Re: [kernel-hardening] Re: [PATCH] make TIOCSTI ioctl require CAP_SYS_ADMIN

[Serge E. Hallyn]

Quoting Serge E. Hallyn (serge@xxxxxxxxxx):
> Quoting Matt Brown (matt@xxxxxxxxx):
> > On 04/19/2017 07:53 PM, Serge E. Hallyn wrote:
> > >Quoting Matt Brown (matt@xxxxxxxxx):
> > >>On 04/19/2017 12:58 AM, Serge E. Hallyn wrote:
> > >>>On Tue, Apr 18, 2017 at 11:45:26PM -0400, Matt Brown wrote:
> > >>>>This patch reproduces GRKERNSEC_HARDEN_TTY functionality from the grsecurity
> > >>>>project in-kernel.
> > >>>>
> > >>>>This will create the Kconfig SECURITY_TIOCSTI_RESTRICT and the corresponding
> > >>>>sysctl kernel.tiocsti_restrict that, when activated, restrict all TIOCSTI
> > >>>>ioctl calls from non CAP_SYS_ADMIN users.
> > >>>>
> > >>>>Possible effects on userland:
> > >>>>
> > >>>>There could be a few user programs that would be effected by this
> > >>>>change.
> > >>>>See: <https://codesearch.debian.net/search?q=ioctl%5C%28.*TIOCSTI>
> > >>>>notable programs are: agetty, csh, xemacs and tcsh
> > >>>>
> > >>>>However, I still believe that this change is worth it given that the
> > >>>>Kconfig defaults to n. This will be a feature that is turned on for the
> > >>>
> > >>>It's not worthless, but note that for instance before this was fixed
> > >>>in lxc, this patch would not have helped with escapes from privileged
> > >>>containers.
> > >>>
> > >>
> > >>I assume you are talking about this CVE:
> > >>https://bugzilla.redhat.com/show_bug.cgi?id=1411256
> > >>
> > >>In retrospect, is there any way that an escape from a privileged
> > >>container with the this bug could have been prevented?
> > >
> > >I don't know, that's what I was probing for.  Detecting that the pgrp
> > >or session - heck, the pid namespace - has changed would seem like a
> > >good indicator that it shouldn't be able to push.
> > >
> > 
> > pgrp and session won't do because in the case we are discussing
> > current->signal->tty is the same as tty.
> > 
> > This is the current check that is already in place:
> >  | if ((current->signal->tty != tty) && !capable(CAP_SYS_ADMIN))
> >  | 	return -EPERM;
> 
> Yeah...
> 
> > The only thing I could find to detect the tty message coming from a
> > container is as follows:
> >  | task_active_pid_ns(current)->level
> > 
> > This will be zero when run on the host, but 1 when run inside a
> > container. However this is very much a hack and could probably break
> > some userland stuff where there are multiple levels of namespaces.
> 
> Yes.  This is also however why I don't like the current patch, because
> capable() will never be true in a container, so nested containers break.
> 
> What does current->signal->tty->pgrp actually point to?  Can we take
> it to be the pgrp which opened it?  Could we check
> ns_capable(current_pid_ns()->user_ns, CAP_whatever) and get a meaningful
> answer?
> 

Ok I see that's meaningless, you can't get pidns from pid.  We could
instead add a user_ns *owner to the struct tty and store the user_ns
of the task which opened it.  It's more invasive, but also more
meaningful.