Re: [PATCH net v3] sctp: fix use-after-free after association migration
From: Xin Long
Date: Thu Nov 19 2026 - 11:53:00 EST
On Sun, Nov 15, 2026 at 10:40 PM Yong Liu <pkfxxxing@xxxxxxxxx> wrote:
>
> SCTP timers and ICMP error handling can cache asoc->base.sk before
> locking it. If accept or peeloff migrates the association in between,
> processing under the old socket lock can race with an abort on the new
> socket that frees the association's queues. Transport and association
> references do not keep those queues alive. With mdelay() instrumentation
> widening the race, KASAN reported a slab-use-after-free in
> sctp_chunk_abandoned() from the T3 retransmit path.
>
> Recheck the association's socket after taking its BH lock and defer
> mismatched timer callbacks through the existing retry path. Pass the
> locked socket to the IPv4 and IPv6 ICMP handlers instead of rereading it,
> and treat a mismatch like a busy socket before processing ICMP errors.
>
> Use READ_ONCE() for socket pointer reads in these paths and receive/backlog
> processing, paired with WRITE_ONCE() in sctp_assoc_migrate(). Take the new
> socket reference before publishing it. The cached socket is protected by
> BH context and the endpoint's RCU-deferred sock_put().
>
> Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2")
> Cc: stable@xxxxxxxxxxxxxxx
> Signed-off-by: Yong Liu <pkfxxxing@xxxxxxxxx>
> ---
> Changes in v3:
> - Update the commit message and subject as requested by Jakub Kicinski.
>
> v2: https://lore.kernel.org/all/20261113032805.1447885-1-pkfxxxing@xxxxxxxxx/
> Review: https://lore.kernel.org/all/20261115100933.75ae0534@xxxxxxxxxx/
>
> Changes in v2:
> - Use Yong Liu for the author and Signed-off-by, as requested by Jakub
> Kicinski.
> - Address Xin Long's review and the Sashiko ICMP finding: pass the locked
> socket to the IPv4/IPv6 error handlers and their UDP encapsulation
> callers, and revalidate it before immediate ICMP processing.
> - Use READ_ONCE() for socket pointer reads in sctp_err_lookup(), sctp_rcv()
> and sctp_backlog_rcv().
> - Broaden the subject and commit message to cover the ICMP paths.
>
> v1: https://lore.kernel.org/all/20261110120212.1110283-2-pkfxxxing@xxxxxxxxx/
>
> Based on b3dfbde62cf1981c4638939c159dfd087b7a72e9 (Linux 7.4.0-rc2).
>
> net/sctp/associola.c | 4 ++--
> net/sctp/input.c | 32 ++++++++++++++++---------------
> net/sctp/ipv6.c | 11 ++++++-----
> net/sctp/sm_sideeffect.c | 41 ++++++++++++++++++++--------------------
> 4 files changed, 45 insertions(+), 43 deletions(-)
>
> diff --git a/net/sctp/associola.c b/net/sctp/associola.c
> index 4521be3bd8..c5d520c8d7 100644
> --- a/net/sctp/associola.c
> +++ b/net/sctp/associola.c
> @@ -1088,8 +1088,8 @@ void sctp_assoc_migrate(struct sctp_association *assoc, struct sock *newsk)
> sctp_endpoint_hold(assoc->ep);
>
> /* Get a reference to the new sock. */
> - assoc->base.sk = newsk;
> - sock_hold(assoc->base.sk);
> + sock_hold(newsk);
> + WRITE_ONCE(assoc->base.sk, newsk);
>
> /* Add the association to the new endpoint's list of associations. */
> sctp_endpoint_add_asoc(newsp->ep, assoc);
> diff --git a/net/sctp/input.c b/net/sctp/input.c
> index 9494cfa511..e25c94ad08 100644
> --- a/net/sctp/input.c
> +++ b/net/sctp/input.c
> @@ -169,7 +169,7 @@ int sctp_rcv(struct sk_buff *skb)
>
> /* Retrieve the common input handling substructure. */
> rcvr = asoc ? &asoc->base : &ep->base;
> - sk = rcvr->sk;
> + sk = READ_ONCE(rcvr->sk);
>
> /*
> * RFC 2960, 8.4 - Handle "Out of the blue" Packets.
> @@ -217,7 +217,7 @@ int sctp_rcv(struct sk_buff *skb)
> */
> bh_lock_sock(sk);
>
> - if (sk != rcvr->sk) {
> + if (sk != READ_ONCE(rcvr->sk)) {
> /* Our cached sk is different from the rcvr->sk. This is
> * because migrate()/accept() may have moved the association
> * to a new socket and released all the sockets. So now we
> @@ -226,7 +226,7 @@ int sctp_rcv(struct sk_buff *skb)
> * of the current sk.
> */
> bh_unlock_sock(sk);
> - sk = rcvr->sk;
> + sk = READ_ONCE(rcvr->sk);
> bh_lock_sock(sk);
> }
>
> @@ -292,7 +292,7 @@ int sctp_backlog_rcv(struct sock *sk, struct sk_buff *skb)
> goto done;
> }
>
> - if (unlikely(rcvr->sk != sk)) {
> + if (unlikely(READ_ONCE(rcvr->sk) != sk)) {
> /* In this case, the association moved from one socket to
> * another. We are currently sitting on the backlog of the
> * old socket, so we need to move.
> @@ -304,7 +304,7 @@ int sctp_backlog_rcv(struct sock *sk, struct sk_buff *skb)
> * Otherwise, we can safely push the chunk on the inqueue.
> */
>
> - sk = rcvr->sk;
> + sk = READ_ONCE(rcvr->sk);
> local_bh_disable();
> bh_lock_sock(sk);
>
> @@ -377,7 +377,7 @@ void sctp_icmp_frag_needed(struct sock *sk, struct sctp_association *asoc,
> t->pl.probe_size + sctp_transport_pl_hlen(t) <= pmtu))
> return;
As the clashiko [1] reported, there's asoc->base.sk access in
sctp_transport_pl_hlen(), You can delete the !t check (t can't be NULL),
and move the rest check to sctp_transport_update_pmtu(), which is called
under sock lock, like:
pmtu = SCTP_TRUNC4(pmtu);
+ if (t->pathmtu <= pmtu &&
+ t->pl.probe_size + sctp_transport_pl_hlen(t) <= pmtu)
+ return false;
+
As for sctp_addrs_lookup_transport() in the report, I think you need
to do something like:
- bound_dev_if = READ_ONCE(t->asoc->base.sk->sk_bound_dev_if);
- if (sctp_sk_bound_dev_eq(net, bound_dev_if, dif, sdif) &&
+ sk = READ_ONCE(t->asoc->base.sk);
+ if (sctp_sk_bound_dev_eq(net,
READ_ONCE(sk->sk_bound_dev_if), dif, sdif) &&
sctp_bind_addr_match(&t->asoc->base.bind_addr,
- laddr, sctp_sk(t->asoc->base.sk)))
+ laddr, sctp_sk(sk)))
return t;
>
> - if (sock_owned_by_user(sk)) {
> + if (sock_owned_by_user(sk) || sk != READ_ONCE(asoc->base.sk)) {
> atomic_set(&t->mtu_info, pmtu);
> asoc->pmtu_pending = 1;
> t->pmtu_pending = 1;
> @@ -410,7 +410,8 @@ void sctp_icmp_redirect(struct sock *sk, struct sctp_transport *t,
> {
> struct dst_entry *dst;
>
> - if (sock_owned_by_user(sk) || !t)
> + if (sock_owned_by_user(sk) || !t ||
> + sk != READ_ONCE(t->asoc->base.sk))
This '!t' check can be deleted, as the t passed to
sctp_icmp_redirect() can never be NULL.
> return;
> dst = sctp_transport_dst_check(t);
> if (dst)
> @@ -432,7 +433,7 @@ void sctp_icmp_proto_unreachable(struct sock *sk,
> struct sctp_association *asoc,
> struct sctp_transport *t)
> {
> - if (sock_owned_by_user(sk)) {
> + if (sock_owned_by_user(sk) || sk != READ_ONCE(asoc->base.sk)) {
> if (timer_pending(&t->proto_unreach_timer))
> return;
> else {
> @@ -492,7 +493,7 @@ struct sock *sctp_err_lookup(struct net *net, int family, struct sk_buff *skb,
> if (!asoc)
> return NULL;
>
> - sk = asoc->base.sk;
> + sk = READ_ONCE(asoc->base.sk);
>
> /* RFC 4960, Appendix C. ICMP Handling
> *
> @@ -526,7 +527,7 @@ struct sock *sctp_err_lookup(struct net *net, int family, struct sk_buff *skb,
> /* If too many ICMPs get dropped on busy
> * servers this needs to be solved differently.
> */
> - if (sock_owned_by_user(sk))
> + if (sock_owned_by_user(sk) || sk != READ_ONCE(asoc->base.sk))
> __NET_INC_STATS(net, LINUX_MIB_LOCKDROPPEDICMPS);
>
> *app = asoc;
> @@ -546,11 +547,11 @@ void sctp_err_finish(struct sock *sk, struct sctp_transport *t)
> sctp_transport_put(t);
> }
>
> -static void sctp_v4_err_handle(struct sctp_transport *t, struct sk_buff *skb,
> +static void sctp_v4_err_handle(struct sock *sk, struct sctp_transport *t,
> + struct sk_buff *skb,
> __u8 type, __u8 code, __u32 info)
> {
> struct sctp_association *asoc = t->asoc;
> - struct sock *sk = asoc->base.sk;
> int err = 0;
>
> switch (type) {
> @@ -582,7 +583,8 @@ static void sctp_v4_err_handle(struct sctp_transport *t, struct sk_buff *skb,
> default:
> return;
> }
> - if (!sock_owned_by_user(sk) && inet_test_bit(RECVERR, sk)) {
> + if (!sock_owned_by_user(sk) && sk == READ_ONCE(asoc->base.sk) &&
> + inet_test_bit(RECVERR, sk)) {
> sk->sk_err = err;
> sk_error_report(sk);
> } else { /* Only an error on timeout */
As the clashiko [1] reported, this 'else' need to become:
} else if (sk == READ_ONCE(asoc->base.sk)) {
> @@ -630,7 +632,7 @@ int sctp_v4_err(struct sk_buff *skb, __u32 info)
> return -ENOENT;
> }
>
> - sctp_v4_err_handle(transport, skb, type, code, info);
> + sctp_v4_err_handle(sk, transport, skb, type, code, info);
> sctp_err_finish(sk, transport);
>
> return 0;
> @@ -660,7 +662,7 @@ int sctp_udp_v4_err(struct sock *sk, struct sk_buff *skb)
> }
> if (hdr->type == ICMP_DEST_UNREACH && hdr->code == ICMP_FRAG_NEEDED)
> info = ntohs(hdr->un.frag.mtu);
> - sctp_v4_err_handle(t, skb, hdr->type, hdr->code, info);
> + sctp_v4_err_handle(sk, t, skb, hdr->type, hdr->code, info);
>
> sctp_err_finish(sk, t);
> return 1;
> diff --git a/net/sctp/ipv6.c b/net/sctp/ipv6.c
> index 97390569c4..f96e79ca7f 100644
> --- a/net/sctp/ipv6.c
> +++ b/net/sctp/ipv6.c
> @@ -123,11 +123,11 @@ static struct notifier_block sctp_inet6addr_notifier = {
> .notifier_call = sctp_inet6addr_event,
> };
>
> -static void sctp_v6_err_handle(struct sctp_transport *t, struct sk_buff *skb,
> +static void sctp_v6_err_handle(struct sock *sk, struct sctp_transport *t,
> + struct sk_buff *skb,
> __u8 type, __u8 code, __u32 info)
> {
> struct sctp_association *asoc = t->asoc;
> - struct sock *sk = asoc->base.sk;
> int err = 0;
>
> switch (type) {
> @@ -149,7 +149,8 @@ static void sctp_v6_err_handle(struct sctp_transport *t, struct sk_buff *skb,
> }
>
> icmpv6_err_convert(type, code, &err);
> - if (!sock_owned_by_user(sk) && inet6_test_bit(RECVERR6, sk)) {
> + if (!sock_owned_by_user(sk) && sk == READ_ONCE(asoc->base.sk) &&
> + inet6_test_bit(RECVERR6, sk)) {
> sk->sk_err = err;
> sk_error_report(sk);
> } else {
Same here, change to:
} else if (sk == READ_ONCE(asoc->base.sk)) {
[1] https://netdev-ai.bots.linux.dev/sashiko/#/patchset/20261116033855.1792036-1-pkfxxxing%40gmail.com
Thanks.