1.2. TCP/IP Networks
Modern networking applications require a sophisticated approach to carrying data from one machine to another. If you are managing a Linux machine that has many users, each of whom may wish to simultaneously connect to remote hosts on a network, you need a way of allowing them to share your network connection without interfering with each other. The approach that a large number of modern networking protocols uses is called packet-switching. A packet is a small chunk of data that is transferred from one machine to another across the network. The switching occurs as the datagram is carried across each link in the network. A packet-switched network shares a single network link among many users by alternately sending packets from one user to another across that link.
The solution that Unix systems, and subsequently many non-Unix systems, have adopted is known as TCP/IP. When talking about TCP/IP networks you will hear the term datagram, which technically has a special meaning but is often used interchangeably with packet. In this section, we will have a look at underlying concepts of the TCP/IP protocols.
1.2.1. Introduction to TCP/IP Networks
TCP/IP traces its origins to a research project funded by the United States Defense Advanced Research Projects Agency (DARPA) in 1969. The ARPANET was an experimental network that was converted into an operational one in 1975 after it had proven to be a success.
In 1983, the new protocol suite TCP/IP was adopted as a standard, and all hosts on the network were required to use it. When ARPANET finally grew into the Internet (with ARPANET itself passing out of existence in 1990), the use of TCP/IP had spread to networks beyond the Internet itself. Many companies have now built corporate TCP/IP networks, and the Internet has grown to a point at which it could almost be considered a mainstream consumer technology. It is difficult to read a newspaper or magazine now without seeing reference to the Internet; almost everyone can now use it.
For something concrete to look at as we discuss TCP/IP throughout the following sections, we will consider Groucho Marx University (GMU), situated somewhere in Fredland, as an example. Most departments run their own Local Area Networks, while some share one and others run several of them. They are all interconnected and hooked to the Internet through a single high-speed link.
Suppose your Linux box is connected to a LAN of Unix hosts at the Mathematics department, and its name is erdos. To access a host at the Physics department, say quark, you enter the following command:
$ rlogin quark.physics Welcome to the Physics Department at GMU (ttyq2) login: |
At the prompt, you enter your login name, say andres, and your password. You are then given a shell[1] on quark, to which you can type as if you were sitting at the system's console. After you exit the shell, you are returned to your own machine's prompt. You have just used one of the instantaneous, interactive applications that TCP/IP provides: remote login.
While being logged into quark, you might also want to run a graphical user interface application, like a word processing program, a graphics drawing program, or even a World Wide Web browser. The X windows system is a fully network-aware graphical user environment, and it is available for many different computing systems. To tell this application that you want to have its windows displayed on your host's screen, you have to set the DISPLAY environment variable:
$ DISPLAY=erdos.maths:0.0 $ export DISPLAY |
If you now start your application, it will contact your X server instead of quark's, and display all its windows on your screen. Of course, this requires that you have X11 runnning on erdos. The point here is that TCP/IP allows quark and erdos to send X11 packets back and forth to give you the illusion that you're on a single system. The network is almost transparent here.
Another very important application in TCP/IP networks is NFS, which stands for Network File System. It is another form of making the network transparent, because it basically allows you to treat directory hierarchies from other hosts as if they were local file systems and look like any other directories on your host. For example, all users' home directories can be kept on a central server machine from which all other hosts on the LAN mount them. The effect is that users can log in to any machine and find themselves in the same home directory. Similarly, it is possible to share large amounts of data (such as a database, documentation or application programs) among many hosts by maintaining one copy of the data on a server and allowing other hosts to access it. We will come back to NFS in Chapter 14.
Of course, these are only examples of what you can do with TCP/IP networks. The possibilities are almost limitless, and we'll introduce you to more as you read on through the book.
We will now have a closer look at the way TCP/IP works. This information will help you understand how and why you have to configure your machine. We will start by examining the hardware, and slowly work our way up.
1.2.2. Ethernets
The most common type of LAN hardware is known as Ethernet. In its simplest form, it consists of a single cable with hosts attached to it through connectors, taps, or transceivers. Simple Ethernets are relatively inexpensive to install, which together with a net transfer rate of 10, 100, or even 1,000 Megabits per second, accounts for much of its popularity.
Ethernets come in three flavors: thick, thin, and twisted pair. Thin and thick Ethernet each use a coaxial cable, differing in diameter and the way you may attach a host to this cable. Thin Ethernet uses a T-shaped “BNC” connector, which you insert into the cable and twist onto a plug on the back of your computer. Thick Ethernet requires that you drill a small hole into the cable, and attach a transceiver using a “vampire tap.” One or more hosts can then be connected to the transceiver. Thin and thick Ethernet cable can run for a maximum of 200 and 500 meters respectively, and are also called 10base-2 and 10base-5. The “base” refers to “baseband modulation” and simply means that the data is directly fed onto the cable without any modem. The number at the start refers to the speed in Megabits per second, and the number at the end is the maximum length of the cable in hundreds of metres. Twisted pair uses a cable made of two pairs of copper wires and usually requires additional hardware known as active hubs. Twisted pair is also known as 10base-T, the “T” meaning twisted pair. The 100 Megabits per second version is known as 100base-T.
To add a host to a thin Ethernet installation, you have to disrupt network service for at least a few minutes because you have to cut the cable to insert the connector. Although adding a host to a thick Ethernet system is a little complicated, it does not typically bring down the network. Twisted pair Ethernet is even simpler. It uses a device called a “hub,” which serves as an interconnection point. You can insert and remove hosts from a hub without interrupting any other users at all.
Many people prefer thin Ethernet for small networks because it is very inexpensive; PC cards come for as little as US $30 (many companies are literally throwing them out now), and cable is in the range of a few cents per meter. However, for large-scale installations, either thick Ethernet or twisted pair is more appropriate. For example, the Ethernet at GMU's Mathematics Department originally chose thick Ethernet because it is a long route that the cable must take so traffic will not be disrupted each time a host is added to the network. Twisted pair installations are now very common in a variety of installations. The Hub hardware is dropping in price and small units are now available at a price that is attractive to even small domestic networks. Twisted pair cabling can be significantly cheaper for large installations, and the cable itself is much more flexible than the coaxial cables used for the other Ethernet systems. The network administrators in GMU's mathematics department are planning to replace the existing network with a twisted pair network in the coming finanical year because it will bring them up to date with current technology and will save them significant time when installing new host computers and moving existing computers around.
One of the drawbacks of Ethernet technology is its limited cable length, which precludes any use of it other than for LANs. However, several Ethernet segments can be linked to one another using repeaters, bridges, or routers. Repeaters simply copy the signals between two or more segments so that all segments together will act as if they are one Ethernet. Due to timing requirements, there may not be more than four repeaters between any two hosts on the network. Bridges and routers are more sophisticated. They analyze incoming data and forward it only when the recipient host is not on the local Ethernet.
Ethernet works like a bus system, where a host may send packets (or frames) of up to 1,500 bytes to another host on the same Ethernet. A host is addressed by a six-byte address hardcoded into the firmware of its Ethernet network interface card (NIC). These addresses are usually written as a sequence of two-digit hex numbers separated by colons, as in aa:bb:cc:dd:ee:ff.
A frame sent by one station is seen by all attached stations, but only the destination host actually picks it up and processes it. If two stations try to send at the same time, a collision occurs. Collisions on an Ethernet are detected very quickly by the electronics of the interface cards and are resolved by the two stations aborting the send, each waiting a random interval and re-attempting the transmission. You'll hear lots of stories about collisions on Ethernet being a problem and that utilization of Ethernets is only about 30 percent of the available bandwidth because of them. Collisions on Ethernet are a normal phenomenon, and on a very busy Ethernet network you shouldn't be surprised to see collision rates of up to about 30 percent. Utilization of Ethernet networks is more realistically limited to about 60 percent before you need to start worrying about it.[2]
1.2.3. Other Types of Hardware
In larger installations, such as Groucho Marx University, Ethernet is usually not the only type of equipment used. There are many other data communications protocols available and in use. All of the protocols listed are supported by Linux, but due to space constraints we'll describe them briefly. Many of the protocols have HOWTO documents that describe them in detail, so you should refer to those if you're interested in exploring those that we don't describe in this book.
At Groucho Marx University, each department's LAN is linked to the campus high-speed “backbone” network, which is a fiber optic cable running a network technology called Fiber Distributed Data Interface (FDDI). FDDI uses an entirely different approach to transmitting data, which basically involves sending around a number of tokens, with a station being allowed to send a frame only if it captures a token. The main advantage of a token-passing protocol is a reduction in collisions. Therefore, the protocol can more easily attain the full speed of the transmission medium, up to 100 Mbps in the case of FDDI. FDDI, being based on optical fiber, offers a significant advantage because its maximum cable length is much greater than wire-based technologies. It has limits of up to around 200 km, which makes it ideal for linking many buildings in a city, or as in GMU's case, many buildings on a campus.
Similarly, if there is any IBM computing equipment around, an IBM Token Ring network is quite likely to be installed. Token Ring is used as an alternative to Ethernet in some LAN environments, and offers the same sorts of advantages as FDDI in terms of achieving full wire speed, but at lower speeds (4 Mbps or 16 Mbps), and lower cost because it is based on wire rather than fiber. In Linux, Token Ring networking is configured in almost precisely the same way as Ethernet, so we don't cover it specifically.
Although it is much less likely today than in the past, other LAN technologies, such as ArcNet and DECNet, might be installed. Linux supports these too, but we don't cover them here.
Many national networks operated by Telecommunications companies support packet switching protocols. Probably the most popular of these is a standard named X.25. Many Public Data Networks, like Tymnet in the U.S., Austpac in Australia, and Datex-P in Germany offer this service. X.25 defines a set of networking protocols that describes how data terminal equipment, such as a host, communicates with data communications equipment (an X.25 switch). X.25 requires a synchronous data link, and therefore special synchronous serial port hardware. It is possible to use X.25 with normal serial ports if you use a special device called a PAD (Packet Assembler Disassembler). The PAD is a standalone device that provides asynchronous serial ports and a synchronous serial port. It manages the X.25 protocol so that simple terminal devices can make and accept X.25 connections. X.25 is often used to carry other network protocols, such as TCP/IP. Since IP datagrams cannot simply be mapped onto X.25 (or vice versa), they are encapsulated in X.25 packets and sent over the network. There is an experimental implementation of the X.25 protocol available for Linux.
A more recent protocol commonly offered by telecommunications companies is called Frame Relay. The Frame Relay protocol shares a number of technical features with the X.25 protocol, but is much more like the IP protocol in behavior. Like X.25, Frame Relay requires special synchronous serial hardware. Because of their similarities, many cards support both of these protocols. An alternative is available that requires no special internal hardware, again relying on an external device called a Frame Relay Access Device (FRAD) to manage the encapsulation of Ethernet packets into Frame Relay packets for transmission across a network. Frame Relay is ideal for carrying TCP/IP between sites. Linux provides drivers that support some types of internal Frame Relay devices.
If you need higher speed networking that can carry many different types of data, such as digitized voice and video, alongside your usual data, ATM (Asynchronous Transfer Mode) is probably what you'll be interested in. ATM is a new network technology that has been specifically designed to provide a manageable, high-speed, low-latency means of carrying data, and provide control over the Quality of Service (Q.S.). Many telecommunications companies are deploying ATM network infrastructure because it allows the convergence of a number of different network services into one platform, in the hope of achieving savings in management and support costs. ATM is often used to carry TCP/IP. The Networking-HOWTO offers information on the Linux support available for ATM.
Frequently, radio amateurs use their radio equipment to network their computers; this is commonly called packet radio. One of the protocols used by amateur radio operators is called AX.25 and is loosely derived from X.25. Amateur radio operators use the AX.25 protocol to carry TCP/IP and other protocols, too. AX.25, like X.25, requires serial hardware capable of synchronous operation, or an external device called a “Terminal Node Controller” to convert packets transmitted via an asynchronous serial link into packets transmitted synchronously. There are a variety of different sorts of interface cards available to support packet radio operation; these cards are generally referred to as being “Z8530 SCC based,” and are named after the most popular type of communications controller used in the designs. Two of the other protocols that are commonly carried by AX.25 are the NetRom and Rose protocols, which are network layer protocols. Since these protocols run over AX.25, they have the same hardware requirements. Linux supports a fully featured implementation of the AX.25, NetRom, and Rose protocols. The AX25-HOWTO is a good source of information on the Linux implementation of these protocols.
Other types of Internet access involve dialing up a central system over slow but cheap serial lines (telephone, ISDN, and so on). These require yet another protocol for transmission of packets, such as SLIP or PPP, which will be described later.
1.2.4. The Internet Protocol
Of course, you wouldn't want your networking to be limited to one Ethernet or one point-to-point data link. Ideally, you would want to be able to communicate with a host computer regardless of what type of physical network it is connected to. For example, in larger installations such as Groucho Marx University, you usually have a number of separate networks that have to be connected in some way. At GMU, the Math department runs two Ethernets: one with fast machines for professors and graduates, and another with slow machines for students. Both are linked to the FDDI campus backbone network.
This connection is handled by a dedicated host called a gateway that handles incoming and outgoing packets by copying them between the two Ethernets and the FDDI fiber optic cable. For example, if you are at the Math department and want to access quark on the Physics department's LAN from your Linux box, the networking software will not send packets to quark directly because it is not on the same Ethernet. Therefore, it has to rely on the gateway to act as a forwarder. The gateway (named sophus) then forwards these packets to its peer gateway niels at the Physics department, using the backbone network, with niels delivering it to the destination machine. Data flow between erdos and quark is shown in Figure 1-1.
This scheme of directing data to a remote host is called routing, and packets are often referred to as datagrams in this context. To facilitate things, datagram exchange is governed by a single protocol that is independent of the hardware used: IP, or Internet Protocol. In Chapter 2, we will cover IP and the issues of routing in greater detail.
The main benefit of IP is that it turns physically dissimilar networks into one apparently homogeneous network. This is called internetworking, and the resulting “meta-network” is called an internet. Note the subtle difference here between an internet and the Internet. The latter is the official name of one particular global internet.
Of course, IP also requires a hardware-independent addressing scheme. This is achieved by assigning each host a unique 32-bit number called the IP address. An IP address is usually written as four decimal numbers, one for each 8-bit portion, separated by dots. For example, quark might have an IP address of 0x954C0C04, which would be written as 149.76.12.4. This format is also called dotted decimal notation and sometimes dotted quad notation. It is increasingly going under the name IPv4 (for Internet Protocol, Version 4) because a new standard called IPv6 offers much more flexible addressing, as well as other modern features. It will be at least a year after the release of this edition before IPv6 is in use.
You will notice that we now have three different types of addresses: first there is the host's name, like quark, then there are IP addresses, and finally, there are hardware addresses, like the 6-byte Ethernet address. All these addresses somehow have to match so that when you type rlogin quark, the networking software can be given quark's IP address; and when IP delivers any data to the Physics department's Ethernet, it somehow has to find out what Ethernet address corresponds to the IP address.
We will deal with these situations in Chapter 2. For now, it's enough to remember that these steps of finding addresses are called hostname resolution, for mapping hostnames onto IP addresses, and address resolution, for mapping the latter to hardware addresses.
1.2.5. IP Over Serial Lines
On serial lines, a “de facto” standard exists known as SLIP, or Serial Line IP. A modification of SLIP known as CSLIP, or Compressed SLIP, performs compression of IP headers to make better use of the relatively low bandwidth provided by most serial links. Another serial protocol is PPP, or the Point-to-Point Protocol. PPP is more modern than SLIP and includes a number of features that make it more attractive. Its main advantage over SLIP is that it isn't limited to transporting IP datagrams, but is designed to allow just about any protocol to be carried across it.
1.2.6. The Transmission Control Protocol
Sending datagrams from one host to another is not the whole story. If you log in to quark, you want to have a reliable connection between your rlogin process on erdos and the shell process on quark. Thus, the information sent to and fro must be split up into packets by the sender and reassembled into a character stream by the receiver. Trivial as it seems, this involves a number of complicated tasks.
A very important thing to know about IP is that, by intent, it is not reliable. Assume that ten people on your Ethernet started downloading the latest release of Netscape's web browser source code from GMU's FTP server. The amount of traffic generated might be too much for the gateway to handle, because it's too slow and it's tight on memory. Now if you happen to send a packet to quark, sophus might be out of buffer space for a moment and therefore unable to forward it. IP solves this problem by simply discarding it. The packet is irrevocably lost. It is therefore the responsibility of the communicating hosts to check the integrity and completeness of the data and retransmit it in case of error.
This process is performed by yet another protocol, Transmission Control Protocol (TCP), which builds a reliable service on top of IP. The essential property of TCP is that it uses IP to give you the illusion of a simple connection between the two processes on your host and the remote machine, so you don't have to care about how and along which route your data actually travels. A TCP connection works essentially like a two-way pipe that both processes may write to and read from. Think of it as a telephone conversation.
TCP identifies the end points of such a connection by the IP addresses of the two hosts involved and the number of a port on each host. Ports may be viewed as attachment points for network connections. If we are to strain the telephone example a little more, and you imagine that cities are like hosts, one might compare IP addresses to area codes (where numbers map to cities), and port numbers to local codes (where numbers map to individual people's telephones). An individual host may support many different services, each distinguished by its own port number.
In the rlogin example, the client application (rlogin) opens a port on erdos and connects to port 513 on quark, to which the rlogind server is known to listen. This action establishes a TCP connection. Using this connection, rlogind performs the authorization procedure and then spawns the shell. The shell's standard input and output are redirected to the TCP connection, so that anything you type to rlogin on your machine will be passed through the TCP stream and be given to the shell as standard input.
1.2.7. The User Datagram Protocol
Of course, TCP isn't the only user protocol in TCP/IP networking. Although suitable for applications like rlogin, the overhead involved is prohibitive for applications like NFS, which instead uses a sibling protocol of TCP called UDP, or User Datagram Protocol. Just like TCP, UDP allows an application to contact a service on a certain port of the remote machine, but it doesn't establish a connection for this. Instead, you use it to send single packets to the destination service—hence its name.
Assume you want to request a small amount of data from a database server. It takes at least three datagrams to establish a TCP connection, another three to send and confirm a small amount of data each way, and another three to close the connection. UDP provides us with a means of using only two datagrams to achieve almost the same result. UDP is said to be connectionless, and it doesn't require us to establish and close a session. We simply put our data into a datagram and send it to the server; the server formulates its reply, puts the data into a datagram addressed back to us, and transmits it back. While this is both faster and more efficient than TCP for simple transactions, UDP was not designed to deal with datagram loss. It is up to the application, a name server for example, to take care of this.
1.2.8. More on Ports
Ports may be viewed as attachment points for network connections. If an application wants to offer a certain service, it attaches itself to a port and waits for clients (this is also called listening on the port). A client who wants to use this service allocates a port on its local host and connects to the server's port on the remote host. The same port may be open on many different machines, but on each machine only one process can open a port at any one time.
An important property of ports is that once a connection has been established between the client and the server, another copy of the server may attach to the server port and listen for more clients. This property permits, for instance, several concurrent remote logins to the same host, all using the same port 513. TCP is able to tell these connections from one another because they all come from different ports or hosts. For example, if you log in twice to quark from erdos, the first rlogin client will use the local port 1023, and the second one will use port 1022. Both, however, will connect to the same port 513 on quark. The two connections will be distinguished by use of the port numbers used at erdos.
This example shows the use of ports as rendezvous points, where a client contacts a specific port to obtain a specific service. In order for a client to know the proper port number, an agreement has to be reached between the administrators of both systems on the assignment of these numbers. For services that are widely used, such as rlogin, these numbers have to be administered centrally. This is done by the IETF (Internet Engineering Task Force), which regularly releases an RFC titled Assigned Numbers (RFC-1700). It describes, among other things, the port numbers assigned to well-known services. Linux uses a file called /etc/services that maps service names to numbers.
It is worth noting that although both TCP and UDP connections rely on ports, these numbers do not conflict. This means that TCP port 513, for example, is different from UDP port 513. In fact, these ports serve as access points for two different services, namely rlogin (TCP) and rwho (UDP).
1.2.9. The Socket Library
In Unix operating systems, the software performing all the tasks and protocols described above is usually part of the kernel, and so it is in Linux. The programming interface most common in the Unix world is the Berkeley Socket Library. Its name derives from a popular analogy that views ports as sockets and connecting to a port as plugging in. It provides the bind call to specify a remote host, a transport protocol, and a service that a program can connect or listen to (using connect, listen, and accept). The socket library is somewhat more general in that it provides not only a class of TCP/IP-based sockets (the AF_INET sockets), but also a class that handles connections local to the machine (the AF_UNIX class). Some implementations can also handle other classes, like the XNS (Xerox Networking System) protocol or X.25.
In Linux, the socket library is part of the standard libc C library. It supports the AF_INET and AF_INET6 sockets for TCP/IP and AF_UNIX for Unix domain sockets. It also supports AF_IPX for Novell's network protocols, AF_X25 for the X.25 network protocol, AF_ATMPVC and AF_ATMSVC for the ATM network protocol and AF_AX25, AF_NETROM, and AF_ROSE sockets for Amateur Radio protocol support. Other protocol families are being developed and will be added in time.