runC can't kill containers containing cgroup-freezer frozen processes

As I understand it, runC's procedure for killing a container is to

1. Freeze the container's cgroup
2. Issue SIGKILL to the process tree
3. Unfreeze the container's cgroup
4. Wait for the processes to terminate

This occurs in signalAllProcesses() of init_linux.go.

If at the time of this procedure there is a non-empty freezer sub-cgroup with freezer.state==FROZEN, runC will block forever on step 4, because the tasks in that sub-cgroup are still frozen.

This prevents runC containers from safely using freezer groups; e.g. if you try nesting runC, attempting to killing the outer container runs the risk of deadlocking.

I believe I encountered this issue in practice when nesting runC. Specifically, I was attempting to kill a k8s container that itself contained a runC "sub"-container. The issue should also in theory be reproducible by running make test on this repo, and then attempting to docker kill at the wrong moment (specifically, during testing when a nested runC is in the middle the send-signal loop inside signalAllProcesses()).