Configuring the Networking Hardware

We've been talking quite a bit about network interfaces and general TCP/IP issues, but we haven't really covered what happens when the "networking code" in the kernel accesses a piece of hardware. In order to describe this accurately, we have to talk a little about the concept of interfaces and drivers.

First, of course, there's the hardware itself, for example an Ethernet, FDDI or Token Ring card: this is a slice of Epoxy cluttered with lots of tiny chips with strange numbers on them, sitting in a slot of your PC. This is what we generally call a physical device.

For you to use a network card, special functions have to be present in your Linux kernel that understand the particular way this device is accessed. The software that implements these functions is called a device driver. Linux has device drivers for many different types of network interface cards: ISA, PCI, MCA, EISA, Parallel port, PCMCIA, and more recently, USB.

But what do we mean when we say a driver "handles" a device? Let's consider an Ethernet card. The driver has to be able to communicate with the peripheral's on-card logic somehow: it has to send commands and data to the card, while the card should deliver any data received to the driver.

In IBM-style personal computers, this communication takes place through a cluster of I/O addresses that are mapped to registers on the card and/or through shared or direct memory transfers. All commands and data the kernel sends to the card have to go to these addresses. I/O and memory addresses are generally described by providing the starting or base address. Typical base addresses for ISA bus Ethernet cards are 0x280 or 0x300. PCI bus network cards generally have their I/O address automatically assigned.

Usually you don't have to worry about any hardware issues such as the base address because the kernel makes an attempt at boot time to detect a card's location. This is called auto probing, which means that the kernel reads several memory or I/O locations and compares the data it reads there with what it would expect to see if a certain network card were installed at that location. However, there may be network cards it cannot detect automatically; this is sometimes the case with cheap network cards that are not-quite clones of standard cards from other manufacturers. Also, the kernel will normally attempt to detect only one network device when booting. If you're using more than one card, you have to tell the kernel about the other cards explicitly.

Another parameter that you might have to tell the kernel about is the interrupt request line. Hardware components usually interrupt the kernel when they need to be taken care of-for example, when data has arrived or a special condition occurs. In an ISA bus PC, interrupts may occur on one of 15 interrupt channels numbered 0, 1, and 3 through 15. The interrupt number assigned to a hardware component is called its interrupt request number (IRQ).[1]

As described in Chapter 2, the kernel accesses a piece of network hardware through a software construct called an interface. Interfaces offer an abstract set of functions that are the same across all types of hardware, such as sending or receiving a datagram.

Interfaces are identified by means of names. In many other Unix-like operating systems, the network interface is implemented as a special device file in the /dev/ directory. If you type the ls -las /dev/ command, you will see what these device files look like. In the file permissions (second) column you will see that device files begin with a letter rather than the hyphen seen for normal files. This character indicates the device type. The most common device types are b, which indicates the device is a block device and handles whole blocks of data with each read and write, and c, which indicates the device is a character device and handles data one character at a time. Where you would normally see the file length in the ls output, you instead see two numbers, called the major and minor device numbers. These numbers indicate the actual device with which the device file is associated.

Each device driver registers a unique major number with the kernel. Each instance of that device registers a unique minor number for that major device. The tty interfaces, /dev/tty*, are a character mode device indicated by the "c", and each have a major number of , but /dev/tty1 has a minor number of , and /dev/tty2 has a minor number of . Device files are very useful for many types of devices, but can be clumsy to use when trying to find an unused device to open.

Linux interface names are defined internally in the kernel and are not device files in the /dev directory. Some typical device names are listed later in Section 2." The assignment of interfaces to devices usually depends on the order in which devices are configured. For instance, the first Ethernet card installed will become eth0, and the next will be eth1. SLIP interfaces are handled differently from others because they are assigned dynamically. Whenever a SLIP connection is established, an interface is assigned to the serial port.

Figure 3-1 illustrates the relationship between the hardware, device drivers, and interfaces.

Figure 3-1. The relationship between drivers, interfaces, and hardware

When booting, the kernel displays the devices it detects and the interfaces it installs. The following is an excerpt from typical boot messages:

This example shows that the kernel has been compiled with TCP/IP enabled, and it includes drivers for SLIP, CSLIP, and PPP. The third line from the bottom says that a 3C509 Ethernet card was detected and installed as interface eth0. If you have some other type of network card-perhaps a D-Link pocket adaptor, for example-the kernel will usually print a line starting with its device name-dl0 in the D-Link example-followed by the type of card detected. If you have a network card installed but don't see any similar message, the kernel is unable to detect your card properly. This situation will be discussed later in the section "Ethernet Autoprobing."

Notes

1. Kernel Configuration

Most Linux distributions are supplied with boot disks that work for all common types of PC hardware. Generally, the supplied kernel is highly modularized and includes nearly every possible driver. This is a great idea for boot disks, but is probably not what you'd want for long-term use. There isn't much point in having drivers cluttering up your disk that you will never use. Therefore, you will generally roll your own kernel and include only those drivers you actually need or want; that way you save a little disk space and reduce the time it takes to compile a new kernel.

In any case, when running a Linux system, you should be familiar with building a kernel. Think of it as a right of passage, an affirmation of the one thing that makes free software as powerful as it is-you have the source. It isn't a case of, "I have to compile a kernel," rather it's a case of, "I can compile a kernel." The basics of compiling a Linux kernel are explained in Matt Welsh's book, Running Linux (O'Reilly). Therefore, we will discuss only configuration options that affect networking in this section.

One important point that does bear repeating here is the way the kernel version numbering scheme works. Linux kernels are numbered in the following format: . The first digit indicates the major version number. This digit changes when there are large and significant changes to the kernel design. For example, the kernel changed from major 1 to 2 when the kernel obtained support for machines other than Intel machines. The second number is the minor version number. In many respects, this number is the most important number to look at. The Linux development community has adopted a standard at which even minor version numbers indicate production, or stable, kernels and odd minor version numbers indicate development, or unstable, kernels. The stable kernels are what you should use on a machine that is important to you, as they have been more thoroughly tested. The development kernels are what you should use if you are interested in experimenting with the newest features of Linux, but they may have problems that haven't yet been found and fixed. The third number is simply incremented for each release of a minor version.[1]

When running make menuconfig, you are presented with a text-based menu that offers lists of configuration questions, such as whether you want kernel math emulation. One of these queries asks you whether you want TCP/IP networking support. You must answer this with y to get a kernel capable of networking.

1.1. Kernel Options in Linux 2.0 and Higher

After the general option section is complete, the configuration will go on to ask whether you want to include support for various features, such as SCSI drivers or sound cards. The prompt will indicate what options are available. You can press to obtain a description of what the option is actually offering. You'll always have the option of yes (y) to statically include the component in the kernel, or no (n) to exclude the component completely. You'll also see the module (m) option for those components that may be compiled as a run-time loadable module. Modules need to be loaded before they can be used, and are useful for drivers of components that you use infrequently.

The subsequent list of questions deal with networking support. The exact set of configuration options is in constant flux due to ongoing development. A typical list of options offered by most kernel versions around 2.0 and 2.1 looks like this:

You must answer this question with y if you want to use any type of networking devices, whether they are Ethernet, SLIP, PPP, or whatever. When you answer the question with y, support for Ethernet-type devices is enabled automatically. You must answer additional questions if you want to enable support for other types of network drivers:

These questions concern the various link layer protocols that Linux supports. Both PPP and SLIP allow you to transport IP datagrams across serial lines. PPP is actually a suite of protocols used to send network traffic across serial lines. Some of the protocols that form PPP manage the way that you authenticate yourself to the dial-in server, while others manage the way certain protocols are carried across the link-PPP is not limited to carrying TCP/IP datagrams; it may also carry other protocol such as IPX.

If you answer y or m to SLIP support, you will be prompted to answer the three questions that appear below it. The compressed header option provides support for CSLIP, a technique that compresses TCP/IP headers to as little as three bytes. Note that this kernel option does not turn on CSLIP automatically; it merely provides the necessary kernel functions for it. The Keepalive and linefill option causes the SLIP support to periodically generate activity on the SLIP line to avoid it being dropped by an inactivity timer. The Six bit SLIP encapsulation option allows you to run SLIP over lines and circuits that are not capable of transmitting the whole 8-bit data set cleanly. This is similar to the uuencoding or binhex technique used to send binary files by electronic mail.

PLIP provides a way to send IP datagrams across a parallel port connection. It is mostly used to communicate with PCs running DOS. On typical PC hardware, PLIP can be faster than PPP or SLIP, but it requires much more CPU overhead to perform, so while the transfer rate might be good, other tasks on the machine may be slow.

The following questions address network cards from various vendors. As more drivers are being developed, you are likely to see questions added to this section. If you want to build a kernel you can use on a number of different machines, or if your machine has more than one type of network card installed, you can enable more than one driver:

Finally, in the file system section, the configuration script will ask you whether you want support for NFS, the networking file system. NFS lets you export file systems to several hosts, which makes the files appear as if they were on an ordinary hard disk attached to the host:

We describe NFS in detail in Chapter 14.

1.2. Kernel Networking Options in Linux 2.0.0 and Higher

Linux 2.0.0 marked a significant change in Linux Networking. Many features were made a standard part of the Kernel, such as support for IPX. A number of options were also added and made configurable. Many of these options are used only in very special circumstances and we won't cover them in detail. The Networking HOWTO probably addresses what is not covered here. We'll list a number of useful options in this section, and explain when you'd want to use each one:

Basics

To use TCP/IP networking, you must answer this question with y. If you answer with n, however, you will still be able to compile the kernel with IPX support:

Gateways

You have to enable this option if your system acts as a gateway between two networks or between a LAN and a SLIP link, etc. It doesn't hurt to enable this by default, but you may want to disable it to configure a host as a so-called firewall. Firewalls are hosts that are connected to two or more networks, but don't route traffic between them. They're commonly used to provide users with Internet access at minimal risk to the internal network. Users are allowed to log in to the firewall and use Internet services, but the company's machines are protected from outside attacks because incoming connections can't cross the firewall (firewalls are covered in detail in Chapter 9 ):

Virtual hosting

These options together allow to you configure more than one IP address onto an interface. This is sometimes useful if you want to do "virtual hosting," through which a single machine can be configured to look and act as though it were actually many separate machines, each with its own network personality. We'll talk more about IP aliasing in a moment:

Accounting

This option enables you to collect data on the volume of IP traffic leaving and arriving at your machine (we cover this is detail in Chapter 10 ):

PC hug

This option works around an incompatibility with some versions of PC/TCP, a commercial TCP/IP implementation for DOS-based PCs. If you enable this option, you will still be able to communicate with normal Unix machines, but performance may be hurt over slow links:

Diskless booting

This function enables Reverse Address Resolution Protocol (RARP). RARP is used by diskless clients and X terminals to request their IP address when booting. You should enable RARP if you plan to serve this sort of client. A small program called rarp, included with the standard networking utilities, is used to add entries to the kernel RARP table:

MTU

When sending data over TCP, the kernel has to break up the stream into blocks of data to pass to IP. The size of the block is called the Maximum Transmission Unit, or MTU. For hosts that can be reached over a local network such as an Ethernet, it is typical to use an MTU as large as the maximum length of an Ethernet packet-1,500 bytes. When routing IP over a Wide Area Network like the Internet, it is preferable to use smaller-sized datagrams to ensure that they don't need to be further broken down along the route through a process called IP fragmentation.[2] The kernel is able to automatically determine the smallest MTU of an IP route and to automatically configure a TCP connection to use it. This behavior is on by default. If you answer y to this option this feature will be disabled.

If you do want to use smaller packet sizes for data sent to specific hosts (because, for example, the data goes through a SLIP link), you can do so using the mss option of the route command, which is briefly discussed at the end of this chapter:

Security feature

The IP protocol supports a feature called Source Routing. Source routing allows you to specify the route a datagram should follow by coding the route into the datagram itself. This was once probably useful before routing protocols such as RIP and OSPF became commonplace. But today it's considered a security threat because it can provide clever attackers with a way of circumventing certain types of firewall protection by bypassing the routing table of a router. You would normally want to filter out source routed datagrams, so this option is normally enabled:

Novell support

This option enables support for IPX, the transport protocol Novell Networking uses. Linux will function quite happily as an IPX router and this support is useful in environments where you have Novell fileservers. The NCP filesystem also requires IPX support enabled in your kernel; if you wish to attach to and mount your Novell filesystems you must have this option enabled (we'll dicuss IPX and the NCP filesystem in Chapter 15):

Amateur radio

These three options select support for the three Amateur Radio protocols supported by Linux: AX.25, NetRom and Rose (we don't describe them in this book, but they are covered in detail in the AX25 HOWTO):

Linux supports another driver type: the dummy driver. The following question appears toward the start of the device-driver section:

The dummy driver doesn't really do much, but it is quite useful on standalone or PPP/SLIP hosts. It is basically a masqueraded loopback interface. On hosts that offer PPP/SLIP but have no other network interface, you want to have an interface that bears your IP address all the time. This is discussed in a little more detail in Section 5.7.7' in Chapter 5. Note that today you can achieve the same result by using the IP alias feature and configuring your IP address as an alias on the loopback interface.

Notes

2. A Tour of Linux Network Devices

The Linux kernel supports a number of hardware drivers for various types of equipment. This section gives a short overview of the driver families available and the interface names they use.

There is a number of standard names for interfaces in Linux, which are listed here. Most drivers support more than one interface, in which case the interfaces are numbered, as in eth0 and eth1

lo

This is the local loopback interface. It is used for testing purposes, as well as a couple of network applications. It works like a closed circuit in that any datagram written to it will immediately be returned to the host's networking layer. There's always one loopback device present in the kernel, and there's little sense in having more.

eth0 eth1

These are the Ethernet card interfaces. They are used for most Ethernet cards, including many of the parallel port Ethernet cards.

tr0 tr1

These are the Token Ring card interfaces. They are used for most Token Ring cards, including non-IBM manufactured cards.

sl0 sl1

These are the SLIP interfaces. SLIP interfaces are associated with serial lines in the order in which they are allocated for SLIP.

ppp0 ppp1

These are the PPP interfaces. Just like SLIP interfaces, a PPP interface is associated with a serial line once it is converted to PPP mode.

plip0 plip1

These are the PLIP interfaces. PLIP transports IP datagrams over parallel lines. The interfaces are allocated by the PLIP driver at system boot time and are mapped onto parallel ports. In the 2.0.x kernels there is a direct relationship between the device name and the I/O port of the parallel port, but in later kernels the device names are allocated sequentially, just as for SLIP and PPP devices.

ax0 ax1

These are the AX.25 interfaces. AX.25 is the primary protocol used by amateur radio operators. AX.25 interfaces are allocated and mapped in a similar fashion to SLIP devices.

There are many other types of interfaces available for other network drivers. We've listed only the most common ones.

During the next few sections, we will discuss the details of using the drivers described previously. The Networking HOWTO provides details on how to configure most of the others, and the AX25 HOWTO explains how to configure the Amateur Radio network devices.

Ethernet Installation

The current Linux network code supports a large variety of Ethernet cards. Most drivers were written by Donald Becker, who authored a family of drivers for cards based on the National Semiconductor 8390 chip; these have become known as the Becker Series Drivers. Many other developers have contributed drivers, and today there are few common Ethernet cards that aren't supported by Linux. The list of supported Ethernet cards is growing all the time, so if your card isn't supported yet, chances are it will be soon.

Sometime earlier in Linux's history we would have attempted to list all supported Ethernet cards, but that would now take too much time and space. Fortunately, Paul Gortmaker maintains the Ethernet HOWTO, which lists each of the supported cards and provides useful information about getting each of them running under Linux.[1] It is posted monthly to the comp.os.linux.answers newsgroup, and is also available on any of the Linux Documentation Project mirror sites.

Even if you are confident you know how to install a particular type of Ethernet card in your machine, it is often worthwhile taking a look at what the Ethernet HOWTO has to say about it. You will find information that extends beyond simple configuration issues. For example, it could save you a lot of headaches to know the behavior of some DMA-based Ethernet cards that use the same DMA channel as the Adaptec 1542 SCSI controller by default. Unless you move one of them to a different DMA channel, you will wind up with the Ethernet card writing packet data to arbitrary locations on your hard disk.

To use any of the supported Ethernet cards with Linux, you may use a precompiled kernel from one of the major Linux distributions. These generally have modules available for all of the supported drivers, and the installation process usually allows you to select which drivers you want loaded. In the long term, however, it's better to build your own kernel and compile only those drivers you actually need; this saves disk space and memory.

1. Ethernet Autoprobing

Many of the Linux Ethernet drivers are smart enough to know how to search for the location of your Ethernet card. This saves you having to tell the kernel where it is manually. The Ethernet HOWTO lists whether a particular driver uses autoprobing and in which order it searches the I/O address for the card.

There are three limitations to the autoprobing code. First, it may not recognize all cards properly. This is especially true for some of the cheaper clones of common cards. Second, the kernel won't autoprobe for more than one card unless specifically instructed. This was a conscious design decision, as it is assumed you will want to have control over which card is assigned to which interface. The best way to do this reliably is to manually configure the Ethernet cards in your machine. Third, the driver may not probe at the address that your card is configured for. Generally speaking, the drivers will autoprobe at the addresses that the particular device is capable of being configured for, but sometimes certain addresses are ignored to avoid hardware conflicts with other types of cards that commonly use that same address.

PCI network cards should be reliably detected. But if you are using more than one card, or if the autoprobe should fail to detect your card, you have a way to explicitly tell the kernel about the card's base address and name.

At boot time you can supply arguments and information to the kernel that any of the kernel components may read. This mechanism allows you to pass information to the kernel that Ethernet drivers can use to locate your Ethernet hardware without making the driver probe.

If you use lilo to boot your system, you can pass parameters to the kernel by specifying them through the append option in the lilo.conf file. To inform the kernel about an Ethernet device, you can pass the following parameters:

The first four parameters are numeric, while the last is the device name. The irq, base_addr, and name parameters are required, but the two param parameters are optional. Any of the numeric values may be set to zero, which causes the kernel to determine the value by probing.

The first parameter sets the IRQ assigned to the device. By default, the kernel will try to autodetect the device's IRQ channel. The 3c503 driver, for example, has a special feature that selects a free IRQ from the list 5, 9, 3, 4 and configures the card to use this line. The base_addr parameter gives the I/O base address of the card; a value of zero tells the kernel to probe the addresses listed above.

Different drivers use the next two parameters differently. For shared-memory cards, such as the WD80x3, they specify starting and ending addresses of the shared memory area. Other cards commonly use param1 to set the level at which debugging information is displayed. Values of 1 through 7 denote increasing levels of verbosity, while 8 turns them off altogether; 0 denotes the default. The 3c503 driver uses param2 to choose between the internal transceiver (default) or an external transceiver (a value of 1). The former uses the card's BNC connector; the latter uses its AUI port. The param arguments need not be included at all if you don't have anything special to configure.

The first non-numeric argument is interpreted by the kernel as the device name. You must specify a device name for each Ethernet card you describe.

If you have two Ethernet cards, you can have Linux autodetect one card and pass the second card's parameters with lilo, but you'll probably want to manually configure both cards. If you decide to have the kernel probe for one and manually configure the second, you must make sure the kernel doesn't accidentally find the second card first, or else the other one won't be registered at all. You do this by passing lilo a reserve option, which explicitly tells the kernel to avoid probing the I/O space taken up by the second card. For instance, to make Linux install a second Ethernet card at 0x300 as eth1, you would pass the following parameters to the kernel:

The reserve option makes sure no driver accesses the second card's I/O space when probing for some device. You may also use the kernel parameters to override autoprobing for eth0 :

You can turn off autoprobing altogether. You might do this, for example, to stop a kernel probing for an Ethernet card you might have temporarily removed. Disabling autoprobing is as simple as specifying a base_addr argument of -1:

To supply these parameters to the kernel at boot time, you enter the parameters at the lilo 'boot:' prompt. To have lilo give you the 'boot:' at the prompt, you must press any one of the Control, Alt or Shift keys while lilo is booting. If you press the Tab key at the prompt, you will be presented with a list of kernels that you may boot. To boot a kernel with parameters supplied, enter the name of the kernel you wish to boot, followed by a space, then followed by the parameters you wish to supply. When you press the Enter key, lilo will load that kernel and boot it with the parameters you've supplied.

To make this change occur automatically on each reboot, enter the parameters into the /etc/lilo.conf using the append= keyword. An example might look like this:

After you've edited lilo.conf, you must rerun the lilo command to activate the change.

4. The PLIP Driver

Parallel Line IP (PLIP) is a cheap way to network when you want to connect only two machines. It uses a parallel port and a special cable, achieving speeds of 10 kilobytes per second to 20 kilobytes per second.

PLIP was originally developed by Crynwr, Inc. Its design at the time was rather ingenious (or, if you prefer, a hack), because the original parallel ports on IBM PCs were designed to spend their time being unidirectional printer ports; the eight data lines could be used only to send data from the PC to the peripheral device, but not the other way around. The Cyrnwr PLIP design worked around this limitation by using the port's five status lines for input, which limited it to transferring all data as nibbles (half bytes) only, but allowed for bidirectional transfer. This mode of operation was called PLIP "mode 0." Today, the parallel ports supplied on PC hardware cater to full bidirectional 8-bit data transfer, and PLIP has been extended to accomodate this with the addition of PLIP "mode 1."

Linux kernels up to and including Version 2.0 support PLIP mode 0 only, and an enhanced parallel port driver exists as a patch against the 2.0 kernel and as a standard part of the 2.2 kernel code to provide PLIP mode 1 operation, too. Unlike earlier versions of the PLIP code, the driver now attempts to be compatible with the PLIP implementations from Crynwr, as well as the PLIP driver in NCSA telnet. To connect two machines using PLIP, you need a special cable sold at some shops as a Null Printer or Turbo Laplink cable. You can, however, make one yourself fairly easily; Appendix B shows you how.

The PLIP driver for Linux is the work of almost countless persons. It is currently maintained by Niibe Yutaka. If compiled into the kernel, it sets up a network interface for each of the possible printer ports, with plip0 corresponding to parallel port lp0 plip1 corresponding to lp1, etc. The mapping of interfaces to ports differs in the 2.0 kernels and the 2.2 kernels. In the 2.0 kernels, the mapping was hardwired in the drivers/net/Spacd.c file in the kernel source. The default mappings in this file are:

If you configured your printer port in a different way, you must change these values in drivers/net/Space.c in the Linux kernel source and build a new kernel.

In the 2.2 kernels, the PLIP driver uses the "parport" parallel port sharing driver developed by Philip Blundell. The new driver allocates the PLIP network device names serially, just as for the Ethernet or PPP drivers, so the first PLIP device created is plip0, the second is plip1, and so on. The physical parallel port hardware is also allocated serially. By default, the parallel port driver will attempt to detect your parallel port hardware with an autoprobe routine, recording the physical device information in the order found. It is better practice to explicitly tell the kernel the physical I/O parameters. You can do this by supplying arguments to the parport_pc.o module as you load it, or if you have compiled the driver into your kernel, using lilo to supply arguments to the kernel at boot time. The IRQ setting of any device may be changed later by writing the new IRQ value to the related /proc/parport/*/irq file.

Configuring the physical I/O parameters in a 2.2 kernel when loading the module is straightforward. For instance, to tell the driver that you have two PC-style parallel ports at I/O addresses 0x278 and 0c378 and IRQs 5 and 7, respectively, you would load the module with the following arguments:

The corresponding arguments to pass to the kernel for a compiled-in driver are:

You would use the lilo append keyword to have these arguments passed to the kernel automatically at boot time.

When the PLIP driver is initialized, either at boot time if it is built-in, or when the plip.o module is loaded, each of the parallel ports will have a plip network device associated with it. plip0 will be assigned to the first parallel port device, plip1 the second, and so on. You can manually override this automatic assignment using another set of kernel arguments. For instance, to assign parport0 to network device plip0, and parport1 to network device plip1, you would use kernel arguments of:

This mapping does not mean, however, that you cannot use these parallel ports for printing or other purposes. The physical parallel port devices are used by the PLIP driver only when the corresponding interface is configured up.

5. The PPP and SLIP Drivers

Point-to-Point Protocol (PPP) and Serial Line IP (SLIP) are widely used protocols for carrying IP packets over a serial link. A number of institutions offer dialup PPP and SLIP access to machines that are on the Internet, thus providing IP connectivity to private persons (something that's otherwise hardly affordable).

No hardware modifications are necessary to run PPP or SLIP; you can use any serial port. Since serial port configuration is not specific to TCP/IP networking, we have devoted a separate chapter to this. Please refer to Chapter 4, for more information. We cover PPP in detail in Chapter 8, and SLIP in Chapter 7.

6. Other Network Types

Most other network types are configured similarly to Ethernet. The arguments passed to the loadable modules will be different and some drivers may not support more than one card, but just about everything else is the same. Documentation for these cards is generally available in the /usr/src/linux /Documentation/networking/ directory of the Linux kernel source.