On 22/12/2016 01:57, Al Viro wrote:
> On Thu, Dec 22, 2016 at 01:01:39AM +0100, Mickaël Salaün wrote:
> 
>> SELinux should be interested. This is useful to create sandboxes so
>> other LSM may be interested too
>>
>> I'm working on a new LSM and I would like this kind of hook to create a
>> real read-only environment.
> 
> What the...?  Have you noticed
>         if (!sb_start_write_trylock(inode->i_sb))
>                 return;
> 
>         if (__mnt_want_write(mnt) != 0)
>                 goto skip_update;
> in touch_atime()?  Just mount them read-only in your sandbox (on either
> level - both per-mountpoint and per-fs r/o will do) and be done
> with that; why bother with LSM when regular tools would suffice?
> 

Of course a read-only mount point can do the trick (except for anonymous
inodes). However, a security policy (e.g. for SELinux) should not (and
can't always) rely on mount options. For example, a security policy can
come from a distro but they may not want to tie mount options with this
policy. We may also not want a sandbox to being able to change mount
options (even with user namespaces).

Being able to write (meta-)data, whereas a security policy said that
it's not allowed, seems like a flaw in this policy. Moreover, modifying
access time is an easy way to create cover-channels without any LSM
being able to notice it.