在 2025/9/23 08:45, Kuniyuki Iwashima 写道:

On Sat, Sep 20, 2025 at 4:00 AM [off-list ref] wrote:

quoted

From: Xuanqiang Luo <redacted>

Since ehash lookups are lockless, if another CPU is converting sk to tw
concurrently, fetching the newly inserted tw with tw->tw_refcnt == 0 cause
lookup failure.

The call trace map is drawn as follows:
    CPU 0                                CPU 1
    -----                                -----
                                      inet_twsk_hashdance_schedule()
                                      spin_lock()
                                      inet_twsk_add_node_rcu(tw, ...)
__inet_lookup_established()
(find tw, failure due to tw_refcnt = 0)
                                      __sk_nulls_del_node_init_rcu(sk)
                                      refcount_set(&tw->tw_refcnt, 3)
                                      spin_unlock()

By replacing sk with tw atomically via hlist_nulls_replace_init_rcu() after
setting tw_refcnt, we ensure that tw is either fully initialized or not
visible to other CPUs, eliminating the race.

It's worth noting that we replace under lock_sock(), so no need to check if sk
is hashed. Thanks to Kuniyuki Iwashima!

Fixes: 3ab5aee7fe84 ("net: Convert TCP & DCCP hash tables to use RCU / hlist_nulls")
Suggested-by: Kuniyuki Iwashima <kuniyu@google.com>

This is not needed.  A pure review does not deserve Suggested-by.
This is used when someone suggests changing the core idea of
the patch.

Got it, but still really appreciate your detailed
and patient review!

quoted

Signed-off-by: Xuanqiang Luo <redacted>
---
  net/ipv4/inet_timewait_sock.c | 13 ++++---------
  1 file changed, 4 insertions(+), 9 deletions(-)

diff --git a/net/ipv4/inet_timewait_sock.c b/net/ipv4/inet_timewait_sock.c
index 5b5426b8ee92..bb98888584a8 100644
--- a/net/ipv4/inet_timewait_sock.c
+++ b/net/ipv4/inet_timewait_sock.c

@@ -116,7 +116,7 @@ void inet_twsk_hashdance_schedule(struct inet_timewait_sock *tw,
         spinlock_t *lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
         struct inet_bind_hashbucket *bhead, *bhead2;

-       /* Step 1: Put TW into bind hash. Original socket stays there too.
+       /* Put TW into bind hash. Original socket stays there too.
            Note, that any socket with inet->num != 0 MUST be bound in
            binding cache, even if it is closed.
          */

@@ -140,14 +140,6 @@ void inet_twsk_hashdance_schedule(struct inet_timewait_sock *tw,

         spin_lock(lock);

-       /* Step 2: Hash TW into tcp ehash chain */
-       inet_twsk_add_node_rcu(tw, &ehead->chain);
-
-       /* Step 3: Remove SK from hash chain */
-       if (__sk_nulls_del_node_init_rcu(sk))
-               sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
-
-
         /* Ensure above writes are committed into memory before updating the
          * refcount.
          * Provides ordering vs later refcount_inc().

@@ -162,6 +154,9 @@ void inet_twsk_hashdance_schedule(struct inet_timewait_sock *tw,
          */
         refcount_set(&tw->tw_refcnt, 3);

I discussed this series with Eric last week, and he pointed out
(thanks!) that we need to be careful here about memory barrier.

refcount_set() is just WRITE_ONCE() and thus can be reordered,
and twsk could be published with 0 refcnt, resulting in another RST.

Thanks for Eric's pointer!

Could you let me know if my modification here works?

That is, moving smp_wmb() to after the refcount update:

@@ -140,19 +140,6 @@ void inet_twsk_hashdance_schedule(struct inet_timewait_sock *tw,

         spin_lock(lock);

-       /* Step 2: Hash TW into tcp ehash chain */
-       inet_twsk_add_node_rcu(tw, &ehead->chain);
-
-       /* Step 3: Remove SK from hash chain */
-       if (__sk_nulls_del_node_init_rcu(sk))
-               sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
-
-
-       /* Ensure above writes are committed into memory before updating the
-        * refcount.
-        * Provides ordering vs later refcount_inc().
-        */
-       smp_wmb();
         /* tw_refcnt is set to 3 because we have :
          * - one reference for bhash chain.
          * - one reference for ehash chain.

@@ -162,6 +149,14 @@ void inet_twsk_hashdance_schedule(struct inet_timewait_sock *tw,
          */
         refcount_set(&tw->tw_refcnt, 3);

+       /* Ensure tw_refcnt has been set before tw is published by
+        * necessary memory barrier.
+        */
+       smp_wmb();
+
+       hlist_nulls_replace_init_rcu(&sk->sk_nulls_node, &tw->tw_node);
+       sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
+
         inet_twsk_schedule(tw, timeo);

         spin_unlock(lock);

Thanks!
Xuanqiang

quoted

+       hlist_nulls_replace_init_rcu(&sk->sk_nulls_node, &tw->tw_node);
+       sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
+
         inet_twsk_schedule(tw, timeo);

         spin_unlock(lock);
--
2.25.1