per-interface packet filters, design approach

[Andre Oppermann]

Bruce M Simpson wrote:
> 
> Hi,
> 
> At this point I'm tempted to explicitly *not* roll support for IPFW1
> in XORP's ACL manager precisely because of its limitations; see below.

I'd say IPFW1 is dead.  Just ignore it and require IPFW2 on 4.x.

> On Tue, Dec 14, 2004 at 03:19:28PM +0100, Andre Oppermann wrote:
> > IPFW2 does have this functionality.  It's called "sets" of rules which
> > you can atomically swap, enable, disable, flush and intermix with each
> > other.  It's there, read ipfw(8), it's on the 15th line.
> 
> OK. This is probably something I can deal with. Basically XORP has an
> intermediate rule representation which tries to be generic enough to
> deal with divergent packet filter implementations across various OS
> platforms, and yet specific enough to be useful.
> 
> A XORP rule tuple looks like this:
>  ifname, vifname, src, dst, proto, sport, dport, action
> Rule matches take place on all fields but the 'action' part of the tuple.

Can you provide examples of a XORP packet filter rule set?

> The interface to XORP's packet ACL manager is transaction driven to ensure
> atomicity. Where this atomicity can't be guaranteed by the underlying
> back-end, the back-end should attempt to mimic it using whatever tricks
> are necessary.
> 
> Snapshots get taken at two levels: XORP's intermediate representation
> described above, and the back-end's state. These snapshots can be taken
> either for the purpose of safely rolling back state when rulesets are
> being changed, or for communicating rulesets between different parts of
> the packet ACL system.
> 
> I would imagine that mapping between an IPFW2 'set' and a PaIpfwBackend
> snapshot on the fly would do the trick.

Perfect match.  You can even keep up to 32 versions in the kernel and
do one-syscall rollback's/forward's to any of them.

> I just committed the core of XORP's ACL manager last week, please feel
> free to have a look at it and give me feedback.

I did take a quick look but my c++ understanding is horribly and I don't
have time to work myself through the XORP framework.

-- 
Andre