Re: [PATCH v2 3/6] Teach SELinux about a new userfaultfd class

[Daniel Colascione <dancol@google.com>]

On Wed, Feb 12, 2020 at 11:10 AM Stephen Smalley <sds@tycho.nsa.gov> wrote:
>
> On 2/12/20 2:04 PM, Daniel Colascione wrote:
> > On Wed, Feb 12, 2020 at 10:59 AM Stephen Smalley <sds@tycho.nsa.gov> wrote:
> >>
> >> On 2/12/20 1:04 PM, Stephen Smalley wrote:
> >>> On 2/12/20 12:19 PM, Daniel Colascione wrote:
> >>>> Thanks for taking a look.
> >>>>
> >>>> On Wed, Feb 12, 2020 at 9:04 AM Stephen Smalley <sds@tycho.nsa.gov>
> >>>> wrote:
> >>>>>
> >>>>> On 2/11/20 5:55 PM, Daniel Colascione wrote:
> >>>>>> Use the secure anonymous inode LSM hook we just added to let SELinux
> >>>>>> policy place restrictions on userfaultfd use. The create operation
> >>>>>> applies to processes creating new instances of these file objects;
> >>>>>> transfer between processes is covered by restrictions on read, write,
> >>>>>> and ioctl access already checked inside selinux_file_receive.
> >>>>>>
> >>>>>> Signed-off-by: Daniel Colascione <dancol@google.com>
> >>>>>
> >>>>> (please add linux-fsdevel and viro to the cc for future versions of this
> >>>>> patch since it changes the VFS)
> >>>>>
> >>>>>> ---
> >>>>>> diff --git a/security/selinux/hooks.c b/security/selinux/hooks.c
> >>>>>> index 1659b59fb5d7..e178f6f40e93 100644
> >>>>>> --- a/security/selinux/hooks.c
> >>>>>> +++ b/security/selinux/hooks.c
> >>>>>> @@ -2915,6 +2919,69 @@ static int selinux_inode_init_security(struct
> >>>>>> inode *inode, struct inode *dir,
> >>>>>> +
> >>>>>> +     /*
> >>>>>> +      * We shouldn't be creating secure anonymous inodes before LSM
> >>>>>> +      * initialization completes.
> >>>>>> +      */
> >>>>>> +     if (unlikely(!selinux_state.initialized))
> >>>>>> +             return -EBUSY;
> >>>>>
> >>>>> I don't think this is viable; any arbitrary actions are possible before
> >>>>> policy is loaded, and a Linux distro can be brought up fully with
> >>>>> SELinux enabled and no policy loaded.  You'll just need to have a
> >>>>> default behavior prior to initialization.
> >>>>
> >>>> We'd have to fail open then, I think, and return an S_PRIVATE inode
> >>>> (the regular anon inode).
> >>>
> >>> Not sure why.  You aren't doing anything in the hook that actually
> >>> relies on selinux_state.initialized being set (i.e. nothing requires a
> >>> policy).  The avc_has_perm() call will just succeed until a policy is
> >>> loaded.  So if these inodes are created prior to policy load, they will
> >>> get assigned the task SID (which would be the kernel SID prior to policy
> >>> load or first exec or write to /proc/self/attr/current afterward) and
> >>> UFFD class (in your current code), be permitted, and then once policy is
> >>> loaded any further access will get checked against the kernel SID.
> >>>
> >>>>>> +     /*
> >>>>>> +      * We only get here once per ephemeral inode.  The inode has
> >>>>>> +      * been initialized via inode_alloc_security but is otherwise
> >>>>>> +      * untouched, so check that the state is as
> >>>>>> +      * inode_alloc_security left it.
> >>>>>> +      */
> >>>>>> +     BUG_ON(isec->initialized != LABEL_INVALID);
> >>>>>> +     BUG_ON(isec->sclass != SECCLASS_FILE);
> >>>>>
> >>>>> I think the kernel discourages overuse of BUG_ON/BUG/...
> >>>>
> >>>> I'm not sure what counts as overuse.
> >>>
> >>> Me either (not my rule) but I'm pretty sure this counts or you'd see a
> >>> lot more of these kinds of BUG_ON() checks throughout.  Try to reserve
> >>> them for really critical cases.
> >>>
> >>>>>> +
> >>>>>> +#ifdef CONFIG_USERFAULTFD
> >>>>>> +     if (fops == &userfaultfd_fops)
> >>>>>> +             isec->sclass = SECCLASS_UFFD;
> >>>>>> +#endif
> >>>>>
> >>>>> Not sure we want or need to introduce a new security class for each user
> >>>>> of anonymous inodes since the permissions should be the same as for
> >>>>> file.
> >>>>
> >>>> The purpose of this change is to apply special policy to userfaultfd
> >>>> FDs in particular. Isn't having a UFFD security class the best way to
> >>>> go about that? (There's no path.) Am I missing something?
> >>>
> >>> It is probably the simplest approach; it just doesn't generalize to all
> >>> users of anonymous inodes. We can distinguish them in one of two ways:
> >>> use a different class like you did (requires a code change every time we
> >>> add a new one and yet another duplicate of the file class) or use a
> >>> different SID/context/type. The latter could be achieved by calling
> >>> security_transition_sid() with the provided name wrapped in a qstr and
> >>> specifying type_transition rules on the name.  Then policy could define
> >>> derived types for each domain, ala
> >>> type_transition init self:file "[userfaultfd]" init_userfaultfd;
> >>> type_transition untrusted_app self:file "[userfaultfd]"
> >>> untrusted_app_userfaultfd;
> >>> ...
> >>>
> >>>>> Also not sure we want to be testing fops for each such case.
> >>>>
> >>>> I was also thinking of just providing some kind of context string
> >>>> (maybe the name), which might be friendlier to modules, but the loose
> >>>> coupling kind of scares me, and for this particular application, since
> >>>> UFFD is always in the core and never in a module, checking the fops
> >>>> seems a bit more robust and doesn't hurt anything.
> >>>
> >>> Yes, not sure how the vfs folks feel about either coupling (the
> >>> name-based one or the fops-based one).  Neither seems great.
> >>>
> >>>>> We
> >>>>> were looking at possibly leveraging the name as a key and using
> >>>>> security_transition_sid() to generate a distinct SID/context/type for
> >>>>> the inode via type_transition rules in policy.  We have some WIP along
> >>>>> those lines.
> >>>>
> >>>> Where? Any chance it would be ready soon? I'd rather not hold up this
> >>>> work for a more general mechanism.
> >>>
> >>> Hopefully will have a patch available soon.  But not saying this
> >>> necessarily has to wait either.
> >>>
> >>>>>> +     /*
> >>>>>> +      * Always give secure anonymous inodes the sid of the
> >>>>>> +      * creating task.
> >>>>>> +      */
> >>>>>> +
> >>>>>> +     isec->sid = tsec->sid;
> >>>>>
> >>>>> This doesn't generalize for other users of anonymous inodes, e.g. the
> >>>>> /dev/kvm case where we'd rather inherit the SID and class from the
> >>>>> original /dev/kvm inode itself.
> >>>>
> >>>> I think someone mentioned on the first version of this patch that we
> >>>> could make it more flexible if the need arose. If we do want to do it
> >>>> now, we could have the anon_inode security hook accept a "parent" or
> >>>> "context" inode that modules could inspect for the purposes of forming
> >>>> the new inode's SID. Does that make sense to you?
> >>>
> >>> Yes, that's the approach in our current WIP, except we call it a
> >>> "related" inode since it isn't necessarily connected to the anon inode
> >>> in any vfs sense.
> >>
> >> The other key difference in our WIP approach is that we assumed that we
> >> couldn't mandate allocating a separate anon inode for each of these fds
> >> and we wanted to cover all anonymous inodes (not opt-in), so we are
> >> storing the SID/class pair as additional fields in the
> >> file_security_struct and have modified file_has_perm() and others to
> >> look there for anonymous inodes.
> >
> > A separate inode seems like the simpler approach for now, because it
> > means that we have fewer places to check for security information ---
> > and it's not as if an inode is particularly expensive. We can always
> > switch later.
>
> We'd prefer having a separate inode if possible but didn't think that
> would fly with the vfs folks,

Let's ask them.

> especially if we try to apply this to all
> anonymous inodes.

For the moment, we're not.

> It might be ok for userfaultfd usage as a specific
> case but there is a reason why anonymous inodes were introduced and
> creating a separate inode each time defeats that purpose IIUC.  It will
> be interesting to see how they respond.

Sort of. Anonymous inodes also free other parts of the kernel from
having to deal with special-purpose filesystems (like pipefs) on which
to hang custom inodes. It's just a generic "just give me an inode and
I don't care about the filesystem" feature, and if we actually get a
new inode each time, we still do the job. Pipe seems to be good with
creating inodes each time.