Re: [PATCH net-next] tcp: switch rtt estimations to usec resolution

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

On Sun, 2014-02-23 at 09:36 +0200, Julian Anastasov wrote:
> 	Hello,
> 
> On Sat, 22 Feb 2014, Eric Dumazet wrote:
> 
> > From: Eric Dumazet <edumazet@google.com>
> > 
> > Upcoming congestion controls for TCP require usec resolution for RTT
> > estimations. Millisecond resolution is simply not enough these days.
> 
> ...
> 
> > --- a/net/ipv4/tcp_metrics.c
> > +++ b/net/ipv4/tcp_metrics.c
> 
> ...
> 
> > @@ -408,10 +395,10 @@ void tcp_update_metrics(struct sock *sk)
> >  	 */
> >  	if (!tcp_metric_locked(tm, TCP_METRIC_RTT)) {
> >  		if (m <= 0)
> > -			rtt = tp->srtt;
> > +			rtt = tp->srtt_us;
> >  		else
> >  			rtt -= (m >> 3);
> > -		tcp_metric_set_msecs(tm, TCP_METRIC_RTT, rtt);
> > +		tcp_metric_set(tm, TCP_METRIC_RTT, rtt);
> 
> 	What is the way to handle this change of
> resolution in TCP_METRICS_ATTR_VALS? The ip tool has
> code in ip/tcp_metrics.c process_msg() that needs to
> know what we get, ms or us. Do we need to add some
> new attribute as flag or version? Or we need to increase
> TCP_METRICS_GENL_VERSION in kernel and to check it in
> process_msg ?

crap... This should have been usec units as we did for tcp_info years
ago.

Solution is to add TCP_METRIC_RTT_US and TCP_METRIC_RTTVAR_US attributes

I'll send a v2

Thanks !