On Thu, 15 Aug 2019 20:46:01 +0200, Eric Dumazet wrote:
> On 8/15/19 6:00 PM, Davide Caratti wrote:
>
> >
> > + if (net_admin) {
> > + const struct tcp_ulp_ops *ulp_ops;
> > +
> > + rcu_read_lock();
> > + ulp_ops = icsk->icsk_ulp_ops;
> > + if (ulp_ops)
> > + err = tcp_diag_put_ulp(skb, sk, ulp_ops);
> > + rcu_read_unlock();
> > + if (err)
> > + return err;
> > + }
> > return 0;
>
>
> Why is rcu_read_lock() and rcu_read_unlock() used at all ?
>
> icsk->icsk_ulp_ops does not seem to be rcu protected ?
>
> If this was, then an rcu_dereference() would be appropriate.
Indeed it's ulp_data not ulp_ops that are protected. Davide,
perhaps we could push the RCU lock into tls_get_info(), after all?
And tls_context has to use rcu_deference there, as Eric points out,
plus we should probably NULL-check it.
