Re: [PATCH net-next 1/2] RTAX_INITRWND should be able to bring the rcv_ssthresh above 64KiB

[Eric Dumazet <edumazet@google.com>]

On Thu, Jul 21, 2022 at 5:10 PM Marek Majkowski <marek@cloudflare.com> wrote:
>
> We already support RTAX_INITRWND / initrwnd path attribute:
>
>  $ ip route change local 127.0.0.0/8 dev lo initrwnd 1024
>
> However normally, the initial advertised receive window is limited to
> 64KiB by rcv_ssthresh, regardless of initrwnd. This patch changes
> that, bumping up rcv_ssthresh to value derived from initrwnd. This
> allows for larger initial advertised receive windows, which is useful
> for specific types of TCP flows: big BDP ones, where there is a lot of
> data to send immediately after the flow is established.
>
> There are three places where we initialize sockets:
>  - tcp_output:tcp_connect_init
>  - tcp_minisocks:tcp_openreq_init_rwin
>  - syncookies
>
> In the first two we already have a call to `tcp_rwnd_init_bpf` and
> `dst_metric(RTAX_INITRWND)` which retrieve the bpf/path initrwnd
> attribute. We use this value to bring `rcv_ssthresh` up, potentially
> above the traditional 64KiB.
>
> With higher initial `rcv_ssthresh` the receiver will open the receive
> window more aggresively, which can improve large BDP flows - large
> throughput and latency.
>
> This patch does not cover the syncookies case.
>
> Signed-off-by: Marek Majkowski <marek@cloudflare.com>
> ---
>  include/net/inet_sock.h  |  1 +
>  net/ipv4/tcp_minisocks.c |  8 ++++++--
>  net/ipv4/tcp_output.c    | 10 ++++++++--
>  3 files changed, 15 insertions(+), 4 deletions(-)
>
> diff --git a/include/net/inet_sock.h b/include/net/inet_sock.h
> index daead5fb389a..bc68c9b70942 100644
> --- a/include/net/inet_sock.h
> +++ b/include/net/inet_sock.h
> @@ -89,6 +89,7 @@ struct inet_request_sock {
>                                 no_srccheck: 1,
>                                 smc_ok     : 1;
>         u32                     ir_mark;
> +       u32                     rcv_ssthresh;

Why do we need to store this value in the request_sock ?

It is derived from a route attribute and MSS, all this should be
available when the full blown socket is created.

It would also work even with syncookies.

>         union {
>                 struct ip_options_rcu __rcu     *ireq_opt;
>  #if IS_ENABLED(CONFIG_IPV6)
> diff --git a/net/ipv4/tcp_minisocks.c b/net/ipv4/tcp_minisocks.c
> index 6854bb1fb32b..89ba2a30a012 100644
> --- a/net/ipv4/tcp_minisocks.c
> +++ b/net/ipv4/tcp_minisocks.c
> @@ -360,6 +360,7 @@ void tcp_openreq_init_rwin(struct request_sock *req,
>         u32 window_clamp;
>         __u8 rcv_wscale;
>         u32 rcv_wnd;
> +       int adj_mss;
>         int mss;
>
>         mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
> @@ -378,15 +379,18 @@ void tcp_openreq_init_rwin(struct request_sock *req,
>         else if (full_space < rcv_wnd * mss)
>                 full_space = rcv_wnd * mss;
>
> +       adj_mss = mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0);
> +
>         /* tcp_full_space because it is guaranteed to be the first packet */
>         tcp_select_initial_window(sk_listener, full_space,
> -               mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
> +               adj_mss,
>                 &req->rsk_rcv_wnd,
>                 &req->rsk_window_clamp,
>                 ireq->wscale_ok,
>                 &rcv_wscale,
>                 rcv_wnd);
>         ireq->rcv_wscale = rcv_wscale;
> +       ireq->rcv_ssthresh = max(req->rsk_rcv_wnd, rcv_wnd * adj_mss);
>  }
>  EXPORT_SYMBOL(tcp_openreq_init_rwin);
>
> @@ -502,7 +506,7 @@ struct sock *tcp_create_openreq_child(const struct sock *sk,
>         newtp->rx_opt.tstamp_ok = ireq->tstamp_ok;
>         newtp->rx_opt.sack_ok = ireq->sack_ok;
>         newtp->window_clamp = req->rsk_window_clamp;
> -       newtp->rcv_ssthresh = req->rsk_rcv_wnd;
> +       newtp->rcv_ssthresh = ireq->rcv_ssthresh;
>         newtp->rcv_wnd = req->rsk_rcv_wnd;
>         newtp->rx_opt.wscale_ok = ireq->wscale_ok;
>         if (newtp->rx_opt.wscale_ok) {
> diff --git a/net/ipv4/tcp_output.c b/net/ipv4/tcp_output.c
> index 18c913a2347a..0f2d4174ea59 100644
> --- a/net/ipv4/tcp_output.c
> +++ b/net/ipv4/tcp_output.c
> @@ -3642,6 +3642,7 @@ static void tcp_connect_init(struct sock *sk)
>         struct tcp_sock *tp = tcp_sk(sk);
>         __u8 rcv_wscale;
>         u32 rcv_wnd;
> +       u32 mss;
>
>         /* We'll fix this up when we get a response from the other end.
>          * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
> @@ -3679,8 +3680,10 @@ static void tcp_connect_init(struct sock *sk)
>         if (rcv_wnd == 0)
>                 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
>
> +       mss = tp->advmss - (tp->rx_opt.ts_recent_stamp ?
> +                           tp->tcp_header_len - sizeof(struct tcphdr) : 0);
>         tcp_select_initial_window(sk, tcp_full_space(sk),
> -                                 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
> +                                 mss,
>                                   &tp->rcv_wnd,
>                                   &tp->window_clamp,
>                                   sock_net(sk)->ipv4.sysctl_tcp_window_scaling,
> @@ -3688,7 +3691,10 @@ static void tcp_connect_init(struct sock *sk)
>                                   rcv_wnd);
>
>         tp->rx_opt.rcv_wscale = rcv_wscale;
> -       tp->rcv_ssthresh = tp->rcv_wnd;
> +       if (rcv_wnd)
> +               tp->rcv_ssthresh = max(tp->rcv_wnd, rcv_wnd * mss);
> +       else
> +               tp->rcv_ssthresh = tp->rcv_wnd;
>
>         sk->sk_err = 0;
>         sock_reset_flag(sk, SOCK_DONE);
> --
> 2.25.1
>