Re: [PATCH RFC net-next] vsock: Reschedule connect_work for O_NONBLOCK connect() requests

[Stefano Garzarella <sgarzare@redhat.com>]

On Thu, Aug 04, 2022 at 04:44:47PM -0700, Peilin Ye wrote:
>Hi Stefano,
>
>On Thu, Aug 04, 2022 at 08:59:23AM +0200, Stefano Garzarella wrote:
>> The last thing I was trying to figure out before sending the patch was
>> whether to set sock->state = SS_UNCONNECTED in vsock_connect_timeout().
>>
>> I think we should do that, otherwise a subsequent to connect() with
>> O_NONBLOCK set would keep returning -EALREADY, even though the timeout has
>> expired.
>>
>> What do you think?
>
>Thanks for bringing this up, after thinking about sock->state, I have 3
>thoughts:
>
>1. I think the root cause of this memleak is, we keep @connect_work
>   pending, even after the 2nd, blocking request times out (or gets
>   interrupted) and sets sock->state back to SS_UNCONNECTED.
>
>   @connect_work is effectively no-op when sk->sk_state is
>   TCP_CLOS{E,ING} anyway, so why not we just cancel @connect_work when
>   blocking requests time out or get interrupted?  Something like:
>
>diff --git a/net/vmw_vsock/af_vsock.c b/net/vmw_vsock/af_vsock.c
>index f04abf662ec6..62628af84164 100644
>--- a/net/vmw_vsock/af_vsock.c
>+++ b/net/vmw_vsock/af_vsock.c
>@@ -1402,6 +1402,9 @@ static int vsock_connect(struct socket *sock, struct sockaddr *addr,
>                lock_sock(sk);
>
>                if (signal_pending(current)) {
>+                       if (cancel_delayed_work(&vsk->connect_work))
>+                               sock_put(sk);
>+
>                        err = sock_intr_errno(timeout);
>                        sk->sk_state = sk->sk_state == TCP_ESTABLISHED ? TCP_CLOSING : TCP_CLOSE;
>                        sock->state = SS_UNCONNECTED;
>@@ -1409,6 +1412,9 @@ static int vsock_connect(struct socket *sock, struct sockaddr *addr,
>                        vsock_remove_connected(vsk);
>                        goto out_wait;
>                } else if (timeout == 0) {
>+                       if (cancel_delayed_work(&vsk->connect_work))
>+                               sock_put(sk);
>+
>                        err = -ETIMEDOUT;
>                        sk->sk_state = TCP_CLOSE;
>                        sock->state = SS_UNCONNECTED;
>
>   Then no need to worry about rescheduling @connect_work, and the state
>   machine becomes more accurate.  What do you think?  I will ask syzbot
>   to test this.

It could work, but should we set `sk->sk_err` and call sk_error_report() 
to wake up thread waiting on poll()?

Maybe the previous version is simpler.

>
>2. About your suggestion of setting sock->state = SS_UNCONNECTED in
>   vsock_connect_timeout(), I think it makes sense.  Are you going to
>   send a net-next patch for this?

If you have time, feel free to send it.

Since it is a fix, I believe you can use the "net" tree. (Also for this 
patch).

Remember to put the "Fixes" tag that should be the same.

>
>3. After a TCP_SYN_SENT sock receives VIRTIO_VSOCK_OP_RESPONSE in
>   virtio_transport_recv_connecting(), why don't we cancel 
>   @connect_work?
>   Am I missing something?

Because when the timeout will fire, vsock_connect_timeout() will just 
call sock_put() since sk->sk_state is changed.

Of course, we can cancel it if we want, but I think it's not worth it.
In the end, this rescheduling patch should solve all the problems.

Thanks,
Stefano