Problem with 9ba09998baa9 ("selinux: Implement the watch_key security hook") in linux-next

[Paul Moore]

On Tue, Apr 21, 2020 at 8:29 AM David Howells <dhowells at redhat.com> wrote:
> Paul Moore <paul at paul-moore.com> wrote:
>
> > ... in particular it is the fifth argument to avc_has_perm(),
> > "KEY_NEED_VIEW" which is a problem.  KEY_NEED_VIEW is not a SELinux
> > permission and would likely result in odd behavior when passed to
> > avc_has_perm().
>
> I think it works because KEY_NEED_VIEW happens to coincide with KEY__VIEW.  It
> should just use KEY__VIEW instead.

Yes, it looks like it.  To be clear, it is dangerous to rely on
permission values from outside SELinux aligning with SELinux
permissions; changing the outside permission values w/o adjusting the
SELinux hook code to do the necessary translation will result in some
scary behavior (wrong permission checks).

> > it probably makes the most sense to pull the permission mapping in
> > selinux_key_permission() out into a separate function, e.g.
> > key_perm_to_av(...)
>
> Agreed.  How about the attached patch?  I wonder if I should do bit-by-bit
> translation rather than using a switch?  But then it's difficult to decide
> what it means if someone passes in two KEY_NEED_* flags OR'd together - is it
> either or both?

Comments inline.

> > and then use this newly created mapping function in [...]
> > selinux_watch_key()
>
> No, I think I should just hard-code KEY__VIEW there.

FWIW, my comment was based on a version of linux-next where you were
making policycap based permission adjustments to KEY_VIEW and I
thought you would want the same adjustments to be applied to both
access control points.  That code appears to now be gone in
linux-next.

> ---
> commit 70d1d82aa014ae4511976b4d80c17138006dddec
> Author: David Howells <dhowells at redhat.com>
> Date:   Tue Apr 21 13:15:16 2020 +0100
>
>     selinux: Fix use of KEY_NEED_* instead of KEY__* perms
>
>     selinux_key_getsecurity() is passing the KEY_NEED_* permissions to
>     security_sid_to_context() instead of the KEY__* values.  It happens to work
>     because the values are all coincident.
>
>     Fixes: d720024e94de ("[PATCH] selinux: add hooks for key subsystem")
>     Reported-by: Paul Moore <paul at paul-moore.com>
>     Signed-off-by: David Howells <dhowells at redhat.com>
>
> diff --git a/security/selinux/hooks.c b/security/selinux/hooks.c
> index 0b4e32161b77..32f7fa538c5f 100644
> --- a/security/selinux/hooks.c
> +++ b/security/selinux/hooks.c
> @@ -6539,12 +6539,27 @@ static void selinux_key_free(struct key *k)
>         kfree(ksec);
>  }
>
> +static unsigned int selinux_keyperm_to_av(unsigned int need_perm)
> +{
> +       switch (need_perm) {
> +       case KEY_NEED_VIEW:     return KEY__VIEW;
> +       case KEY_NEED_READ:     return KEY__READ;
> +       case KEY_NEED_WRITE:    return KEY__WRITE;
> +       case KEY_NEED_SEARCH:   return KEY__SEARCH;
> +       case KEY_NEED_LINK:     return KEY__LINK;
> +       case KEY_NEED_SETATTR:  return KEY__SETATTR;
> +       default:
> +               return 0;
> +       }

Regarding your question of permission translation via switch-statement
as opposed to bit-by-bit comparison, I think it depends on if multiple
permissions are going to be required in a single call to the hook.
The failure mode for the code above if multiple permissions are
requested is not very good, it defaults to *no* permission which means
that if someone requested KEY_NEED_SEARCH|KEY_NEED_VIEW (or some other
combination) then the SELinux check would not check any permissions
... that seems wrong to me.

If we want to stick with a switch statement I think we should have it
return -EPERM for the default case to protect against this.  We don't
need the full 32-bits afforded us by the unsigned int.

> +}
> +
>  static int selinux_key_permission(key_ref_t key_ref,
>                                   const struct cred *cred,
> -                                 unsigned perm)
> +                                 unsigned need_perm)
>  {
>         struct key *key;
>         struct key_security_struct *ksec;
> +       unsigned int perm;
>         u32 sid;
>
>         /* if no specific permissions are requested, we skip the
> @@ -6553,6 +6568,7 @@ static int selinux_key_permission(key_ref_t key_ref,
>         if (perm == 0)
>                 return 0;
>
> +       perm = selinux_keyperm_to_av(need_perm);

... and add a check for (perm < 0) as discussed above if we stick with
the switch statement.

>         sid = cred_sid(cred);
>
>         key = key_ref_to_ptr(key_ref);
>

-- 
paul moore
www.paul-moore.com