[PATCH net 2/4] udp: Keep the bound device on disconnect unless connect() set it
From: Juan Agustin Niccolini
Date: Fri Nov 20 2026 - 12:44:39 EST
__udp_disconnect() clears sk_bound_dev_if unconditionally, without a
bind conflict check.
This is only safe for a device that connect() set, which
__ip6_datagram_connect() does for a link-local or multicast peer. The
device can also come from:
- The user, with SO_BINDTODEVICE, SO_BINDTOIFINDEX or a bind() to a
scoped IPv6 address.
- A cgroup program, as with ip vrf exec.
The bind conflict checks treat sockets bound to different devices as
not overlapping. As a result, an unprivileged user can bind an address
and port that another user's socket holds on another device, clear its
own device, and receive the other socket's datagrams:
victim: SO_BINDTODEVICE(lo), bind(0.0.0.0:P)
attacker: SO_BINDTODEVICE(eth0), bind(127.0.0.1:P)
attacker: connect(AF_UNSPEC) /* clears the device */
Record in sk_userlocks that connect() set the device of a socket that
had none, and only clear the device on disconnect in that case.
SO_BINDTODEVICE and SO_BINDTOIFINDEX drop the record.
This changes connect(AF_UNSPEC) for UDP, ping, raw and L2TP sockets,
which all use __udp_disconnect(): a device that connect() did not set
is now kept, so changing it afterwards needs CAP_NET_RAW, as on a
socket that never disconnected.
Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2")
Cc: stable@xxxxxxxxxxxxxxx
Signed-off-by: Juan Agustin Niccolini <juanagusnicco@xxxxxxxxx>
---
Notes:
- Within a VRF, connect() can replace the master with one of its
interfaces, which is now kept on disconnect, so binding back to the
master needs CAP_NET_RAW. Before, disconnect took the socket out of
its VRF. Restoring the master is not safe: the bind conflict checks
treat a master and its interfaces as different devices, so another
socket may hold the master by then. Fixing that is a separate,
VRF-wide change.
- Ping, raw and L2TP sockets call __udp_disconnect() without the socket
lock, so the sk_userlocks update is unlocked there. That predates
this change: commit 286c72deabaa ("udp: must lock the socket in
udp_disconnect()") only added the lock for UDP. Fixing it means
changing how ping, raw and L2TP handle disconnect, which is separate
from this fix, and a lost update cannot cause a bind overlap anyway.
include/net/sock.h | 4 ++++
net/core/sock.c | 1 +
net/ipv4/udp.c | 13 +++++++++----
net/ipv6/datagram.c | 17 +++++++++++++++--
4 files changed, 29 insertions(+), 6 deletions(-)
diff --git a/include/net/sock.h b/include/net/sock.h
index 173b37b2b8bb..78fc65adbcb8 100644
--- a/include/net/sock.h
+++ b/include/net/sock.h
@@ -1534,6 +1534,10 @@ static inline int __sk_prot_rehash(struct sock *sk)
* define SOCK_CONNECT_BIND - &sock->sk_userlocks flag for auto-bind at connect() time
*/
#define SOCK_CONNECT_BIND 16
+/**
+ * define SOCK_CONNECT_DEV - &sock->sk_userlocks flag for a connect()-set device
+ */
+#define SOCK_CONNECT_DEV 32
struct socket_alloc {
struct socket socket;
diff --git a/net/core/sock.c b/net/core/sock.c
index e2401d735c5a..5ce4aefc514d 100644
--- a/net/core/sock.c
+++ b/net/core/sock.c
@@ -650,6 +650,7 @@ static int sock_bindtoindex_locked(struct sock *sk, int ifindex)
/* Paired with all READ_ONCE() done locklessly. */
WRITE_ONCE(sk->sk_bound_dev_if, ifindex);
+ WRITE_ONCE(sk->sk_userlocks, sk->sk_userlocks & ~SOCK_CONNECT_DEV);
if (sk->sk_prot->rehash)
sk->sk_prot->rehash(sk);
diff --git a/net/ipv4/udp.c b/net/ipv4/udp.c
index 446729b41d66..9f1f45afee53 100644
--- a/net/ipv4/udp.c
+++ b/net/ipv4/udp.c
@@ -2195,6 +2195,7 @@ static int udp_connect(struct sock *sk, struct sockaddr_unsized *uaddr,
int __udp_disconnect(struct sock *sk, int flags)
{
+ u8 userlocks = READ_ONCE(sk->sk_userlocks);
struct inet_sock *inet = inet_sk(sk);
/*
* 1003.1g - break association.
@@ -2204,15 +2205,19 @@ int __udp_disconnect(struct sock *sk, int flags)
WRITE_ONCE(inet->inet_daddr, 0);
inet->inet_dport = 0;
sock_rps_reset_rxhash(sk);
- WRITE_ONCE(sk->sk_bound_dev_if, 0);
- if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK)) {
+ if (userlocks & SOCK_CONNECT_DEV) {
+ WRITE_ONCE(sk->sk_bound_dev_if, 0);
+ WRITE_ONCE(sk->sk_userlocks,
+ sk->sk_userlocks & ~SOCK_CONNECT_DEV);
+ }
+ if (!(userlocks & SOCK_BINDADDR_LOCK)) {
inet_reset_saddr(sk);
if (sk->sk_prot->rehash &&
- (sk->sk_userlocks & SOCK_BINDPORT_LOCK))
+ (userlocks & SOCK_BINDPORT_LOCK))
sk->sk_prot->rehash(sk);
}
- if (!(sk->sk_userlocks & SOCK_BINDPORT_LOCK)) {
+ if (!(userlocks & SOCK_BINDPORT_LOCK)) {
sk->sk_prot->unhash(sk);
inet->inet_sport = 0;
}
diff --git a/net/ipv6/datagram.c b/net/ipv6/datagram.c
index ff7a7ffbf983..0296207dbeb0 100644
--- a/net/ipv6/datagram.c
+++ b/net/ipv6/datagram.c
@@ -139,6 +139,14 @@ void ip6_datagram_release_cb(struct sock *sk)
}
EXPORT_SYMBOL_GPL(ip6_datagram_release_cb);
+static void ip6_datagram_connect_dev(struct sock *sk, int old_dif, int dif)
+{
+ if (!old_dif && dif)
+ WRITE_ONCE(sk->sk_userlocks,
+ sk->sk_userlocks | SOCK_CONNECT_DEV);
+ WRITE_ONCE(sk->sk_bound_dev_if, dif);
+}
+
int __ip6_datagram_connect(struct sock *sk, struct sockaddr_unsized *uaddr,
int addr_len)
{
@@ -229,11 +237,13 @@ int __ip6_datagram_connect(struct sock *sk, struct sockaddr_unsized *uaddr,
err = -EINVAL;
goto out;
}
- WRITE_ONCE(sk->sk_bound_dev_if, usin->sin6_scope_id);
+ ip6_datagram_connect_dev(sk, old_bound_dev_if,
+ usin->sin6_scope_id);
}
if (!sk->sk_bound_dev_if && (addr_type & IPV6_ADDR_MULTICAST))
- WRITE_ONCE(sk->sk_bound_dev_if, READ_ONCE(np->mcast_oif));
+ ip6_datagram_connect_dev(sk, old_bound_dev_if,
+ READ_ONCE(np->mcast_oif));
/* Connect to link-local address requires an interface */
if (!sk->sk_bound_dev_if) {
@@ -265,6 +275,9 @@ int __ip6_datagram_connect(struct sock *sk, struct sockaddr_unsized *uaddr,
np->flow_label = old_fl6_flowlabel;
inet->inet_dport = old_dport;
WRITE_ONCE(sk->sk_bound_dev_if, old_bound_dev_if);
+ if (!old_bound_dev_if)
+ WRITE_ONCE(sk->sk_userlocks,
+ sk->sk_userlocks & ~SOCK_CONNECT_DEV);
goto out;
}
--
2.43.0