A burst from NewReno when a partial ACK is received

[Noritoshi Demizu]

I'm using FreeBSD current for my experiences.
I observed bursts sent by NewReno when a partial ACK is received.
I have two packet traces of such bursts.  One of such bursts is
analyzed at http://www.demizu.org/~noritosi/memo/2005/0623/ .

I think tcp_newreno_partial_ack() in tcp_input.c rev 1.275 has a bug
in calculating a new value of snd_cwnd at the tail of the function.
Currently, snd_cwnd is re-calculated as following:

  L.3113:  tp->snd_cwnd -= (th->th_ack - tp->snd_una - tp->t_maxseg);

Since snd_cwnd is u_long, if snd_cwnd < SEG.ACK - SND.UNA - MSS,
snd_cwnd becomes awfully huge and a burst of data can be sent.

To fix this problem, I'd like to suggest the patch below.

Thanks.

Regards,
Noritoshi Demizu


Index: tcp_input.c
===================================================================
RCS file: /home/cvsup/FreeBSD/ncvs/src/sys/netinet/tcp_input.c,v
retrieving revision 1.275
diff -u -r1.275 tcp_input.c
--- tcp_input.c	1 Jun 2005 12:03:18 -0000	1.275
+++ tcp_input.c	23 Jun 2005 06:20:28 -0000
@@ -3110,7 +3112,11 @@
 	 * Partial window deflation.  Relies on fact that tp->snd_una
 	 * not updated yet.
 	 */
-	tp->snd_cwnd -= (th->th_ack - tp->snd_una - tp->t_maxseg);
+	if (tp->snd_cwnd > th->th_ack - tp->snd_una)
+		tp->snd_cwnd -= th->th_ack - tp->snd_una;
+	else
+		tp->snd_cwnd = 0;
+	tp->snd_cwnd += tp->t_maxseg;
 }
 
 /*