Re: [PATCH] ksmbd: server: avoid busy polling in accept loop

[Qingfang Deng]

Hi Stefan,

On Tue, Nov 11, 2025 at 3:16 PM Stefan Metzmacher <metze@xxxxxxxxx> wrote:
> >> Also remove:
> >>    - TCP_NODELAY, which has no effect on a listening socket.
> >>    - sk_rcvtimeo and sk_sndtimeo assignments, which only caused accept()
> >>      to return -EAGAIN prematurely.
> >
> > Aren't these inherited to the accepted sockets?
> > So we need to apply them to the accepted sockets now
> > instead of dropping them completely?

You're right, TCP_NODELAY of a new accepted socket is inherited from
the listen socket, so it should not be removed.

>
> Actually the timeouts are added to the client connection,
> but not the TCP_NODELAY.
>
> But looking at it more detailed I'm wondering if this might
> introduce a deadlock.
>
> We have this in the accepting thread:
>
>          while (!kthread_should_stop()) {
>                  mutex_lock(&iface->sock_release_lock);
>                  if (!iface->ksmbd_socket) {
>                          mutex_unlock(&iface->sock_release_lock);
>                          break;
>                  }
>                  ret = kernel_accept(iface->ksmbd_socket, &client_sk, 0);
>                  mutex_unlock(&iface->sock_release_lock);
>                  if (ret)
>                          continue;
>
>
> And in the stopping code this:
>
>          case NETDEV_DOWN:
>                  iface = ksmbd_find_netdev_name_iface_list(netdev->name);
>                  if (iface && iface->state == IFACE_STATE_CONFIGURED) {
>                          ksmbd_debug(CONN, "netdev-down event: netdev(%s) is going down\n",
>                                          iface->name);
>                          tcp_stop_kthread(iface->ksmbd_kthread);
>                          iface->ksmbd_kthread = NULL;
>                          mutex_lock(&iface->sock_release_lock);
>                          tcp_destroy_socket(iface->ksmbd_socket);
>                          iface->ksmbd_socket = NULL;
>                          mutex_unlock(&iface->sock_release_lock);
>
>                          iface->state = IFACE_STATE_DOWN;
>                          break;
>                  }
>
>
>
> I guess that now kernel_accept() call waits forever holding iface->sock_release_lock
> and tcp_stop_kthread(iface->ksmbd_kthread); doesn't have any impact anymore
> as we may never reach kthread_should_stop() anymore.
>
> We may want to do a kernel_sock_shutdown(ksmbd_socket, SHUT_RDWR) after
> tcp_stop_kthread(iface->ksmbd_kthread); but before mutex_lock(&iface->sock_release_lock);
> so that kernel_accept() hopefully returns directly.
> And we only call sock_release(ksmbd_socket); under the iface->sock_release_lock mutex.

In kernel v6.1 or later, kthread_stop() in tcp_stop_kthread() will
send a signal to the ksmbd kthread so accept() will return -EINTR.
Before v6.1 it can actually get stuck, as accept() will block forever.

If you're fixing the issue when this patch was backported to versions
before v6.1, this will not work, because kthread_stop() blocks until
the target kthread returns, so shutdown() will never get called. The
sock_release_lock mutex seems redundant because of that.
Instead, shutdown() can be called _before_ kthread_stop() so accept()
will return -EINVAL.

Namjae, should I send a v2 with both issues addressed?

-- Qingfang