security_task_getsecid() and subjective vs objective task creds

[Casey Schaufler]

On 2/18/2021 11:34 AM, Paul Moore wrote:
> Hi all,
>
> When looking into a problem I noticed that audit was recording the
> wrong subject label for a process.  Doing a bit of digging I realized
> this was caused by the SELinux security_task_getsecid() implementation
> returning the objective security label (taken from task->real_cred),
> and not the subjective security label (taken from task->cred).
>
> Looking around at the other LSMs which implement this hook, Smack and
> AppArmor, it appears they both do the same thing: return the objective
> security ID for the process.  Looking quickly at the various non-LSM
> callers of the security_task_getsecid() hook, it unfortunately looks
> like all of them expect the subjective security ID to be returned.
> The only caller I'm not 100% confident in is binder, but from what I
> can tell it looks like they are expecting the subjective ID too.
>
> At least we are consistently wrong :)

We may have come down with a case of helperitis. 

> How do we want to fix this?  The obvious fix is to change the SELinux,
> AppArmor, and Smack security_task_getsecid() implementations to return
> the subjective security ID (->cred), and likely make a note in
> lsm_hooks.h,

That would be my choice.

>  but if someone can see a case where we would need both
> the subjective and objective security labels speak up and we can
> introduce a new hook for the subjective label, and likely add a "_obj"
> to the end of the existing hook to help make it more clear.  If
> neither of those options are acceptable, we could convert all of the
> existing callers to use something like the line below (assumes
> current), but that is the least appealing option as far as I'm
> concerned.
>
>   security_cred_getsecid(current_cred(), &sid);
>
> Opinions?

If the objective cred isn't being used in the access control decision
it seems pointless to add it to the audit record. If there is a case
where the task is being treated as an object, signal delivery comes to
mind, you still want the objective credential. So it seems like care
may be required to ensure that the correct value (sub vs obj) is
used.