Re: [PATCH] tcp: make TCP quick ACK behavior modifiable

[Eric Dumazet <eric.dumazet@gmail.com>]

Le lundi 23 août 2010 à 21:00 +0200, Hagen Paul Pfeifer a écrit :
> The TCP quick ACK mechanism analyze if a connection is interactive or
> not. Per default the quick ACK mechanism is enabled and ACK packets are
> triggered instantly to raise the CWND fast - which is clever for
> bulk data (non-interactive) flows. On the other hand interactive protocols
> like HTTP, SMTP or XMPP will suffer from the quick ACK mechanism
> because one additional packets is generated. A simple heuristic
> detects if a connection is interactive (pingpong) and if so,
> disable the quick ACK. But, the mechanism is not in the ability to
> blindly guess if a connection is interactive, and so it must wait for at
> least one return packet with payload.
> 
> For the server side this requires one additional packet because (packet
> number 5 and 6 can be combined):
> 
> 192.168.1.35.44833 > 78.47.222.210.80: Flags [S], seq 2854340018, win 5840, options [mss 1460,sackOK,TS val 4382726 ecr 0,nop,wscale 7], length 0
> 78.47.222.210.80 > 192.168.1.35.44833: Flags [S.], seq 719041385, ack 2854340019, win 5792, options [mss 1452,sackOK,TS val 2606891996 ecr 4382726,nop,wscale 7], length 0
> 192.168.1.35.44833 > 78.47.222.210.80: Flags [.], ack 1, win 46, options [nop,nop,TS val 4382730 ecr 2606891996], length 0
> 192.168.1.35.44833 > 78.47.222.210.80: Flags [P.], seq 1:682, ack 1, win 46, options [nop,nop,TS val 4382730 ecr 2606891996], length 681
> 78.47.222.210.80 > 192.168.1.35.44833: Flags [.], ack 682, win 56, options [nop,nop,TS val 2606892002 ecr 4382730], length 0
> 78.47.222.210.80 > 192.168.1.35.44833: Flags [.], seq 1:1441, ack 682, win 56, options [nop,nop,TS val 2606892002 ecr 4382730], length 1440
> 192.168.1.35.44833 > 78.47.222.210.80: Flags [.], ack 1441, win 69, options [nop,nop,TS val 4382737 ecr 2606892002], length 0
> 
> This patch provides a sysctl interface for the administrator to globally
> enable or disable TCP quick ACKs. Short lived protocols like HTTP will
> save a non unimportant portion of packets!
> 
> Disable TCP Quick ACK:
> $ echo 0 > /proc/sys/net/ipv4/tcp_quickack
> 
> Enable TCP Quick ACK:
> $ echo 1 > /proc/sys/net/ipv4/tcp_quickack
> 
> Signed-off-by: Hagen Paul Pfeifer <hagen@jauu.net>
> Cc: David S. Miller <davem@davemloft.net>
> Cc: Eric Dumazet <eric.dumazet@gmail.com>
> Cc: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
> ---
>  include/net/inet_connection_sock.h |    3 +++
>  net/ipv4/sysctl_net_ipv4.c         |    7 +++++++
>  net/ipv4/tcp.c                     |    2 ++
>  3 files changed, 12 insertions(+), 0 deletions(-)
> 
> diff --git a/include/net/inet_connection_sock.h b/include/net/inet_connection_sock.h
> index b6d3b55..da2fbaf 100644
> --- a/include/net/inet_connection_sock.h
> +++ b/include/net/inet_connection_sock.h
> @@ -170,9 +170,12 @@ static inline int inet_csk_ack_scheduled(const struct sock *sk)
>  	return inet_csk(sk)->icsk_ack.pending & ICSK_ACK_SCHED;
>  }
>  
> +extern int sysctl_tcp_quickack;
> +
>  static inline void inet_csk_delack_init(struct sock *sk)
>  {
>  	memset(&inet_csk(sk)->icsk_ack, 0, sizeof(inet_csk(sk)->icsk_ack));
> +	inet_csk(sk)->icsk_ack.pingpong = sysctl_tcp_quickack ? 0 : 1;

What about dccp using this function ?


>  }
>  
>  extern void inet_csk_delete_keepalive_timer(struct sock *sk);
> diff --git a/net/ipv4/sysctl_net_ipv4.c b/net/ipv4/sysctl_net_ipv4.c
> index d96c1da..8923ca8 100644
> --- a/net/ipv4/sysctl_net_ipv4.c
> +++ b/net/ipv4/sysctl_net_ipv4.c
> @@ -394,6 +394,13 @@ static struct ctl_table ipv4_table[] = {
>  		.proc_handler	= proc_dointvec
>  	},
>  	{
> +		.procname	= "tcp_quickack",
> +		.data		= &sysctl_tcp_quickack,
> +		.maxlen		= sizeof(int),
> +		.mode		= 0644,
> +		.proc_handler	= proc_dointvec
> +	},
> +	{
>  		.procname	= "tcp_mem",
>  		.data		= &sysctl_tcp_mem,
>  		.maxlen		= sizeof(sysctl_tcp_mem),
> diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
> index 176e11a..5161689 100644
> --- a/net/ipv4/tcp.c
> +++ b/net/ipv4/tcp.c
> @@ -279,6 +279,8 @@
>  
>  int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT;
>  
> +int sysctl_tcp_quickack __read_mostly = 1;
> +
>  struct percpu_counter tcp_orphan_count;
>  EXPORT_SYMBOL_GPL(tcp_orphan_count);
>  


So here is a new undocumented setting ? hint hint...

I thought setsockopt(TCP_QUICKACK) was already available for this
optimization ?