Re: [RESEND][PATCH v4] cgroup: Use CAP_SYS_RESOURCE to allow a process to migrate other tasks between cgroups

From: Andy Lutomirski
Date: Mon Dec 05 2016 - 19:37:27 EST


On Mon, Dec 5, 2016 at 4:28 PM, John Stultz <john.stultz@xxxxxxxxxx> wrote:
> On Tue, Nov 22, 2016 at 4:57 PM, John Stultz <john.stultz@xxxxxxxxxx> wrote:
>> On Tue, Nov 8, 2016 at 4:12 PM, Andy Lutomirski <luto@xxxxxxxxxxxxxx> wrote:
>>> On Tue, Nov 8, 2016 at 4:03 PM, Alexei Starovoitov
>>> <alexei.starovoitov@xxxxxxxxx> wrote:
>>>> On Tue, Nov 08, 2016 at 03:51:40PM -0800, Andy Lutomirski wrote:
>>>>>
>>>>> I hate to say it, but I think I may see a problem. Current
>>>>> developments are afoot to make cgroups do more than resource control.
>>>>> For example, there's Landlock and there's Daniel's ingress/egress
>>>>> filter thing. Current cgroup controllers can mostly just DoS their
>>>>> controlled processes. These new controllers (or controller-like
>>>>> things) can exfiltrate data and change semantics.
>>>>>
>>>>> Does anyone have a security model in mind for these controllers and
>>>>> the cgroups that they're attached to? I'm reasonably confident that
>>>>> CAP_SYS_RESOURCE is not the answer...
>>>>
>>>> and specifically the answer is... ?
>>>> Also would be great if you start with specifying the question first
>>>> and the problem you're trying to solve.
>>>>
>>>
>>> I don't have a good answer right now. Here are some constraints, though:
>>>
>>> 1. An insufficiently privileged process should not be able to move a
>>> victim into a dangerous cgroup.
>>>
>>> 2. An insufficiently privileged process should not be able to move
>>> itself into a dangerous cgroup and then use execve to gain privilege
>>> such that the execve'd program can be compromised.
>>>
>>> 3. An insufficiently privileged process should not be able to make an
>>> existing cgroup dangerous in a way that could compromise a victim in
>>> that cgroup.
>>>
>>> 4. An insufficiently privileged process should not be able to make a
>>> cgroup dangerous in a way that bypasses protections that would
>>> otherwise protect execve() as used by itself or some other process in
>>> that cgroup.
>>>
>>> Keep in mind that "dangerous" may apply to a cgroup's descendents in
>>> addition to the cgroup being controlled.
>>
>> Sorry for taking awhile to get back to you here. I'm a little
>> befuddled as to what next steps I should consider (and honestly, I'm
>> not totally sure I really grok your concern here, particularly what
>> you mean with "dangrous cgroups").
>>
>> So is going back to the CAP_CGROUP_MIGRATE approach (to properly
>> separate "sufficiently" from "insufficiently privileged") better?
>>
>> Or something closer to the original method Android used of each cgroup
>> having an allow_attach() check which could determine what is
>> sufficiently privledged for the respective level of danger the cgroup
>> might poise?
>>
>> Or just stepping back, what method would you imagine to be reasonable
>> to allow a specified task to migrate other tasks between cgroups
>> without it having to be root/suid?
>
> Any suggested feedback here?

I really don't know. The cgroupfs interface is a bit unfortunate in
that it doesn't really express the constraints. To safely migrate a
task, ISTM you ought to have some form of privilege over the task
*and* some form of privilege over the cgroup. cgroupfs only handles
the latter.

CAP_CGROUP_MIGRATE ought to be okay. Or maybe cgroupfs needs to gain
a concept of "dangerous" cgroups and further restrict them and
CAP_SYS_RESOURCE should be fine for non-dangerous cgroups? I think I
favor the latter, but it might be nice to hear from Tejun first.

--Andy