Re: [RFC] refining systemd mountpoints

On Thu, Jan 9, 2020 at 10:42 PM Dominick Grift <[email protected]> wrote:
>
> On Thu, Jan 09, 2020 at 04:06:38PM -0500, Chris PeBenito wrote:
> > I'd like to refine how the policy handles systemd's mounton so that it works
> > similar to how we manage mountpoints for mount_t. Since systemd can be made
> > to mount over just about anything, I'm looking at adding a new conditional
> > that would allow init_t to mounton non_security_file_type, and then an
> > interface like files_mountpoint().
> >
> > The question is for the implementation of the interface; I see two options,
> > either the interface allows mounton for all file-like classes, or the
> > classes are specified as a parameter:
> >
> > --------
> > init.te:
> > attribute init_mountpoint_type;
> > allow init_t init_mountpoint_type:dir_file_class_set mounton;
> >
> > init.if:
> > interface(`init_mountpoint',`
> > typeattribute $1 init_mountpoint_type;
> > ')
> > --------
> >
> > or
> >
> > --------
> > init.if:
> > interface(`init_mountpoint',`
> > allow init_t $1:$2 mounton;
> > ')
> > --------
> >
> > I like the first option because it is clearer since you can see the mounton
> > in init.te, but that is excessive access.  The second option could be made
> > to look like the first option, but it would need several attributes and
> > interfaces, e.g. init_dir_mountpoint_type, init_file_mountpoint_type, etc.
> > which isn't so desirable.
> >
> > Any thoughts on this?
>
> I implemented the former in my policy. ie the dir_file_class_set equiv..
>
> 4163               (allow subj bind_path_obj_type_attribute (dirs (create)))
> 4164               (allow subj bind_path_obj_type_attribute list_dir_perms)
> 4165               (allow subj bind_path_obj_type_attribute (dir (mounton)))
> 4166               (allow subj bind_path_obj_type_attribute create_file_perms)
> 4167               (allow subj bind_path_obj_type_attribute (file (mounton)))
>
> As you can see i even allow systemd to create the mountpoint in case it does not exist. For example if /etc/machine-id does not exist and I have a BindReadOnlyPath=/etc/machine-id then systemd will touch /etc/machine-id and mount it ro
>
> It also generally buggy. Systemd does not (alway's) use setfscreatecon to create the mountpoints. And sometimes it does use setfscreatecon where it shouldnt.
>
> https://github.com/systemd/systemd/issues/13762
>

Using dir_file_class_set in the first option seems excessive. For
example systemd mounts /dev/kmsg as chr_file, as explained in
https://github.com/SELinuxProject/refpolicy/pull/144, and allowing
mounting on kmsg_device_t:dir would not make sense. Nevertheless, as
systemd seems to be the only things I know that mounts over /dev/kmsg,
it seems that dev_mounton_kmsg(init_t) could be replaced by
init_mountpoint(kmsg_device_t, chr_file) or
init_chr_mountpoint(kmsg_device_t).

That being said, I personally prefer the first option (with an
attribute) for types that are regular files and directories, as having
files labeled like directories is quite common.

In short, what about this?
--------
init.te:
attribute init_mountpoint_type;
allow init_t init_mountpoint_type:{dir, file} mounton;

init.if:
interface(`init_mountpoint',`
typeattribute $1 init_mountpoint_type;
')
interface(`init_mounton_chr',` # Or "init_chr_mountpoint"?
allow init_t $1:chr_file mounton; # An attribute could also be used here
')
--------

Cheers,
Nicolas