Re: [PATCH -v3 1/3] Capabilities: move cap_file_mmap to commoncap.c

From: Stephen Smalley
Date: Thu Jul 30 2009 - 16:08:23 EST


On Thu, 2009-07-30 at 15:01 -0500, Serge E. Hallyn wrote:
> Quoting Stephen Smalley (sds@xxxxxxxxxxxxx):
> > On Thu, 2009-07-30 at 13:50 -0400, Eric Paris wrote:
> > > On Thu, 2009-07-30 at 11:58 -0400, Stephen Smalley wrote:
> > > > On Thu, 2009-07-30 at 10:54 -0500, Serge E. Hallyn wrote:
> > > > > Quoting Eric Paris (eparis@xxxxxxxxxx):
> > > > > > On Thu, 2009-07-30 at 00:14 -0500, Serge E. Hallyn wrote:
> > > > > > > Quoting Eric Paris (eparis@xxxxxxxxxx):
> > > > > > > > Currently we duplicate the mmap_min_addr test in cap_file_mmap and in
> > > > > > > > security_file_mmap if !CONFIG_SECURITY. This patch moves cap_file_mmap
> > > > > > > > into commoncap.c and then calls that function directly from
> > > > > > > > security_file_mmap ifndef CONFIG_SECURITY like all of the other capability
> > > > > > > > checks are done.
> > > > > > >
> > > > > > > It also
> > > > > > >
> > > > > > > 1. changes the return value in error case from -EACCES to
> > > > > > > -EPERM
> > > > > > > 2. no onger sets PF_SUPERPRIV in t->flags if the capability
> > > > > > > is used.
> > > > > > >
> > > > > > > Do we care about these?
> > > > > >
> > > > > > Personally, not really, but I'll gladly put them back if you care. #2
> > > > > > seems more interesting to me than number 1. I actually kinda like
> > > > > > getting EPERM from caps rather than EACCES since them I know if I was
> > > > > > denied by selinux or by caps.....
> > > > > >
> > > > > > -Eric
> > > > >
> > > > > Yup, I asked bc I didn't particularly care myself.
> > > > >
> > > > > I think I agree with you about -EPERM being better anyway. However I
> > > > > (now) think in this case PF_SUPERPRIV definately should be set, as this
> > > > > is a clear use of a capability to do something that couldn't have been
> > > > > done without it.
> > > >
> > > > On a related but different note, we should consider all current uses of
> > > > cap_capable(), as they represent capability checks that will not be
> > > > subject to a further restrictive check by other security modules. In
> > > > this case and in the vm_enough_memory case, that is intentional, but not
> > > > so clear for other uses in commoncap.c.
> > >
> > > Most of commoncap.c is called either as a secondary hook from the active
> > > lsm (aka selinux calls the commoncap.c functions) or in the !
> > > CONFIG_SECURITY case.
> > >
> > > I'll audit this afternoon to see which of them might not fit these
> > > rules....
> >
> > That isn't what I meant. Most of the commoncap functions call capable()
> > rather than directly calling cap_capable(), thereby causing:
> > - PF_SUPERPRIV to be set, and
> > - The primary security module (e.g. SELinux) to apply its own
> > restrictive check.
> >
> > That is useful as it allows SELinux or AppArmor or TOMOYO to veto e.g.
> > CAP_SYS_PTRACE without replicating the same logic within its own hook.
> >
> > The current exceptions are:
> > cap_inh_is_capped() called from cap_capset(),
> > cap_task_prctl() in the PR_SET_SECUREBITS case,
> > cap_vm_enough_memory(),
> > cap_file_mmap() after your patch.
> >
> > The latter two are indeed cases where we made a conscious choice that
> > SELinux would not apply its capability check against policy. But the
> > first two are unclear to me.
>
> cap_inh_is_capped:
>
> I'm not sure why it's cap_capable() instead of capable(). However, if we
> switch to using capable(), then we should switch the conditions in the caller
> around, since at the moment just because the capable() check returned true
> doesn't mean that we actually end up needing it.
>
> (CC:ing Andrew Morgan as I believe he wrote this and may have had a reason)
>
> cap_task_prctl: I don't see any reason why that shouldn't be capable().
>
> cap_vm_enough_memory(): I seem to recall we explicitly decided that we
> did not want PF_SUPERPRIV set in this case.

Yes - that one was intentional, as it gets applied to all tasks that
allocate mappings. For the same reason, we don't audit that check in
SELinux.

> cap_file_mmap(): well now that you mention it, it does seem like SELinux
> would want a say in whether the task gets CAP_SYS_RAWIO here, so maybe
> it should be capable() after all?

No, we apply our own specific permission check for it.

--
Stephen Smalley
National Security Agency

--
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/