[lkml]   [2014]   [Oct]   [22]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [PATCHv1 7/8] cgroup: cgroup namespace setns support
On Wed, Oct 22, 2014 at 11:37 AM, Aditya Kali <> wrote:
> On Tue, Oct 21, 2014 at 5:58 PM, Andy Lutomirski <> wrote:
>> On Tue, Oct 21, 2014 at 5:46 PM, Aditya Kali <> wrote:
>>> On Tue, Oct 21, 2014 at 3:42 PM, Andy Lutomirski <> wrote:
>>>> On Tue, Oct 21, 2014 at 3:33 PM, Aditya Kali <> wrote:
>>>>>>> And with explicit permission from
>>>>>>> cgroup subsystem (something like cgroup.may_unshare as you had
>>>>>>> suggested previously), we can make sure that unprivileged processes
>>>>>>> cannot pin themselves. Also, maintaining this invariant (your current
>>>>>>> cgroup is always under your cgroupns-root) keeps the code and the
>>>>>>> semantics simple.
>>>>>> I actually think it makes the semantics more complex. The less policy
>>>>>> you stick in the kernel, the easier it is to understand the impact of
>>>>>> that policy.
>>>>> My inclination is towards keeping things simpler - both in code as
>>>>> well as in configuration. I agree that cgroupns might seem
>>>>> "less-flexible", but in its current form, it encourages consistent
>>>>> container configuration. If you have a process that needs to move
>>>>> around between cgroups belonging to different containers, then that
>>>>> process should probably not be inside any container's cgroup
>>>>> namespace. Allowing that will just make the cgroup namespace
>>>>> pretty-much meaningless.
>>>> The problem with pinning is that preventing it causes problems
>>>> (specifically, either something potentially complex and incompatible
>>>> needs to be added or unprivileged processes will be able to pin
>>>> themselves).
>>>> Unless I'm missing something, a normal cgroupns user doesn't actually
>>>> need kernel pinning support to effectively constrain its members'
>>>> cgroups.
>>> So there are 2 scenarios to consider:
>>> We have 2 containers with cgroups: /container1 and /container2
>>> Assume process P is running under cgroupns-root '/container1'
>>> (1) process P wants to 'write' to cgroup.procs outside its
>>> cgroupns-root (say to /container2/cgroup.procs)
>> This, at least, doesn't have the problem with unprivileged processes
>> pinning themselves.
>>> (2) An admin process running in init_cgroup_ns (or any parent cgroupns
>>> with cgroupns-root above /container1) wants to write pid of process P
>>> to /container2/cgroup.procs (which lies outside of P's cgroupns-root)
>>> For (1), I think its ok to reject such a write. This is consistent
>>> with the restriction in cgroup_file_write added in 'Patch 6' of this
>>> set. I believe this should be independent of visibility of the cgroup
>>> hierarchy for P.
>>> For (2), we may allow the write to succeed if we make sure that the
>>> process doing the write is an admin process (with CAP_SYS_ADMIN in its
>>> userns AND over P's cgroupns->user_ns).
>> Why is its userns relevant?
>> Why not just check whether the target cgroup is in the process doing
>> the write's cgroupns? (NB: you need to check f_cred, here, not
>> current_cred(), but that's orthogonal.) Then the policy becomes: no
>> user of cgroupfs can move any process outside of the cgroupfs's user's
>> cgroupns root.
> Humm .. it doesn't have to be. I think its simpler to not enforce
> artificial permission checks unless there is a security concern (and
> in this case, there doesn't seem to be any). So I will leave the
> capability check out from here.
>> I think I'm okay with this.
>>> If this write succeeds, then:
>>> (a) process P's /proc/<pid>/cgroup does not show anything when viewed
>>> by 'self' or any other process in P's cgrgroupns. I would really like
>>> to avoid showing relative paths or paths outside the cgroupns-root
>> The empty string seems just as problematic to me.
> Actually, there is no right answer here. Our options are:
> * show relative path
> -- this will break userspace as /proc/<pid>/cgroup does not show
> relative paths today. This is also very ambiguous (is it relative to
> cgroupns-root or relative to /proc/<pid>cgroup file reader's cgroup?).

Confused now. If ".." in /proc/pid/group would be ambiguous, then so
would a path relative to cgroupns root, right? Or am I missing

(I'm not saying that ".." is beautiful or that it won't confuse
things. I'm just not sure why it's ambiguous.)

> * show absolute path
> -- this will also wrong as the process won't be able to make sense of
> it unless it has exposure to the global cgroup hierarchy.
> -- worse case is this that the global path also exists under the
> cgroupns-root ... so now the process thinks its in completely wrong
> cgroup
> -- this exposes system
> * show only "/"
> -- this is arguably better, but if the process tires to verify that
> its pid is in cgroup.procs of the cgroupns-root, its in for a
> surprise!
> In either case, whatever we expose, the userspace won't be able to use
> this path correctly (worse yet, it associates wrong cgroup for that
> path). So I think its best to not print out the line for default
> hierarchy at all. This happens today when cgroupfs is not mounted. I
> am open to other suggestions.

I suppose that ".." is a possible security problem. If I can force
you to see lots of ..s in there, then I might be about to get you to
write outside cgroupfs.

Grr. No great solution here. I suppose that the empty string isn't
so bad. We could also write something obviously invalid like
"(unreachable)". As long as no one actually creates a cgroup called
"(unreachable)", then this could result in errors but not actual

>>> * should we then also allow setns() without first entering the
>>> cgroupns-root? setns also checks the same conditions as in (a) plus it
>>> checks that your current cgroup is descendant of target cgroupns-root.
>>> Alternatively we can special-case setns() to own cgroupns so that it
>>> doesn't fail.
>> I think setns should completely ignore the caller's cgroup and should
>> not change it. Userspace can do this.
> All above changes more or less means that tasks cannot pin themselves
> by unsharing cgroupns. Do you agree that we don't need that "explicit
> permission from cgroupfs" anymore (via cgroup.may_unshare file or
> other mechanism)?

Yes, I agree.

>>> * migration for these processes will be tricky, if not impossible. But
>>> the admin trying to do this probably doesn't care about it or will
>>> provision for it.
>> Migration for processes in a mntns that have a current directory
>> outside their mntns is also difficult or impossible. Same with
>> pidnses with an fd pointing at /proc/self from an outside-the-pid-ns
>> procfs. Nothing new here.
>> --Andy
> Thanks for the review!

No problem.

 \ /
  Last update: 2014-10-22 21:01    [W:0.177 / U:2.556 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site