On Mon, Mar 4, 2019 at 8:53 PM Stephen Smalley [off-list ref] wrote:

On 3/4/19 12:01 PM, Mark Salyzyn wrote:

quoted

On 11/29/2018 05:49 AM, Vivek Goyal wrote:

quoted

So will override_creds=off solve the NFS issue also where all access will
happen with the creds of task now? Though it will stil require more
priviliges in task for other operations in overlay to succeed.

NFS problems seems to have ended the discussion, too many stakeholders?
too many outstanding questions?

Do we accept the limitations of the override_creds patch as is, and then
have the folks more familiar with the NFS scenario(s) build on it?

[TL;DR]

After looking at all this discussion, it feels like a larger audited
rewrite of the security model is in order and override_creds=off may be
a disservice (although expediently deals with Android's needs) to a
correct general solution. I admit I have little idea where to go from
here for a general solution.

As far as I see it, the model of creator && caller credentials is a
problem for any non-overlapping (MAC) privilege models. This patch
allows one to drop any creator privilege escalation, re-introducing the
"caller" to the lower layers.

As such I would expect a better model is to _always_ check the caller
credentials again in the lower layers, and only check the creator
credentials, some without caller credentials, for some special cases?
Change an && to an || for some of the checks? What are those special
cases? I must admit _none_ of those special cases need attention in the
Android usage models though making it difficult for me to do the fight
thing for the associated stakeholders.

As I recall, there were multiple problems with using current process'
creds for the operations on the lower/upper/work directories:

- Some overlayfs operations on the lower/upper/work directories required
privilege (capabilities) that the current process might lack, e.g. to
set ownership and the like on upper or work files, to set special xattrs
used internally by overlayfs for whiteouts or similar purposes, to act
on files within the work dir which was inaccessible to the current
process to prevent accessing files in an incomplete state, etc.
Originally that was handled by temporarily elevating the effective
capabilities around the privileged operation in the overlayfs code but
that didn't help with the SELinux or other LSM capability checking that
was triggered upon the capable calls.  Without some change there you'd
have to allow all client process domains all of the relevant
capabilities in policy, greatly increasing their privileges.

- The original logic was checking access to the lower dir/files in the
context of the current process when performing some operation that
modifies the file content or metadata, thereby triggering a SELinux/LSM
write or similar check, even though the actual data or metadata
modification occurs to the copied-up file instead and does not affect
the lower dir/files.  That was preventing making the lower dir/file
labels read-only to the client processes in the policy, which was
desired for the container use case.

You'd need to solve those problems in some way.

This is entire discussion in avoiding the elephant in the room -
Tampering with underlying layers.
We only talk about functionality and enforcement of legit overlayfs
operation, but what about an adversary that writes/changes files on
underlying layers? For example, permission to write new files can
be elevated to effectively modifying or deleting files.

To me the logic in "mounter credentials" is not only functional.
Whoever has access to underlying layer can really cause a lot of
damage to users accessing the overlay mount, so in a way, this
power needs to be translated to some sort of elevated credentials.

The the current model seems solid in that respect - access to
underlying layer can be restricted to the "mounter" who effectively
crafts an entirely new set of files.

Another way to think about this - consider a user space overlayfs
implementation, for example:
https://github.com/containers/fuse-overlayfs

In what way should the requirements to overlayfs security model
differ from the model already imposed on a user space implementation?

I do understand the Android use case for override_creds=off, but
to me, this is a functional configuration that says "turn off overlay
specific security considerations... I know what I am doing".
I am not saying that the Android adb debug use case is insecure.
I am just not seeing how a && or || model for mounter and caller
credential add up to a coherent story (yet...).

Thanks,
Amir.