Re: [PATCH v2 1/2] seccomp: notify user trap about unused filter

[Christian Brauner <christian.brauner@ubuntu.com>]

On Fri, May 29, 2020 at 01:06:59AM -0700, Kees Cook wrote:
> On Fri, May 29, 2020 at 09:56:41AM +0200, Christian Brauner wrote:
> > On Thu, May 28, 2020 at 04:11:00PM -0700, Kees Cook wrote:
> > > void seccomp_filter_release(const struct task_struct *tsk)
> > > {
> > > 	struct seccomp_filter *orig = READ_ONCE(tsk->seccomp.filter);
> > > 
> > > 	smp_store_release(&tsk->seccomp.filter, NULL);
> > 
> > I need to go through the memory ordering requirements before I can say
> > yay or nay confidently to this. :)
> > 
> > > 	__seccomp_filter_release(orig);
> > > }
> 
> The only caller will be release_task() after dethread, so honestly this
> was just me being paranoid. I don't think it actually needs the
> READ_ONCE() nor the store_release. I think I wrote all that before I'd
> convinced myself it was safe to remove a filter then. But I'm still
> suspicious given the various ways release_task() gets called... I just
> know that if mode 2 is set and filter == NULL, seccomp will fail closed,
> so I went the paranoid route. :)

release_task() should only be called once per thread otherwise we'd have
big problems. And every time we call release_task() we're already
EXIT_DEAD iirc. So there should be no way someone else can find us (in a
relevant way and especially not from userspace).
exit_notify() -> if we're autoreaping we're EXIT_DEAD anyway, if we're
not autoreaping we'll wait_task_zombie() eventually -> we're EXIT_DEAD
(parent has reaped us)

find_child_reaper() -> we couldn't find a child reaper for us, i.e. we
were (namespace) init -> unlink all tasks we were ptracing
(exit_ptrace()) and if they're EXIT_DEAD move them to the dead list ->
release_task()'s that are EXIT_DEAD that we ptraced.

and de_thread() relies on EXIT_DEAD too:
/*
 * We are going to release_task()->ptrace_unlink() silently,
 * the tracer can sleep in do_wait(). EXIT_DEAD guarantees
 * the tracer wont't block again waiting for this thread.
*/

(This is a very _rough_ sketch.)
So I think that's safe.

Christian