On Fri, May 18, 2018 at 1:09 PM, Willem de Bruijn <willemdebruijn.ker...@gmail.com> wrote: > On Fri, May 18, 2018 at 11:44 AM, David Miller <da...@davemloft.net> wrote: >> From: Eric Dumazet <eric.duma...@gmail.com> >> Date: Fri, 18 May 2018 08:30:43 -0700 >> >>> We probably need to revert Willem patch >>> (7ce875e5ecb8562fd44040f69bda96c999e38bbc) >> >> Is it really valid to reach ip_recv_err with an ipv6 socket? > > I guess the issue is that setsockopt IPV6_ADDRFORM is not an > atomic operation, so that the socket is neither fully ipv4 nor fully > ipv6 by the time it reaches ip_recv_error. > > sk->sk_socket->ops = &inet_dgram_ops; > < HERE > > sk->sk_family = PF_INET; > > Even calling inet_recv_error to demux would not necessarily help. > > Safest would be to look up by skb->protocol, similar to what > ipv6_recv_error does to handle v4-mapped-v6. > > Or to make that function safe with PF_INET and swap the order > of the above two operations. > > All sound needlessly complicated for this rare socket option, but > I don't have a better idea yet. Dropping on the floor is not nice, > either.
Ensuring that ip_recv_error correctly handles packets from either socket and removing the warning should indeed be good. It is robust against v4-mapped packets from an AF_INET6 socket, but see caveat on reconnect below. The code between ipv6_recv_error for v4-mapped addresses and ip_recv_error is essentially the same, the main difference being whether to return network headers as sockaddr_in with SOL_IP or sockaddr_in6 with SOL_IPV6. There are very few other locations in the stack that explicitly test sk_family in this way and thus would be vulnerable to races with IPV6_ADDRFORM. I'm not sure whether it is possible for a udpv6 socket to queue a real ipv6 packet on the error queue, disconnect, connect to an ipv4 address, call IPV6_ADDRFORM and then call ip_recv_error on a true ipv6 packet. That would return buggy data, e.g., in msg_name.
