f_modown and LSM inconsistency (was [PATCH v2 1/4] Landlock: Add signal control)

[Mickaël Salaün]

On Mon, Aug 12, 2024 at 12:30:03PM -0400, Paul Moore wrote:
> On Mon, Aug 12, 2024 at 11:06 AM Jann Horn <jannh at google.com> wrote:
> > On Mon, Aug 12, 2024 at 4:57 PM Paul Moore <paul at paul-moore.com> wrote:
> > > On Mon, Aug 12, 2024 at 9:09 AM Jann Horn <jannh at google.com> wrote:
> > > > On Mon, Aug 12, 2024 at 12:04 AM Paul Moore <paul at paul-moore.com> wrote:
> > >
> > > ...
> > >
> > > > > From a LSM perspective I suspect we are always going to need some sort
> > > > > of hook in the F_SETOWN code path as the LSM needs to potentially
> > > > > capture state/attributes/something-LSM-specific at that
> > > > > context/point-in-time.
> > > >
> > > > The only thing LSMs currently do there is capture state from
> > > > current->cred. So if the VFS takes care of capturing current->cred
> > > > there, we should be able to rip out all the file_set_fowner stuff.
> > > > Something like this (totally untested):
> > >
> > > I've very hesitant to drop the LSM hook from the F_SETOWN path both
> > > because it is reasonable that other LSMs may want to do other things
> > > here,
> >
> > What is an example for other things an LSM might want to do there? As
> > far as I understand, the whole point of this hook is to record the
> > identity of the sender of signals - are you talking about an LSM that
> > might not be storing credentials in struct cred, or something like
> > that?
> 
> Sure.  The LSM framework is intentionally very vague and limited on
> what restrictions it places on individual LSMs; we want to be able to
> support a wide range of security models and concepts.  I view the
> F_SETOWN hook are important because it is a control point that is used
> to set/copy/transfer/whatever security attributes from the current
> process to a file/fd for the purpose of managing signals on the fd.
> 
> > > and adding a LSM hook to the kernel, even if it is re-adding a
> > > hook that was previously removed, is a difficult and painful process
> > > with an uncertain outcome.
> >
> > Do you mean that even if the LSM hook ends up with zero users
> > remaining, you'd still want to keep it around in case it's needed
> > again later?
> 
> I want the security_file_set_fowner() hook to remain a viable hook for
> capturing the current task's security attributes, regardless of what
> security attributes the LSM is interested in capturing and where those
> attributes are stored.

I don't see the point to keep an unused hook, we could add it back later
if there is a valid use case, but I'll send a v2 without this removal.