From: Juan Agustin Niccolini <juanagusnicco@gmail.com>
To: "David S . Miller" <davem@davemloft.net>,
Eric Dumazet <edumazet@kernel.org>,
Jakub Kicinski <kuba@kernel.org>, Paolo Abeni <pabeni@redhat.com>
Cc: Simon Horman <horms@kernel.org>,
Kuniyuki Iwashima <kuniyu@google.com>,
Willem de Bruijn <willemdebruijn.kernel@gmail.com>,
Willem de Bruijn <willemb@google.com>, Eric Garver <e@erig.me>,
Shuah Khan <shuah@kernel.org>,
linux-kselftest@vger.kernel.org, netdev@vger.kernel.org,
linux-kernel@vger.kernel.org
Subject: [PATCH net 3/4] udp: Ignore a connect()-set device in the bind conflict check
Date: Fri, 20 Nov 2026 12:43:40 -0500 [thread overview]
Message-ID: <20261120174341.35873-4-juanagusnicco@gmail.com> (raw)
In-Reply-To: <20261120174341.35873-1-juanagusnicco@gmail.com>
The bind conflict checks compare sk_bound_dev_if, which connect() also
writes, for a link-local or multicast IPv6 peer.
This is only safe if the device is the one that bind() reserved. A
socket bound with no device holds its port on every device, but while
it is connected to such a peer it looks bound to one device, so a bind
of the same port on another device is admitted.
As a result, an unprivileged user can bind a port that another user's
socket holds, and receive the datagrams that arrive on its device once
that socket disconnects:
victim: bind([::]:P), connect(fe80::1%eth0)
attacker: SO_BINDTODEVICE(eth1), bind([::]:P) /* should fail */
victim: connect(AF_UNSPEC)
Compare the device that bind() reserved instead: none for a socket
whose device connect() set, which SOCK_CONNECT_DEV records. A VRF
master that connect() narrowed to one of its interfaces does not get
the flag and is compared as before.
Note that the checks read the flag without the other socket's lock, so
it is ordered against the device with barriers: seeing the device
before the flag would trust a connect()-set device again.
Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2")
Cc: stable@vger.kernel.org
Signed-off-by: Juan Agustin Niccolini <juanagusnicco@gmail.com>
---
Notes:
- A connect() that sets the device and is undone again, both between
the two reads of the flag, can still be seen as a bound device, so a
conflicting bind can be admitted. Not fixed here because closing it
needs a sequence count in struct sock, a larger change for a window
that the other socket's owner has to open within two loads.
- connect() also narrows the local address of a wildcard-bound
socket, which opens the same window through the address. It is not
fixed here: connect() moves the socket to the secondary hash slot of
the narrowed address, so a specific-address bind on a busy port never
scans it. Closing that needs either a full scan on every bind or no
rehash on connect, which slows down binds or lookups on busy ports,
so I leave it for a separate patch.
net/core/sock.c | 2 ++
net/ipv4/udp.c | 46 +++++++++++++++++++++++++++++++++++----------
net/ipv6/datagram.c | 12 ++++++++++--
3 files changed, 48 insertions(+), 12 deletions(-)
diff --git a/net/core/sock.c b/net/core/sock.c
index 5ce4aefc514d..48ec344a3a8b 100644
--- a/net/core/sock.c
+++ b/net/core/sock.c
@@ -650,6 +650,8 @@ 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);
+ /* Pairs with the first smp_rmb() in udp_bound_dev_if(). */
+ smp_wmb();
WRITE_ONCE(sk->sk_userlocks, sk->sk_userlocks & ~SOCK_CONNECT_DEV);
if (sk->sk_prot->rehash)
diff --git a/net/ipv4/udp.c b/net/ipv4/udp.c
index 9f1f45afee53..a13bf32cbdb9 100644
--- a/net/ipv4/udp.c
+++ b/net/ipv4/udp.c
@@ -131,11 +131,35 @@ EXPORT_PER_CPU_SYMBOL_GPL(udp_memory_per_cpu_fw_alloc);
#define MAX_UDP_PORTS 65536
#define PORTS_PER_CHAIN (MAX_UDP_PORTS / UDP_HTABLE_SIZE_MIN_PERNET)
-static bool udp_lport_same_bind(const struct sock *sk, const struct sock *sk2)
+static int udp_bound_dev_if(const struct sock *sk)
+{
+ int dif;
+
+ if (READ_ONCE(sk->sk_userlocks) & SOCK_CONNECT_DEV)
+ return 0;
+ /* Pairs with the smp_wmb() before SOCK_CONNECT_DEV is cleared. */
+ smp_rmb();
+ dif = READ_ONCE(sk->sk_bound_dev_if);
+ /* Pairs with the smp_wmb() after SOCK_CONNECT_DEV is set. */
+ smp_rmb();
+ if (READ_ONCE(sk->sk_userlocks) & SOCK_CONNECT_DEV)
+ return 0;
+ return dif;
+}
+
+static bool udp_bound_dev_overlap(int dif, const struct sock *sk2)
+{
+ int dif2 = udp_bound_dev_if(sk2);
+
+ return !dif || !dif2 || dif == dif2;
+}
+
+static bool udp_lport_same_bind(const struct sock *sk, int dif,
+ const struct sock *sk2)
{
return sk2->sk_family == sk->sk_family &&
ipv6_only_sock(sk2) == ipv6_only_sock(sk) &&
- sk2->sk_bound_dev_if == sk->sk_bound_dev_if &&
+ udp_bound_dev_if(sk2) == dif &&
sk2->sk_reuse == sk->sk_reuse &&
inet_rcv_saddr_equal(sk, sk2, false);
}
@@ -145,6 +169,7 @@ static int udp_lib_lport_inuse(struct net *net, __u16 num,
unsigned long *bitmap,
struct sock *sk, unsigned int log)
{
+ int dif = udp_bound_dev_if(sk);
kuid_t uid = sk_uid(sk);
struct sock *sk2;
@@ -157,14 +182,13 @@ static int udp_lib_lport_inuse(struct net *net, __u16 num,
sk2 != sk &&
(bitmap || udp_sk(sk2)->udp_port_hash == num) &&
(bitmap || !sk2->sk_reuse || !sk->sk_reuse) &&
- (!sk2->sk_bound_dev_if || !sk->sk_bound_dev_if ||
- sk2->sk_bound_dev_if == sk->sk_bound_dev_if) &&
- inet_rcv_saddr_equal(sk, sk2, true)) {
+ inet_rcv_saddr_equal(sk, sk2, true) &&
+ udp_bound_dev_overlap(dif, sk2)) {
if (!bitmap) {
if (sk2->sk_reuseport && sk->sk_reuseport &&
!rcu_access_pointer(sk->sk_reuseport_cb) &&
uid_eq(uid, sk_uid(sk2))) {
- if (udp_lport_same_bind(sk, sk2))
+ if (udp_lport_same_bind(sk, dif, sk2))
return 0;
continue;
}
@@ -184,6 +208,7 @@ static int udp_lib_lport_inuse2(struct net *net, __u16 num,
struct udp_hslot *hslot2,
struct sock *sk)
{
+ int dif = udp_bound_dev_if(sk);
kuid_t uid = sk_uid(sk);
struct sock *sk2;
int res = 0;
@@ -194,13 +219,12 @@ static int udp_lib_lport_inuse2(struct net *net, __u16 num,
sk2 != sk &&
(udp_sk(sk2)->udp_port_hash == num) &&
(!sk2->sk_reuse || !sk->sk_reuse) &&
- (!sk2->sk_bound_dev_if || !sk->sk_bound_dev_if ||
- sk2->sk_bound_dev_if == sk->sk_bound_dev_if) &&
- inet_rcv_saddr_equal(sk, sk2, true)) {
+ inet_rcv_saddr_equal(sk, sk2, true) &&
+ udp_bound_dev_overlap(dif, sk2)) {
if (sk2->sk_reuseport && sk->sk_reuseport &&
!rcu_access_pointer(sk->sk_reuseport_cb) &&
uid_eq(uid, sk_uid(sk2))) {
- if (!udp_lport_same_bind(sk, sk2))
+ if (!udp_lport_same_bind(sk, dif, sk2))
continue;
res = 0;
} else {
@@ -2207,6 +2231,8 @@ int __udp_disconnect(struct sock *sk, int flags)
sock_rps_reset_rxhash(sk);
if (userlocks & SOCK_CONNECT_DEV) {
WRITE_ONCE(sk->sk_bound_dev_if, 0);
+ /* Pairs with the first smp_rmb() in udp_bound_dev_if(). */
+ smp_wmb();
WRITE_ONCE(sk->sk_userlocks,
sk->sk_userlocks & ~SOCK_CONNECT_DEV);
}
diff --git a/net/ipv6/datagram.c b/net/ipv6/datagram.c
index 0296207dbeb0..37a093f74724 100644
--- a/net/ipv6/datagram.c
+++ b/net/ipv6/datagram.c
@@ -141,9 +141,12 @@ 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)
+ if (!old_dif && dif) {
WRITE_ONCE(sk->sk_userlocks,
sk->sk_userlocks | SOCK_CONNECT_DEV);
+ /* Pairs with the second smp_rmb() in udp_bound_dev_if(). */
+ smp_wmb();
+ }
WRITE_ONCE(sk->sk_bound_dev_if, dif);
}
@@ -275,9 +278,14 @@ 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)
+ if (!old_bound_dev_if) {
+ /* Pairs with the first smp_rmb() in
+ * udp_bound_dev_if().
+ */
+ smp_wmb();
WRITE_ONCE(sk->sk_userlocks,
sk->sk_userlocks & ~SOCK_CONNECT_DEV);
+ }
goto out;
}
--
2.43.0
next prev parent reply other threads:[~2026-11-20 17:43 UTC|newest]
Thread overview: 5+ messages / expand[flat|nested] mbox.gz Atom feed top
2026-11-20 17:43 [PATCH net 0/4] udp: Fix bind conflict check bypasses Juan Agustin Niccolini
2026-11-20 17:43 ` [PATCH net 1/4] udp: Fix bind conflict check stopping at the first reuseport socket Juan Agustin Niccolini
2026-11-20 17:43 ` [PATCH net 2/4] udp: Keep the bound device on disconnect unless connect() set it Juan Agustin Niccolini
2026-11-20 17:43 ` Juan Agustin Niccolini [this message]
2026-11-20 17:43 ` [PATCH net 4/4] selftests: net: Add UDP bind conflict tests Juan Agustin Niccolini
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20261120174341.35873-4-juanagusnicco@gmail.com \
--to=juanagusnicco@gmail.com \
--cc=davem@davemloft.net \
--cc=e@erig.me \
--cc=edumazet@kernel.org \
--cc=horms@kernel.org \
--cc=kuba@kernel.org \
--cc=kuniyu@google.com \
--cc=linux-kernel@vger.kernel.org \
--cc=linux-kselftest@vger.kernel.org \
--cc=netdev@vger.kernel.org \
--cc=pabeni@redhat.com \
--cc=shuah@kernel.org \
--cc=willemb@google.com \
--cc=willemdebruijn.kernel@gmail.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox
all inboxes | Powered by JetHome®