On Thu, 2019-08-15 at 14:38 -0700, Jakub Kicinski wrote:
> On Thu, 15 Aug 2019 20:46:01 +0200, Eric Dumazet wrote:
hello Eric and Jakub, thanks a lot for looking at this.
> > 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.
the goal is to protect execution of 'ss -tni' against concurrent removal
of tls.ko module, similarly to what was done in inet_sk_diag_fill() when
INET_DIAG_CONG is requested [1]. But after reading more carefully, the
assignment of ulp_ops needs to be:
ulp_ops = READ_ONCE(icsk->icsk_ulp_ops);
which I lost in internal reviews, with some additional explanatory
comment. Ok if I correct the above hunk with READ_ONCE() and add a
comment?
> Davide, perhaps we could push the RCU lock into tls_get_info(), after all?
It depends on whether concurrent dump / module removal is an issue for TCP
ULPs, like it was for congestion control schemes [1]. Any advice?
> And tls_context has to use rcu_deference there, as Eric points out,
> plus we should probably NULL-check it.
yes, it makes sense, for patch 3/3, in the assignment of 'ctx'. Instead of
calling tls_get_ctx() in tls_get_info() I will do
ctx = rcu_dereference(inet_csk(sk)->icsk_ulp_data);
and let it return 0 in case of NULL ctx (as it doesn't look like a faulty
situation). Ok?
--
davide
[1] see:
commit 521f1cf1dbb9d5ad858dca5dc75d1b45f64b6589
Author: Eric Dumazet <eduma...@google.com>
Date: Thu Apr 16 18:10:35 2015 -0700
inet_diag: fix access to tcp cc information
