1. Introduction

Welcome, gentle reader.

This document is a journey; some parts are well-travelled, and in other areas you will find yourself almost alone. The best advice I can give you is to grab a large, cozy mug of coffee or hot chocolate, get into a comfortable chair, and absorb the contents before venturing out into the sometimes dangerous world of networking.

1.1 What is netfilter?

netfilter is a framework for packet mangling, outside the normal Berkeley socket interface. It has five parts. Firstly, each protocol defines "hooks" (IPv4 defines 6) which are well-defined points in a packet's traversal of that protocol stack. At each of these points, the protocol will call the netfilter framework with the packet and the hook number.

Secondly, parts of the kernel can register to listen to the different hooks for each protocol. So when a packet is passed to the netfilter framework, it checks to see if anyone has registered for that protocol and hook; if so, they each get a chance to examine (and possibly alter) the packet in order, and discard the packet, allow it to pass, or ask netfilter to queue the packet for userspace.

The third part is that packets which have been queued for userspace are collected (usually by the netfilter_dev driver) for sending to userspace; these packets are handled asynchronously.

The fourth part is that any packet dropped once inside the IP layer can be examined before it is actually dropped. For example, this allows modules to examine IP-spoofed packets (dropped by the routing code).

The final part consists of cool comments in the code and documentation. This is instrumental for any experimental project. The netfilter motto is (stolen shamelessly from Cort Dougan):

        ``So... how is this better than KDE?''

In addition to this raw framework, various modules have been written which provide functionality similar to previous (pre-netfilter) kernels, in particular, an extensible NAT system, and an extensible packet filtering system (iptables).

1.2 What's wrong with what we had in 2.0 and 2.2?

1. No infrastructure established for passing packet to userspace:
   • Kernel coding is hard
   • Kernel coding must be done in C/C++
   • Dynamic filtering policies do not belong in kernel
   • 2.2 introduced copying packets to userspace via netlink, but reinjecting packets is slow, and subject to `sanity' checks (eg. reinjecting packet claiming to come from an existing interface is not possible).
2. Transparent proxying is a crock:
   • We look up every packet to see if there is a socket bound to that address
   • Root is allowed to bind to foreign addresses
   • Can't redirect locally-generated packets
   • REDIRECT doesn't handle UDP replies: redirecting UDP named packets to 1153 doesn't work because some clients don't like replies coming from anything other than port 53.
   • REDIRECT doesn't coordinate with tcp/udp port allocation: a user may get a port shadowed by a REDIRECT rule.
   • Has been broken at least twice during 2.1 series.
   • Code is extremely intrusive. Consider the stats on the number of #ifdef CONFIG_IP_TRANSPARENT_PROXY in 2.2.1: 34 occurances in 11 files. Compare this with CONFIG_IP_FIREWALL, which has 10 occurrances in 5 files.
3. Creating packet filter rules independent of interface addresses is not possible:
   • Must know local interface addresses to distinguish locally-generated or locally-terminating packets from through packets.
   • Even that is not enough in cases of redirection or masquerading.
   • Forward chain only has information on outgoing interface, meaning you have to figure where a packet came from using knowledge of the network topography.
4. Masquerading is tacked onto packet filtering:

   Interactions between packet filtering and masquerading make firewalling complex:

   • At input filtering, reply packets appear to be destined for box itself
   • At forward filtering, demasqueraded packets are not seen at all
   • At output filtering, packets appear to come from local box

5. TOS manipulation, redirect, ICMP unreachable and mark (which can effect port forwarding, routing, and QoS) are tacked onto packet filter code as well.
6. ipchains code is neither modular, nor extensible (eg. MAC address filtering, options filtering, etc).
7. Lack of sufficient infrastructure has led to a profusion of different techniques:
   • Masquerading, plus per-protocol modules
   • Fast static NAT by routing code (doesn't have per-protocol handling)
   • Port forwarding, redirect, auto forwarding
   • The Linux NAT and Virtual Server Projects.
8. Incompatibility between CONFIG_NET_FASTROUTE and packet filtering:
   • Forwarded packets traverse three chains anyway
   • No way to tell if these chains can be bypassed
9. Inspection of packets dropped due to routing protection (eg. Source Address Verification) not possible.
10. No way of atomically reading counters on packet filter rules.

1.3 Who are you?

I'm the only one foolish enough to do this. As ipchains co-author and current Linux Kernel IP Firewall maintainer, I see many of the problems that people have with the current system, as well as getting exposure to what they are trying to do.

1.4 Why does it crash?

Woah! You should have seen it last week!

Because I'm not as great a programmer as we might all wish, and I certainly haven't tested all scenarios, because of lack of time, equipment and/or inspiration. I have two Intel laptops connected by PLIP and a NetWinder connected to one laptop via 10baseT. (Cort, are you listening? Where's my PowerPC box?)