Re: [PATCH v19 0/4] overlayfs override_creds=off & nested get xattr fix

[Paul Moore <paul@paul-moore.com>]

On Wed, Mar 22, 2023 at 3:28 AM Johannes Segitz <jsegitz@suse.com> wrote:
> On Fri, Mar 11, 2022 at 03:52:54PM -0500, Paul Moore wrote:
> > On Fri, Mar 11, 2022 at 9:01 AM Vivek Goyal <vgoyal@redhat.com> wrote:
> > > Agreed. After going through the patch set, I was wondering what's the
> > > overall security model and how to visualize that.
> > >
> > > So probably there needs to be a documentation patch which explains
> > > what's the new security model and how does it work.
> >
> > Yes, of course.  I'll be sure to add a section to the existing docs.
> >
> > > Also think both in terms of DAC and MAC. (Instead of just focussing too
> > > hard on SELinux).
> >
> > Definitely.  Most of what I've been thinking about the past day or so
> > has been how to properly handle some of the DAC/capability issues; I
> > have yet to start playing with the code, but for the most part I think
> > the MAC/SELinux bits are already working properly.
> >
> > > My understanding is that in current model, some of the overlayfs
> > > operations require priviliges. So mounter is supposed to be priviliged
> > > and does the operation on underlying layers.
> > >
> > > Now in this new model, there will be two levels of check. Both overlay
> > > level and underlying layer checks will happen in the context of task
> > > which is doing the operation. So first of all, all tasks will need
> > > to have enough priviliges to be able to perform various operations
> > > on lower layer.
> > >
> > > If we do checks at both the levels in with the creds of calling task,
> > > I guess that probably is fine. (But will require a closer code inspection
> > > to make sure there is no privilege escalation both for mounter as well
> > > calling task).
> >
> > I have thoughts on this, but I don't think I'm yet in a position to
> > debate this in depth just yet; I still need to finish poking around
> > the code and playing with a few things :)
> >
> > It may take some time before I'm back with patches, but I appreciate
> > all of the tips and insight - thank you!
>
> Let me resurrect this discussion. With
> https://github.com/fedora-selinux/selinux-policy/commit/1e8688ea694393c9d918939322b72dfb44a01792
> the Fedora policy changed kernel_t to a confined domain. This means that
> many overlayfs setups that are created in initrd will now run into issues,
> as it will have kernel_t as part of the saved credentials.

Regardless of any overlayfs cred work, it seems like it would also be
worth spending some time to see if the kernel_t mounter creds
situation can also be improved.  I'm guessing this is due to mounts
happening before the SELinux policy is loaded?  Has anyone looked into
mounting the SELinux policy even earlier in these cases (may not be
possible) and/or umount/mount/remounting the affected overlayfs-based
filesystems after the policy has been loaded?

I can't say I'm the best person to comment on how the Fedora SELinux
policy is structured, but I do know a *little* about SELinux and I
think that accepting kernel_t as an overlayfs mounter cred is a
mistake.

> So while the
> original use case that inspired the patch set was probably not very common
> that now changed.
>
> It's tricky to work around this. Loading a policy in initrd causes a lot of
> issues now that kernel_t isn't unconfined anymore. Once the policy is
> loaded by systemd changing the mounts is tough since we use it for /etc and
> at this time systemd already has open file handles for policy files in
> /etc.

It's been a while since I worked on this, but I pretty much had to
give up on the read-write case, the overlayfs copy-up/work-dir
approach made this impractical, or at least I couldn't think of a sane
way to handle this without some sort of credential override.  However,
I did have a quick-and-dirty prototype that appeared to work well in
the read-only/no-work-dir case; I think I still have it in a
development branch somewhere, I can dig it back up and get it ported
to a modern kernel if there is any interest.

However, when discussing the prototype with Christian Brauner off-list
(added to the CC line) he still objected to the no-cred-override
approach and said it wasn't something he could support, so I dropped
it and focused on the other piles of fire lying about my desk (my
apologies to Christian if I'm mis-remembering/understanding the
conversation).  I still think there is value in supporting a
no-creds-override option, and if there is basic support for getting
this upstream I'm happy to pick the work back up, but I can't invest a
lot more time in this if there isn't an agreement from the
overlayfs/VFS maintainers that this is something that would consider.

--
paul-moore.com