2025-02-03, 10:52:41 +0100, Sabrina Dubroca wrote:
> 2025-01-13, 10:31:26 +0100, Antonio Quartulli wrote:
> > +static void ovpn_encrypt_post(struct sk_buff *skb, int ret)
> > +{
> > + struct ovpn_peer *peer = ovpn_skb_cb(skb)->peer;
> > +
> > + if (unlikely(ret < 0))
> > + goto err;
> > +
> > + skb_mark_not_on_list(skb);
> > +
> > + switch (peer->sock->sock->sk->sk_protocol) {
>
> We have a ref on the peer, but not on the ovpn_sock. DEL_PEER could
> have detached the sock by the time the crypto completes.
>
> (unfortunately I don't have any idea to fix this yet)
Maybe an idea:
Since ovpn_sock itself lives under RCU (because of sk_user_data),
peer->sock should be an RCU pointer and also follow RCU rules. For
most parts (io.c, netlink.c) the conversion is not too
problematic. TCP is more difficult.
I still need to think about whether this works, and whether this is
worth the complexity, or if we could solve this in another way.
--
Sabrina
