On 05.10.23 20:14, Wenjia Zhang wrote:
On 26.09.23 11:06, D. Wythe wrote:
On 9/26/23 3:18 PM, Alexandra Winter wrote:
On 26.09.23 05:00, D. Wythe wrote:
You are right. The key point is how to ensure the valid of smc sock
during the life time of clc sock, If so, READ_ONCE is good
enough. Unfortunately, I found that there are no such guarantee, so
it's still a life-time problem.
Did you discover a scenario, where clc sock could live longer than
smc sock?
Wouldn't that be a dangerous scenario in itself? I still have some
hope that the lifetime of an smc socket is by design longer
than that of the corresponding tcp socket.
Hi Alexandra,
Yes there is. Considering scenario:
tcp_v4_rcv(skb)
/* req sock */
reqsk = _inet_lookup_skb(skb)
/* listen sock */
sk = reqsk(reqsk)->rsk_listener;
sock_hold(sk);
tcp_check_req(sk)
smc_release /*
release smc listen sock */
__smc_release
smc_close_active() /* smc_sk->sk_state = SMC_CLOSED; */
if
(smc_sk->sk_state == SMC_CLOSED)
smc_clcsock_release();
sock_release(clcsk); /* close clcsock */
sock_put(sk); /* might not the final refcnt */
sock_put(smc_sk) /* might be the final refcnt of smc_sock */
syn_recv_sock(sk...)
/* might be the final refcnt of tcp listen sock */
sock_put(sk);
Fortunately, this scenario only affects smc_syn_recv_sock and
smc_hs_congested, as other callbacks already have locks to protect smc,
which can guarantee that the sk_user_data is either NULL (set in
smc_close_active) or valid under the lock.
I'm kind of confused with this scenario. How could the
smc_clcsock_release()->sock_release(clcsk) happen?
Because the syn_recv_sock happens short prior to accept(), that means
that the &smc->tcp_listen_work is already triggered but the real
accept() is still not happening. At this moment, the incoming connection
is being added into the accept queue. Thus, if the sk->sk_state is
changed from SMC_LISTEN to SMC_CLOSED in smc_close_active(), there is
still "flush_work(&smc->tcp_listen_work);" after that. That ensures the
smc_clcsock_release() should not happen, if smc_clcsock_accept() is not
finished. Do you think that the execution of the &smc->tcp_listen_work
is already done? Or am I missing something?
> Hi wenjia,
>
> Sorry for late reply, we have just returned from vacation.
>
> The smc_clcsock_release here release the listen clcsock rather than
> the child clcsock.
> So the flush_work might not be helpful for this scenario.
>
> Best wishes,
> D. Wythe
It seems like that I lost some mails these days :-( Just saw your answer.
Maybe I didn't describe my thought clearly. Following data flow is your
scenario, right?
–
(sk_state == SMC_LISTEN)|
tcp_check_req() | smc_release()
| ->__smc_release()
| -> smc_close_active()
| -> sk->sk_state = SMC_CLOSED;
| -> ...
| -> smc->clcsock->sk->sk_user_data = NULL;
| -> ...
|*1) -> flush_work(&smc->tcp_listen_work);
|*4)
| -> smc_clcsock_accept()
| -> kernel_accept()
| -> inet_csk_accept()
|*5)
| if (sk->sk_state == SMC_CLOSED)
|*3)-> smc_clcsock_release()
-> syn_recv_sock() *2)|
|
v
My question is how the smc_clcsock_release() could happen after the
syn_recv_sock()?
IMO, the syn_recv_sock() should be called during the
&smc->tcp_listen_work, which is corresponding to lsmc (listen smc). And
in smc_clcsock_accept(), the lsmc->clcsock as the listening socket goes
on to be used to accept a new connection. If the &smc->tcp_listen_work
is not finished, *1) will wait for its finishing. It can only happen in
following situation:
*4) sk_state is SMC_CLOSED, then no connection is accepted.
*5) old sk_state is SMC_LISTEN, TCP accept is successful. But current
sk_state is SMC_CLOSED. Thus, no new smc connection.
What do you think? Please let me know if I have any lapse of thought.
Thanks,
Wenjia
