[PATCH 15/18] nvmet-tcp: enable TLS handshake upcall

[Chuck Lever III]

> On Mar 28, 2023, at 4:44 AM, Sagi Grimberg <sagi at grimberg.me> wrote:
> 
> 
>>>> 
>>>> Nearly there.
>>>> 
>>>> The actual patch would be:
>>>> 
>>>> @@ -2031,10 +1988,16 @@ static void nvmet_tcp_listen_data_ready(struct sock *sk)
>>>>          trace_sk_data_ready(sk);
>>>> 
>>>>          read_lock_bh(&sk->sk_callback_lock);
>>>> +       /* Ignore if the callback has been changed */
>>>> +       if (sk->sk_data_ready != nvmet_tcp_listen_data_ready)
>>>> +               goto out;
>>>>          port = sk->sk_user_data;
>>>>          if (!port)
>>>>                  goto out;
>>>> 
>>>> +       if (port->data_ready)
>>>> +               port->data_ready(sk);
>>>> +
>>>>          if (sk->sk_state == TCP_LISTEN)
>>>>                  queue_work(nvmet_wq, &port->accept_work);
>>>>   out:
>>>> 
>>>> As the callbacks will be changed once TLS is activated, and we really should not attempt to run if sk_data_ready() points to another function,
>>>> as then the sk_user_data pointer will most likely be changed, too,
>>>> causing all sorts of issues.
>>> 
>>> Umm, something is unclear to me. if nvmet_tcp_listen_data_ready is
>>> called doesn't it by definition mean that sk->sk_data_ready ==
>>> nvmet_tcp_listen_data_ready ?
>>> 
>>> Are you talking about a situation where between
>>> nvmet_tcp_listen_data_ready is starting and until the
>>> sk->sk_callback_lock the data_ready cb (and the user data
>>> pointer) is changed?
>> No. Far simpler:
>> Starting kTLS will change the callbacks.
>> So sk->sk_data_ready will point to our callback before
>> the upcall, but to the kTLS version _after_ the upcall.
>> It typically doesn't matter, as we're setting it to
>> nvmet_tcp_data_ready() anyway.
> 
> For ktls won't we set it to nvmet_tcp_data_ready only when
> the handshake is done?
> 
>> But there might be the odd case where the data_ready
>> callback is invoked after kTLS has been started but before
>> we're setting it to nvmet_tcp_data_ready().
> 
> What does nvmet_tcp_data_ready has to do with it?
> You are changing nvmet_tcp_listen_data_ready.
> 
>> Then we cannot guarantee that sk_user_data really is set
>> to the 'queue' pointer, so we should skip the function.
> 
> nvme_tcp_listen_data_ready was invoked, then sk->sk_data_ready
> is nvme_tcp_listen_data_ready. Are you referring to the case
> where the callback has changed before the read lock ?
> 
> I would like to understand why svc_tcp_listen_data_ready doesn't
> have this race as well. Chuck?

svc doesn't alter the pointer address in the listener's
sk_data_ready field, and neither does kTLS.

For a connected socket, when the handshake has completed,
svc examines the socket for more work one last time before
deciding it's idle (a kind of simulated data_ready). That
should pick up work that arrived undetected.

Maybe I'm missing something?

--
Chuck Lever