Re: [PATCH v4 0/2] cgroup: allow management of subtrees by new cgroup namespaces

[Aleksa Sarai]

> This is an updated (and rewritten) version of v3 of this patchset[1].
>
> The main difference is that I changed how we the "allow management" is
> implemented. Rather than just chmod-ing the cgroup directory (which
> everyone agreed was quite an odd way of doing it),
> unshare(CLONE_NEWCGROUP) will create a new subtree in every cgroup the
> task is associated with. The task will then be migrated to those
> subtrees (which form the root cset of the cgroup namespace). This change
> will be transparent to namespaced processes, and they'll gain a new
> ability (the ability to create cgroups).
>
> The name of the cgroup is randomly generated to ensure we don't get
> conflicts (but maybe this should be dealt with in a nicer way). In
> addition, I've updated the cgroup.procs write permission checks to be
> user namespace aware, but I also added an additional "permitted" case
> (where all of the tasks are in the same cgroup namespace and %current
> has CAP_SYS_ADMIN in all of the relevant user namespaces).
>
> I'm not _completely_ convinced about the addition of that case, and
> maybe we should drop it (but I might be biased since this all comes from
> the requirements of rootless containers).
>
> Also, I haven't added a way to disable the functionality on a per-cgroup
> (or even global) basis. Maybe there should be a way to do that, but I'm
> not sure how it should be done (a cgroup.ns_subtrees file that allows
> administrators to change it on a per-cgroup basis, or just a sysctl?).
>
> PTAL.
>
> [1]: https://lkml.org/lkml/2016/5/2/280
>
> Aleksa Sarai (2):
>    cgroup: make cgroup.procs permissions userns-aware
>    cgroup: implement subtree creation on copy_cgroup_ns()
>
>   kernel/cgroup.c | 149 +++++++++++++++++++++++++++++++++++++++++++++++---------
>   1 file changed, 126 insertions(+), 23 deletions(-)

Are there any comments on this version of the patchset? I thought we had 
reached an agreement that the underlying feature (allowing a process to 
manage its own cgroups) was useful. Is there a better way of solving 
this problem, that I don't know of?

--
Aleksa Sarai
Software Engineer (Containers)
SUSE Linux GmbH
https://www.cyphar.com/