On Mon, Mar 9, 2026 at 8:26 AM Eric Dumazet [off-list ref] wrote:

On Mon, Mar 9, 2026 at 3:38 AM Jiayuan Chen [off-list ref] wrote:

quoted

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@syzkaller.appspotmail.com
Closes: https://lore.kernel.org/all/67eaf9b8.050a0220.3c3d88.004a.GAE@google.com/T/ (local)
Suggested-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Jiayuan Chen <jiayuan.chen@linux.dev>
---
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.

This looks better, but please note there is something missing in your
RCU implementation.

You still need to ensure s->sk_user_data is read / written with
load/store tearing prevention.
Standard rcu_dereference()/rcu_assign() are dealing with this aspect.

For instance, the following helper is assuming a lock was held :

static inline struct smc_sock *smc_clcsock_user_data(const struct sock *clcsk)
{
return (struct smc_sock *)
       ((uintptr_t)clcsk->sk_user_data & ~SK_USER_DATA_NOCOPY);
}


One possible way is to use
rcu_dereference_sk_user_data()/rcu_assign_sk_user_data() or similar
helpers.

More context in this old commit

commit 559835ea7292e2f09304d81eda16f4209433245e
Author: Pravin B Shelar [off-list ref]
Date:   Tue Sep 24 10:25:40 2013 -0700

    vxlan: Use RCU apis to access sk_user_data.

    Use of RCU api makes vxlan code easier to understand.  It also
    fixes bug due to missing ACCESS_ONCE() on sk_user_data dereference.
    In rare case without ACCESS_ONCE() compiler might omit vs on
    sk_user_data dereference.
    Compiler can use vs as alias for sk->sk_user_data, resulting in
    multiple sk_user_data dereference in rcu read context which
    could change.

    CC: Jesse Gross [off-list ref]
    Signed-off-by: Pravin B Shelar [off-list ref]
    Signed-off-by: David S. Miller [off-list ref]