Re: [RFC] Implement ambient capability set.

[Andy Lutomirski]

On Wed, Feb 4, 2015 at 2:02 PM, Serge E. Hallyn <serge@xxxxxxxxxx> wrote:
> Quoting Serge E. Hallyn (serge@xxxxxxxxxx):
>> Quoting Andy Lutomirski (luto@xxxxxxxxxxxxxx):
>> > On Wed, Feb 4, 2015 at 1:27 PM, Serge E. Hallyn <serge@xxxxxxxxxx> wrote:
>> > > Quoting Andy Lutomirski (luto@xxxxxxxxxxxxxx):
>> > >> On Wed, Feb 4, 2015 at 1:16 PM, Serge E. Hallyn <serge@xxxxxxxxxx> wrote:
>> > >> > Quoting Andy Lutomirski (luto@xxxxxxxxxxxxxx):
>> > >> >> On Wed, Feb 4, 2015 at 10:49 AM, Christoph Lameter <cl@xxxxxxxxx> wrote:
>> > >> >> > +
>> > >> >> > +               if (!cap_valid(arg2))
>> > >> >> > +                       return -EINVAL;
>> > >> >> > +
>> > >> >> > +               new =prepare_creds();
>> > >> >> > +               if (arg3 == 0)
>> > >> >> > +                       cap_lower(new->cap_ambient, arg2);
>> > >> >> > +               else
>> > >> >> > +                       cap_raise(new->cap_ambient, arg2);
>> > >> >> > +               return commit_creds(new);
>> > >> >> > +
>> > >> >>
>> > >> >> This let you add capabilities you don't even have to cap_ambient.  I'm
>> > >> >> fine with that as long as the cap evolution rule changes, as above.
>> > >> >
>> > >> > How about if instead we do restrict it to what's in pP?  I don't
>> > >> > want CAP_SETPCAP to become a cheap way to get all caps back.  With
>> > >> > or without NNP.
>> > >>
>> > >> We'd also have to modify everything that can change pP to change pA as
>> > >> well if we went this route.  I'd be okay with that, but it would make
>> > >> the patch much larger, and I'm not entirely sure I see the benefit.
>> > >> It would keep the number of possible states smaller, which could be
>> > >> nice.
>> > >
>> > > Do you mean if we didn't require NNP?  I'm suggesting that even if
>> > > we require NNP we should restrict any new bits added to pA to be
>> > > in pP at the prctl call.  Then whether or not to drop them from
>> > > pA when they are dropped from pP, I'm not yet certain.
>> >
>> > I mean regardless of whether we require NNP.
>> >
>> > I think that, unless we change the evolution rule, we would need to
>> > drop from pA when bits are dropped from pP to preserve the idea that
>> > dropping bits from pP drops them for good (as long as ruid != 0 or
>> > some securebit is set).
>>
>> Ok, so iiuc the rules would be:
>>
>> 1. must set nnp and have ns_capable(CAP_SETPCAP) to
>> call prctl(PR_SET_AMBIENT_WHATEVER)
>>
>> 2. adding bits to pA requires they be in pP at prctl time
>>
>> 3. dropping bits from pP drops them also from pA

I'm still unconvinced that 2 and 3 is better than using pP & pA in
execve, but I could go either way on that.

>>
>> 4. at exec, fP |= pA;  pA' = pA
>
> Actually I'm tempted to say that if fP is not empty, then we stick with current
> rules for fP/pP and clear out pA.  If fP is empty, then fP = pA
>
> Then, if fP is not empty on the file, we either drop nnp at exec (or
> don't use nnp at all), or we refuse exec if fP > pA.

We can't drop nnp at exec.

>
> We definately do not want to run a file which has capability X
> in fP, when X is not in pA.

Confused.  This will break everything, including Christoph's use case.
The status quo for running ping from bash has pA == 0 and fP != 0.

--Andy
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/