Re: [PATCH 2/2] groups: Allow unprivileged processes to use setgroups to drop groups

[Josh Triplett <josh-iaAMLnmF4UmaiuxdJuQwMA@public.gmane.org>]

On Sat, Nov 15, 2014 at 10:40:06PM -0500, Theodore Ts'o wrote:
> On Sat, Nov 15, 2014 at 06:35:05PM -0800, Josh Triplett wrote:
> > >So arbitrarily anyone to drop groups from their supplemental group
> > >list will result in a change from both existing practice and legacy
> > >Unix systems, and it could potentially lead to a security exposure.
> > 
> > As Andy pointed out, you can already do that with a user namespace,
> > for any case not involving a setuid or setgid (or otherwise
> > privilege-gaining) program.  And requiring no_new_privs handles
> > that.
> 
> Well, it's no worse than what we can do already with the user
> namespace, yes.  I'm still worried it's going to come as a surprise
> for some configurations because it's a change from what was allowed
> historically.  Then again, pretty much all of the tripwire and rootkit
> scanners won't notice a "setuid" program that uses capabilities
> instead of the traditional setuid bit, and most sysadmins won't think
> to check for an executable with a forced capability mask, so this
> isn't exactly a new problem....

We certainly have introduced orthogonal APIs in various areas before,
such that applications written prior to those APIs may interact
interestingly with them; we don't allow *breaking* those applications,
or introducing security holes, but the existence of applications
designed to block one particular way to do something doesn't
*automatically* rule out the possibility of adding another way to do it.
It does require some careful thought, though.

When we introduced seccomp filters for syscalls, we tied them to
no_new_privs to prevent any possible security holes caused by selective
syscall denial/filtration.

In this case, I'm indifferent about whether unprivileged setgroups works
without no_new_privs; if people are comfortable with that, fine, and if
people would prefer no_new_privs (or for that matter a sysctl, a
compile-time option, or any other means of making the behavior
optional), I can do that too.  The security model of "having a group
gives you less privilege than not having it" seems crazy, but
nonetheless I can see a couple of easy ways that we can avoid breaking
that pattern, no_new_privs being one of them.  I'd like to make sure
that nobody sees any other real-world corner case that unprivileged
setgroups would break.

- Josh Triplett