[PATCH v2 1/4] Landlock: Add signal control

[Jann Horn]

On Fri, Aug 9, 2024 at 12:59 PM Mickaël Salaün <mic at digikod.net> wrote:
> On Thu, Aug 08, 2024 at 04:42:23PM +0200, Jann Horn wrote:
> > On Thu, Aug 8, 2024 at 4:09 PM Mickaël Salaün <mic at digikod.net> wrote:
> > > On Thu, Aug 08, 2024 at 03:10:54AM +0200, Jann Horn wrote:
> > > > On Thu, Aug 8, 2024 at 1:36 AM Tahera Fahimi <fahimitahera at gmail.com> wrote:
> > > > > On Wed, Aug 07, 2024 at 08:16:47PM +0200, Mickaël Salaün wrote:
> > > > > > On Tue, Aug 06, 2024 at 11:55:27PM +0200, Jann Horn wrote:
> > > > > > > On Tue, Aug 6, 2024 at 8:56 PM Jann Horn <jannh at google.com> wrote:
> > > > > > > > On Tue, Aug 6, 2024 at 8:11 PM Tahera Fahimi <fahimitahera at gmail.com> wrote:
> > > > > > > > > Currently, a sandbox process is not restricted to send a signal
> > > > > > > > > (e.g. SIGKILL) to a process outside of the sandbox environment.
> > > > > > > > > Ability to sending a signal for a sandboxed process should be
> > > > > > > > > scoped the same way abstract unix sockets are scoped. Therefore,
> > > > > > > > > we extend "scoped" field in a ruleset with
> > > > > > > > > "LANDLOCK_SCOPED_SIGNAL" to specify that a ruleset will deny
> > > > > > > > > sending any signal from within a sandbox process to its
> > > > > > > > > parent(i.e. any parent sandbox or non-sandboxed procsses).
> > > > > > > [...]
> > > > > > > > > +       if (is_scoped)
> > > > > > > > > +               return 0;
> > > > > > > > > +
> > > > > > > > > +       return -EPERM;
> > > > > > > > > +}
> > > > > > > > > +
> > > > > > > > > +static int hook_file_send_sigiotask(struct task_struct *tsk,
> > > > > > > > > +                                   struct fown_struct *fown, int signum)
> > > > > >
> > > > > > I was wondering if we should handle this case, but I guess it makes
> > > > > > sense to have a consistent policy for all kind of user-triggerable
> > > > > > signals.
> > > > > >
> > > > > > > > > +{
> > > > > > > > > +       bool is_scoped;
> > > > > > > > > +       const struct landlock_ruleset *dom, *target_dom;
> > > > > > > > > +       struct task_struct *result = get_pid_task(fown->pid, fown->pid_type);
> > > > > > > >
> > > > > > > > I'm not an expert on how the fowner stuff works, but I think this will
> > > > > > > > probably give you "result = NULL" if the file owner PID has already
> > > > > > > > exited, and then the following landlock_get_task_domain() would
> > > > > > > > probably crash? But I'm not entirely sure about how this works.
> > > > > > > >
> > > > > > > > I think the intended way to use this hook would be to instead use the
> > > > > > > > "file_set_fowner" hook to record the owning domain (though the setup
> > > > > > > > for that is going to be kind of a pain...), see the Smack and SELinux
> > > > > > > > definitions of that hook. Or alternatively maybe it would be even
> > > > > > > > nicer to change the fown_struct to record a cred* instead of a uid and
> > > > > > > > euid and then use the domain from those credentials for this hook...
> > > > > > > > I'm not sure which of those would be easier.
> > > > > > >
> > > > > > > (For what it's worth, I think the first option would probably be
> > > > > > > easier to implement and ship for now, since you can basically copy
> > > > > > > what Smack and SELinux are already doing in their implementations of
> > > > > > > these hooks. I think the second option would theoretically result in
> > > > > > > nicer code, but it might require a bit more work, and you'd have to
> > > > > > > include the maintainers of the file locking code in the review of such
> > > > > > > refactoring and have them approve those changes. So if you want to get
> > > > > > > this patchset into the kernel quickly, the first option might be
> > > > > > > better for now?)
> > > > > > >
> > > > > >
> > > > > > I agree, let's extend landlock_file_security with a new "fown" pointer
> > > > > > to a Landlock domain. We'll need to call landlock_get_ruleset() in
> > > > > > hook_file_send_sigiotask(), and landlock_put_ruleset() in a new
> > > > > > hook_file_free_security().
> > > > > I think we should add a new hook (hook_file_set_owner()) to initialize
> > > > > the "fown" pointer and call landlock_get_ruleset() in that?
> > > >
> > > > Yeah. Initialize the pointer in the file_set_fowner hook, and read the
> > > > pointer in the file_send_sigiotask hook.
> > > >
> > > > Note that in the file_set_fowner hook, you'll probably need to use
> > > > both landlock_get_ruleset() (to take a reference on the ruleset you're
> > > > storing in the fown pointer) and landlock_put_ruleset() (to drop the
> > > > reference to the ruleset that the fown pointer was pointing to
> > > > before). And you'll need to use some kind of lock to protect the fown
> > > > pointer - either by adding an appropriate lock next to your fown
> > > > pointer or by using some appropriate existing lock in "struct file".
> > > > Probably it's cleanest to have your own lock for this? (This lock will
> > > > have to be something like a spinlock, not a mutex, since you need to
> > > > be able to acquire it in the file_set_fowner hook, which runs inside
> > > > an RCU read-side critical section, where sleeping is forbidden -
> > > > acquiring a mutex can sleep and therefore is forbidden in this
> > > > context, acquiring a spinlock can't sleep.)
> > >
> > > Yes, I think this should work for file_set_fowner:
> > >
> > > struct landlock_ruleset *prev_dom, *new_dom;
> > >
> > > new_dom = landlock_get_current_domain();
> > > landlock_get_ruleset(new_dom);
> > >
> > > /* Cf. f_modown() */
> > > write_lock_irq(&filp->f_owner.lock);
> > > prev_dom = rcu_replace_pointer(&landlock_file(file)->fown_domain,
> > >         new_dom, lockdep_is_held(&filp->f_owner.lock));
> > > write_unlock_irq(&filp->f_owner.lock);
> > >
> > > landlock_put_ruleset_rcu(prev_dom);
> > >
> > >
> > > With landlock_put_ruleset_rcu() define with this:
> > >
> > > diff --git a/security/landlock/ruleset.c b/security/landlock/ruleset.c
> > > index a93bdbf52fff..897116205520 100644
> > > --- a/security/landlock/ruleset.c
> > > +++ b/security/landlock/ruleset.c
> > > @@ -524,6 +524,20 @@ void landlock_put_ruleset_deferred(struct landlock_ruleset *const ruleset)
> > >         }
> > >  }
> > >
> > > +static void free_ruleset_rcu(struct rcu_head *const head)
> > > +{
> > > +       struct landlock_ruleset *ruleset;
> > > +
> > > +       ruleset = container_of(head, struct landlock_ruleset, rcu);
> > > +       free_ruleset(ruleset);
> > > +}
> >
> > free_ruleset() can block but RCU callbacks aren't allowed to block,
> > that's why landlock_put_ruleset_deferred() exists.
>
> Yes, but landlock_put_ruleset_deferred() doesn't wait for RCU read-side
> critical sections.

Ah, I phrased that badly - I didn't mean to suggest that you should
use landlock_put_ruleset_deferred() as a replacement for call_rcu().

[...]
> > So if you want to use RCU lifetime for this, I think you'll have to
> > turn landlock_put_ruleset() and landlock_put_ruleset_deferred() into
> > one common function that always, when reaching refcount 0, schedules
> > an RCU callback which then schedules a work_struct which then does
> > free_ruleset().
> >
> > I think that would be a little ugly, and it would look nicer to just
> > use normal locking in the file_send_sigiotask hook?
>
> I don't see how we can do that without delaying the free_ruleset() call
> to after the RCU read-side critical section in f_setown().

It should work if you used landlock_put_ruleset_deferred() instead of
landlock_put_ruleset().

> What about calling refcount_dec_and_test() in free_ruleset_rcu()?  That
> would almost always queue this call but it looks safe.

Every queued RCU invocation needs to have its own rcu_head - I think
the approach you're suggesting could end up queuing the same rcu_head
multiple times?

> An alternative might be to call synchronize_rcu() in free_ruleset(), but
> it's a big ugly too.
>
> BTW, I don't understand why neither SELinux nor Smack use (explicit)
> atomic operations nor lock.

Yeah, I think they're sloppy and kinda wrong - but it sorta works in
practice mostly because they don't have to do any refcounting around
this?

> And it looks weird that
> security_file_set_fowner() isn't called by f_modown() with the same
> locking to avoid races.

True. I imagine maybe the thought behind this design could have been
that LSMs should have their own locking, and that calling an LSM hook
with IRQs off is a little weird? But the way the LSMs actually use the
hook now, it might make sense to call the LSM with the lock held and
IRQs off...