Packet filtering refers to the technique of conditionally allowing or denying packets entering or exiting a network or host based on the characteristics of that packet. There are two fundamental types of packet filters. A static packet filter is a set of rules against which every packet is checked, and allowed or denied.
A dynamic packet filter keeps track of the connections currently passing the firewall. This is usually described as a stateful or dynamic packet filtering engine. Netfilter provides the capability for linux (2.4+) to operate as a stateful packet filtering device.
The network layer portion of a firewall solution, packet filtering is one part of a good security stance. As the embodiment and manifestation of an organizational security policy for network layer traffic, the packet filter restricts traffic flows between networks and hosts. There is tremendous value from a security perspective in enforcing these traffic flows, instead of allowing arbitrary traffic flow.
The use of packet filtering to enforce these traffic flows is not restricted to routers and firewalls alone. Standalone servers and workstations can use these very same tools to protect themselves. There are a couple of common approaches to packet filtering. Generally, network security professionals subscribe to the notion that the filtering policy should deny or drop all traffic and selectively allow desired traffic. An alternate, more open, policy suggests allowing everything, selectively blocking undesirable traffic.
The languages used in most packet filtering tools for describing IP packets allow for a great deal of specifity when identifying traffic. This specifity enables an administrator a great deal of flexibility for protecting resources and limiting traffic flows.
Packet filtering under linux has a long history, punctuated by major alterations in the packet filtering systems included in the kernel. In the mid- and late-1990s, ipfwadm exposed the three packet filtering chains of kernel 2.0 to the user: in, forward, and out. Individual entries added to these chains would be traversed in order in the ruleset. The first matching rule in each chain would be used, and every packet passing through a router would traverse each of these chains.
With the advent of linux 2.2, users could create their own chains and chain structures. The kernel architecture was different, but from the user's perspective, the manner in which the rules were written was only slightly different. Rule chains, traversed rather like subroutines and manipulated with ipchains, could be arbitrarily complex and nested. The built-in packet filtering chains are input, output and forward. The first matching rule in any chain called from one of the built-in chains would be used. Every packet passing through a router would traverse (at least) the three built-in rule chains. There is backward compatible support for ipfwadm syntax via a wrapper shell script which converts the command to an ipchains syntax.
In kernel 2.4, the netfilter architecture which provides functionality other than packet filtering, allows users to create the arbitrary chains and chain structures similar to those supported by linux 2.2. The built in chains are INPUT, FORWARD, and OUTPUT. A major difference in the use of chains was introduced in linux 2.4; packets passing through a router will traverse the FORWARD chain only. User-defined iptables chains resemble branches rather than subroutines. Under linux 2.4, ipchains compatibility is maintained with a kernel module. For ipfwadm compatibility, the kernel module and the aforementioned wrapper shell script function adequately.
The packet filtering support under linux has grown increasingly complex and mature with successive kernels and development efforts on the user space tools. The netfilter architecture of linux 2.4 represented a tremendous step forward in the packet filtering capabilities of linux with support for stateful packet filtering.
Although the functionality offered by linux kernels for protecting network resources with packet filtering allows tremendously specific network layer access control and auditing capability, it alone cannot successfully and completely protect network resources. There are limits to the usefulness of packet filters.
In cases where a packet filter restricts access to a resource based on the source IP address attempting to access that resource, the packet filter cannot verify whether the packets originate from the real device or from a host or router spoofing this source address. A transparent proxy illustrates this problem perfectly. A transparent proxy frequently runs on a masquerading or NAT host which is connected to the Internet. This machine intercepts outbound connections for a particular protocol (e.g, HTTP), and simulates the real server to the client. The client may have a packet filter limiting outbound connections to a single IP and port pair, but the transparent proxy will still operate on the outbound connection.
This is an innocuous example, indeed. A potentially more threatening example is an ssh server which accepts connections only from an IP range. Any router between the two endpoints which can spoof IP packets will be able to pass the packet filter, whether it is a stateful or a static packet filter. This should underscore the importance of solid application layer security in addition to the need for judiciously employed packet filtering.
A packet filter makes no effort to validate the contents of a data stream, so data passed over a packet filter may be bogus, invalid or otherwise incorrect. The packet filter only verifies that the network layer datagrams are correctly addressed and well-formed . Many security devices, such as firewalls, include support for proxies, which are application aware. These are security mechanisms which can validate data streams. Proxies are often integrated with packet filters for a tight network layer and application layer firewall.
Another area of network security which is not addressed by packet filtering is encryption. Encryption can be used at a number of different layers in a networked environment. Compare IPSec, encrypted packets, with Secure Sockets Layer (SSL), which encrypts a single application layer session. IPSec operates at layer 3, while SSL operates above layer 4. Packet filtering does not directly address the issue of encryption in any way.
There are a few good starting place for users wishing for some guidelines on securing machines. First, the Security Quickstart HOWTO is a good place to begin. There is also the Security HOWTO. These and several other good general security resources are also available via linuxsecurity.com's documentation area.
Much of the previous discussion applies to packet filtering in general, and linux suffers from the same limitations of packet filtering. It is folly to assume that a good packet filter makes a network immune from security issues.
 In truth, there is some examination of data inside the network layer datagram. Almost all packet filtering engines allow the user to distinguish between the different IP protocol types, such as GRE, TCP, UDP, ICMP, and even attributes of these datagrams and segments. The important thing to realize is that a packet filter makes no effort to examine the data stream.