[altq 1135] Re: parent classes with filters and priorities ?

[Kenjiro Cho <kjc@csl.sony.co.jp>]

Peter Galbavy wrote:
> This is a genaral question about how filters are prioritised and selected
> and then specifically for what I am doing;
> 
> This is a section of my manually generated config; soon to be automatic:
> 
> ---
> class cbq nge0 customers root_class priority 2 pbandwidth 10 default
>         filter nge0 customers 213.170.0.0 netmask 255.255.248.0 0 0 0 0
>         filter nge0 customers 192.88.83.0 netmask 255.255.255.0 0 0 0 0
>         filter nge0 customers 193.195.141.0 netmask 255.255.255.0 0 0 0 0
> 
> class cbq nge0 knowtion customers borrow priority 3 exactbandwidth 2M red
>         filter nge0 knowtion 0 0 213.170.0.0 netmask 255.255.248.0 0 0
>         filter nge0 knowtion 0 0 192.88.83.0 netmask 255.255.255.0 0 0
>         filter nge0 knowtion 0 0 193.195.141.0 netmask 255.255.255.0 0 0
> ---
> 
> nge0 is a 1000BaseT interface from an edge router to the core network. What
> I want to achieve is:
> 
> 1. define traffic queues from the customer to the rest of the world
> 2. ignore restrictions if the destination is for another customer / customer
> network.

I'm not sure if I understand your intension but
 - in the altq filter syntax, <dest addr> comes first.
   (this might be confusing but there are historical reasons...)
 - if nge0 is an interface from edge to core, it shouldn't have
   outgoing traffic both from and to the same customer...

> When I use this config, altqd warns that the later filters may override the
> earlier ones. I also tried placing both these classes under the same parent
> (rather than one under the other), but with similar warnings.
> 
> To help me clarify whether or not I can ignore these warnings, could someone
> with more understanding please explain what the decision process is in
> matchinng filters ? First matched ? Last matched ? etc.

It is described in altq-3.0/docs/TIPS.txt (or
http://www.csl.sony.co.jp/~kjc/software/TIPS.txt) section 2.2.

> Do filters in parent classes limit what gets checked in lower level classes
> ? Should they maybe ?

The filter order has nothing to do with the relation of classes.
Only the order of the filter rules matters.

> Also, in general, how does a large number of filters effect performance ? Is
> the matching done via a hash or other fast look up or is it a linear search
> ?

ALTQ uses a linear search for wildcard filters.
Fixed filters which has all the fields specified are hashed.

> Can I suggest that whoever is working on filters looks at the work done
> recently on OpenBSD pf in case there are things there that could be used in
> the classifiers in altq ?

Probably, we need a classifier version of ALTQ, say ALTCF :)

-Kenjiro