[PATCH v1 net-next 0/2] net: Disallow buggy TCP transitions and IPV6_ADDRFORM.

[PATCH v1 net-next 0/2] net: Disallow buggy TCP transitions and IPV6_ADDRFORM.

[Kuniyuki Iwashima]

The following operations have recently received a lot of AI-driven
bug reports:

  * listen() + shutdown() + connect()
  * connect() + connect(AF_UNSPEC) + listen()

  * setsockopt(IPV6_ADDRFORM)

They are not worth fixing by churning the fast path, so this series
disallows the transitions and removes IPV6_ADDRFORM.

  https://lore.kernel.org/netdev/CANn89i+px52TtJy3S9=uHxGj3s-WueguRo1Z_4FxO=02KLmwmQ@mail.gmail.com/
  https://lore.kernel.org/netdev/5832ba17-c096-4f07-aa73-2e2bb6f79856@redhat.com/

I did not add Fixes tags, but if we want to backport, the tag
would be:

  Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2")


Kuniyuki Iwashima (2):
  tcp: Do not allow buggy transitions between ehash and lhash2.
  ipv6: Remove IPV6_ADDRFORM.

 include/linux/net.h                |  2 +-
 include/net/inet_connection_sock.h |  1 +
 net/core/sock.c                    |  4 +-
 net/ipv4/af_inet.c                 |  3 --
 net/ipv4/inet_connection_sock.c    |  1 +
 net/ipv4/inet_hashtables.c         | 11 ++++
 net/ipv6/af_inet6.c                |  4 --
 net/ipv6/ipv6_sockglue.c           | 87 ------------------------------
 8 files changed, 15 insertions(+), 98 deletions(-)

-- 
2.55.0.1003.g10538fe699-goog

[PATCH v1 net-next 1/2] tcp: Do not allow buggy transitions between ehash and lhash2.

[Kuniyuki Iwashima]

The following state transitions have long been a playground for
syzbot, and recently AI joined in, reporting a lot more bugs.

  * listen() + shutdown() + connect()
  * connect() + connect(AF_UNSPEC) + listen()

All the fix attempts would add more code to the fast path, which
is not worth it.

Instead of playing whack-a-mole with these edge-case bugs,
let's disallow these transitions.

Note that unhashed_state is placed in the 4-byte hole after
icsk_pmtu_cookie.

  $ pahole -C inet_connection_sock vmlinux
  struct inet_connection_sock {
  ...
	__u32                      icsk_pmtu_cookie;     /*  1208     4 */
	unsigned char              unhashed_state;       /*  1212     1 */

	/* XXX 3 bytes hole, try to pack */

Reported-by: Kyle Zeng <kylebot@openai.com>
Closes: https://lore.kernel.org/netdev/20260731140512.566464-1-david.lee@trailofbits.com/
Reported-by: Michal Luczaj <mhal@rbox.co>
Closes: https://lore.kernel.org/netdev/20260803-sockmap-lookup-tcp-leak-v2-0-306e025bfe66@rbox.co/
Reported-by: Hyunwoo Kim <imv4bel@gmail.com>
Closes: https://lore.kernel.org/netdev/20260824033331.1084971-1-imv4bel@gmail.com/
Signed-off-by: Kuniyuki Iwashima <kuniyu@google.com>
---
 include/net/inet_connection_sock.h |  1 +
 net/ipv4/inet_connection_sock.c    |  1 +
 net/ipv4/inet_hashtables.c         | 11 +++++++++++
 3 files changed, 13 insertions(+)

diff --git a/include/net/inet_connection_sock.h b/include/net/inet_connection_sock.h
index 433c2df23076..903499581db7 100644
--- a/include/net/inet_connection_sock.h
+++ b/include/net/inet_connection_sock.h
@@ -94,6 +94,7 @@ struct inet_connection_sock {
 	u32			  icsk_rto_max;
 	__u32                     icsk_delack_max;
 	__u32			  icsk_pmtu_cookie;
+	unsigned char		  unhashed_state;
 	const struct tcp_congestion_ops *icsk_ca_ops;
 	const struct inet_connection_sock_af_ops *icsk_af_ops;
 	const struct tcp_ulp_ops  *icsk_ulp_ops;
diff --git a/net/ipv4/inet_connection_sock.c b/net/ipv4/inet_connection_sock.c
index 6257459bcee2..560ca861b6d0 100644
--- a/net/ipv4/inet_connection_sock.c
+++ b/net/ipv4/inet_connection_sock.c
@@ -1253,6 +1253,7 @@ struct sock *inet_csk_clone_lock(const struct sock *sk,
 	memset(&newicsk->icsk_accept_queue, 0,
 	       sizeof(newicsk->icsk_accept_queue));
 
+	newicsk->unhashed_state = 0;
 	inet_sk_set_state(newsk, TCP_SYN_RECV);
 
 	inet_clone_ulp(req, newsk, priority);
diff --git a/net/ipv4/inet_hashtables.c b/net/ipv4/inet_hashtables.c
index ba0faa9ae2bb..5abcb0debb27 100644
--- a/net/ipv4/inet_hashtables.c
+++ b/net/ipv4/inet_hashtables.c
@@ -803,6 +803,11 @@ int inet_hash(struct sock *sk)
 	inet_init_ehash_secret();
 
 	WARN_ON(!sk_unhashed(sk));
+
+	if (unlikely(inet_csk(sk)->unhashed_state &&
+		     inet_csk(sk)->unhashed_state != TCP_LISTEN))
+		return -EINVAL;
+
 	ilb2 = inet_lhash2_bucket_sk(hashinfo, sk);
 
 	spin_lock(&ilb2->lock);
@@ -832,6 +837,9 @@ void inet_unhash(struct sock *sk)
 		return;
 
 	sock_rps_delete_flow(sk);
+
+	inet_csk(sk)->unhashed_state = sk->sk_state;
+
 	if (sk->sk_state == TCP_LISTEN) {
 		struct inet_listen_hashbucket *ilb2;
 
@@ -1058,6 +1066,9 @@ int __inet_hash_connect(struct inet_timewait_death_row *death_row,
 	int ret, i, low, high;
 	bool local_ports;
 
+	if (unlikely(inet_csk(sk)->unhashed_state == TCP_LISTEN))
+		return -EINVAL;
+
 	if (port) {
 		local_bh_disable();
 		ret = check_established(death_row, sk, port, NULL, false,
-- 
2.55.0.1003.g10538fe699-goog

[PATCH v1 net-next 2/2] ipv6: Remove IPV6_ADDRFORM.

[Kuniyuki Iwashima]

Recently, IPV6_ADDRFORM has received many AI-driven bug reports.
Fixing them properly would needlessly churn the fast paths in TCP
and UDP.

IPV6_ADDRFORM was initially introduced in RFC 2133 in 1997,
but only two years later, it was removed from RFC 2553 in 1999.

In 2026, modern applications natively support dual-stack sockets;
notably, systemd's socket activation does not use IPV6_ADDRFORM.

Also, getsockopt(IPV6_ADDRFORM) can be replaced with SO_DOMAIN.

Let's remove IPV6_ADDRFORM.

Later, we can remove sk->sk_prot_creator and revert commit
c26c192c3d48 ("udp: properly deal with xfrm encap and ADDRFORM").

Reported-by: Daehyeon Ko <4ncienth@gmail.com>
Closes: https://lore.kernel.org/netdev/20260902010408.1057857-1-4ncienth@gmail.com/
Reported-by: Hyunwoo Kim <imv4bel@gmail.com>
Closes: https://lore.kernel.org/netdev/20260824033331.1084971-1-imv4bel@gmail.com/
Signed-off-by: Kuniyuki Iwashima <kuniyu@google.com>
---
 include/linux/net.h      |  2 +-
 net/core/sock.c          |  4 +-
 net/ipv4/af_inet.c       |  3 --
 net/ipv6/af_inet6.c      |  4 --
 net/ipv6/ipv6_sockglue.c | 87 ----------------------------------------
 5 files changed, 2 insertions(+), 98 deletions(-)

diff --git a/include/linux/net.h b/include/linux/net.h
index 3d82966e2243..470100ae7107 100644
--- a/include/linux/net.h
+++ b/include/linux/net.h
@@ -166,7 +166,7 @@ struct socket {
 
 	struct file		*file;
 	struct sock		*sk;
-	const struct proto_ops	*ops; /* Might change with IPV6_ADDRFORM or MPTCP. */
+	const struct proto_ops	*ops; /* Might change with MPTCP. */
 
 	struct socket_wq	wq;
 };
diff --git a/net/core/sock.c b/net/core/sock.c
index fa60b7494c58..1d5927cd49a1 100644
--- a/net/core/sock.c
+++ b/net/core/sock.c
@@ -770,7 +770,7 @@ bool sk_mc_loop(const struct sock *sk)
 		return false;
 	if (!sk)
 		return true;
-	/* IPV6_ADDRFORM can change sk->sk_family under us. */
+
 	switch (READ_ONCE(sk->sk_family)) {
 	case AF_INET:
 		return inet_test_bit(MC_LOOP, sk);
@@ -4010,7 +4010,6 @@ int sock_common_getsockopt(struct socket *sock, int level, int optname,
 {
 	struct sock *sk = sock->sk;
 
-	/* IPV6_ADDRFORM can change sk->sk_prot under us. */
 	return READ_ONCE(sk->sk_prot)->getsockopt(sk, level, optname, optval, optlen);
 }
 EXPORT_SYMBOL(sock_common_getsockopt);
@@ -4032,7 +4031,6 @@ int sock_common_setsockopt(struct socket *sock, int level, int optname,
 {
 	struct sock *sk = sock->sk;
 
-	/* IPV6_ADDRFORM can change sk->sk_prot under us. */
 	return READ_ONCE(sk->sk_prot)->setsockopt(sk, level, optname, optval, optlen);
 }
 EXPORT_SYMBOL(sock_common_setsockopt);
diff --git a/net/ipv4/af_inet.c b/net/ipv4/af_inet.c
index 32d006c1a8ee..d9421ac38d78 100644
--- a/net/ipv4/af_inet.c
+++ b/net/ipv4/af_inet.c
@@ -582,7 +582,6 @@ int inet_dgram_connect(struct socket *sock, struct sockaddr_unsized *uaddr,
 	if (addr_len < sizeof(uaddr->sa_family))
 		return -EINVAL;
 
-	/* IPV6_ADDRFORM can change sk->sk_prot under us. */
 	prot = READ_ONCE(sk->sk_prot);
 
 	if (uaddr->sa_family == AF_UNSPEC)
@@ -789,7 +788,6 @@ int inet_accept(struct socket *sock, struct socket *newsock,
 {
 	struct sock *sk1 = sock->sk, *sk2;
 
-	/* IPV6_ADDRFORM can change sk->sk_prot under us. */
 	arg->err = -EINVAL;
 	sk2 = READ_ONCE(sk1->sk_prot)->accept(sk1, arg);
 	if (!sk2)
@@ -875,7 +873,6 @@ void inet_splice_eof(struct socket *sock)
 	if (unlikely(inet_send_prepare(sk)))
 		return;
 
-	/* IPV6_ADDRFORM can change sk->sk_prot under us. */
 	prot = READ_ONCE(sk->sk_prot);
 	if (prot->splice_eof)
 		prot->splice_eof(sock);
diff --git a/net/ipv6/af_inet6.c b/net/ipv6/af_inet6.c
index 282912a11999..f0efdc13baf4 100644
--- a/net/ipv6/af_inet6.c
+++ b/net/ipv6/af_inet6.c
@@ -429,7 +429,6 @@ int inet6_bind_sk(struct sock *sk, struct sockaddr_unsized *uaddr, int addr_len)
 	const struct proto *prot;
 	int err = 0;
 
-	/* IPV6_ADDRFORM can change sk->sk_prot under us. */
 	prot = READ_ONCE(sk->sk_prot);
 	/* If the socket has its own bind function then use it. */
 	if (prot->bind)
@@ -567,7 +566,6 @@ int inet6_ioctl(struct socket *sock, unsigned int cmd, unsigned long arg)
 	case SIOCSIFDSTADDR:
 		return addrconf_set_dstaddr(net, argp);
 	default:
-		/* IPV6_ADDRFORM can change sk->sk_prot under us. */
 		prot = READ_ONCE(sk->sk_prot);
 		if (!prot->ioctl)
 			return -ENOIOCTLCMD;
@@ -636,7 +634,6 @@ int inet6_sendmsg(struct socket *sock, struct msghdr *msg, size_t size)
 	if (unlikely(inet_send_prepare(sk)))
 		return -EAGAIN;
 
-	/* IPV6_ADDRFORM can change sk->sk_prot under us. */
 	prot = READ_ONCE(sk->sk_prot);
 	return INDIRECT_CALL_2(prot->sendmsg, tcp_sendmsg, udpv6_sendmsg,
 			       sk, msg, size);
@@ -651,7 +648,6 @@ int inet6_recvmsg(struct socket *sock, struct msghdr *msg, size_t size,
 	if (likely(!(flags & MSG_ERRQUEUE)))
 		sock_rps_record_flow(sk);
 
-	/* IPV6_ADDRFORM can change sk->sk_prot under us. */
 	prot = READ_ONCE(sk->sk_prot);
 	return INDIRECT_CALL_2(prot->recvmsg, tcp_recvmsg, udpv6_recvmsg,
 			       sk, msg, size, flags);
diff --git a/net/ipv6/ipv6_sockglue.c b/net/ipv6/ipv6_sockglue.c
index b4c977434c2e..1f68fb64a43e 100644
--- a/net/ipv6/ipv6_sockglue.c
+++ b/net/ipv6/ipv6_sockglue.c
@@ -547,86 +547,7 @@ int do_ipv6_setsockopt(struct sock *sk, int level, int optname,
 
 	sockopt_lock_sock(sk);
 
-	/* Another thread has converted the socket into IPv4 with
-	 * IPV6_ADDRFORM concurrently.
-	 */
-	if (unlikely(sk->sk_family != AF_INET6))
-		goto unlock;
-
 	switch (optname) {
-
-	case IPV6_ADDRFORM:
-		if (optlen < sizeof(int))
-			goto e_inval;
-		if (val == PF_INET) {
-			if (sk->sk_type == SOCK_RAW)
-				break;
-
-			if (sk->sk_protocol == IPPROTO_UDP) {
-				if (udp_sk(sk)->pending == AF_INET6) {
-					retv = -EBUSY;
-					break;
-				}
-			} else if (sk->sk_protocol == IPPROTO_TCP) {
-				if (sk->sk_prot != &tcpv6_prot) {
-					retv = -EBUSY;
-					break;
-				}
-			} else {
-				break;
-			}
-
-			if (sk->sk_state != TCP_ESTABLISHED) {
-				retv = -ENOTCONN;
-				break;
-			}
-
-			if (ipv6_only_sock(sk) ||
-			    !ipv6_addr_v4mapped(&sk->sk_v6_daddr)) {
-				retv = -EADDRNOTAVAIL;
-				break;
-			}
-
-			__ipv6_sock_mc_close(sk);
-			__ipv6_sock_ac_close(sk);
-
-			if (sk->sk_protocol == IPPROTO_TCP) {
-				struct inet_connection_sock *icsk = inet_csk(sk);
-
-				sock_prot_inuse_add(net, sk->sk_prot, -1);
-				sock_prot_inuse_add(net, &tcp_prot, 1);
-
-				/* Paired with READ_ONCE(sk->sk_prot) in inet6_stream_ops */
-				WRITE_ONCE(sk->sk_prot, &tcp_prot);
-				/* Paired with READ_ONCE() in tcp_(get|set)sockopt() */
-				WRITE_ONCE(icsk->icsk_af_ops, &ipv4_specific);
-				WRITE_ONCE(sk->sk_socket->ops, &inet_stream_ops);
-				WRITE_ONCE(sk->sk_family, PF_INET);
-				tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
-			} else {
-				sock_prot_inuse_add(net, sk->sk_prot, -1);
-				sock_prot_inuse_add(net, &udp_prot, 1);
-
-				/* Paired with READ_ONCE(sk->sk_prot) in inet6_dgram_ops */
-				WRITE_ONCE(sk->sk_prot, &udp_prot);
-				WRITE_ONCE(sk->sk_socket->ops, &inet_dgram_ops);
-				WRITE_ONCE(sk->sk_family, PF_INET);
-			}
-
-			/* Disable all options not to allocate memory anymore,
-			 * but there is still a race.  See the lockless path
-			 * in udpv6_sendmsg() and ipv6_local_rxpmtu().
-			 */
-			np->rxopt.all = 0;
-
-			inet6_cleanup_sock(sk);
-
-			module_put(THIS_MODULE);
-			retv = 0;
-			break;
-		}
-		goto e_inval;
-
 	case IPV6_V6ONLY:
 		if (optlen < sizeof(int) ||
 		    inet_sk(sk)->inet_num)
@@ -1088,14 +1009,6 @@ int do_ipv6_getsockopt(struct sock *sk, int level, int optname,
 	if (copy_from_sockptr(&len, optlen, sizeof(int)))
 		return -EFAULT;
 	switch (optname) {
-	case IPV6_ADDRFORM:
-		if (sk->sk_protocol != IPPROTO_UDP &&
-		    sk->sk_protocol != IPPROTO_TCP)
-			return -ENOPROTOOPT;
-		if (sk->sk_state != TCP_ESTABLISHED)
-			return -ENOTCONN;
-		val = sk->sk_family;
-		break;
 	case MCAST_MSFILTER:
 		if (in_compat_syscall())
 			return compat_ipv6_get_msfilter(sk, optval, optlen, len);
-- 
2.55.0.1003.g10538fe699-goog

Re: [PATCH v1 net-next 2/2] ipv6: Remove IPV6_ADDRFORM.

[Paolo Abeni]

On 9/4/26 5:35 AM, Kuniyuki Iwashima wrote:
> Recently, IPV6_ADDRFORM has received many AI-driven bug reports.
> Fixing them properly would needlessly churn the fast paths in TCP
> and UDP.
> 
> IPV6_ADDRFORM was initially introduced in RFC 2133 in 1997,
> but only two years later, it was removed from RFC 2553 in 1999.
> 
> In 2026, modern applications natively support dual-stack sockets;
> notably, systemd's socket activation does not use IPV6_ADDRFORM.
> 
> Also, getsockopt(IPV6_ADDRFORM) can be replaced with SO_DOMAIN.
> 
> Let's remove IPV6_ADDRFORM.

Yes please!

Acked-by: Paolo Abeni <pabeni@redhat.com>

> Later, we can remove sk->sk_prot_creator and revert commit
> c26c192c3d48 ("udp: properly deal with xfrm encap and ADDRFORM").

I'm not 110% sure syzkaller or others will find a way to still trigger
the race addressed there, but I guess we will see.

Also I guess we can get rid of several READ_ONCE(sk->sk_family).

/P

Re: [PATCH v1 net-next 1/2] tcp: Do not allow buggy transitions between ehash and lhash2.

[Jakub Sitnicki]

On Fri, Sep 04, 2026 at 03:35 AM GMT, Kuniyuki Iwashima wrote:
> The following state transitions have long been a playground for
> syzbot, and recently AI joined in, reporting a lot more bugs.
>
>   * listen() + shutdown() + connect()
>   * connect() + connect(AF_UNSPEC) + listen()
>
> All the fix attempts would add more code to the fast path, which
> is not worth it.
>
> Instead of playing whack-a-mole with these edge-case bugs,
> let's disallow these transitions.
>
> Note that unhashed_state is placed in the 4-byte hole after
> icsk_pmtu_cookie.
>
>   $ pahole -C inet_connection_sock vmlinux
>   struct inet_connection_sock {
>   ...
> 	__u32                      icsk_pmtu_cookie;     /*  1208     4 */
> 	unsigned char              unhashed_state;       /*  1212     1 */
>
> 	/* XXX 3 bytes hole, try to pack */
>
> Reported-by: Kyle Zeng <kylebot@openai.com>
> Closes: https://lore.kernel.org/netdev/20260731140512.566464-1-david.lee@trailofbits.com/
> Reported-by: Michal Luczaj <mhal@rbox.co>
> Closes: https://lore.kernel.org/netdev/20260803-sockmap-lookup-tcp-leak-v2-0-306e025bfe66@rbox.co/
> Reported-by: Hyunwoo Kim <imv4bel@gmail.com>
> Closes: https://lore.kernel.org/netdev/20260824033331.1084971-1-imv4bel@gmail.com/
> Signed-off-by: Kuniyuki Iwashima <kuniyu@google.com>
> ---
>  include/net/inet_connection_sock.h |  1 +
>  net/ipv4/inet_connection_sock.c    |  1 +
>  net/ipv4/inet_hashtables.c         | 11 +++++++++++
>  3 files changed, 13 insertions(+)
>
> diff --git a/include/net/inet_connection_sock.h b/include/net/inet_connection_sock.h
> index 433c2df23076..903499581db7 100644
> --- a/include/net/inet_connection_sock.h
> +++ b/include/net/inet_connection_sock.h
> @@ -94,6 +94,7 @@ struct inet_connection_sock {
>  	u32			  icsk_rto_max;
>  	__u32                     icsk_delack_max;
>  	__u32			  icsk_pmtu_cookie;
> +	unsigned char		  unhashed_state;
>  	const struct tcp_congestion_ops *icsk_ca_ops;
>  	const struct inet_connection_sock_af_ops *icsk_af_ops;
>  	const struct tcp_ulp_ops  *icsk_ulp_ops;

Glad we went in that direction in the end. Makes like easier.

Nit: Could be a flag, like CAN_LISTEN or CAN_CONNECT?
Either INET_FLAGS_* or maybe we need ICSK_FLAGS_*?

Reviewed-by: Jakub Sitnicki <jakub@cloudflare.com>