Re: [PATCH net] tcp: initialize send sequence numbers before deferring TFO

[Eric Dumazet]

Le lun. 2 nov. 2026 à 09:13, Qingfang Deng <qingfang.deng@xxxxxxxxx> a écrit :
>
> On 2026/11/2 16:00, Eric Dumazet wrote:
> > Le lun. 2 nov. 2026 à 08:41, Qingfang Deng <qingfang.deng@xxxxxxxxx> a écrit :
> >> On 2026/11/2 14:11, Jiayuan Chen wrote:
> >>> On 11/2/26 11:59 AM, Qingfang Deng wrote:
> >>>> TCP_FASTOPEN_CONNECT can return from connect() before tcp_connect_init()
> >>>> when a cached cookie or cookieless Fast Open allows deferring the SYN.
> >>>> At this point write_seq contains the initial sequence number, while
> >>>> snd_una and snd_nxt have not been initialized to that sequence number.
> >>>>
> >>>> Consequently, inet_diag and /proc/net/tcp{,6} report the initial
> >>>> sequence number as the send queue length despite no SYN or payload being
> >>>> queued.
> >>>>
> >>>> Initialize snd_una and snd_nxt to write_seq before marking the
> >>>> connection as deferred. Both sequence-based queue lengths then remain
> >>>> zero until the first write starts the connection and runs
> >>>> tcp_connect_init().
> >>> This patch looks good to me. But idiag_wqueue can still go negative
> >>> on the normal connect path, since the socket is already hashed and
> >>> diag does not take the socket lock. We always set write_seq first,
> >>> then init snd_una later:
> >>>
> >>>     tcp_v4_connect
> >>>       WRITE_ONCE(tp->write_seq, st.seq)
> >>>       tcp_connect
> >>>         tcp_connect_init
> >>>           WRITE_ONCE(tp->snd_una, tp->write_seq)
> >>>
> >>> Can you try this instead?
> >>>
> >>>     r->idiag_wqueue = READ_ONCE(sk->sk_wmem_queued) ?
> >>>                       READ_ONCE(tp->write_seq) - READ_ONCE(tp->snd_una) : 0;
> >> This does not guarantee the load order on a weakly-ordered CPU. LLM
> >> suggests using this load-acquire/store-release sequence:
> >>
> >> r->idiag_wqueue = smp_load_acquire(&sk->sk_wmem_queued) ?
> >> READ_ONCE(tp->write_seq) - READ_ONCE(tp->snd_una) : 0;
> >>
> >> Then in tcp_connect_queue_skb(), replace sk_wmem_queued_add()
> >> with smp_store_release():
> >>
> >> /* Paired with smp_load_acquire() in tcp_diag_get_info(). */
> >> smp_store_release(&sk->sk_wmem_queued, sk->sk_wmem_queued + skb->truesize);
> >>
> >> What do you think?
> > Please do not do this.
> >
> > /proc/net/tcp and inet_diag dump sockets without holding the socket
> > lock. They provide a best effort snapshot, and transient inconsistent
> > values are expected (see the comment about rx_queue in get_tcp4_sock()).
> >
> > The window in tcp_v4_connect()/tcp_v6_connect() is a few usec, with the
> > socket lock held by the caller. Adding memory barriers in the TCP stack,
> > on a field updated in many places, to cover this is not worth it.
>
> tcp_connect_queue_skb(), as the name suggests, only queues the
> connecting SYN. So this adds a small setup cost on weakly-ordered CPUs
> (on x86, smp_store_release degrades to a compiler barrier).
>
> If strong memory ordering is not required, do you prefer Jiayuan's approach?

Neither, let's keep v1 as is.

I know tcp_connect_queue_skb() only queues the SYN. The cost is not the
problem; sk_wmem_queued has many other writers, and one release store
among them is a special ordering rule that nobody will remember the
next time this code is changed. This is too much complexity to hide a
transient value in a best-effort dump.

Without barriers, Jiayuan's version is only a heuristic: the loads are
not ordered, so it guarantees nothing. It also ties a sequence-space
value to memory accounting, and it would have to be copied into
get_tcp4_sock() and get_tcp6_sock() to be consistent.

Thanks.