On Mon, Dec 7, 2020 at 4:37 PM Eric Dumazet <edumazet@xxxxxxxxxx> wrote:
>
> On Mon, Dec 7, 2020 at 12:41 PM Hazem Mohamed Abuelfotoh
> <abuehaze@xxxxxxxxxx> wrote:
> >
> > Previously receiver buffer auto-tuning starts after receiving
> > one advertised window amount of data.After the initial
> > receiver buffer was raised by
> > commit a337531b942b ("tcp: up initial rmem to 128KB
> > and SYN rwin to around 64KB"),the receiver buffer may
> > take too long for TCP autotuning to start raising
> > the receiver buffer size.
> > commit 041a14d26715 ("tcp: start receiver buffer autotuning sooner")
> > tried to decrease the threshold at which TCP auto-tuning starts
> > but it's doesn't work well in some environments
> > where the receiver has large MTU (9001) especially with high RTT
> > connections as in these environments rcvq_space.space will be the same
> > as rcv_wnd so TCP autotuning will never start because
> > sender can't send more than rcv_wnd size in one round trip.
> > To address this issue this patch is decreasing the initial
> > rcvq_space.space so TCP autotuning kicks in whenever the sender is
> > able to send more than 5360 bytes in one round trip regardless the
> > receiver's configured MTU.
> >
> > Fixes: a337531b942b ("tcp: up initial rmem to 128KB and SYN rwin to around 64KB")
> > Fixes: 041a14d26715 ("tcp: start receiver buffer autotuning sooner")
> >
> > Signed-off-by: Hazem Mohamed Abuelfotoh <abuehaze@xxxxxxxxxx>
> > ---
> > net/ipv4/tcp_input.c | 3 ++-
> > 1 file changed, 2 insertions(+), 1 deletion(-)
> >
> > diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c
> > index 389d1b340248..f0ffac9e937b 100644
> > --- a/net/ipv4/tcp_input.c
> > +++ b/net/ipv4/tcp_input.c
> > @@ -504,13 +504,14 @@ static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
> > static void tcp_init_buffer_space(struct sock *sk)
> > {
> > int tcp_app_win = sock_net(sk)->ipv4.sysctl_tcp_app_win;
> > + struct inet_connection_sock *icsk = inet_csk(sk);
> > struct tcp_sock *tp = tcp_sk(sk);
> > int maxwin;
> >
> > if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
> > tcp_sndbuf_expand(sk);
> >
> > - tp->rcvq_space.space = min_t(u32, tp->rcv_wnd, TCP_INIT_CWND * tp->advmss);
> > + tp->rcvq_space.space = min_t(u32, tp->rcv_wnd, TCP_INIT_CWND * icsk->icsk_ack.rcv_mss);
>
> I find using icsk->icsk_ack.rcv_mss misleading.
>
> I would either use TCP_MSS_DEFAULT , or maybe simply 0, since we had
> no samples yet, there is little point to use a magic value.
0 will not work, since we use a do_div(grow, tp->rcvq_space.space)
>
> Note that if a driver uses 16KB of memory to hold a 1500 bytes packet,
> then a 10 MSS GRO packet is consuming 160 KB of memory,
> which is bigger than tcp_rmem[1]. TCP could decide to drop these fat packets.
>
> I wonder if your patch does not work around a more fundamental issue,
> I am still unable to reproduce the issue.
