Re: [PATCH net-next] netem: fix classful handling

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

Le jeudi 29 décembre 2011 à 17:52 +0100, Hagen Paul Pfeifer a écrit :
> * Eric Dumazet | 2011-12-29 10:12:02 [+0100]:
> 
> >> Also, the whole tfifo idea is only to support the wierd idea that
> >> if doing random delay that packets should get reordered based on the
> >> results of the random value; it was an behavior some users wanted
> >> because that is what NISTnet did.
> >
> >tfifo supports a time ordered queuing, wich mimics some jitter in the
> >network. This seems quite useful.
> >
> >I see what you suggest : adding 'time_to_send' in the generic qdisc cb.
> >
> >But it makes no sense if we attach a reordering qdisc, like SFQ :
> >A 'high prio' packet will block the whole netem because we'll have to
> >throttle since this packet time_to_send will be in the future, while
> >many other elligible packets are in queue.
> 
> In other words netem jitter and a qdisc !tfifo will not work. Correct? The
> rate extension also peak the last packet to get the reference time (assuming a
> strict ordering):
> 

Yep, current situation is borked. It assumes we _use_ tfifo, for delay
jitters but also for rate extension.

> [...]
> now = netem_skb_cb(skb_peek_tail(list))->time_to_send;
> [...]
> 
> 
> We should avoid a different (unseeable) behavior depending on the queue
> (tfifo, SFQ).  Another point: operate netem and qdisc on the same computer can
> lead to timing abnormalities. In our test setups we operate qdisc/tcp/whatever
> setups and netem on more then on computer.
> 

After my patch you could use netem as a delay module before a complex
qdisc setup for example.

[ Simulating a 10ms delay on a 10Gigabit link is expensive, since you
need to allow up to ~150.000 packets in tfifo. Maybe we should switch to
tbfifo [giving a limit in bytes, not packets ] ]