Re: [PATCH] namei: permit linking with CAP_FOWNER in userns

[Dirk Steinmetz]

On Tue, 27 Oct 2015 09:33:44 -0500, Seth Forshee wrote:
> On Tue, Oct 20, 2015 at 04:09:19PM +0200, Dirk Steinmetz wrote:
> > Attempting to hardlink to an unsafe file (e.g. a setuid binary) from
> > within an unprivileged user namespace fails, even if CAP_FOWNER is held
> > within the namespace. This may cause various failures, such as a gentoo
> > installation within a lxc container failing to build and install specific
> > packages.
> > 
> > This change permits hardlinking of files owned by mapped uids, if
> > CAP_FOWNER is held for that namespace. Furthermore, it improves consistency
> > by using the existing inode_owner_or_capable(), which is aware of
> > namespaced capabilities as of 23adbe12ef7d3 ("fs,userns: Change
> > inode_capable to capable_wrt_inode_uidgid").
> > 
> > Signed-off-by: Dirk Steinmetz <public@xxxxxxxxxxxxxxxxxx>
> 
> Tested-by: Seth Forshee <seth.forshee@xxxxxxxxxxxxx>
> 
> This is hitting us in Ubuntu during some dpkg upgrades in containers.
> When upgrading a file dpkg creates a hard link to the old file to back
> it up before overwriting it. When packages upgrade suid files owned by a
> non-root user the link isn't permitted, and the package upgrade fails.
> This patch fixes our problem.
> 
> I did want to point what seems to be an inconsistency in how
> capabilities in user namespaces are handled with respect to inodes. When
> I started looking at this my initial thought was to replace
> capable(CAP_FOWNER) with capable_wrt_inode_uidgid(inode, CAP_FOWNER). On
> the face of it this should be equivalent to what's done here, but it
> turns out that capable_wrt_inode_uidgid requires that the inode's uid
> and gid are both mapped into the namespace whereas
> inode_owner_or_capable only requires the uid be mapped. I'm not sure how
> significant that is, but it seems a bit odd.

I agree that this seems odd. I've chosen inode_owner_or_capable over
capable_wrt_inode_uidgid(inode, CAP_FOWNER) as it seemed consistent:
a privileged user (with CAP_SETUID) can impersonate the owner UID and thus
bypass the check completely; this also matches the documented behavior of
CAP_FOWNER: "Bypass permission checks on operations that normally require
the filesystem UID of the process to match the UID of the file".

However, thinking about it I get the feeling that checking the gid seems
reasonable as well. This is, however, independently of user namespaces.
Consider the following scenario in any namespace, including the init one:
- A file has the setgid and user/group executable bits set, and is owned
  by user:group.
- The user 'user' is not in the group 'group', and does not have any
  capabilities.
- The user 'user' hardlinks the file. The permission check will succeed,
  as the user is the owner of the file.
- The file is replaced with a newer version (for example fixing a security
  issue)
- Now user can still use the hardlink-pinned version to execute the file
  as 'user:group' (and for example exploit the security issue).
I would have expected the user to not be able to hardlink, as he lacks
CAP_FSETID, and thus is not allowed to chmod, change or move the file
without loosing the setgid bit. So it is impossible for him to make a non-
hardlink copy with the setgid bit set -- why should he be able to make a
hardlinked one?

It seems to me as if may_linkat would additionally require a check
verifying that either
- not both setgid and group executable bit set
- fsgid == owner gid
- capable_wrt_inode_uidgid(CAP_FSETID) -- or CAP_FOWNER, depending on
  whether to adapt chmod's behavior or keeping everything hardlink-
  related in CAP_FOWNER; I don't feel qualified enough to pick ;)
This would change documented behavior (at least man proc.5's description
of /proc/sys/fs/protected_hardlinks), and I'd consider it a separate
issue, if any (as I'm unsure how realistic that scenario is). I'd
appreciate comments on that.

For other situations than setgid-executable files I do not see issues with
not checking the group id's mapping, as linking would be permitted without
privileges outside of the user namespace (disregarding namespace-internal
setuid bits).

Independently of that, it might be reasonable to consider switching
inode_owner_or_capable towards checking the gid as well and define
something along "uid checks in user namespaces with uid/gid maps require
the file's uid and gid to be mapped, else they will fail" for consistency.

Dirk

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/