Re: [PATCH RESEND 2/5] seccomp: Add wait_killable semantic to seccomp user notifier

From: Tycho Andersen
Date: Tue Apr 27 2021 - 09:50:03 EST


On Mon, Apr 26, 2021 at 10:15:28PM +0000, Sargun Dhillon wrote:
> On Mon, Apr 26, 2021 at 01:02:29PM -0600, Tycho Andersen wrote:
> > On Mon, Apr 26, 2021 at 11:06:07AM -0700, Sargun Dhillon wrote:
> > > @@ -1103,11 +1111,31 @@ static int seccomp_do_user_notification(int this_syscall,
> > > * This is where we wait for a reply from userspace.
> > > */
> > > do {
> > > + interruptible = notification_interruptible(&n);
> > > +
> > > mutex_unlock(&match->notify_lock);
> > > - err = wait_for_completion_interruptible(&n.ready);
> > > + if (interruptible)
> > > + err = wait_for_completion_interruptible(&n.ready);
> > > + else
> > > + err = wait_for_completion_killable(&n.ready);
> > > mutex_lock(&match->notify_lock);
> > > - if (err != 0)
> > > +
> > > + if (err != 0) {
> > > + /*
> > > + * There is a race condition here where if the
> > > + * notification was received with the
> > > + * SECCOMP_USER_NOTIF_FLAG_WAIT_KILLABLE flag, but a
> > > + * non-fatal signal was received before we could
> > > + * transition we could erroneously end our wait early.
> > > + *
> > > + * The next wait for completion will ensure the signal
> > > + * was not fatal.
> > > + */
> > > + if (interruptible && !notification_interruptible(&n))
> > > + continue;
> >
> > I'm trying to understand how one would hit this race,
> >
>
> I'm thinking:
> P: Process that "generates" notification
> S: Supervisor
> U: User
>
> P: Generated notification
> S: ioctl(RECV...) // With wait_killable flag.
> ...complete is called in the supervisor, but the P may not be woken up...
> U: kill -SIGTERM $P
> ...signal gets delivered to p and causes wakeup and
> wait_for_completion_interruptible returns 1...
>
> Then you need to check the race

I see, thanks. This seems like a consequence of having the flag be
per-RECV-call vs. per-filter. Seems like it might be simpler to have
it be per-filter?

Tycho