Re: Possible burst from TCP NewReno when a partial ACK is received

[Noritoshi Demizu <demizu@xxxxxxxxxxxxxx>]

I'd like to commit the patch below if no objection is raised
until August 28, 2005.

Regards,
Noritoshi Demizu


> From: Noritoshi Demizu <demizu@xxxxxxxxxxxxxx>
> To: bugs@xxxxxxxxxxxxxxxxxxxxxxx
> Subject: Possible burst from TCP NewReno when a partial ACK is received
> Date: Mon, 04 Jul 2005 19:11:34 +0900 (JST)
>
> I think TCP NewReno in DragonFlyBSD has a bug which may send
> a burst of data when a partial ACK is received.
>
> More precisely, when a partial ACK is received, if acked bytes is
> larger than congestion window, new congestion window becomes awfully
> huge value.
>
> This bug is the same as the one reported in the following mail:
> http://lists.freebsd.org/pipermail/freebsd-net/2005-June/007723.html
>
> I think tcp_newreno_partial_ack() should be modified as below.
> Through my experiences, I observed that in somecases (ocwnd > acked)
> is false (I inserted printf() there for test).  Without the change
> below, snd_cwnd would become awfully huge value.
>
> Regards,
> Noritoshi Demizu
>
>
> Index: netinet/tcp_input.c
> ===================================================================
> RCS file: /home/cvsup/DragonFlyBSD/dcvs/src/sys/netinet/tcp_input.c,v
> retrieving revision 1.60
> diff -u -r1.60 tcp_input.c
> --- netinet/tcp_input.c	10 May 2005 15:48:10 -0000	1.60
> +++ netinet/tcp_input.c	4 Jul 2005 10:02:57 -0000
> @@ -3108,7 +3108,10 @@
>  	if (SEQ_GT(old_snd_nxt, tp->snd_nxt))
>  		tp->snd_nxt = old_snd_nxt;
>  	/* partial window deflation */
> -	tp->snd_cwnd = ocwnd - acked + tp->t_maxseg;
> +	if (ocwnd > acked)
> +		tp->snd_cwnd = ocwnd - acked + tp->t_maxseg;
> +	else
> +		tp->snd_cwnd = tp->t_maxseg;
>  }
> 
>  /*