introducing the sd_notify_barrier mechanism.

[indyjo]

My PR #3291 implements a fix for this issue.