Don't error out when deleting parent cgroup dirs

[evanphx]

The current logic says "if we created a cgroup directory, we remove it". But the flaw in that logic is that if you run in /sys/fs/cgroup/ns1/container1 you'll end up creating ns1 and container1. But if another container launches as /sys/fs/cgroup/ns1/container2, now while the first run created ns1, it can't delete it because container2 is in there.

The best fix I can think of, other than changing the logic full scale, is to just allow for a removal error on all but the top Own'd directory.

[EtiennePerot]

Idea seems fine, but please direct changes to the master branch.

[evanphx]

@EtiennePerot No problem.

[avagin]

The current logic says "if we created a cgroup directory, we remove it". But the flaw in that logic is that if you run in /sys/fs/cgroup/ns1/container1 you'll end up creating ns1 and container1. But if another container launches as /sys/fs/cgroup/ns1/container2, now while the first run created ns1, it can't delete it because container2 is in there.

The best fix I can think of, other than changing the logic full scale, is to just allow for a removal error on all but the top Own'd directory.

I think we should only remove the container cgroup. All parent cgroups should remain untouched; otherwise, it could race with the creation of other containers.

-------------------------------------------------------------------------------------
ct2                                                      | ct1
-------------------------------------------------------------------------------------
access(/sys/fs/cgroup/ns1) = 0          |
                                                           | rmdir(/sys/fs/cgroup/ns1)
mkdir(/sys/fs/cgroup/ns1/container2) |
-------------------------------------------------------------------------------------