Closing the BPF map permission loophole

[Roberto Sassu]

On Thu, 2022-09-29 at 08:27 -0700, Casey Schaufler wrote:
> On 9/29/2022 12:54 AM, Roberto Sassu wrote:
> > On Wed, 2022-09-28 at 20:24 -0400, Paul Moore wrote:
> > > On Wed, Sep 28, 2022 at 7:24 AM Roberto Sassu
> > > <roberto.sassu at huaweicloud.com> wrote
> > > > On Wed, 2022-09-28 at 12:33 +0200, Toke Høiland-Jørgensen
> > > > wrote:
> > > > > Roberto Sassu <roberto.sassu at huaweicloud.com> writes:
> > > > > 
> > > > > > On Wed, 2022-09-28 at 09:52 +0100, Lorenz Bauer wrote:
> > > > > > > On Mon, 26 Sep 2022, at 17:18, Roberto Sassu wrote:
> > > > > > > > Uhm, if I get what you mean, you would like to add DAC
> > > > > > > > controls
> > > > > > > > to
> > > > > > > > the
> > > > > > > > pinned map to decide if you can get a fd and with which
> > > > > > > > modes.
> > > > > > > > 
> > > > > > > > The problem I see is that a map exists regardless of
> > > > > > > > the
> > > > > > > > pinned
> > > > > > > > path
> > > > > > > > (just by ID).
> > > > > > > Can you spell this out for me? I imagine you're talking
> > > > > > > about
> > > > > > > MAP_GET_FD_BY_ID, but that is CAP_SYS_ADMIN only, right?
> > > > > > > Not
> > > > > > > great
> > > > > > > maybe, but no gaping hole IMO.
> > > > > > +linux-security-module ML (they could be interested in this
> > > > > > topic
> > > > > > as
> > > > > > well)
> > > > > > 
> > > > > > Good to know! I didn't realize it before.
> > > > > > 
> > > > > > I figured out better what you mean by escalating
> > > > > > privileges.
> > > > > > 
> > > > > > Pin a read-only fd, get a read-write fd from the pinned
> > > > > > path.
> > > > > > 
> > > > > > What you want to do is, if I pin a read-only fd, I should
> > > > > > get
> > > > > > read-
> > > > > > only
> > > > > > fds too, right?
> > > > > > 
> > > > > > I think here there could be different views. From my
> > > > > > perspective,
> > > > > > pinning is just creating a new link to an existing object.
> > > > > > Accessing
> > > > > > the link does not imply being able to access the object
> > > > > > itself
> > > > > > (the
> > > > > > same happens for files).
> > > > > > 
> > > > > > I understand what you want to achieve. If I have to choose
> > > > > > a
> > > > > > solution,
> > > > > > that would be doing something similar to files, i.e. add
> > > > > > owner
> > > > > > and
> > > > > > mode
> > > > > > information to the bpf_map structure (m_uid, m_gid,
> > > > > > m_mode). We
> > > > > > could
> > > > > > add the MAP_CHMOD and MAP_CHOWN operations to the bpf()
> > > > > > system
> > > > > > call
> > > > > > to
> > > > > > modify the new fields.
> > > > > > 
> > > > > > When you pin the map, the inode will get the owner and mode
> > > > > > from
> > > > > > bpf_map. bpf_obj_get() will then do DAC-style verification
> > > > > > similar
> > > > > > to
> > > > > > MAC-style verification (with security_bpf_map()).
> > > > > As someone pointed out during the discussing at LPC, this
> > > > > will
> > > > > effectively allow a user to create files owned by someone
> > > > > else,
> > > > > which
> > > > > is
> > > > > probably not a good idea either from a security PoV. (I.e.,
> > > > > user
> > > > > A
> > > > > pins
> > > > > map owned by user B, so A creates a file owned by B).
> > > > Uhm, I see what you mean. Right, it is not a good idea, the
> > > > owner
> > > > of
> > > > the file should the one that pinned the map.
> > > > 
> > > > Other than that, DAC verification on the map would be still
> > > > correct, as
> > > > it would be independent from the DAC verification of the file.
> > > I only became aware of this when the LSM list was CC'd so I'm a
> > > little
> > > behind on what is going on here ... looking quickly through the
> > > mailing list archive it looks like there is an issue with BPF map
> > > permissions not matching well with their associated fd
> > > permissions,
> > > yes?  From a LSM perspective, there are a couple of hooks that
> > > currently use the fd's permissions (read/write) to determine the
> > > appropriate access control check.
> > > From what I understood, access control on maps is done in two
> > > steps.
> > First, whenever someone attempts to get a fd to a map
> > security_bpf_map() is called. LSM implementations could check
> > access if
> > the current process has the right to access the map (whose label
> > can be
> > assigned at map creation time with security_bpf_map_alloc()).
> > 
> > Second, whenever the holder of the obtained fd wants to do an
> > operation
> > on the map (lookup, update, delete, ...), eBPF checks if the fd
> > modes
> > are compatible with the operation to perform (e.g. lookup requires
> > FMODE_CAN_READ).
> > 
> > One problem is that the second part is missing for some operations
> > dealing with the map fd:
> > 
> > Map iterators:
> > https://lore.kernel.org/bpf/20220906170301.256206-1-roberto.sassu@huaweicloud.com/
> > 
> > Map fd directly used by eBPF programs without system call:
> > https://lore.kernel.org/bpf/20220926154430.1552800-1-roberto.sassu@huaweicloud.com/
> > 
> > Another problem is that there is no DAC, only MAC (work in
> > progress). I
> > don't know exactly the status of enabling unprivileged eBPF.
> > 
> > Apart from this, now the discussion is focusing on the following
> > problem. A map (kernel object) can be referenced in two ways: by ID
> > or
> > by path. By ID requires CAP_ADMIN, so we can consider by path for
> > now.
> > 
> > Given a map fd, the holder of that fd can create a new reference
> > (pinning) to the map in the bpf filesystem (a new file whose
> > private
> > data contains the address of the kernel object).
> > 
> > Pinning a map does not have a corresponding permission. Any fd mode
> > is
> > sufficient to do the operation. Furthermore, subsequent requests to
> > obtain a map fd by path could result in receiving a read-write fd,
> > while at the time of pinning the fd was read-only.
> > 
> > While this does not seem to me a concern from MAC perspective, as
> > attempts to get a map fd still have to pass through
> > security_bpf_map(),
> > in general this should be fixed without relying on LSMs.
> > 
> > > Is the plan to ensure that the map and fd permissions are correct
> > > at
> > > the core BPF level, or do we need to do some additional checks in
> > > the
> > > LSMs (currently only SELinux)?
> > Should we add a new map_pin permission in SELinux?
> > 
> > Should we have DAC to restrict pinnning without LSMs?
> 
> As you've hinted above, DAC hasn't been an issue because there isn't
> unprivileged eBPF. Even with privileged eBPF I expect that there are
> going to be cases where not having DAC controls will surprise
> someone.
> The less BPF looks like low level kernel internals and the more it
> looks
> like general userspace code, the more likely this is to be an issue.
> 
> Or ...
> 
> If you are treating maps as kernel internal data structures you don't
> need DAC. If you are treating them as user accessible named objects
> you
> do need DAC. Security modules that implement MAC may chose to control
> kernel internal data access (e.g. SElinux) in addition to named
> objects,
> so you may want to accommodate that as well. If you do decide that
> maps
> are named objects Smack (and possibly AppArmor) needs significant
> work.
> Probably audit and IMA, too.

To me a map seems more than just a kernel object. User space gets a
reference to it through a fd, similarly to files. It seems a named
object because a map can be accessed by user space through a path
(pinned map) or by ID (user space can provide the ID to get a map fd
through a system call).

I'm more familiar with SELinux and Smack. But following what has been
done for SELinux, doing the same for Smack seems straightforward (even
easier, as you have more simple permissions).

Roberto