[PATCH 2/2] security.capability: fix conversions on getxattr

[Eric W. Biederman]

"Serge E. Hallyn" <serge at hallyn.com> writes:

> On Fri, Jan 29, 2021 at 04:55:29PM -0600, Eric W. Biederman wrote:
>> "Serge E. Hallyn" <serge at hallyn.com> writes:
>> 
>> > On Thu, Jan 28, 2021 at 02:19:13PM -0600, Eric W. Biederman wrote:
>> >> "Serge E. Hallyn" <serge at hallyn.com> writes:
>> >> 
>> >> > On Tue, Jan 19, 2021 at 07:34:49PM -0600, Eric W. Biederman wrote:
>> >> >> Miklos Szeredi <mszeredi at redhat.com> writes:
>> >> >> 
>> >> >> > If a capability is stored on disk in v2 format cap_inode_getsecurity() will
>> >> >> > currently return in v2 format unconditionally.
>> >> >> >
>> >> >> > This is wrong: v2 cap should be equivalent to a v3 cap with zero rootid,
>> >> >> > and so the same conversions performed on it.
>> >> >> >
>> >> >> > If the rootid cannot be mapped v3 is returned unconverted.  Fix this so
>> >> >> > that both v2 and v3 return -EOVERFLOW if the rootid (or the owner of the fs
>> >> >> > user namespace in case of v2) cannot be mapped in the current user
>> >> >> > namespace.
>> >> >> 
>> >> >> This looks like a good cleanup.
>> >> >
>> >> > Sorry, I'm not following.  Why is this a good cleanup?  Why should
>> >> > the xattr be shown as faked v3 in this case?
>> >> 
>> >> If the reader is in &init_user_ns.  If the filesystem was mounted in a
>> >> user namespace.   Then the reader looses the information that the
>> >
>> > Can you be more precise about "filesystem was mounted in a user namespace"?
>> > Is this a FUSE thing, the fs is marked as being mounted in a non-init userns?
>> > If that's a possible case, then yes that must be represented as v3.  Using
>> > is_v2header() may be the simpler way to check for that, but the more accurate
>> > check would be "is it v2 header and mounted by init_user_ns".
>> 
>> I think the filesystems current relevant are fuse,overlayfs,ramfs,tmpfs.
>> 
>> > Basically yes, in as many cases as possible we want to just give a v2
>> > cap because more userspace knows what to do with that, but a non-init-userns
>> > mounted fs which provides a v2 fscap should have it represented as v3 cap
>> > with rootid being the kuid that owns the userns.
>> 
>> That is the case we that is being fixed in the patch.
>> 
>> > Or am I still thinking wrongly?  Wouldn't be entirely surprised :)
>> 
>> No you got it.
>
> So then can we make faking a v3 gated on whether
>     sb->s_user_ns != &init_user_ns ?

Sort of.

What Miklos's patch implements is always treating a v2 cap xattr on disk
as v3 internally.

>  	if (is_v2header((size_t) ret, cap)) {
>  		root = 0;
>  	} else if (is_v3header((size_t) ret, cap)) {
>  		nscap = (struct vfs_ns_cap_data *) tmpbuf;
>  		root = le32_to_cpu(nscap->rootid);
>  	} else {
>  		size = -EINVAL;
>  		goto out_free;
>  	}

Then v3 is returned if:
>  	/* If the root kuid maps to a valid uid in current ns, then return
>  	 * this as a nscap. */
>  	mappedroot = from_kuid(current_user_ns(), kroot);
>  	if (mappedroot != (uid_t)-1 && mappedroot != (uid_t)0) {

After that we verify that the fs capability can be seen by the caller
as a v2 cap xattr with:

> >  	if (!rootid_owns_currentns(kroot)) {
> > 		size = -EOVERFLOW;
> > 		goto out_free;

Anything that passes that test and does not encounter a memory
allocation error is returned as a v2.

...

Which in practice does mean that if sb->s_user_ns != &init_user_ns, 
then mappedroot != 0, and is returned as a v3.

The rest of the logic takes care of all of the other crazy silly
combinations.  Like a user namespace that identity maps uid 0,
and then mounts a filesystem.

Eric