Re: [PATCH 3/7] fs: Ignore file caps in mounts from other user namespaces

[ebiederm@xmission.com (Eric W. Biederman)]

Seth Forshee <seth.forshee@canonical.com> writes:

> On Thu, Jul 16, 2015 at 12:44:49AM -0500, Eric W. Biederman wrote:
>> Andy Lutomirski <luto@amacapital.net> writes:
>> 
>> > On Wed, Jul 15, 2015 at 10:04 PM, Eric W. Biederman
>> > <ebiederm@xmission.com> wrote:
>> >> Andy Lutomirski <luto@amacapital.net> writes:
>> >>
>> >>>
>> >>> So here's the semantic question:
>> >>>
>> >>> Suppose an unprivileged user (uid 1000) creates a user namespace and a
>> >>> mount namespace.  They stick a file (owned by uid 1000 as seen by
>> >>> init_user_ns) in there and mark it setuid root and give it fcaps.
>> >>
>> >> To make this make sense I have to ask, is this file on a filesystem
>> >> where uid 1000 as seen by the init_user_ns stored as uid 1000 on
>> >> the filesystem?  Or is this uid 0 as seen by the filesystem?
>> >>
>> >> I assume this is uid 0 on the filesystem in question or else your
>> >> unprivileged user would not have sufficient privileges over the
>> >> filesystem to setup fcaps.
>> >
>> > I was thinking uid 0 as seen by the filesystem.  But even if it were
>> > uid 1000, the unprivileged user can still set whatever mode and xattrs
>> > they want -- they control the backing store.
>> 
>> Yes.   And that is what I was really asking.  Are we taking about a
>> filesystem where the user controls the backing store?
>> 
>> >>> Then global root gets an fd to this filesystem.  If they execve the
>> >>> file directly, then, with my patch 4, it won't act as setuid 1000 and
>> >>> the fcaps will be ignored.  Even with my patch 4, though, if they bind
>> >>> mount the fs and execve the file from their bind mount, it will act as
>> >>> setuid 1000.  Maybe this is odd.  However, with Seth's patch 3, the
>> >>> fcaps will (correctly) not be honored.
>> >>
>> >> With patch 3 you can also think of it as fcaps being honored and you
>> >> get all the caps in the appropriate user namespace, but since you are
>> >> not in that user namespace and so don't have a place to store them
>> >> in struct cred you don't get the file caps.
>> >>
>> >> From the philosophy of interpreting the file as defined by the
>> >> filesystem in principle we could extend struct cred so you actually
>> >> get the creds just in uid 1000s user namespace, but that is very
>> >> unlikely to be worth it.
>> >
>> > I agree.
>> >
>> >>
>> >>> I tend to thing that, if we're not honoring the fcaps, we shouldn't be
>> >>> honoring the setuid bit either.  After all, it's really not a trusted
>> >>> file, even though the only user who could have messed with it really
>> >>> is the apparent owner.
>> >>
>> >> For the file caps we can't honor them because you don't have the bits
>> >> in struct cred.
>> >>
>> >> For setuid we can honor it, and setuid is something that the user
>> >> namespace allows.
>> >>
>> >
>> > We certainly *can* honor it.  But why should we?  I'd be more
>> > comfortable with this if the contents of an untrusted filesystem were
>> > really treated as just data.
>> 
>> In these weird bleed through situtations I don't know that we should.
>> But extending nosuid protections in this way is a bit like yama
>> a bit gratuitious stomping don't care cases in the semantics to
>> make bugs harder to exploit.
>> 
>> >>> And, if we're going to say we don't trust the file and shouldn't honor
>> >>> setuid or fcaps, then merging all the functionality into mnt_may_suid
>> >>> could make sense.  Yes, these two things do different things, but they
>> >>> could hook in to the same place.
>> >>
>> >> There are really two separate questions:
>> >> - Do we trust this filesystem?
>> >> - Do you have the bits to implement this concept?
>> >>
>> >> Even if in this specific context the two questions wind up looking
>> >> exactly the same. I think it makes a lot of sense to ask the two
>> >> questions separately.  As future maintenance changes may cause the
>> >> implementation of the questions to diverge.
>> >>
>> >
>> > Agreed.
>> >
>> > Unless someone thinks of an argument to the contrary, I'd say "no, we
>> > don't trust this filesystem".  I could be convinced otherwise.
>> 
>> But this is context dependent.  From the perspective of the container
>> we really do want to trust the filesystem.  As the container root set it
>> up, and if he isn't being hostile likely has a use for setfcaps files
>> and setuid files and all of the rest.
>> 
>> Perhaps I should phrase it as:
>> - In this context do we trust the code?   AKA mnt_may_suid?
>> - What do these bits mean in this context?  (Usually something more complicated).
>> 
>> Which says to me we want both patches 3 and 4 (even if 4 uses s_user_ns)
>> because 3 is different than 4.
>
> So what I'll do is:
>
>  - Add a s_user_ns check to mnt_may_suid
>  - Keep the (now redundant) s_user_ns check in get_file_caps
>
> I'm on the fence about having both the mnt and user ns checks in
> mnt_may_suid - it might be overkill, but it still adds the protection
> against clearing MNT_NOSUID in a bind mount. So I guess I'll keep the
> mnt ns check.

That sounds like a plan.

Eric