Le jeudi 29 décembre 2011 à 19:25 +0100, Hagen Paul Pfeifer a écrit :

* Eric Dumazet | 2011-12-29 19:10:36 [+0100]:

quoted

I dont quite understand the question. The patch I posted is supposed to
fix the problem. What do you want to tell to the user ?

I assumed that the patch makes it possible to replace standard tfifo?! Tfifo
provides strict ordering, other qdisc's do not. So I thought if someone use
netem rate|jitter with e.g. SFQ then this should be mentioned somewhere. E.g.
"rate|jitter can only be used with tfifo qdisc". Correct me if I am wrong.

Current netem uses a single queue, default tfifo.

Then if you change this tfifo by SFQ, you lose tfifo, and netem doesnt
work at all.

I claim we cannot remove tfifo.

After my patch, you now have :

1) An internal mandatory tfifo queue, to fulfill time_to_send
requirements.

2) An optional qdisc (SFQ in your example), where packets are queued
once dequeued from tfifo at the right time (after netem delay/rate
respected), eventually a packet can finaly be delivered to device with a
_bigger_ delay than the one predicted in tfifo, because of trafic
shaping happening in this optional qdisc.


Full netem block :

              +---------+                   +--------+
-->>-enqueue--| tfifo   |---->>>>--enqueue--|  SFQ   |-dequeue--->>>>> 
              +---------+                   +--------+