[PATCH net v2] net/smc: fix NULL dereference and UAF in smc_tcp_syn_recv_sock()

[Jiayuan Chen]

Syzkaller reported a panic in smc_tcp_syn_recv_sock() [1].

smc_tcp_syn_recv_sock() is called in the TCP receive path
(softirq) via icsk_af_ops->syn_recv_sock on the clcsock (TCP
listening socket). It reads sk_user_data to get the smc_sock
pointer. However, when the SMC listen socket is being closed
concurrently, smc_close_active() sets clcsock->sk_user_data
to NULL under sk_callback_lock, and then the smc_sock itself
can be freed via sock_put() in smc_release().

This leads to two issues:

1) NULL pointer dereference: sk_user_data is NULL when
   accessed.
2) Use-after-free: sk_user_data is read as non-NULL, but the
   smc_sock is freed before its fields (e.g., queued_smc_hs,
   ori_af_ops) are accessed.

The race window looks like this:

  CPU A (softirq)              CPU B (process ctx)

  tcp_v4_rcv()
    TCP_NEW_SYN_RECV:
    sk = req->rsk_listener
    sock_hold(sk)
    /* No lock on listener */
                               smc_close_active():
                                 write_lock_bh(cb_lock)
                                 sk_user_data = NULL
                                 write_unlock_bh(cb_lock)
                                 ...
                                 smc_clcsock_release()
                                 sock_put(smc->sk) x2
                                   -> smc_sock freed!
    tcp_check_req()
      smc_tcp_syn_recv_sock():
        smc = user_data(sk)
          -> NULL or dangling
        smc->queued_smc_hs
          -> crash!

Note that the clcsock and smc_sock are two independent objects
with separate refcounts. TCP stack holds a reference on the
clcsock, which keeps it alive, but this does NOT prevent the
smc_sock from being freed.

Fix this by using RCU and refcount_inc_not_zero() to safely
access smc_sock. Since smc_tcp_syn_recv_sock() is called in
the TCP three-way handshake path, taking read_lock_bh on
sk_callback_lock is too heavy and would not survive a SYN
flood attack. Using rcu_read_lock() is much more lightweight.

- Set SOCK_RCU_FREE on the SMC listen socket so that
  smc_sock freeing is deferred until after the RCU grace
  period. This guarantees the memory is still valid when
  accessed inside rcu_read_lock().
- Use rcu_read_lock() to protect reading sk_user_data.
- Use refcount_inc_not_zero(&smc->sk.sk_refcnt) to pin the
  smc_sock. If the refcount has already reached zero (close
  path completed), it returns false and we bail out safely.

Note: smc_hs_congested() has a similar lockless read of
sk_user_data without rcu_read_lock(), but it only checks for
NULL and accesses the global smc_hs_wq, never dereferencing
any smc_sock field, so it is not affected.

Reproducer was verified with mdelay injection and smc_run,
the issue no longer occurs with this patch applied.

[1] https://syzkaller.appspot.com/bug?extid=827ae2bfb3a3529333e9

Fixes: 8270d9c21041 ("net/smc: Limit backlog connections")
Reported-by: syzbot+827ae2bfb3a3529333e9@xxxxxxxxxxxxxxxxxxxxxxxxx
Closes: https://lore.kernel.org/all/67eaf9b8.050a0220.3c3d88.004a.GAE@xxxxxxxxxx/T/
Suggested-by: Eric Dumazet <edumazet@xxxxxxxxxx>
Signed-off-by: Jiayuan Chen <jiayuan.chen@xxxxxxxxx>
---
v2:
- Use rcu_read_lock() + refcount_inc_not_zero() instead of
  read_lock_bh(sk_callback_lock) + sock_hold(), since this
  is the TCP handshake hot path and read_lock_bh is too
  expensive under SYN flood.
- Set SOCK_RCU_FREE on SMC listen socket to ensure
  RCU-deferred freeing.

v1: https://lore.kernel.org/netdev/20260307032158.372165-1-jiayuan.chen@xxxxxxxxx/
---
 net/smc/af_smc.c | 9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/net/smc/af_smc.c b/net/smc/af_smc.c
index d0119afcc6a1..72ac1d8c62d4 100644
--- a/net/smc/af_smc.c
+++ b/net/smc/af_smc.c
@@ -131,7 +131,13 @@ static struct sock *smc_tcp_syn_recv_sock(const struct sock *sk,
 	struct smc_sock *smc;
 	struct sock *child;
 
+	rcu_read_lock();
 	smc = smc_clcsock_user_data(sk);
+	if (!smc || !refcount_inc_not_zero(&smc->sk.sk_refcnt)) {
+		rcu_read_unlock();
+		return NULL;
+	}
+	rcu_read_unlock();
 
 	if (READ_ONCE(sk->sk_ack_backlog) + atomic_read(&smc->queued_smc_hs) >
 				sk->sk_max_ack_backlog)
@@ -153,11 +159,13 @@ static struct sock *smc_tcp_syn_recv_sock(const struct sock *sk,
 		if (inet_csk(child)->icsk_af_ops == inet_csk(sk)->icsk_af_ops)
 			inet_csk(child)->icsk_af_ops = smc->ori_af_ops;
 	}
+	sock_put(&smc->sk);
 	return child;
 
 drop:
 	dst_release(dst);
 	tcp_listendrop(sk);
+	sock_put(&smc->sk);
 	return NULL;
 }
 
@@ -2691,6 +2699,7 @@ int smc_listen(struct socket *sock, int backlog)
 		write_unlock_bh(&smc->clcsock->sk->sk_callback_lock);
 		goto out;
 	}
+	sock_set_flag(sk, SOCK_RCU_FREE);
 	sk->sk_max_ack_backlog = backlog;
 	sk->sk_ack_backlog = 0;
 	sk->sk_state = SMC_LISTEN;
-- 
2.43.0