[PATCH net 1/1] net: Don't deliver IPv6 packets to IPv4 sockets

[PATCH net 1/1] net: Don't deliver IPv6 packets to IPv4 sockets

[Shihuang Liu]

bpf_sk_assign() allows a TC ingress program to attach an arbitrary
hashed socket to an skb, without checking that the socket family
matches the packet's network layer. As a result, an IPv6 UDP packet
can be assigned to an AF_INET UDP socket: udpv6_rcv() steals the
socket via inet6_steal_sock(), which also lacks a family check, and
queues the IPv6 skb on the AF_INET socket.

recvmsg() on that socket then runs the IPv4 udp_recvmsg(), which
interprets the IPv6 skb control block as IPv4 IP options. When
IP_RETOPTS is enabled on the target socket, __ip_options_echo()
copies up to 153 bytes of attacker-controlled data from the IPv6
Destination Options extension header into the 40-byte option-data
area of the stack-allocated optbuf in ip_cmsg_recv_retopts():

    BUG: KASAN: stack-out-of-bounds in __ip_options_echo
    Write of size 153
    ...
    __ip_options_echo
    ip_cmsg_recv_offset
    udp_recvmsg

Reject sockets whose family is not AF_INET6 in inet6_steal_sock().
An IPv6 packet can never be legitimately delivered to an AF_INET
socket, so drop the stolen socket and return NULL, letting the
callers continue with the regular IPv6 lookup. When the socket is
refcounted, release it with sock_gen_put(), the same type-safe
helper sock_pfree() and sock_edemux() use, so the release stays
correct no matter which kind of socket a future BPF helper allows
to be assigned. This covers both the UDPv6 and TCPv6 receive
paths. The opposite direction, an IPv4 packet assigned to a
dual-stack AF_INET6 socket, remains allowed since that is a
supported configuration.

Fixes: cf7fbe660f2d ("bpf: Add socket assign support")
Cc: stable@vger.kernel.org
Assisted-by: GLM:GLM-5.3
Signed-off-by: Shihuang Liu <shlomojune6@gmail.com>
---
 include/net/inet6_hashtables.h | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/include/net/inet6_hashtables.h b/include/net/inet6_hashtables.h
index 2cc5d416bbb5..b39e59efac8d 100644
--- a/include/net/inet6_hashtables.h
+++ b/include/net/inet6_hashtables.h
@@ -115,6 +115,12 @@ struct sock *inet6_steal_sock(struct net *net, struct sk_buff *skb, int doff,
 	if (!sk)
 		return NULL;
 
+	if (unlikely(sk->sk_family != AF_INET6)) {
+		if (*refcounted)
+			sock_gen_put(sk);
+		return NULL;
+	}
+
 	if (!prefetched || !sk_fullsock(sk))
 		return sk;
 
-- 
2.43.0

Re: [PATCH net 1/1] net: Don't deliver IPv6 packets to IPv4 sockets

[Eric Dumazet]

On Sun, Aug 23, 2026 at 12:18 PM Shihuang Liu <shlomojune6@gmail.com> wrote:
>
> bpf_sk_assign() allows a TC ingress program to attach an arbitrary
> hashed socket to an skb, without checking that the socket family
> matches the packet's network layer. As a result, an IPv6 UDP packet
> can be assigned to an AF_INET UDP socket: udpv6_rcv() steals the
> socket via inet6_steal_sock(), which also lacks a family check, and
> queues the IPv6 skb on the AF_INET socket.
>
> recvmsg() on that socket then runs the IPv4 udp_recvmsg(), which
> interprets the IPv6 skb control block as IPv4 IP options. When
> IP_RETOPTS is enabled on the target socket, __ip_options_echo()
> copies up to 153 bytes of attacker-controlled data from the IPv6
> Destination Options extension header into the 40-byte option-data
> area of the stack-allocated optbuf in ip_cmsg_recv_retopts():
>
>

Really you need to ask your LLM to not slow down linux fast paths.

Fix bpf_sk_assign() and bpf_sk_assign_tcp_reqsk() instead.

Thank you.

pw-bot: cr