Praised by reviewers and practicing TCP/IP programmers alike, the TCP/IP Illustrated series examines the many facets of the TCP/IP protocol suite using a unique and highly-effective visual approach that describes the inner workings of TCP/IP with detail, insight, and clarity.
Volume 3 provides detailed coverage of four essential topics with which today's TCP/IP programmers and network administrators must be thoroughly familiar:
As in the previous two volumes, the book is filled with examples and implementation details within the 4.4BSD-Lite networking code.
The TCP/IP Illustrated series provides a complete picture of the protocol suite that drives the Internet, and gives programmers, system administrators, and serious users the information, understanding, and skills they need to remain at the forefront of networking technology.
(Most chapters open with "Introduction" and conclude with "Summary".)
I. TCP FOR TRANSACTIONS.1. T/TCP Introduction.
Summary.2. T/TCP Protocol.
New TCP Options for T/TCP.
T/TCP Implementation Variables.
State Transition Diagram.
T/TCP Extended States.
Summary.3. T/TCP Examples.
Normal T/TCP Transaction.
Server Receives Old Duplicate SYN.
Request or Reply Exceeds MSS.
Summary.4. T/TCP Protocol (Continued).
Client Port Numbers and TIME_WAIT State.
Purpose of the TIME_WAIT State.
TIME_WAIT State Truncation.
Avoiding the Three-Way Handshake with TAO.
Summary.5. T/TCP Implementation: Socket Layer.
Summary.6. T/TCP Implementation: Routing Table.
Summary.7. T/TCP Implementation: Protocol Control Blocks
Summary.8. T/TCP Implementation: TCP Overview
TCP protosw Structure.
TCP Control Block.
Summary.9. T/TCP Implementation: TCP Output
Summary.10. T/TCP Implementation: TCP Functions
Retransmission Timeout Calculations.
Summary.11. T/TCP Implementation: TCP Input
Initiation of Passive Open.
Initiation of Active Open.
PAWS: Protection Against Wrapped Sequence Numbers.
Completion of Passive Opens and Simultaneous Opens.
ACK Processing (Continued).
Summary.12. T/TCP Implementation: TCP User Requests.
PRU_SEND and PRU_SEND_EOF Requests.
II. ADDITIONAL TCP APPLICATIONS.13. HTTP: Hypertext Transfer Protocol.
Introduction to HTTP and HTML.
Summary.14. Packets Found on an HTTP Server.
Multiple HTTP Servers.
Client SYN Interarrival Time.
listen Backlog Queue.
Client SYN Options.
Client SYN Retransmissions.
Timing Out Persist Probes.
Simulation of T/TCP Routing Table Size.
TCP PCB Cache and Header Prediction.
Summary.15. NNTP: Network News Transfer Protocol.
A Simple News Client.
A More Sophisticated News Client.
III. THE UNIX DOMAIN PROTOCOLS.16. Unix Domain Protocols: Introduction.
Summary.17. Unix Domain Protocols: Implementation.
Unix domain and protosw Structures.
Unix Domain Socket Address Structures.
Unix Domain Protocol Control Blocks.
PRU_ATTACH Request and unp_attach Function.
PRU_DETACH Request and unp_detach Function.
PRU_BIND Request and unp_bind Function.
PRU_CONNECT Request and unp_connect Function.
PRU_CONNECT2 Request and unp_connect2 Function.
socketpair System Call.
pipe System Call.
PRU_DISCONNECT Request and unp_disconnect Function.
PRU_SHUTDOWN Request and unp_shutdown Function.
PRU_ABORT Request and unp_drop Function.
Summary.18. Unix Domain Protocols: I/O and Descriptor Passing.
RTT Measurements Using Ping.
Protocol Stack Measurements.
Latency and Bandwidth.Appendix B. Coding Applications for T/TCP.
This book is divided into three parts,each covering a different topic:
TCP for transactions, commonly called T/TCP. This is an extension to TCP designed to make client-server transactions faster, more efficient, and reliable. This is done by omitting TCP's three-way handshake at the beginning of a connection and shortening the TIME_WAIT state at the end of a connection. We'll see that T/TCP can match UDP's performance for a client-server transaction and that T/TCP provides reliability and adaptability, both major improvements over UDP.
A transaction is defined to be a client request to a server, followed by the server's reply. (The term transaction does not mean a database transaction, with locking, two-phase commit, and backout.)
TCP/IP applications, specifically HTTP (the Hypertext Transfer Protocol, the foundation of the World Wide Web) and NNTP (the Network News Transfer Protocol, the basis for the Usenet news system).
The Unix domain protocols. These protocols are provided by all Unix TCP/IP implementations and on many non-Unix implementations. They provide a form of interprocess communication (IPC) and use the same sockets interface used with TCP/IP. When the client and server are on the same host, the Unix domain protocols are often twice as fast as TCP/IP.
Part 1, the presentation of T/TCP, is in two pieces. Chapters 1-4 describe the protocol and provide numerous examples of how it works. This material is a major expansion of the brief presentation of T/TCP in Section 24.7 of Volume 1. The second piece, Chapters 5-12, describes the actual implementation of T/TCP within the 4.4BSD-Lite networking code (i.e., the code presented in Volume 2). Since the first T/TCP implementation was not released until September 1994, about one year after Volume 1 was published and right as Volume 2 was being completed, the detailed presentation of T/TCP, with examples and all the implementation details, had to wait for another volume in the series.
Part 2, the HTTP and NNTP applications, are a continuation of the TCP/IP applications presented in Chapters 25-30 of Volume 1. In the two years since Volume 1 was published, the popularity of HTTP has grown enormously, as the Internet has exploded, and the use of NNTP has been growing about 75% per year for more than 10 years. HTTP is also a wonderful candidate for T/TCP, given its typical use of TCP: short connections with small amounts of data transferred, where the total time is often dominated by the connection setup and teardown. The heavy use of HTTP (and therefore TCP) on a busy Web server by thousands of different and varied clients also provides a unique opportunity to examine the actual packets at the server (Chapter 14) and look at many features of TCP/IP that were presented in Volumes 1 and 2.
The Unix domain protocols in Part 3 were originally considered for Volume 2 but omitted when its size reached 1200 pages. While it may seem odd to cover protocols other than TCP/IP in a series titled TCP/IP Illustrated, the Unix domain protocols were implemented almost 15 years ago in 4.2BSD alongside the first implementation of BSD TCP/IP. They are used heavily today in any Berkeley-derived kernel, but their use is typically "under the covers," and most users are unaware of their presence. Besides being the foundation for Unix pipes on a Berkeley-derived kernel, another heavy user is the X Window System, when the client and server are on the same host (i.e., on typical workstations). Unix domain sockets are also used to pass descriptors between processes, a powerful technique for interprocess communication. Since the sockets API (application program interface) used with the Unix domain protocols is nearly identical to the sockets API used with TCP/IP, the Unix domain protocols provide an easy way to enhance the performance of local applications with minimal code changes.
Each of the three parts can be read by itself.
Parts 1 and 2 assume a basic understanding of how the TCP/IP protocols work. Readers unfamiliar with TCP/IP should consult the first volume in this series, Stevens 1994, for a thorough description of the TCP/IP protocol suite. The first half of Part 1 (Chapters 1-4, the concepts behind T/TCP along with examples) can be read independent of Volume 2, but the remainder of Part 1 (Chapters 5-12, the implementation of T/TCP) assumes familiarity with the 4.4BSD-Lite networking code, as provided with Volume 2.Many forward and backward references are provided throughout the text, to both topics within this text, and to relevant sections of Volumes 1 and 2 for readers interested in more details. A thorough index is provided, and a list of all the acronyms used throughout the text, along with the compound term for the acronym, appears on the inside front covers. The inside back covers contain an alphabetical cross-reference of all the structures, functions, and macros described in the book and the starting page number of the description. This cross-reference also refers to definitions in Volume 2, when that object is referenced from the code in this volume.
The routing table code in Chapter 6 contains the following copyright notice:
I thank the technical reviewers who read the manuscript and provided important feedback on a tight timetable: Sami Boulos, Alan Cox, Tony DeSimone, Pete Haverlock, Chris Heigham, Mukesh Kacker, Brian Kernighan, Art Mellor, Jeff Mogul, Marianne Mueller, Andras Olah, Craig Partridge, Vern Paxson, Keith Sklower, Ian Lance Taylor, and Gary Wright. A special thanks to the consulting editor, Brian Kernighan, for his rapid, thorough, and helpful reviews throughout the course of the book, and for his continued encouragement and support.
Special thanks are also due Vern Paxson and Andras Olah for their incredibly detailed reviews of the entire manuscript, finding many errors and providing valuable technical suggestions. My thanks also to Vern Paxson for making available his software for analyzing Tcpdump traces, and to Andras Olah for his help with T/TCP over the past year. My thanks also to Bob Braden, the designer of T/TCP, who provided the reference source code implementation on which Part 1 of this book is based.
Others helped in significant ways. Gary Wright and Jim Hogue provided the system on which the data for Chapter 14 was collected. Doug Schmidt provided a copy of the public domain TTCP program that uses Unix domain sockets, for the timing measurements in Chapter 16. Craig Partridge provided a copy of the RDP source code to examine. Mike Karels answered lots of questions.
My thanks once again to the National Optical Astronomy Observatories (NOAO), Sidney Wolff, Richard Wolff, and Steve Grandi, for providing access to their networks and hosts.
Finally, my thanks to all the staff at Addison-Wesley, who have helped over the past years, especially my editor John Wait.
As usual, camera-ready copy of the book was produced by the author, a Troff die-hard, using the Groff package written by James Clark. I welcome electronic mail from any readers with comments, suggestions, or bug fixes.