per-interface packet filters

[Vladimir Grebenschikov]

В вт, 14/12/2004 в 13:54 +0100, Andre Oppermann пишет:
> Vladimir Grebenschikov wrote:
>  
> > > I know this. We have a well commented firewall scripts, we store them at RCS,
> > > we do many things to make our life easier. But my practice (and my collegues)
> > > shows that per interface filters are easier to understand and maintain when
> > > number of interfaces grows up to 20 and more, and they all are logically
> > > different - clients, servers, DMZs, hardware, nated networks, etc.
> > >
> > > Again, this feature is not for all. This is for people who build complicated
> > > routers on FreeBSD. It is not going to hurt standard host setups.
> > 
> > Frankly speaking, I think ppl who runs real-life router with firewall on
> > fbsd will vote for this feature by both hands.
> > 
> > I sometime, some years ago I had freebsd router with near to 100
> > interfaces  (mostly VLANs and FrameRelay customers connections, and
> > about 10 physical media interfaces). This router transfers some
> > thousands packets per second. It was real trouble to rearrange ipfw
> > table with large (very large) number of jumps (especially in case when
> > some number range was exceeded and renumbering required). Also most of
> > router interrupt time was spent in going through client multiplexer part
> > of ipfw ruleset.
> > 
> > Gleb, please do the feature.
> > 
> > Why we do not avoid bottlenecks where they can be avoided ?
> > With that feature we can select right rules for specific interface
> > without do linear search by ruleset.
> 
> It's about HOW to implement it.  I think the ways proposed so far are
> hackish, too complex and outside of our framework which was very well
> designed and allows this kind of feature without any of the hacks and
> extentions discussed here.
> 
> We have to properly DESIGN these feature instead of just hacking them
> in.

Well, I agree, that it is about how to design it.
But I do not think that proposed solution is hackish, and I not alone
with it.

Let's imagine our firewall framework as general firewall, able to be
plugged on different layers, after that you can get following:

1. Plug firewall (dedicated chain) between netgraph nodes
2. Plug firewall on any specific interface 
3. Plug firewall on any network packet processing input/output (current)
4. Plug it into bridging code

In this list interface looks very reasonable as place to plug, for me it
looks even more reasonable then our usual input/output, because packet
on ether output gives you no idea where it come from - local or remote,
especially in complex setups, with often changing interface names and
indexes (pptp server for example). It is not clear how to write rules
that affects only local traffic and transit traffic (I do not mean loop-
back when talk about local traffic)

> > Do we what FreeBSD be used on large scale of setups or we have think
> > targeting ?
> 
> As long as there is a sufficient large base we are not opposed to it.
> What we are opposed at is tradeoffs which favor one particular minority
> special interest over the general average interest set.
> 
> > -- off-topic --
> > Days ago FreeBSD was only OS flexible and stable enough to be use in
> > complex, customized network environments, but now-days it is not so :(,
> > and you know why.
> > -- off-topic -- (not for flame or advocacy, just emotion)
> 
> No, I don't know why and this isn't helping any.
> 
-- 
Vladimir B. Grebenchikov
vova at fbsd.ru