Re: [PATCH 1/1] nbd/server: push pending frames after sending reply

[Kevin Wolf]

Am 24.03.2023 um 11:47 hat Florian Westphal geschrieben:
> qemu-nbd doesn't set TCP_NODELAY on the tcp socket.
> 
> Kernel waits for more data and avoids transmission of small packets.
> Without TLS this is barely noticeable, but with TLS this really shows.
> 
> Booting a VM via qemu-nbd on localhost (with tls) takes more than
> 2 minutes on my system.  tcpdump shows frequent wait periods, where no
> packets get sent for a 40ms period.
> 
> Add explicit (un)corking when processing (and responding to) requests.
> "TCP_CORK, &zero" after earlier "CORK, &one" will flush pending data.
> 
> VM Boot time:
> main:    no tls:  23s, with tls: 2m45s
> patched: no tls:  14s, with tls: 15s
> 
> VM Boot time, qemu-nbd via network (same lan):
> main:    no tls:  18s, with tls: 1m50s
> patched: no tls:  17s, with tls: 18s
> 
> Future optimization: if we could detect if there is another pending
> request we could defer the uncork operation because more data would be
> appended.
> 
> Signed-off-by: Florian Westphal <fw@strlen.de>
> ---
>  nbd/server.c | 3 +++
>  1 file changed, 3 insertions(+)
> 
> diff --git a/nbd/server.c b/nbd/server.c
> index a4750e41880a..848836d41405 100644
> --- a/nbd/server.c
> +++ b/nbd/server.c
> @@ -2667,6 +2667,8 @@ static coroutine_fn void nbd_trip(void *opaque)
>          goto disconnect;
>      }
>  
> +    qio_channel_set_cork(client->ioc, true);
> +
>      if (ret < 0) {
>          /* It wasn't -EIO, so, according to nbd_co_receive_request()
>           * semantics, we should return the error to the client. */
> @@ -2692,6 +2694,7 @@ static coroutine_fn void nbd_trip(void *opaque)
>          goto disconnect;
>      }
>  
> +    qio_channel_set_cork(client->ioc, false);
>  done:
>      nbd_request_put(req);
>      nbd_client_put(client);

In the error paths, we never call set_cork(false) again. I suppose the
reason that this is okay is because the next thing is actually that we
close the socket?

Kevin