net.inet.tcp.slowstart_flightsize in 8-STABLE

[Andre Oppermann]

On 13.07.2010 16:01, Maxim Dounin wrote:
> Hello!
>
> On Wed, May 12, 2010 at 04:47:02PM +0400, Igor Sysoev wrote:
>
>> It seems that net.inet.tcp.slowstart_flightsize does not work in 8-STABLE.
>> For a long time I used slowstart_flightsize=2 on FreeBSD 4, 6, and 7 hosts.
>> However, FreeBSD-8 always starts with the single packet.
>> I saw this on different versions of 8-STABLE since 8 Oct 2009 till
>> 04 Apr 2010.
>
> Finally I had some time to look into it (sorry for long delay).
>
> 1. Slow start isn't used on recent FreeBSD versions for initial snd_cwnd
> calculations as long as you have rfc3390 support switched on (default since
> Jan 06 23:29:46 2004, at least in 7.*).  It effectively sets initial
> snd_cwnd to 3*MSS on common networks and shouldn't cause any problems.
> Slowstart_flightsize only affects connection restarts.
>
> 2. Due to bug in syncache code (patch below) all accepted connections has
> their snd_cwnd reset to 1*MSS (since r171639, 7.0+ AFAIR).
>
> 3. Support for rfc3465 introduced in r187289 uncovered (2) as
> ACK to SYN/ACK no longer causes snd_cwnd increase by MSS (actually, this
> increase shouldn't happen as it's explicitly forbidden by rfc 3390, but
> it's another issue).  Snd_cwnd remains really small (1*MSS + 1) and this
> causes really bad interaction with delayed acks on other side.
>
> As a workaround to delayed acks interaction problems you may disable
> rfc3465 by setting net.inet.tcp.rfc3465 to 0.  Correct fix would be to apply
> the patch below.
>
> To Andre Oppermann: could you please take a look at the patch and
> commit it if found appropriate?

Thanks for your debugging.  I will look into it and commit if there
are no further questions.

-- 
Andre

> # HG changeset patch
> # User Maxim Dounin<mdounin at mdounin.ru>
> # Date 1279028684 -14400
> # Node ID 93699203f408fa8e22c971769bde9c26bd14d410
> # Parent  e2cf8c51a6294a0d09d5628d96c6ab3ab626957e
> Fix cwnd resetting problem introduced in r171639.
>
> Sc_rxmits counts timer engagements, not actual retransmits, and value 1
> corresponds to no retransmits.  Revert check to correct one as present
> before r171639.
>
> diff --git a/netinet/tcp_syncache.c b/netinet/tcp_syncache.c
> --- a/netinet/tcp_syncache.c
> +++ b/netinet/tcp_syncache.c
> @@ -804,8 +804,10 @@ syncache_socket(struct syncache *sc, str
>
>   	/*
>   	 * If the SYN,ACK was retransmitted, reset cwnd to 1 segment.
> +	 * Note that sc_rxmits counts timer engagements, not actual
> +	 * retransmits.
>   	 */
> -	if (sc->sc_rxmits)
> +	if (sc->sc_rxmits>  1)
>   		tp->snd_cwnd = tp->t_maxseg;
>   	tcp_timer_activate(tp, TT_KEEP, tcp_keepinit);
>
>
> Maxim Dounin
>