Re: [PATCH v2 net-next] tcp: fix ABC in tcp_slow_start()

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

On Fri, 2012-07-20 at 09:03 -0700, Neal Cardwell wrote:
> On Fri, Jul 20, 2012 at 8:07 AM, Yuchung Cheng <ycheng@google.com> wrote:
> > On Fri, Jul 20, 2012 at 8:02 AM, Eric Dumazet <eric.dumazet@gmail.com> wrote:
> >>         tp->snd_cwnd_cnt += cnt;
> >>         while (tp->snd_cwnd_cnt >= tp->snd_cwnd) {
> 
> Nice catch, Eric.
> 
> One thing that's always bothered me about the tp->snd_cwnd_cnt code is
> that the slow start and congestion avoidance use different criteria
> for incrementing snd_cwnd_cnt. tcp_slow_start() increments
> snd_cwnd_cnt by snd_cwnd for each ACKed packet, and congestion
> avoidance increases snd_cwnd_cnt by just 1 for each packet.
> 
> This means that if we exit slow start and enter congestion avoidance,
> then we think we can have a "credit" for a bunch of ACKs that never
> happened (up to snd_cwnd-1), so we can conceivably do our first
> additive increase in congestion avoidance up to almost 1RTT too
> early. Can we just get rid of the use of snd_cwnd_cnt in slow start,
> and just use local variables in tcp_slow_start() rather than trying to
> carry state between ACKs?

Apparently tcp_slow_start() needs the snd_cwnd_cnt in case 
"limited slow start"  is used :

cnt = sysctl_tcp_max_ssthresh >> 1;

So to address your point, maybe we should clear  snd_cwnd_cnt
when leaving slow start for congestion avoidance phase ?