tcp_ao_prepare_reset() looks up the key for a reset that is sent on behalf of a listener by the address of the peer. It takes that address from the IPv6 header whenever the listener is an AF_INET6 socket, also when the segment is an IPv4 one that arrived on a dual-stack listener. The address is then read from the middle of the IPv4 header, the key of the peer is not found and the reset is not sent. The v4-mapped check that follows does not help: it looks at sk->sk_v6_daddr, which is the unspecified address for a listener, and it runs after the address has been picked. Use the protocol of the segment to tell the family, like tcp_v6_do_rcv() does. Fixes: 06b22ef29591 ("net/tcp: Wire TCP-AO to request sockets") Cc: stable@vger.kernel.org Signed-off-by: Hungyu Lin --- net/ipv4/tcp_ao.c | 6 ++---- 1 file changed, 2 insertions(+), 4 deletions(-) diff --git a/net/ipv4/tcp_ao.c b/net/ipv4/tcp_ao.c index bb7bbc20ba3f..656dbf805cfc 100644 --- a/net/ipv4/tcp_ao.c +++ b/net/ipv4/tcp_ao.c @@ -813,14 +813,12 @@ int tcp_ao_prepare_reset(const struct sock *sk, struct sk_buff *skb, sisn = th->seq; disn = 0; } + if (skb->protocol == htons(ETH_P_IP)) + family = AF_INET; if (IS_ENABLED(CONFIG_IPV6) && family == AF_INET6) addr = (union tcp_md5_addr *)&ipv6_hdr(skb)->saddr; else addr = (union tcp_md5_addr *)&ip_hdr(skb)->saddr; -#if IS_ENABLED(CONFIG_IPV6) - if (family == AF_INET6 && ipv6_addr_v4mapped(&sk->sk_v6_daddr)) - family = AF_INET; -#endif sk = sk_const_to_full_sk(sk); ao_info = rcu_dereference(tcp_sk(sk)->ao_info); -- 2.43.0