Re: [RFC] capabilities: Ambient capabilities

[Andy Lutomirski <luto-kltTT9wpgjJwATOyAt5JVQ@public.gmane.org>]

On Sat, Mar 14, 2015 at 3:17 PM, Andrew G. Morgan <morgan-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org> wrote:
> On Sat, Mar 14, 2015 at 2:45 PM, Andy Lutomirski <luto-kltTT9wpgjJwATOyAt5JVQ@public.gmane.org> wrote:
>> On Sat, Mar 14, 2015 at 2:09 PM, Andrew G. Morgan <morgan-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org> wrote:
>>> On Fri, Mar 13, 2015 at 10:57 AM, Andy Lutomirski <luto-kltTT9wpgjJwATOyAt5JVQ@public.gmane.org> wrote:
>>>> On Mar 13, 2015 6:24 AM, "Andrew G. Morgan" <morgan-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org> wrote:
>>>>>
>>>>> > It's to preserve the invariant that pA is always a subset of pI.
>>>>>
>>>>> But since a user can always raise a bit in pI if it is present in pP,
>>>>> what does this invariant add to your model other than inconvenience?
>>>>
>>>> The useful part is that dropping a bit from pI also drops it from pA.
>>>> To keep the model consistent, I also require that you add the bit to
>>>> pI before adding it to pA.
>>>
>>> So you are saying that pA is always a strict subset of pI (and pP)?
>>> Then why not explicitly implement it as:
>>>
>>>   pA' = (file caps or setuid or setgid ? 0 : pA)
>>>   pP' = (fP & X) | (pI & [fI | (pA' & pP)] )
>>>
>>> As it is you have so distributed these constraints that it is hard to
>>> be sure it will remain that way.
>>
>> That would be insecure.  If an attacker had pA = CAP_SYS_ADMIN, pI =
>> 0, pP = 0 (i.e. no privs but pA is set somehow) then, unless that's
>> there's some other protection implemented, they could run some setuid
>> program, and that program could switch back to non-root, set pI = 0,
>> and call execve.  Unexpectedly, CAP_SYS_ADMIN would be inherited.
>
> Forgive me, but which bit of
>
>    pI & [fI | (pA' & pP)]
>
> with pI = 0 makes that so?

Oh, I misread that.

I think it's already unnecessarily confusing that you can inherit
nonzero pI without inheriting the corresponding bits in pP, and I
don't want to add yet more degrees of freedom in non-permitted caps.

>>>>
>>>> I don't know what you mean here by naive privilege inheritance.  The
>>>> examples you're taking about aren't inheritance at all; they're
>>>> exploring privilege *grants* during execve.  My patch deliberately
>>>> leaves grants like that alone.
>>>
>>> The pI set is inherited through this exec unmolested.
>>
>> This is flat-out useless.  Having pI = CAP_NET_BIND_SERVICE doesn't
>> let me bind low-numbered ports, full stop.
>
> Even in your proposed model, neither pI nor pA does this. It is what
> is in pE that counts.

Let me state it more precisely.  If your parent puts
CAP_NET_BIND_SERVICE in pI and execs you, you can't bind low-numbered
ports.  If your parent puts CAP_NET_BIND_SERVICE in pA, you can.

>
>>> My Nack remains that you are eliminating the explicit enforcement of
>>> selective inheritance. A lockable secure bit protecting access to your
>>> prctl() function would address this concern.
>>
>> Would a sysctl or securebit that *optionally* allows pA to be disabled
>> satisfy you?
>>
>> I don't understand why lockable is at all useful.  You'd need
>> CAP_SETPCAP to flip it regardless.
>
> Because it means one can create process trees in which this behavior
> is guaranteed to be allowed and/or disallowed.
>

I don't see why guaranteeing that it's allowed is particularly useful.
I also don't see why any of the securebits lock bits are useful.  I
don't specifically object; I just don't see the point.  If you give a
subtree CAP_SETPCAP but you don't trust them enough to leave
securebits unlocked, then I think your threat model is confused.

--Andy