Re: [PATCH 1/4] seccomp: don't use semaphore and wait_queue together

[Andrei Vagin]

On Tue, Aug 30, 2022 at 2:57 AM Christian Brauner <brauner@xxxxxxxxxx> wrote:

> >
> > +static bool notify_wakeup(struct seccomp_filter *filter)
> > +{
> > +     bool ret;
> > +
> > +     rcu_read_lock();
> > +     ret = atomic_add_unless(&filter->notif->requests, -1, 0);
>
> Can you please spell out why the change to wait_event_interruptible()
> below that calls notify_wakeup() makes it necessary to rcu protect
> ->notif?

This is my mistake. rcu is used here when I tried to implement notify_wakeup
without introducing notif->request. The idea was to enumerate all elements of
notif->notifications. Now, it doesn't matter. In this context, filter->notif
can be dereferenced without any additional locks. Thanks for catching this.

>
> Given that you're using rcu_read_lock() here and the
> WRITE_ONCE(fitler->notify, NULL) + kfree_rcu() sequence in
> seccomp_notify_free() it seems to imply that filter->notif could be NULL
> here and so would need a non-NULL check on ->notif before incrementing?
>
> > +     rcu_read_unlock();
> > +
> > +     return ret;
> > +}
>
>  static long seccomp_notify_recv(struct seccomp_filter *filter,
>                               void __user *buf)
> @@ -1467,7 +1479,7 @@ static long seccomp_notify_recv(struct seccomp_filter *filter,
>
>       memset(&unotif, 0, sizeof(unotif));
>
> -     ret = down_interruptible(&filter->notif->request);
> +     ret = wait_event_interruptible(filter->wqh, notify_wakeup(filter));
>       if (ret < 0)
>               return ret;
>

Thanks,
Andrei