Networking Stack Enhancements
Now we finally come to a topic near and dear to my heart. Initially this article was going to focus only on the improvements to the networking stack, but on consideration I decided that it wouldn't be fair to members of the audience who weren't using Linux as a router, firewall, packet-shaper, what-have-you… (not to mention the fact that the other new features are important in their own right). That being said, get ready for some very interesting developments.
Packet Filtering
First, the packet-filtering and masquerading code has been changed again. For the third time in as many kernel series, administrators will have to learn yet another interface to configure their firewall rules and masquerading. The new mechanism is called netfilter and uses the tool iptables to administer the rulesets, as compared to ipfw functionality (using the tool ipfwadm) in the 2.0 kernel and ipchains functionality (using the tool ipchains) in the 2.2 kernel. Frankly, I was a bit skeptical at first; not too long after taking the time to convert all my ipfwadm commands over to ipchains (taking advantage of a few of the improvements along the way), everything has to be redone just so that I can upgrade to the 2.4 kernel!
Well, that's not entirely true. One of the first features I encountered is that netfilter (I will use the term netfilter to refer to the kernel facility and iptables to refer to the explicit command) is that it provides modules (or tables) that implement the ipfwadm or ipchains functionality. The only caveat is that you can't mix the three facilities together, but this shouldn't pose much of a problem. The backward compatibility is a nice option for those who want to take advantage of some of the previously mentioned features of the 2.4 kernel, but can't yet risk reconfiguring the firewall to use netfilter. Because the facilities are implemented as modules, you can switch between them simply by loading and unloading modules and using the appropriate command-line tool for the enabled facility. This certainly eases the migration to netfilter, although it's unlikely once you get a taste of netfilter's features that you'll stick with ipchains for long.
Network Address Translation (NAT)
The first major netfilter feature is a biggie: Linux finally supports true Network Address Translation (NAT), just like Cisco PIX or any other NAT solution. The difference between this and just masquerading is subtle, but fantastic. With masquerading, all client packets being forwarded by the router are massaged to carry the source address of the masquerading router, and this massaging process involves remapping the source port number to a free (normally different) port on the masquerading router. When the destination responds, it sends the packet to the router at the remapped port, and the router then "unmasquerades" the packet and forwards it to the client.
This works perfectly well for standard (really, I should say "simple") client/server protocols where the client chooses any port (typically above 1024) and talks to a single, well-known server port. This scheme shows its limitations, however, when the client doesn't always initiate the connection, or when the server would like to stream UDP packets back to the client. (Examples include IRC, FTP, X, and RealAudio.) The 2.2 world has workarounds for given protocols—kernel modules and user-space daemons—but they all have the limitation that only one of any given well-known port is available on the masquerading machine. Imagine configuring H.323 for a family of workstations behind a masquerading firewall. This protocol can be described as "peer-to-peer," meaning that each side of the connection has both client and server components. Once you map the server port to one of your workstations, you're all out of server ports. X windows suffers the same fate, but fortunately OpenSSH comes to the rescue.
NAT takes care of this issue for you, given that you have a pool of addresses for the network on the external side of the router. For each client, the NAT module can allocate a "real" address to it from this pool, meaning that neither the system on the other end nor the client on your network has any idea that address translation is occurring. Of course, this isn't limited to just clients; you can use this in conjunction with packet filters to enforce very strict firewalling policies by routing all traffic for, say, your SMTP server to the NAT address on the firewall, and then filtering out everything except TCP traffic to and from port 25. To do this, you would merely need to advertise the MX record for that SMTP server to be the address on the NAT, and then configure the NAT to forward this traffic to the mail server. It doesn't take much imagination to conceive a plan for load balancing based on this feature (similar to the Big-IP products from F5 Networks).
What Else Can Netfilter Do?
Let's move on from NAT and talk about the rest of netfilter. One nice thing about netfilter is that it's conceptually more similar to ipchains than ipchains is to ipfwadm. There is still the concept of chains of rules that a packet traverses, as well as the idea of both kernel chains and user-defined chains. Depending on matching criteria, a packet will either match and thus traverse a chain, or pass by it. At the end of the chain, it will be passed on to another chain, dropped, etc. The details of the kernel chains have changed somewhat, so be sure to read the relevant documentation.
The architecture of netfilter is such that it's extensible; extensions for a particular protocol or to add a given functionality are available as modules that be loaded into the kernel. Some of these extensions are familiar to ipchains users, such as the TCP and UDP match extensions. Others are brand new, such as the limit module, which can be used to set the maximum rate at which packets can traverse a particular chain. This can be used to suppress duplicate log entries, prevent DoS (Denial of Service) attacks, and limit bursts of a given type of service. Another extension allows matching based on the owner or group of locally generated packets. As an example, this could be used on multi-user machines to prevent users from sending mail directly out to an external host, forcing them to use the local MTA (Mail Transfer Agent), which would ensure logging as well as prevent spoofing and internally generated spam.
Before moving on, let me mention a couple of other goodies. A LOG extension allows logging of rule matches via the standard syslog facility, and a QUEUE extension sends packets up to user space to be massaged, reviewed, and so on before further traversing the stack. Both of these features promise very sophisticated security tools, such as adaptive packet filters that change their configuration based on traffic volume or type of attack.
QoS/Fair-Queueing
If netfilter doesn't inspire you to rush out and upgrade, perhaps your interests lie more at the heart of the network than on the periphery. Quality of Service (QoS) was all the buzz at the last Networld+INTEROP I attended. It deals with one of the core limitations of TCP/IP—the sticky problem of needing to prioritize network traffic based on interactive versus batch, and important customers versus Joe Nobody. The kernel does this through the QoS facility, which offers a variety of algorithms for the packet scheduler. Some of these algorithms allow special prioritization for real-time scheduling; others can be used to throttle certain types of traffic (on certain interfaces, if desired). Of significance is the fact that the standards-based diffserv (Differentiated Services) and RSVP (Resource Reservation Protocol) protocols are supported, which means that Linux can coordinate with other non–Linux network infrastructures to provide quality of service features to your clients.
Various other gizmos and doodads make the Linux 2.4 kernel attractive. Policy-based routing replaces the iproute2 functionality in version 2.2 to allow routing decisions to be made based on packet source address, TOS field, or a netfilter rule (based on the mark feature). An experimental HTTP server is implemented directly in kernel space for the ultimate in Web server performance, and an Ethernet "frame diverter" provides transparent proxying/redirection all the way down at layer 2. Of all the goodies, I still feel that netfilter is the first and foremost. NAT is vitally important to Linux's future as a network appliance, so I'm glad to see it in there. As Rusty Russell (the Linux IP Firewall maintainer and author of netfilter) states in his Linux 2.4 Packet Filtering HOWTO, "I predicted 6 months, and it took 12, but I felt by the end that it had been done Right." I'm glad that he took the time. You can read up on netfilter at http://netfilter.kernelnotes.org/.