The Art of Unix Programming

Eric Steven Raymond
Thyrsus Enterprises <>
Copyright © 2003 Eric S. Raymond

This document is formatted by Martin Tournoij <>, and is otherwise unchanged from the original.


To Ken Thompson and Dennis Ritchie, because you inspired me.


Table of Contents

  1. Preface
    1. Who Should Read This Book
    2. How to Use This Book
    3. Related References
    4. Conventions Used in This Book
    5. Our Case Studies
    6. Author's Acknowledgements
  2. Context
    1. Philosophy
      1. Culture? What Culture?
      2. The Durability of Unix
      3. The Case against Learning Unix Culture
      4. What Unix Gets Wrong
      5. What Unix Gets Right
        1. Open-Source Software
        2. Cross-Platform Portability and Open Standards
        3. The Internet and the World Wide Web
        4. The Open-Source Community
        5. Flexibility All the Way Down
        6. Unix Is Fun to Hack
        7. The Lessons of Unix Can Be Applied Elsewhere
      6. Basics of the Unix Philosophy
        1. Rule of Modularity: Write simple parts connected by clean interfaces.
        2. Rule of Clarity: Clarity is better than cleverness.
        3. Rule of Composition: Design programs to be connected with other programs.
        4. Rule of Separation: Separate policy from mechanism; separate interfaces from engines.
        5. Rule of Simplicity: Design for simplicity; add complexity only where you must.
        6. Rule of Parsimony: Write a big program only when it is clear by demonstration that nothing else will do.
        7. Rule of Transparency: Design for visibility to make inspection and debugging easier.
        8. Rule of Robustness: Robustness is the child of transparency and simplicity.
        9. Rule of Representation: Fold knowledge into data, so program logic can be stupid and robust.
        10. Rule of Least Surprise: In interface design, always do the least surprising thing.
        11. Rule of Silence: When a program has nothing surprising to say, it should say nothing.
        12. Rule of Repair: Repair what you can — but when you must fail, fail noisily and as soon as possible.
        13. Rule of Economy: Programmer time is expensive; conserve it in preference to machine time.
        14. Rule of Generation: Avoid hand-hacking; write programs to write programs when you can.
        15. Rule of Optimization: Prototype before polishing. Get it working before you optimize it.
        16. Rule of Diversity: Distrust all claims for one true way.
        17. Rule of Extensibility: Design for the future, because it will be here sooner than you think.
      7. The Unix Philosophy in One Lesson
      8. Applying the Unix Philosophy
      9. Attitude Matters Too
    2. History
      1. Origins and History of Unix, 1969-1995
        1. Genesis: 1969–1971
        2. Exodus: 1971–1980
        3. TCP/IP and the Unix Wars: 1980-1990
        4. Blows against the Empire: 1991-1995
      2. Origins and History of the Hackers, 1961-1995
        1. At Play in the Groves of Academe: 1961-1980
        2. Internet Fusion and the Free Software Movement: 1981-1991
        3. Linux and the Pragmatist Reaction: 1991-1998
      3. The Open-Source Movement: 1998 and Onward
      4. The Lessons of Unix History
    3. Contrasts
      1. The Elements of Operating-System Style
        1. What Is the Operating System's Unifying Idea?
        2. Multitasking Capability
        3. Cooperating Processes
        4. Internal Boundaries
        5. File Attributes and Record Structures
        6. Binary File Formats
        7. Preferred User Interface Style
        8. Intended Audience
        9. Entry Barriers to Development
      2. Operating-System Comparisons
        1. VMS
        2. MacOS
        3. OS/2
        4. Windows NT
        5. BeOS
        6. MVS
        7. VM/CMS
        8. Linux
      3. What Goes Around, Comes Around
  3. Design
    1. Modularity
      1. Encapsulation and Optimal Module Size
      2. Compactness and Orthogonality
        1. Compactness
        2. Orthogonality
        3. The SPOT Rule
        4. Compactness and the Strong Single Center
        5. The Value of Detachment
      3. Software Is a Many-Layered Thing
        1. Top-Down versus Bottom-Up
        2. Glue Layers
        3. Case Study: C Considered as Thin Glue
      4. Libraries
        1. Case Study: GIMP Plugins
      5. Unix and Object-Oriented Languages
      6. Coding for Modularity
    2. Textuality
      1. The Importance of Being Textual
        1. Case Study: Unix Password File Format
        2. Case Study: .newsrc Format
        3. Case Study: The PNG Graphics File Format
      2. Data File Metaformats
        1. DSV Style
        2. RFC 822 Format
        3. Cookie-Jar Format
        4. Record-Jar Format
        5. XML
        6. Windows INI Format
        7. Unix Textual File Format Conventions
        8. The Pros and Cons of File Compression
      3. Application Protocol Design
        1. Case Study: SMTP, the Simple Mail Transfer Protocol
        2. Case Study: POP3, the Post Office Protocol
        3. Case Study: IMAP, the Internet Message Access Protocol
      4. Application Protocol Metaformats
        1. The Classical Internet Application Metaprotocol
        2. HTTP as a Universal Application Protocol
        3. BEEP: Blocks Extensible Exchange Protocol
        4. XML-RPC, SOAP, and Jabber
    3. Transparency
      1. Studying Cases
        1. Case Study: audacity
        2. Case Study: fetchmail's -v option
        3. Case Study: GCC
        4. Case Study: kmail
        5. Case Study: SNG
        6. Case Study: The Terminfo Database
        7. Case Study: Freeciv Data Files
      2. Designing for Transparency and Discoverability
        1. The Zen of Transparency
        2. Coding for Transparency and Discoverability
        3. Transparency and Avoiding Overprotectiveness
        4. Transparency and Editable Representations
        5. Transparency, Fault Diagnosis, and Fault Recovery
      3. Designing for Maintainability
    4. Multiprogramming
      1. Separating Complexity Control from Performance Tuning
      2. Taxonomy of Unix IPC Methods
        1. Handing off Tasks to Specialist Programs
        2. Pipes, Redirection, and Filters
        3. Wrappers
        4. Security Wrappers and Bernstein Chaining
        5. Slave Processes
        6. Peer-to-Peer Inter-Process Communication
      3. Problems and Methods to Avoid
        1. Obsolescent Unix IPC Methods
        2. Remote Procedure Calls
        3. Threads — Threat or Menace?
      4. Process Partitioning at the Design Level
    5. Minilanguages
      1. Understanding the Taxonomy of Languages
      2. Applying Minilanguages
        1. Case Study: sng
        2. Case Study: Regular Expressions
        3. Case Study: Glade
        4. Case Study: m4
        5. Case Study: XSLT
        6. Case Study: The Documenter's Workbench Tools
        7. Case Study: fetchmail Run-Control Syntax
        8. Case Study: awk
        9. Case Study: PostScript
        10. Case Study: bc and dc
        11. Case Study: Emacs Lisp
        12. Case Study: JavaScript
      3. Designing Minilanguages
        1. Choosing the Right Complexity Level
        2. Extending and Embedding Languages
        3. Writing a Custom Grammar
        4. Macros — Beware!
        5. Language or Application Protocol?
    6. Generation
      1. Data-Driven Programming
        1. Case Study: ascii
        2. Case Study: Statistical Spam Filtering
        3. Case Study: Metaclass Hacking in fetchmailconf
      2. Ad-hoc Code Generation
        1. Case Study: Generating Code for the ascii Displays
        2. Case Study: Generating HTML Code for a Tabular List
    7. Configuration
      1. What Should Be Configurable?
      2. Where Configurations Live
      3. Run-Control Files
        1. Case Study: The .netrc File
        2. Portability to Other Operating Systems
      4. Environment Variables
        1. System Environment Variables
        2. User Environment Variables
        3. When to Use Environment Variables
        4. Portability to Other Operating Systems
      5. Command-Line Options
        1. The -a to -z of Command-Line Options
        2. Portability to Other Operating Systems
      6. How to Choose among the Methods
        1. Case Study: fetchmail
        2. Case Study: The XFree86 Server
      7. On Breaking These Rules
    8. Interfaces
      1. Applying the Rule of Least Surprise
      2. History of Interface Design on Unix
      3. Evaluating Interface Designs
        1. Tradeoffs between CLI and Visual Interfaces
        2. Case Study: Two Ways to Write a Calculator Program
      4. Transparency, Expressiveness, and Configurability
      5. Unix Interface Design Patterns
        1. The Filter Pattern
        2. The Cantrip Pattern
        3. The Source Pattern
        4. The Sink Pattern
        5. The Compiler Pattern
        6. The ed pattern
        7. The Roguelike Pattern
        8. The ‘Separated Engine and Interface’ Pattern
        9. The CLI Server Pattern
        10. Language-Based Interface Patterns
      6. Applying Unix Interface-Design Patterns
        1. The Polyvalent-Program Pattern
      7. The Web Browser as a Universal Front End
      8. Silence Is Golden
    9. Optimization
      1. Don't Just Do Something, Stand There!
      2. Measure before Optimizing
      3. Nonlocality Considered Harmful
      4. Throughput vs. Latency
        1. Batching Operations
        2. Overlapping Operations
        3. Caching Operation Results
    10. Complexity
      1. Speaking of Complexity
        1. The Three Sources of Complexity
        2. Tradeoffs between Interface and Implementation Complexity
        3. Essential, Optional, and Accidental Complexity
        4. Mapping Complexity
        5. When Simplicity Is Not Enough
      2. A Tale of Five Editors
        1. ed
        2. vi
        3. Sam
        4. Emacs
        5. Wily
      3. The Right Size for an Editor
        1. Identifying the Complexity Problems
        2. Compromise Doesn't Work
        3. Is Emacs an Argument against the Unix Tradition?
      4. The Right Size of Software
  4. Implementation
    1. Languages
      1. Unix's Cornucopia of Languages
      2. Why Not C?
      3. Interpreted Languages and Mixed Strategies
      4. Language Evaluations
        1. C
        2. C++
        3. Shell
        4. Perl
        5. Tcl
        6. Python
        7. Java
        8. Emacs Lisp
      5. Trends for the Future
      6. Choosing an X Toolkit
    2. Tools
      1. A Developer-Friendly Operating System
      2. Choosing an Editor
        1. Useful Things to Know about vi
        2. Useful Things to Know about Emacs
        3. The Antireligious Choice: Using Both
      3. Special-Purpose Code Generators
        1. yacc and lex
        2. Case Study: Glade
      4. make: Automating Your Recipes
        1. Basic Theory of make
        2. make in Non-C/C++ Development
        3. Utility Productions
        4. Generating Makefiles
      5. Version-Control Systems
        1. Why Version Control?
        2. Version Control by Hand
        3. Automated Version Control
        4. Unix Tools for Version Control
      6. Runtime Debugging
      7. Profiling
      8. Combining Tools with Emacs
        1. Emacs and make
        2. Emacs and Runtime Debugging
        3. Emacs and Version Control
        4. Emacs and Profiling
        5. Like an IDE, Only Better
    3. Reuse
      1. The Tale of J. Random Newbie
      2. Transparency as the Key to Reuse
      3. From Reuse to Open Source
      4. The Best Things in Life Are Open
      5. Where to Look?
      6. Issues in Using Open-Source Software
      7. Licensing Issues
        1. What Qualifies as Open Source
        2. Standard Open-Source Licenses
        3. When You Need a Lawyer
  5. Community
    1. Portability
      1. Evolution of C
        1. Early History of C
        2. C Standards
      2. Unix Standards
        1. Standards and the Unix Wars
        2. The Ghost at the Victory Banquet
        3. Unix Standards in the Open-Source World
      3. IETF and the RFC Standards Process
      4. Specifications as DNA, Code as RNA
      5. Programming for Portability
        1. Portability and Choice of Language
        2. Avoiding System Dependencies
        3. Tools for Portability
      6. Internationalization
      7. Portability, Open Standards, and Open Source
    2. Documentation
      1. Documentation Concepts
      2. The Unix Style
        1. The Large-Document Bias
        2. Cultural Style
      3. The Zoo of Unix Documentation Formats
        1. troff and the Documenter's Workbench Tools
        2. TeX
        3. Texinfo
        4. POD
        5. HTML
        6. DocBook
      4. The Present Chaos and a Possible Way Out
      5. DocBook
        1. Document Type Definitions
        2. Other DTDs
        3. The DocBook Toolchain
        4. Migration Tools
        5. Editing Tools
        6. Related Standards and Practices
        7. SGML
        8. XML-DocBook References
      6. Best Practices for Writing Unix Documentation
    3. Open Source
      1. Unix and Open Source
      2. Best Practices for Working with Open-Source Developers
        1. Good Patching Practice
        2. Good Project- and Archive-Naming Practice
        3. Good Development Practice
        4. Good Distribution-Making Practice
        5. Good Communication Practice
      3. The Logic of Licenses: How to Pick One
      4. Why You Should Use a Standard License
      5. Varieties of Open-Source Licensing
        1. MIT or X Consortium License
        2. BSD Classic License
        3. Artistic License
        4. General Public License
        5. Mozilla Public License
    4. Futures
      1. Essence and Accident in Unix Tradition
      2. Plan 9: The Way the Future Was
      3. Problems in the Design of Unix
        1. A Unix File Is Just a Big Bag of Bytes
        2. Unix Support for GUIs Is Weak
        3. File Deletion Is Forever
        4. Unix Assumes a Static File System
        5. The Design of Job Control Was Badly Botched
        6. The Unix API Doesn't Use Exceptions
        7. ioctl2 and fcntl2 Are an Embarrassment
        8. The Unix Security Model May Be Too Primitive
        9. Unix Has Too Many Different Kinds of Names
        10. File Systems Might Be Considered Harmful
        11. Towards a Global Internet Address Space
      4. Problems in the Environment of Unix
      5. Problems in the Culture of Unix
      6. Reasons to Believe
  6. Appendix
    1. Glossary of Abbreviations
    2. References
    3. Contributors
    4. Revision History
    5. Rootless Root
      1. Editor's Introduction
      2. Master Foo and the Ten Thousand Lines
      3. Master Foo and the Script Kiddie
      4. Master Foo Discourses on the Two Paths
      5. Master Foo and the Methodologist
      6. Master Foo Discourses on the Graphical User Interface
      7. Master Foo and the Unix Zealot
      8. Master Foo Discourses on the Unix-Nature
      9. Master Foo and the End User


Unix is not so much an operating system as an oral history.

– Neal Stephenson

There is a vast difference between knowledge and expertise. Knowledge lets you deduce the right thing to do; expertise makes the right thing a reflex, hardly requiring conscious thought at all.

This book has a lot of knowledge in it, but it is mainly about expertise. It is going to try to teach you the things about Unix development that Unix experts know, but aren't aware that they know. It is therefore less about technicalia and more about shared culture than most Unix books — both explicit and implicit culture, both conscious and unconscious traditions. It is not a ‘how-to’ book, it is a ‘why-to’ book.

The why-to has great practical importance, because far too much software is poorly designed. Much of it suffers from bloat, is exceedingly hard to maintain, and is too difficult to port to new platforms or extend in ways the original programmers didn't anticipate. These problems are symptoms of bad design. We hope that readers of this book will learn something of what Unix has to teach about good design.

This book is divided into four parts: Context, Design, Tools, and Community. The first part (Context) is philosophy and history, to help provide foundation and motivation for what follows. The second part (Design) unfolds the principles of the Unix philosophy into more specific advice about design and implementation. The third part (Tools) focuses on the software Unix provides for helping you solve problems. The fourth part (Community) is about the human-to-human transactions and agreements that make the Unix culture so effective at what it does.

Because this is a book about shared culture, I never planned to write it alone. You will notice that the text includes guest appearances by prominent Unix developers, the shapers of the Unix tradition. The book went through an extended public review process during which I invited these luminaries to comment on and argue with the text. Rather than submerging the results of that review process in the final version, these guests were encouraged to speak with their own voices, amplifying and developing and even disagreeing with the main line of the text.

In this book, when I use the editorial ‘we’ it is not to pretend omniscience but to reflect the fact that it attempts to articulate the expertise of an entire community.

Because this book is aimed at transmitting culture, it includes much more in the way of history and folklore and asides than is normal for a technical book. Enjoy; these things, too, are part of your education as a Unix programmer. No single one of the historical details is vital, but the gestalt of them all is important. We think it makes a more interesting story this way. More importantly, understanding where Unix came from and how it got the way it is will help you develop an intuitive feel for the Unix style.

For the same reason, we refuse to write as if history is over. You will find an unusually large number of references to the time of writing in this book. We do not wish to pretend that current practice reflects some sort of timeless and perfectly logical outcome of preordained destiny. References to time of writing are meant as an alert to the reader two or three or five years hence that the associated statements of fact may have become dated and should be double-checked.

Other things this book is not is neither a C tutorial, nor a guide to the Unix commands and API. It is not a reference for sed or yacc or Perl or Python. It's not a network programming primer, nor an exhaustive guide to the mysteries of X. It's not a tour of Unix's internals and architecture, either. Other books cover these specifics better, and this book points you at them as appropriate.

Beyond all these technical specifics, the Unix culture has an unwritten engineering tradition that has developed over literally millions of man-years[1] of skilled effort. This book is written in the belief that understanding that tradition, and adding its design patterns to your toolkit, will help you become a better programmer and designer.

Cultures consist of people, and the traditional way to learn Unix culture is from other people and through the folklore, by osmosis. This book is not a substitute for person-to-person acculturation, but it can help accelerate the process by allowing you to tap the experience of others.


[1] The three and a half decades between 1969 and 2003 is a long time. Going by the historical trend curve in number of Unix sites during that period, probably somewhere upwards of fifty million man-years have been plowed into Unix development worldwide.

Who Should Read This Book

You should read this book if you are an experienced Unix programmer who is often in the position of either educating novice programmers or debating partisans of other operating systems, and you find it hard to articulate the benefits of the Unix approach.

You should read this book if you are a C, C++, or Java programmer with experience on other operating systems and you are about to start a Unix-based project.

You should read this book if you are a Unix user with novice-level up to middle-level skills in the operating system, but little development experience, and want to learn how to design software effectively under Unix.

You should read this book if you are a non-Unix programmer who has figured out that the Unix tradition might have something to teach you. We believe you're right, and that the Unix philosophy can be exported to other operating systems. So we will pay more attention to non-Unix environments (especially Microsoft operating systems) than is usual in a Unix book; and when tools and case studies are portable, we say so.

You should read this book if you are an application architect considering platforms or implementation strategies for a major general-market or vertical application. It will help you understand the strengths of Unix as a development platform, and of the Unix tradition of open source as a development method.

You should not read this book if what you are looking for is the details of C coding or how to use the Unix kernel API. There are many good books on these topics; Advanced Programming in the Unix Environment [Stevens92] is classic among explorations of the Unix API, and The Practice of Programming [Kernighan-Pike99] is recommended reading for all C programmers (indeed for all programmers in any language).

How to Use This Book

This book is both practical and philosophical. Some parts are aphoristic and general, others will examine specific case studies in Unix development. We will precede or follow general principles and aphorisms with examples that illustrate them: examples drawn not from toy demonstration programs but rather from real working code that is in use every day.

We have deliberately avoided filling the book with lots of code or specification-file examples, even though in many places this might have made it easier to write (and in some places perhaps easier to read!). Most books about programming give too many low-level details and examples, but fail at giving the reader a high-level feel for what is really going on. In this book, we prefer to err in the opposite direction.

Therefore, while you will often be invited to read code and specification files, relatively few are actually included in the book. Instead, we point you at examples on the Web.

Absorbing these examples will help solidify the principles you learn into semi-instinctive working knowledge. Ideally, you should read this book near the console of a running Unix system, with a Web browser handy. Any Unix will do, but the software case studies are more likely to be preinstalled and immediately available for inspection on a Linux system. The pointers in the book are invitations to browse and experiment. Introduction of these pointers is paced so that wandering off to explore for a while won't break up exposition that has to be continuous.

Note: While we have made every effort to cite URLs that should remain stable and usable, there is no way we can guarantee this. If you find that a cited link has gone stale, use common sense and do a phrase search with your favorite Web search engine. Where possible we suggest ways to do this near the URLs we cite.

Most abbreviations used in this book are expanded at first use. For convenience, we have also provided a glossary in an appendix.

References are usually by author name. Numbered footnotes are for URLs that would intrude on the text or that we suspect might be perishable; also for asides, war stories, and jokes.[2]

To make this book more accessible to less technical readers, we invited some non-programmers to read it and identify terms that seemed both obscure and necessary to the flow of exposition. We also use footnotes for definitions of elementary terms that an experienced programmer is unlikely to need.


[2] This particular footnote is dedicated to Terry Pratchett, whose use of footnotes is quite...inspiring.

Related References

Some famous papers and a few books by Unix's early developers have mined this territory before. Kernighan and Pike's The Unix Programming Environment [Kernighan-Pike84] stands out among these and is rightly considered a classic. But today it shows its age a bit; it doesn't cover the Internet, and the World Wide Web or the new wave of interpreted languages like Perl, Tcl, and Python.

About halfway into the composition of this book, we learned of Mike Gancarz's The Unix Philosophy [Gancarz]. This book is excellent within its range, but did not attempt to cover the full spectrum of topics we felt needed to be addressed. Nevertheless we are grateful to the author for the reminder that the very simplest Unix design patterns have been the most persistent and successful ones.

The Pragmatic Programmer [Hunt-Thomas] is a witty and wise disquisition on good design practice pitched at a slightly different level of the software-design craft (more about coding, less about higher-level partitioning of problems) than this book. The authors' philosophy is an outgrowth of Unix experience, and it is an excellent complement to this book.

The Practice of Programming [Kernighan-Pike99] covers some of the same ground as The Pragmatic Programmer from a position deep within the Unix tradition.

Finally (and with admitted intent to provoke) we recommend Zen Flesh, Zen Bones [Reps-Senzaki], an important collection of Zen Buddhist primary sources. References to Zen are scattered throughout this book. They are included because Zen provides a vocabulary for addressing some ideas that turn out to be very important for software design but are otherwise very difficult to hold in the mind. Readers with religious attachments are invited to consider Zen not as a religion but as a therapeutic form of mental discipline — which, in its purest non-theistic forms, is exactly what Zen is.

Conventions Used in This Book

The term “UNIX” is technically and legally a trademark of The Open Group, and should formally be used only for operating systems which are certified to have passed The Open Group's elaborate standards-conformance tests. In this book we use “Unix” in the looser sense widely current among programmers, to refer to any operating system (whether formally Unix-branded or not) that is either genetically descended from Bell Labs's ancestral Unix code or written in close imitation of its descendants. In particular, Linux (from which we draw most of our examples) is a Unix under this definition.

This book employs the Unix manual page convention of tagging Unix facilities with a following manual section in parentheses, usually on first introduction when we want to emphasize that this is a Unix command. Thus, for example, read “munger(1)” as “the ‘munger’ program, which will be documented in section 1 (user tools) of the Unix manual pages, if it's present on your system”. Section 2 is C system calls, section 3 is C library calls, section 5 is file formats and protocols, section 8 is system administration tools. Other sections vary among Unixes but are not cited in this book. For more, type man 1 man at your Unix shell prompt (older System V Unixes may require man -s 1 man ).

Sometimes we mention a Unix application (such as Emacs), without a manual-section suffix and capitalized. This is a clue that the name actually represents a well-established family of Unix programs with essentially the same function, and we are discussing generic properties of all of them. Emacs, for example, includes xemacs.

At various points later in this book we refer to ‘old school’ and ‘new school’ methods. As with rap music, new-school starts about 1990. In this context, it's associated with the rise of scripting languages, GUIs, open-source Unixes, and the Web. Old-school refers to the pre-1990 (and especially pre-1985) world of expensive (shared) computers, proprietary Unixes, scripting in shell, and C everywhere. This difference is worth pointing out because cheaper and less memory-constrained machines have wrought some significant changes on the Unix programming style.

Our Case Studies

A lot of books on programming rely on toy examples constructed specifically to prove a point. This one won't. Our case studies will be real, pre-existing pieces of software that are in production use every day. Here are some of the major ones:

cdrtools / xcdroast

These two separate projects are usually used together. The cdrtools package is a set of CLI tools for writing CD-ROMs; Web search for “cdrtools”. The xcdroast application is a GUI front end for cdrtools; see the xcdroast project site.


The fetchmail program retrieves mail from remote-mail servers using the POP3 or IMAP post-office protocols. See the fetchmail home page (or search for “fetchmail” on the Web).


The GIMP (GNU Image Manipulation Program) is a full-featured paint, draw, and image-manipulation program that can edit a huge variety of graphical formats in sophisticated ways. Sources are available from the GIMP home page (or search for "GIMP" on the Web).


The mutt mail user agent is the current best-of-breed among text-based Unix electronic mail agents, with notably good support for MIME (Multipurpose Internet Mail Extensions) and the use of privacy aids such as PGP (Pretty Good Privacy) and GPG (GNU Privacy Guard). Source code and executable binaries are available at the Mutt project site.


The xmlto command renders DocBook and other XML documents in various output formats, including HTML and text and PostScript. For sources and documentation, see the xmlto project site.

To minimize the amount of code the user needs to read to understand the examples, we have tried to choose case studies that can be used more than once, ideally to illustrate several different design principles and practices. For this same reason, many of the examples are from my projects. No claim that these are the best possible ones is implied, merely that I find them sufficiently familiar to be useful for multiple expository purposes.

Author's Acknowledgements

The guest contributors (Ken Arnold, Steven M. Bellovin, Stuart Feldman, Jim Gettys, Steve Johnson, Brian Kernighan, David Korn, Mike Lesk, Doug McIlroy, Marshall Kirk McKusick, Keith Packard, Henry Spencer, and Ken Thompson) added a great deal of value to this book. Doug McIlroy, in particular, went far beyond the call of duty in the thoroughness of his critique and the depth of his contributions, displaying the same care and dedication to excellence which he brought to managing the original Unix research group thirty years ago.

Special thanks go to Rob Landley and to my wife Catherine Raymond, both of whom delivered intensive line-by-line critiques of manuscript drafts. Rob's insightful and attentive commentary actually inspired more than one entire chapter in the final manuscript, and he had a lot to do with its present organization and range; if he had written all the text he pushed me to improve, I would have to call him a co-author. Cathy was my test audience representing non-technical readers; to the extent this book is accessible to people who aren't already programmers, that's largely her doing.

This book benefited from discussions with many other people over the five years it took me to write it. Mark M. Miller helped me achieve enlightenment about threads. John Cowan supplied some insights about interface design patterns and drafted the case studies of wily and VM/CMS, and Jef Raskin showed me where the Rule of Least Surprise comes from. The UIUC System Architecture Group contributed useful feedback on early chapters. The sections on What Unix Gets Wrong and Flexibility in Depth were directly inspired by their review. Russell J. Nelson contributed the material on Bernstein chaining in Chapter 7. Jay Maynard contributed most of the material in the MVS case study in Chapter 3. Les Hatton provided many helpful comments on the Languages chapter and motivated the portion of Chapter 4 on Optimal Module Size. David A. Wheeler contributed many perceptive criticisms and some case-study material, especially in the Design part. Russ Cox helped develop the survey of Plan 9. Dennis Ritchie corrected me on some historical points about C.

Hundreds of Unix programmers, far too many to list here, contributed advice and comments during the book's public review period between January and June of 2003. As always, I found the process of open peer review over the Web both intensely challenging and intensely rewarding. Also as always, responsibility for any errors in the resulting work remains my own.

The expository style and some of the concerns of this book have been influenced by the design patterns movement; indeed, I flirted with the idea of titling the book Unix Design Patterns. I didn't, because I disagree with some of the implicit central dogmas of the movement and don't feel the need to use all its formal apparatus or accept its cultural baggage. Nevertheless, my approach has certainly been influenced by Christopher Alexander's work[3] (especially The Timeless Way of Building and A Pattern Language), and I owe the Gang of Four and other members of their school a large debt of gratitude for showing me how it is possible to use Alexander's insights to talk about software design at a high level without merely uttering vague and useless generalities. Interested readers should see Design Patterns: Elements of Reusable Object-Oriented Software[GangOfFour] for an introduction to design patterns.

The title of this book is, of course, a reference to Donald Knuth's The Art of Computer Programming. While not specifically associated with the Unix tradition, Knuth has been an influence on us all.

Editors with vision and imagination aren't as common as they should be. Mark Taub is one; he saw merit in a stalled project and skillfully nudged me into finishing it. Copy editors with a good ear for prose style and enough ability to improve writing that isn't like theirs are even less common, but Mary Lou Nohr makes that grade. Jerry Votta seized on my concept for the cover and made it look better than I had imagined. The whole crew at Addison-Wesley gets high marks for making the editorial and production process as painless as possible, and for cheerfully accommodating my control-freak tendencies not just over the text but deep into the details of the book's visual design, art, and marketing.


[3] An appreciation of Alexander's work, with links to on-line versions of significant portions, may be found at Some Notes on Christopher Alexander.


Chapter 1. Philosophy

Those who do not understand Unix are condemned to reinvent it, poorly.

– Henry Spencer, Usenet signature, November 1987

Culture? What Culture?

This is a book about Unix programming, but in it we're going to toss around the words ‘culture’, ‘art’, and ‘philosophy’ a lot. If you are not a programmer, or you are a programmer who has had little contact with the Unix world, this may seem strange. But Unix has a culture; it has a distinctive art of programming; and it carries with it a powerful design philosophy. Understanding these traditions will help you build better software, even if you're developing for a non-Unix platform.

Every branch of engineering and design has technical cultures. In most kinds of engineering, the unwritten traditions of the field are parts of a working practitioner's education as important as (and, as experience grows, often more important than) the official handbooks and textbooks. Senior engineers develop huge bodies of implicit knowledge, which they pass to their juniors by (as Zen Buddhists put it) “a special transmission, outside the scriptures”.

Software engineering is generally an exception to this rule; technology has changed so rapidly, software environments have come and gone so quickly, that technical cultures have been weak and ephemeral. There are, however, exceptions to this exception. A very few software technologies have proved durable enough to evolve strong technical cultures, distinctive arts, and an associated design philosophy transmitted across generations of engineers.

The Unix culture is one of these. The Internet culture is another — or, in the twenty-first century, arguably the same one. The two have grown increasingly difficult to separate since the early 1980s, and in this book we won't try particularly hard.

The Durability of Unix

Unix was born in 1969 and has been in continuous production use ever since. That's several geologic eras by computer-industry standards — older than the PC or workstations or microprocessors or even video display terminals, and contemporaneous with the first semiconductor memories. Of all production timesharing systems today, only IBM's VM/CMS can claim to have existed longer, and Unix machines have provided hundreds of thousands of times more service hours; indeed, Unix has probably supported more computing than all other timesharing systems put together.

Unix has found use on a wider variety of machines than any other operating system can claim. From supercomputers to handhelds and embedded networking hardware, through workstations and servers and PCs and minicomputers, Unix has probably seen more architectures and more odd hardware than any three other operating systems combined.

Unix has supported a mind-bogglingly wide spectrum of uses. No other operating system has shone simultaneously as a research vehicle, a friendly host for technical custom applications, a platform for commercial-off-the-shelf business software, and a vital component technology of the Internet.

Confident predictions that Unix would wither away, or be crowded out by other operating systems, have been made yearly since its infancy. And yet Unix, in its present-day avatars as Linux and BSD and Solaris and MacOS X and half a dozen other variants, seems stronger than ever today.

Robert Metcalf [the inventor of Ethernet] says that if something comes along to replace Ethernet, it will be called “Ethernet”, so therefore Ethernet will never die.[4] Unix has already undergone several such transformations.

– Ken Thompson

At least one of Unix's central technologies — the C language — has been widely naturalized elsewhere. Indeed it is now hard to imagine doing software engineering without C as a ubiquitous common language of systems programming. Unix also introduced both the now-ubiquitous tree-shaped file namespace with directory nodes and the pipeline for connecting programs.

Unix's durability and adaptability have been nothing short of astonishing. Other technologies have come and gone like mayflies. Machines have increased a thousandfold in power, languages have mutated, industry practice has gone through multiple revolutions — and Unix hangs in there, still producing, still paying the bills, and still commanding loyalty from many of the best and brightest software technologists on the planet.

One of the many consequences of the exponential power-versus-time curve in computing, and the corresponding pace of software development, is that 50% of what one knows becomes obsolete over every 18 months. Unix does not abolish this phenomenon, but does do a good job of containing it. There's a bedrock of unchanging basics — languages, system calls, and tool invocations — that one can actually keep using for years, even decades. Elsewhere it is impossible to predict what will be stable; even entire operating systems cycle out of use. Under Unix, there is a fairly sharp distinction between transient knowledge and lasting knowledge, and one can know ahead of time (with about 90% certainty) which category something is likely to fall in when one learns it. Thus the loyalty Unix commands.

Much of Unix's stability and success has to be attributed to its inherent strengths, to design decisions Ken Thompson, Dennis Ritchie, Brian Kernighan, Doug McIlroy, Rob Pike and other early Unix developers made back at the beginning; decisions that have been proven sound over and over. But just as much is due to the design philosophy, art of programming, and technical culture that grew up around Unix in the early days. This tradition has continuously and successfully propagated itself in symbiosis with Unix ever since.


[4] In fact, Ethernet has already been replaced by a different technology with the same name — twice. Once when coax was replaced with twisted pair, and a second time when gigabit Ethernet came in.

The Case against Learning Unix Culture

Unix's durability and its technical culture are certainly of interest to people who already like Unix, and perhaps to historians of technology. But Unix's original application as a general-purpose timesharing system for mid-sized and larger computers is rapidly receding into the mists of history, killed off by personal workstations. And there is certainly room for doubt that it will ever achieve success in the mainstream business-desktop market now dominated by Microsoft.

Outsiders have frequently dismissed Unix as an academic toy or a hacker's sandbox. One well-known polemic, the Unix Hater's Handbook[Garfinkel], follows an antagonistic line nearly as old as Unix itself in writing its devotees off as a cult religion of freaks and losers. Certainly the colossal and repeated blunders of AT&T, Sun, Novell, and other commercial vendors and standards consortia in mispositioning and mismarketing Unix have become legendary.

Even from within the Unix world, Unix has seemed to be teetering on the brink of universality for so long as to raise the suspicion that it will never actually get there. A skeptical outside observer's conclusion might be that Unix is too useful to die but too awkward to break out of the back room; a perpetual niche operating system.

What confounds the skeptics' case is, more than anything else, the rise of Linux and other open-source Unixes (such as the modern BSD variants). Unix's culture proved too vital to be smothered even by a decade of vendor mismanagement. Today the Unix community itself has taken control of the technology and marketing, and is rapidly and visibly solving Unix's problems (in ways we'll examine in more detail in Chapter 20).

What Unix Gets Wrong

For a design that dates from 1969, it is remarkably difficult to identify design choices in Unix that are unequivocally wrong. There are several popular candidates, but each is still a subject of spirited debate not merely among Unix fans but across the wider community of people who think about and design operating systems.

Unix files have no structure above byte level. File deletion is irrevocable. The Unix security model is arguably too primitive. Job control is botched. There are too many different kinds of names for things. Having a file system at all may have been the wrong choice. We will discuss these technical issues in Chapter 20.

But perhaps the most enduring objections to Unix are consequences of a feature of its philosophy first made explicit by the designers of the X windowing system. X strives to provide “mechanism, not policy”, supporting an extremely general set of graphics operations and deferring decisions about toolkits and interface look-and-feel (the policy) up to application level. Unix's other system-level services display similar tendencies; final choices about behavior are pushed as far toward the user as possible. Unix users can choose among multiple shells. Unix programs normally provide many behavior options and sport elaborate preference facilities.

This tendency reflects Unix's heritage as an operating system designed primarily for technical users, and a consequent belief that users know better than operating-system designers what their own needs are.

This tenet was firmly established at Bell Labs by Dick Hamming[5] who insisted in the 1950s when computers were rare and expensive, that open-shop computing, where customers wrote their own programs, was imperative, because “it is better to solve the right problem the wrong way than the wrong problem the right way”.

– Doug McIlroy

But the cost of the mechanism-not-policy approach is that when the user can set policy, the user must set policy. Nontechnical end-users frequently find Unix's profusion of options and interface styles overwhelming and retreat to systems that at least pretend to offer them simplicity.

In the short term, Unix's laissez-faire approach may lose it a good many nontechnical users. In the long term, however, it may turn out that this ‘mistake’ confers a critical advantage — because policy tends to have a short lifetime, mechanism a long one. Today's fashion in interface look-and-feel too often becomes tomorrow's evolutionary dead end (as people using obsolete X toolkits will tell you with some feeling!). So the flip side of the flip side is that the “mechanism, not policy” philosophy may enable Unix to renew its relevance long after competitors more tied to one set of policy or interface choices have faded from view.[6]


[5] Yes, the Hamming of ‘Hamming distance’ and ‘Hamming code’.

[6] Jim Gettys, one of the architects of X (and a contributor to this book), has meditated in depth on how X's laissez-faire style might be productively carried forward in The Two-Edged Sword [Gettys]. This essay is well worth reading, both for its specific proposals and for its expression of the Unix mindset.

What Unix Gets Right

The explosive recent growth of Linux, and the increasing importance of the Internet, give us good reasons to suppose that the skeptics' case is wrong. But even supposing the skeptical assessment is true, Unix culture is worth learning because there are some things that Unix and its surrounding culture clearly do better than any competitors.

Open-Source Software

Though the term “open source” and the Open Source Definition were not invented until 1998, peer-review-intensive development of freely shared source code was a key feature of the Unix culture from its beginnings.

For its first ten years AT&T's original Unix, and its primary variant Berkeley Unix, were normally distributed with source code. This enabled most of the other good things that follow here.

Cross-Platform Portability and Open Standards

Unix is still the only operating system that can present a consistent, documented application programming interface (API) across a heterogeneous mix of computers, vendors, and special-purpose hardware. It is the only operating system that can scale from embedded chips and handhelds, up through desktop machines, through servers, and all the way to special-purpose number-crunching behemoths and database back ends.

The Unix API is the closest thing to a hardware-independent standard for writing truly portable software that exists. It is no accident that what the IEEE originally called the Portable Operating System Standard quickly got a suffix added to its acronym and became POSIX. A Unix-equivalent API was the only credible model for such a standard.

Binary-only applications for other operating systems die with their birth environments, but Unix sources are forever. Forever, at least, given a Unix technical culture that polishes and maintains them across decades.

The Internet and the World Wide Web

The Defense Department's contract for the first production TCP/IP stack went to a Unix development group because the Unix in question was largely open source. Besides TCP/IP, Unix has become the one indispensable core technology of the Internet Service Provider industry. Ever since the demise of the TOPS family of operating systems in the mid-1980s, most Internet server machines (and effectively all above the PC level) have relied on Unix.

Not even Microsoft's awesome marketing clout has been able to dent Unix's lock on the Internet. While the TCP/IP standards (on which the Internet is based) evolved under TOPS-10 and are theoretically separable from Unix, attempts to make them work on other operating systems have been bedeviled by incompatibilities, instabilities, and bugs. The theory and specifications are available to anyone, but the engineering tradition to make them into a solid and working reality exists only in the Unix world.[7]

The Internet technical culture and the Unix culture began to merge in the early 1980s, and are now inseparably symbiotic. The design of the World Wide Web, the modern face of the Internet, owes as much to Unix as it does to the ancestral ARPANET. In particular, the concept of the Uniform Resource Locator (URL) so central to the Web is a generalization of the Unix idea of one uniform file namespace everywhere. To function effectively as an Internet expert, an understanding of Unix and its culture are indispensable.

The Open-Source Community

The community that originally formed around the early Unix source distributions never went away — after the great Internet explosion of the early 1990s, it recruited an entire new generation of eager hackers on home machines.

Today, that community is a powerful support group for all kinds of software development. High-quality open-source development tools abound in the Unix world (we'll examine many in this book). Open-source Unix applications are usually equal to, and are often superior to, their proprietary equivalents [Fuzz]. Entire Unix operating systems, with complete toolkits and basic applications suites, are available for free over the Internet. Why code from scratch when you can adapt, reuse, recycle, and save yourself 90% of the work?

This tradition of code-sharing depends heavily on hard-won expertise about how to make programs cooperative and reusable. And not by abstract theory, but through a lot of engineering practice — unobvious design rules that allow programs to function not just as isolated one-shot solutions but as synergistic parts of a toolkit. A major purpose of this book is to elucidate those rules.

Today, a burgeoning open-source movement is bringing new vitality, new technical approaches, and an entire generation of bright young programmers into the Unix tradition. Open-source projects including the Linux operating system and symbionts such as Apache and Mozilla have brought the Unix tradition an unprecedented level of mainstream visibility and success. The open-source movement seems on the verge of winning its bid to define the computing infrastructure of tomorrow — and the core of that infrastructure will be Unix machines running on the Internet.

Flexibility All the Way Down

Many operating systems touted as more ‘modern’ or ‘user friendly’ than Unix achieve their surface glossiness by locking users and developers into one interface policy, and offer an application-programming interface that for all its elaborateness is rather narrow and rigid. On such systems, tasks the designers have anticipated are very easy — but tasks they have not anticipated are often impossible or at best extremely painful.

Unix, on the other hand, has flexibility in depth. The many ways Unix provides to glue together programs mean that components of its basic toolkit can be combined to produce useful effects that the designers of the individual toolkit parts never anticipated.

Unix's support of multiple styles of program interface (often seen as a weakness because it increases the perceived complexity of the system to end users) also contributes to flexibility; no program that wants to be a simple piece of data plumbing is forced to carry the complexity overhead of an elaborate GUI.

Unix tradition lays heavy emphasis on keeping programming interfaces relatively small, clean, and orthogonal — another trait that produces flexibility in depth. Throughout a Unix system, easy things are easy and hard things are at least possible.

Unix Is Fun to Hack

People who pontificate about Unix's technical superiority often don't mention what may ultimately be its most important strength, the one that underlies all its successes. Unix is fun to hack.

Unix boosters seem almost ashamed to acknowledge this sometimes, as though admitting they're having fun might damage their legitimacy somehow. But it's true; Unix is fun to play with and develop for, and always has been.

There are not many operating systems that anyone has ever described as ‘fun’. Indeed, the friction and labor of development under most other environments has been aptly compared to kicking a dead whale down the beach.[8] The kindest adjectives one normally hears are on the order of “tolerable” or “not too painful”. In the Unix world, by contrast, the operating system rewards effort rather than frustrating it. People programming under Unix usually come to see it not as an adversary to be clubbed into doing one's bidding by main effort but rather as an actual positive help.

This has real economic significance. The fun factor started a virtuous circle early in Unix's history. People liked Unix, so they built more programs for it that made it nicer to use. Today people build entire, production-quality open-source Unix systems as a hobby. To understand how remarkable this is, ask yourself when you last heard of anybody cloning OS/360 or VAX VMS or Microsoft Windows for fun.

The ‘fun’ factor is not trivial from a design point of view, either. The kind of people who become programmers and developers have ‘fun’ when the effort they have to put out to do a task challenges them, but is just within their capabilities. ‘Fun’ is therefore a sign of peak efficiency. Painful development environments waste labor and creativity; they extract huge hidden costs in time, money, and opportunity.

If Unix were a failure in every other way, the Unix engineering culture would be worth studying for the ways it keeps the fun in development — because that fun is a sign that it makes developers efficient, effective, and productive.

The Lessons of Unix Can Be Applied Elsewhere

Unix programmers have accumulated decades of experience while pioneering operating-system features we now take for granted. Even non-Unix programmers can benefit from studying that Unix experience. Because Unix makes it relatively easy to apply good design principles and development methods, it is an excellent place to learn them.

Other operating systems generally make good practice rather more difficult, but even so some of the Unix culture's lessons can transfer. Much Unix code (including all its filters, its major scripting languages, and many of its code generators) will port directly to any operating system supporting ANSI C (for the excellent reason that C itself was a Unix invention and the ANSI C library embodies a substantial chunk of Unix's services!).


[7] Other operating systems have generally copied or cloned Unix TCP/IP implementations. It is their loss that they have not generally adopted the robust tradition of peer review that goes with it, exemplified by documents like RFC 1025 ( TCP and IP Bake Off).

[8] This was originally said of the IBM MVS TSO facility by Stephen C. Johnson, perhaps better known as the author of yacc.

Basics of the Unix Philosophy

The ‘Unix philosophy’ originated with Ken Thompson's early meditations on how to design a small but capable operating system with a clean service interface. It grew as the Unix culture learned things about how to get maximum leverage out of Thompson's design. It absorbed lessons from many sources along the way.

The Unix philosophy is not a formal design method. It wasn't handed down from the high fastnesses of theoretical computer science as a way to produce theoretically perfect software. Nor is it that perennial executive's mirage, some way to magically extract innovative but reliable software on too short a deadline from unmotivated, badly managed, and underpaid programmers.

The Unix philosophy (like successful folk traditions in other engineering disciplines) is bottom-up, not top-down. It is pragmatic and grounded in experience. It is not to be found in official methods and standards, but rather in the implicit half-reflexive knowledge, the expertise that the Unix culture transmits. It encourages a sense of proportion and skepticism — and shows both by having a sense of (often subversive) humor.

Doug McIlroy, the inventor of Unix pipes and one of the founders of the Unix tradition, had this to say at the time [McIlroy78]:

(i) Make each program do one thing well. To do a new job, build afresh rather than complicate old programs by adding new features.

(ii) Expect the output of every program to become the input to another, as yet unknown, program. Don't clutter output with extraneous information. Avoid stringently columnar or binary input formats. Don't insist on interactive input.

(iii) Design and build software, even operating systems, to be tried early, ideally within weeks. Don't hesitate to throw away the clumsy parts and rebuild them.

(iv) Use tools in preference to unskilled help to lighten a programming task, even if you have to detour to build the tools and expect to throw some of them out after you've finished using them.

He later summarized it this way (quoted in A Quarter Century of Unix[Salus]):

This is the Unix philosophy: Write programs that do one thing and do it well. Write programs to work together. Write programs to handle text streams, because that is a universal interface.

Rob Pike, who became one of the great masters of C, offers a slightly different angle in Notes on C Programming [Pike]:

Rule 1. You can't tell where a program is going to spend its time. Bottlenecks occur in surprising places, so don't try to second guess and put in a speed hack until you've proven that's where the bottleneck is.

Rule 2. Measure. Don't tune for speed until you've measured, and even then don't unless one part of the code overwhelms the rest.

Rule 3. Fancy algorithms are slow when n is small, and n is usually small. Fancy algorithms have big constants. Until you know that n is frequently going to be big, don't get fancy. (Even if n does get big, use Rule 2 first.)

Rule 4. Fancy algorithms are buggier than simple ones, and they're much harder to implement. Use simple algorithms as well as simple data structures.

Rule 5. Data dominates. If you've chosen the right data structures and organized things well, the algorithms will almost always be self-evident. Data structures, not algorithms, are central to programming.[9]

Rule 6. There is no Rule 6.

Ken Thompson, the man who designed and implemented the first Unix, reinforced Pike's rule 4 with a gnomic maxim worthy of a Zen patriarch:

When in doubt, use brute force.

More of the Unix philosophy was implied not by what these elders said but by what they did and the example Unix itself set. Looking at the whole, we can abstract the following ideas:

  1. Rule of Modularity: Write simple parts connected by clean interfaces.
  2. Rule of Clarity: Clarity is better than cleverness.
  3. Rule of Composition: Design programs to be connected to other programs.
  4. Rule of Separation: Separate policy from mechanism; separate interfaces from engines.
  5. Rule of Simplicity: Design for simplicity; add complexity only where you must.
  6. Rule of Parsimony: Write a big program only when it is clear by demonstration that nothing else will do.
  7. Rule of Transparency: Design for visibility to make inspection and debugging easier.
  8. Rule of Robustness: Robustness is the child of transparency and simplicity.
  9. Rule of Representation: Fold knowledge into data so program logic can be stupid and robust.
  10. Rule of Least Surprise: In interface design, always do the least surprising thing.
  11. Rule of Silence: When a program has nothing surprising to say, it should say nothing.
  12. Rule of Repair: When you must fail, fail noisily and as soon as possible.
  13. Rule of Economy: Programmer time is expensive; conserve it in preference to machine time.
  14. Rule of Generation: Avoid hand-hacking; write programs to write programs when you can.
  15. Rule of Optimization: Prototype before polishing. Get it working before you optimize it.
  16. Rule of Diversity: Distrust all claims for “one true way”.
  17. Rule of Extensibility: Design for the future, because it will be here sooner than you think.

If you're new to Unix, these principles are worth some meditation. Software-engineering texts recommend most of them; but most other operating systems lack the right tools and traditions to turn them into practice, so most programmers can't apply them with any consistency. They come to accept blunt tools, bad designs, overwork, and bloated code as normal — and then wonder what Unix fans are so annoyed about.

Rule of Modularity: Write simple parts connected by clean interfaces.

As Brian Kernighan once observed, “Controlling complexity is the essence of computer programming” [Kernighan-Plauger]. Debugging dominates development time, and getting a working system out the door is usually less a result of brilliant design than it is of managing not to trip over your own feet too many times.

Assemblers, compilers, flowcharting, procedural programming, structured programming, “artificial intelligence”, fourth-generation languages, object orientation, and software-development methodologies without number have been touted and sold as a cure for this problem. All have failed as cures, if only because they ‘succeeded’ by escalating the normal level of program complexity to the point where (once again) human brains could barely cope. As Fred Brooks famously observed [Brooks], there is no silver bullet.

The only way to write complex software that won't fall on its face is to hold its global complexity down — to build it out of simple parts connected by well-defined interfaces, so that most problems are local and you can have some hope of upgrading a part without breaking the whole.

Rule of Clarity: Clarity is better than cleverness.

Because maintenance is so important and so expensive, write programs as if the most important communication they do is not to the computer that executes them but to the human beings who will read and maintain the source code in the future (including yourself).

In the Unix tradition, the implications of this advice go beyond just commenting your code. Good Unix practice also embraces choosing your algorithms and implementations for future maintainability. Buying a small increase in performance with a large increase in the complexity and obscurity of your technique is a bad trade — not merely because complex code is more likely to harbor bugs, but also because complex code will be harder to read for future maintainers.

Code that is graceful and clear, on the other hand, is less likely to break — and more likely to be instantly comprehended by the next person to have to change it. This is important, especially when that next person might be yourself some years down the road.

Never struggle to decipher subtle code three times. Once might be a one-shot fluke, but if you find yourself having to figure it out a second time — because the first was too long ago and you've forgotten details — it is time to comment the code so that the third time will be relatively painless.

– Henry Spencer

Rule of Composition: Design programs to be connected with other programs.

It's hard to avoid programming overcomplicated monoliths if none of your programs can talk to each other.

Unix tradition strongly encourages writing programs that read and write simple, textual, stream-oriented, device-independent formats. Under classic Unix, as many programs as possible are written as simple filters, which take a simple text stream on input and process it into another simple text stream on output.

Despite popular mythology, this practice is favored not because Unix programmers hate graphical user interfaces. It's because if you don't write programs that accept and emit simple text streams, it's much more difficult to hook the programs together.

Text streams are to Unix tools as messages are to objects in an object-oriented setting. The simplicity of the text-stream interface enforces the encapsulation of the tools. More elaborate forms of inter-process communication, such as remote procedure calls, show a tendency to involve programs with each others' internals too much.

To make programs composable, make them independent. A program on one end of a text stream should care as little as possible about the program on the other end. It should be made easy to replace one end with a completely different implementation without disturbing the other.

GUIs can be a very good thing. Complex binary data formats are sometimes unavoidable by any reasonable means. But before writing a GUI, it's wise to ask if the tricky interactive parts of your program can be segregated into one piece and the workhorse algorithms into another, with a simple command stream or application protocol connecting the two. Before devising a tricky binary format to pass data around, it's worth experimenting to see if you can make a simple textual format work and accept a little parsing overhead in return for being able to hack the data stream with general-purpose tools.

When a serialized, protocol-like interface is not natural for the application, proper Unix design is to at least organize as many of the application primitives as possible into a library with a well-defined API. This opens up the possibility that the application can be called by linkage, or that multiple interfaces can be glued on it for different tasks.

(We discuss these issues in detail in Chapter 7.)

Rule of Separation: Separate policy from mechanism; separate interfaces from engines.

In our discussion of what Unix gets wrong, we observed that the designers of X made a basic decision to implement “mechanism, not policy” —to make X a generic graphics engine and leave decisions about user-interface style to toolkits and other levels of the system. We justified this by pointing out that policy and mechanism tend to mutate on different timescales, with policy changing much faster than mechanism. Fashions in the look and feel of GUI toolkits may come and go, but raster operations and compositing are forever.

Thus, hardwiring policy and mechanism together has two bad effects: It makes policy rigid and harder to change in response to user requirements, and it means that trying to change policy has a strong tendency to destabilize the mechanisms.

On the other hand, by separating the two we make it possible to experiment with new policy without breaking mechanisms. We also make it much easier to write good tests for the mechanism (policy, because it ages so quickly, often does not justify the investment).

This design rule has wide application outside the GUI context. In general, it implies that we should look for ways to separate interfaces from engines.

One way to effect that separation is, for example, to write your application as a library of C service routines that are driven by an embedded scripting language, with the application flow of control written in the scripting language rather than C. A classic example of this pattern is the Emacs editor, which uses an embedded Lisp interpreter to control editing primitives written in C. We discuss this style of design in Chapter 11.

Another way is to separate your application into cooperating front-end and back-end processes communicating through a specialized application protocol over sockets; we discuss this kind of design in Chapter 5 and Chapter 7. The front end implements policy; the back end, mechanism. The global complexity of the pair will often be far lower than that of a single-process monolith implementing the same functions, reducing your vulnerability to bugs and lowering life-cycle costs.

Rule of Simplicity: Design for simplicity; add complexity only where you must.

Many pressures tend to make programs more complicated (and therefore more expensive and buggy). One such pressure is technical machismo. Programmers are bright people who are (often justly) proud of their ability to handle complexity and juggle abstractions. Often they compete with their peers to see who can build the most intricate and beautiful complexities. Just as often, their ability to design outstrips their ability to implement and debug, and the result is expensive failure.

The notion of “intricate and beautiful complexities” is almost an oxymoron. Unix programmers vie with each other for “simple and beautiful” honors — a point that's implicit in these rules, but is well worth making overt.

– Doug McIlroy

Even more often (at least in the commercial software world) excessive complexity comes from project requirements that are based on the marketing fad of the month rather than the reality of what customers want or software can actually deliver. Many a good design has been smothered under marketing's pile of “checklist features” — features that, often, no customer will ever use. And a vicious circle operates; the competition thinks it has to compete with chrome by adding more chrome. Pretty soon, massive bloat is the industry standard and everyone is using huge, buggy programs not even their developers can love.

Either way, everybody loses in the end.

The only way to avoid these traps is to encourage a software culture that knows that small is beautiful, that actively resists bloat and complexity: an engineering tradition that puts a high value on simple solutions, that looks for ways to break program systems up into small cooperating pieces, and that reflexively fights attempts to gussy up programs with a lot of chrome (or, even worse, to design programs around the chrome).

That would be a culture a lot like Unix's.

Rule of Parsimony: Write a big program only when it is clear by demonstration that nothing else will do.

‘Big’ here has the sense both of large in volume of code and of internal complexity. Allowing programs to get large hurts maintainability. Because people are reluctant to throw away the visible product of lots of work, large programs invite overinvestment in approaches that are failed or suboptimal.

(We'll examine the issue of the right size of software in more detail in Chapter 13.)

Rule of Transparency: Design for visibility to make inspection and debugging easier.

Because debugging often occupies three-quarters or more of development time, work done early to ease debugging can be a very good investment. A particularly effective way to ease debugging is to design for transparency and discoverability.

A software system is transparent when you can look at it and immediately understand what it is doing and how. It is discoverable when it has facilities for monitoring and display of internal state so that your program not only functions well but can be seen to function well.

Designing for these qualities will have implications throughout a project. At minimum, it implies that debugging options should not be minimal afterthoughts. Rather, they should be designed in from the beginning — from the point of view that the program should be able to both demonstrate its own correctness and communicate to future developers the original developer's mental model of the problem it solves.

For a program to demonstrate its own correctness, it needs to be using input and output formats sufficiently simple so that the proper relationship between valid input and correct output is easy to check.

The objective of designing for transparency and discoverability should also encourage simple interfaces that can easily be manipulated by other programs — in particular, test and monitoring harnesses and debugging scripts.

Rule of Robustness: Robustness is the child of transparency and simplicity.

Software is said to be robust when it performs well under unexpected conditions which stress the designer's assumptions, as well as under normal conditions.

Most software is fragile and buggy because most programs are too complicated for a human brain to understand all at once. When you can't reason correctly about the guts of a program, you can't be sure it's correct, and you can't fix it if it's broken.

It follows that the way to make robust programs is to make their internals easy for human beings to reason about. There are two main ways to do that: transparency and simplicity.

For robustness, designing in tolerance for unusual or extremely bulky inputs is also important. Bearing in mind the Rule of Composition helps; input generated by other programs is notorious for stress-testing software (e.g., the original Unix C compiler reportedly needed small upgrades to cope well with Yacc output). The forms involved often seem useless to humans. For example, accepting empty lists/strings/etc., even in places where a human would seldom or never supply an empty string, avoids having to special-case such situations when generating the input mechanically.

– Henry Spencer

One very important tactic for being robust under odd inputs is to avoid having special cases in your code. Bugs often lurk in the code for handling special cases, and in the interactions among parts of the code intended to handle different special cases.

We observed above that software is transparent when you can look at it and immediately see what is going on. It is simple when what is going on is uncomplicated enough for a human brain to reason about all the potential cases without strain. The more your programs have both of these qualities, the more robust they will be.

Modularity (simple parts, clean interfaces) is a way to organize programs to make them simpler. There are other ways to fight for simplicity. Here's another one.

Rule of Representation: Fold knowledge into data, so program logic can be stupid and robust.

Even the simplest procedural logic is hard for humans to verify, but quite complex data structures are fairly easy to model and reason about. To see this, compare the expressiveness and explanatory power of a diagram of (say) a fifty-node pointer tree with a flowchart of a fifty-line program. Or, compare an array initializer expressing a conversion table with an equivalent switch statement. The difference in transparency and clarity is dramatic. See Rob Pike's Rule 5.

Data is more tractable than program logic. It follows that where you see a choice between complexity in data structures and complexity in code, choose the former. More: in evolving a design, you should actively seek ways to shift complexity from code to data.

The Unix community did not originate this insight, but a lot of Unix code displays its influence. The C language's facility at manipulating pointers, in particular, has encouraged the use of dynamically-modified reference structures at all levels of coding from the kernel upward. Simple pointer chases in such structures frequently do duties that implementations in other languages would instead have to embody in more elaborate procedures.

(We also cover these techniques in Chapter 9.)

Rule of Least Surprise: In interface design, always do the least surprising thing.

(This is also widely known as the Principle of Least Astonishment.)

The easiest programs to use are those that demand the least new learning from the user — or, to put it another way, the easiest programs to use are those that most effectively connect to the user's pre-existing knowledge.

Therefore, avoid gratuitous novelty and excessive cleverness in interface design. If you're writing a calculator program, ‘+’ should always mean addition! When designing an interface, model it on the interfaces of functionally similar or analogous programs with which your users are likely to be familiar.

Pay attention to your expected audience. They may be end users, they may be other programmers, or they may be system administrators. What is least surprising can differ among these groups.

Pay attention to tradition. The Unix world has rather well-developed conventions about things like the format of configuration and run-control files, command-line switches, and the like. These traditions exist for a good reason: to tame the learning curve. Learn and use them.

(We'll cover many of these traditions in Chapter 5 and Chapter 10.)

The flip side of the Rule of Least Surprise is to avoid making things superficially similar but really a little bit different. This is extremely treacherous because the seeming familiarity raises false expectations. It's often better to make things distinctly different than to make them almost the same.

– Henry Spencer

Rule of Silence: When a program has nothing surprising to say, it should say nothing.

One of Unix's oldest and most persistent design rules is that when a program has nothing interesting or surprising to say, it should shut up. Well-behaved Unix programs do their jobs unobtrusively, with a minimum of fuss and bother. Silence is golden.

This “silence is golden” rule evolved originally because Unix predates video displays. On the slow printing terminals of 1969, each line of unnecessary output was a serious drain on the user's time. That constraint is gone, but excellent reasons for terseness remain.

I think that the terseness of Unix programs is a central feature of the style. When your program's output becomes another's input, it should be easy to pick out the needed bits. And for people it is a human-factors necessity — important information should not be mixed in with verbosity about internal program behavior. If all displayed information is important, important information is easy to find.

– Ken Arnold

Well-designed programs treat the user's attention and concentration as a precious and limited resource, only to be claimed when necessary.

(We'll discuss the Rule of Silence and the reasons for it in more detail at the end of Chapter 11.)

Rule of Repair: Repair what you can — but when you must fail, fail noisily and as soon as possible.

Software should be transparent in the way that it fails, as well as in normal operation. It's best when software can cope with unexpected conditions by adapting to them, but the worst kinds of bugs are those in which the repair doesn't succeed and the problem quietly causes corruption that doesn't show up until much later.

Therefore, write your software to cope with incorrect inputs and its own execution errors as gracefully as possible. But when it cannot, make it fail in a way that makes diagnosis of the problem as easy as possible.

Consider also Postel's Prescription:[10] “Be liberal in what you accept, and conservative in what you send”. Postel was speaking of network service programs, but the underlying idea is more general. Well-designed programs cooperate with other programs by making as much sense as they can from ill-formed inputs; they either fail noisily or pass strictly clean and correct data to the next program in the chain.

However, heed also this warning:

The original HTML documents recommended “be generous in what you accept”, and it has bedeviled us ever since because each browser accepts a different superset of the specifications. It is the specifications that should be generous, not their interpretation.

– Doug McIlroy

McIlroy adjures us to design for generosity rather than compensating for inadequate standards with permissive implementations. Otherwise, as he rightly points out, it's all too easy to end up in tag soup.

Rule of Economy: Programmer time is expensive; conserve it in preference to machine time.

In the early minicomputer days of Unix, this was still a fairly radical idea (machines were a great deal slower and more expensive then). Nowadays, with every development shop and most users (apart from the few modeling nuclear explosions or doing 3D movie animation) awash in cheap machine cycles, it may seem too obvious to need saying.

Somehow, though, practice doesn't seem to have quite caught up with reality. If we took this maxim really seriously throughout software development, most applications would be written in higher-level languages like Perl, Tcl, Python, Java, Lisp and even shell — languages that ease the programmer's burden by doing their own memory management (see [Ravenbrook]).

And indeed this is happening within the Unix world, though outside it most applications shops still seem stuck with the old-school Unix strategy of coding in C (or C++). Later in this book we'll discuss this strategy and its tradeoffs in detail.

One other obvious way to conserve programmer time is to teach machines how to do more of the low-level work of programming. This leads to...

Rule of Generation: Avoid hand-hacking; write programs to write programs when you can.

Human beings are notoriously bad at sweating the details. Accordingly, any kind of hand-hacking of programs is a rich source of delays and errors. The simpler and more abstracted your program specification can be, the more likely it is that the human designer will have gotten it right. Generated code (at every level) is almost always cheaper and more reliable than hand-hacked.

We all know this is true (it's why we have compilers and interpreters, after all) but we often don't think about the implications. High-level-language code that's repetitive and mind-numbing for humans to write is just as productive a target for a code generator as machine code. It pays to use code generators when they can raise the level of abstraction — that is, when the specification language for the generator is simpler than the generated code, and the code doesn't have to be hand-hacked afterwards.

In the Unix tradition, code generators are heavily used to automate error-prone detail work. Parser/lexer generators are the classic examples; makefile generators and GUI interface builders are newer ones.

(We cover these techniques in Chapter 9.)

§ Rule of Optimization: Prototype before polishing. Get it working before you optimize it.

The most basic argument for prototyping first is Kernighan & Plauger's; “90% of the functionality delivered now is better than 100% of it delivered never”. Prototyping first may help keep you from investing far too much time for marginal gains.

For slightly different reasons, Donald Knuth (author of The Art Of Computer Programming, one of the field's few true classics) popularized the observation that “Premature optimization is the root of all evil”.[11] And he was right.

Rushing to optimize before the bottlenecks are known may be the only error to have ruined more designs than feature creep. From tortured code to incomprehensible data layouts, the results of obsessing about speed or memory or disk usage at the expense of transparency and simplicity are everywhere. They spawn innumerable bugs and cost millions of man-hours — often, just to get marginal gains in the use of some resource much less expensive than debugging time.

Disturbingly often, premature local optimization actually hinders global optimization (and hence reduces overall performance). A prematurely optimized portion of a design frequently interferes with changes that would have much higher payoffs across the whole design, so you end up with both inferior performance and excessively complex code.

In the Unix world there is a long-established and very explicit tradition (exemplified by Rob Pike's comments above and Ken Thompson's maxim about brute force) that says: Prototype, then polish. Get it working before you optimize it . Or: Make it work first, then make it work fast. ‘Extreme programming' guru Kent Beck, operating in a different culture, has usefully amplified this to: “Make it run, then make it right, then make it fast”.

The thrust of all these quotes is the same: get your design right with an un-optimized, slow, memory-intensive implementation before you try to tune. Then, tune systematically, looking for the places where you can buy big performance wins with the smallest possible increases in local complexity.

Prototyping is important for system design as well as optimization — it is much easier to judge whether a prototype does what you want than it is to read a long specification. I remember one development manager at Bellcore who fought against the “requirements” culture years before anybody talked about “rapid prototyping” or “agile development”. He wouldn't issue long specifications; he'd lash together some combination of shell scripts and awk code that did roughly what was needed, tell the customers to send him some clerks for a few days, and then have the customers come in and look at their clerks using the prototype and tell him whether or not they liked it. If they did, he would say “you can have it industrial strength so-many-months from now at such-and-such cost”. His estimates tended to be accurate, but he lost out in the culture to managers who believed that requirements writers should be in control of everything.

– Mike Lesk

Using prototyping to learn which features you don't have to implement helps optimization for performance; you don't have to optimize what you don't write. The most powerful optimization tool in existence may be the delete key.

One of my most productive days was throwing away 1000 lines of code.

– Ken Thompson

(We'll go into a bit more depth about related ideas in Chapter 12.)

Rule of Diversity: Distrust all claims for “one true way”.

Even the best software tools tend to be limited by the imaginations of their designers. Nobody is smart enough to optimize for everything, nor to anticipate all the uses to which their software might be put. Designing rigid, closed software that won't talk to the rest of the world is an unhealthy form of arrogance.

Therefore, the Unix tradition includes a healthy mistrust of “one true way” approaches to software design or implementation. It embraces multiple languages, open extensible systems, and customization hooks everywhere.

Rule of Extensibility: Design for the future, because it will be here sooner than you think.

If it is unwise to trust other people's claims for “one true way”, it's even more foolish to believe them about your own designs. Never assume you have the final answer. Therefore, leave room for your data formats and code to grow; otherwise, you will often find that you are locked into unwise early choices because you cannot change them while maintaining backward compatibility.

When you design protocols or file formats, make them sufficiently self-describing to be extensible. Always, always either include a version number, or compose the format from self-contained, self-describing clauses in such a way that new clauses can be readily added and old ones dropped without confusing format-reading code. Unix experience tells us that the marginal extra overhead of making data layouts self-describing is paid back a thousandfold by the ability to evolve them forward without breaking things.

When you design code, organize it so future developers will be able to plug new functions into the architecture without having to scrap and rebuild the architecture. This rule is not a license to add features you don't yet need; it's advice to write your code so that adding features later when you do need them is easy. Make the joints flexible, and put “If you ever need to...” comments in your code. You owe this grace to people who will use and maintain your code after you.

You'll be there in the future too, maintaining code you may have half forgotten under the press of more recent projects. When you design for the future, the sanity you save may be your own.


[9] Pike's original adds “(See Brooks p. 102.)” here. The reference is to an early edition of The Mythical Man-Month [Brooks]; the quote is “Show me your flow charts and conceal your tables and I shall continue to be mystified, show me your tables and I won't usually need your flow charts; they'll be obvious”.

[10] Jonathan Postel was the first editor of the Internet RFC series of standards, and one of the principal architects of the Internet. A tribute page is maintained by the Postel Center for Experimental Networking.

[11] In full: “We should forget about small efficiencies, say about 97% of the time: premature optimization is the root of all evil”. Knuth himself attributes the remark to C. A. R. Hoare.

The Unix Philosophy in One Lesson

All the philosophy really boils down to one iron law, the hallowed ‘KISS principle’ of master engineers everywhere:

KISS: Keep It Simple Stupid

Unix gives you an excellent base for applying the KISS principle. The remainder of this book will help you learn how.

Applying the Unix Philosophy

These philosophical principles aren't just vague generalities. In the Unix world they come straight from experience and lead to specific prescriptions, some of which we've already developed above. Here's a by no means exhaustive list:

We'll see the Unix design rules, and the prescriptions that derive from them, applied over and over again in the remainder of this book. Unsurprisingly, they tend to converge with the very best practices from software engineering in other traditions.[12]


[12] One notable example is Butler Lampson's Hints for Computer System Design [Lampson], which I discovered late in the preparation of this book. It not only expresses a number of Unix dicta in forms that were clearly discovered independently, but uses many of the same tag lines to illustrate them.

Attitude Matters Too

When you see the right thing, do it — this may look like more work in the short term, but it's the path of least effort in the long run. If you don't know what the right thing is, do the minimum necessary to get the job done, at least until you figure out what the right thing is.

To do the Unix philosophy right, you have to be loyal to excellence. You have to believe that software design is a craft worth all the intelligence, creativity, and passion you can muster. Otherwise you won't look past the easy, stereotyped ways of approaching design and implementation; you'll rush into coding when you should be thinking. You'll carelessly complicate when you should be relentlessly simplifying — and then you'll wonder why your code bloats and debugging is so hard.

To do the Unix philosophy right, you have to value your own time enough never to waste it. If someone has already solved a problem once, don't let pride or politics suck you into solving it a second time rather than re-using. And never work harder than you have to; work smarter instead, and save the extra effort for when you need it. Lean on your tools and automate everything you can.

Software design and implementation should be a joyous art, a kind of high-level play. If this attitude seems preposterous or vaguely embarrassing to you, stop and think; ask yourself what you've forgotten. Why do you design software instead of doing something else to make money or pass the time? You must have thought software was worthy of your passion once....

To do the Unix philosophy right, you need to have (or recover) that attitude. You need to care. You need to play. You need to be willing to explore.

We hope you'll bring this attitude to the rest of this book. Or, at least, that this book will help you rediscover it.

Chapter 2. History

Those who cannot remember the past are condemned to repeat it.

– George Santayana The Life of Reason (1905)

The past informs practice. Unix has a long and colorful history, much of which is still live as folklore, assumptions, and (too often) battle scars in the collective memory of Unix programmers. In this chapter we'll survey the history of Unix, with an eye to explaining why, in 2003, today's Unix culture looks the way it does.

Origins and History of Unix, 1969-1995

A notorious ‘second-system effect‘ often afflicts the successors of small experimental prototypes. The urge to add everything that was left out the first time around all too frequently leads to huge and overcomplicated design. Less well known, because less common, is the ‘third-system effect’; sometimes, after the second system has collapsed of its own weight, there is a chance to go back to simplicity and get it really right.

The original Unix was a third system. Its grandfather was the small and simple Compatible Time-Sharing System (CTSS), either the first or second timesharing system ever deployed (depending on some definitional questions we are going to determinedly ignore). Its father was the pioneering Multics project, an attempt to create a feature-packed ‘information utility’ that would gracefully support interactive timesharing of mainframe computers by large communities of users. Multics, alas, did collapse of its own weight. But Unix was born from that collapse.

Genesis: 1969–1971

Unix was born in 1969 out of the mind of a computer scientist at Bell Laboratories, Ken Thompson. Thompson had been a researcher on the Multics project, an experience which spoiled him for the primitive batch computing that was the rule almost everywhere else. But the concept of timesharing was still a novel one in the late 1960s; the first speculations on it had been uttered barely ten years earlier by computer scientist John McCarthy (also the inventor of the Lisp language), the first actual deployment had been in 1962, seven years earlier, and timesharing operating systems were still experimental and temperamental beasts.

Computer hardware was at that time more primitive than even people who were there to see it can now easily recall. The most powerful machines of the day had less computing power and internal memory than a typical cellphone of today.[13] Video display terminals were in their infancy and would not be widely deployed for another six years. The standard interactive device on the earliest timesharing systems was the ASR-33 teletype — a slow, noisy device that printed upper-case-only on big rolls of yellow paper. The ASR-33 was the natural parent of the Unix tradition of terse commands and sparse responses.

When Bell Labs withdrew from the Multics research consortium, Ken Thompson was left with some Multics-inspired ideas about how to build a file system. He was also left without a machine on which to play a game he had written called Space Travel, a science-fiction simulation that involved navigating a rocket through the solar system. Unix began its life on a scavenged PDP-7 minicomputer[14] like the one shown in Figure 2.1, as a platform for the Space Travel game and a testbed for Thompson's ideas about operating system design.

Figure 2.1. The PDP-7.

The PDP-7.

The full origin story is told in [Ritchie79] from the point of view of Thompson's first collaborator, Dennis Ritchie, the man who would become known as the co-inventor of Unix and the inventor of the C language. Dennis Ritchie, Doug McIlroy, and a few colleagues had become used to interactive computing under Multics and did not want to lose that capability. Thompson's PDP-7 operating system offered them a lifeline.

Ritchie observes: “What we wanted to preserve was not just a good environment in which to do programming, but a system around which a fellowship could form. We knew from experience that the essence of communal computing, as supplied by remote-access, time-shared machines, is not just to type programs into a terminal instead of a keypunch, but to encourage close communication”. The theme of computers being viewed not merely as logic devices but as the nuclei of communities was in the air; 1969 was also the year the ARPANET (the direct ancestor of today's Internet) was invented. The theme of “fellowship” would resonate all through Unix's subsequent history.

Thompson and Ritchie's Space Travel implementation attracted notice. At first, the PDP-7's software had to be cross-compiled on a GE mainframe. The utility programs that Thompson and Ritchie wrote to support hosting game development on the PDP-7 itself became the core of Unix — though the name did not attach itself until 1970. The original spelling was “UNICS” (UNiplexed Information and Computing Service), which Ritchie later described as “a somewhat treacherous pun on Multics”, which stood for MULTiplexed Information and Computing Service.

Even at its earliest stages, PDP-7 Unix bore a strong resemblance to today's Unixes and provided a rather more pleasant programming environment than was available anywhere else in those days of card-fed batch mainframes. Unix was very close to being the first system under which a programmer could sit down directly at a machine and compose programs on the fly, exploring possibilities and testing while composing. All through its lifetime Unix has had a pattern of growing more capabilities by attracting highly skilled volunteer efforts from programmers impatient with the limitations of other operating systems. This pattern was set early, within Bell Labs itself.

The Unix tradition of lightweight development and informal methods also began at its beginning. Where Multics had been a large project with thousands of pages of technical specifications written before the hardware arrived, the first running Unix code was brainstormed by three people and implemented by Ken Thompson in two days — on an obsolete machine that had been designed to be a graphics terminal for a ‘real’ computer.

Unix's first real job, in 1971, was to support what would now be called word processing for the Bell Labs patent department; the first Unix application was the ancestor of the nroff(1) text formatter. This project justified the purchase of a PDP-11, a much more capable minicomputer. Management remained blissfully unaware that the word-processing system that Thompson and colleagues were building was incubating an operating system. Operating systems were not in the Bell Labs plan — AT&T had joined the Multics consortium precisely to avoid doing an operating system on its own. Nevertheless, the completed system was a rousing success. It established Unix as a permanent and valued part of the computing ecology at Bell Labs, and began another theme in Unix's history — a close association with document-formatting, typesetting, and communications tools. The 1972 manual claimed 10 installations.

Later, Doug McIlroy would write of this period [McIlroy91]: “Peer pressure and simple pride in workmanship caused gobs of code to be rewritten or discarded as better or more basic ideas emerged. Professional rivalry and protection of turf were practically unknown: so many good things were happening that nobody needed to be proprietary about innovations”. But it would take another quarter century for all the implications of that observation to come home.

Exodus: 1971–1980

The original Unix operating system was written in assembler, and the applications in a mix of assembler and an interpreted language called B, which had the virtue that it was small enough to run on the PDP-7. But B was not powerful enough for systems programming, so Dennis Ritchie added data types and structures to it. The resulting C language evolved from B beginning in 1971; in 1973 Thompson and Ritchie finally succeeded in rewriting Unix in their new language. This was quite an audacious move; at the time, system programming was done in assembler in order to extract maximum performance from the hardware, and the very concept of a portable operating system was barely a gleam in anyone's eye. As late as 1979, Ritchie could write: “It seems certain that much of the success of Unix follows from the readability, modifiability, and portability of its software that in turn follows from its expression in high-level languages”, in the knowledge that this was a point that still needed making.

Ken (seated) and Dennis (standing) at a PDP-11 in 1972.

A 1974 paper in Communications of the ACM [Ritchie-Thompson] gave Unix its first public exposure. In that paper, its authors described the unprecedentedly simple design of Unix, and reported over 600 Unix installations. All were on machines underpowered even by the standards of that day, but (as Ritchie and Thompson wrote) “constraint has encouraged not only economy, but also a certain elegance of design”.

After the CACM paper, research labs and universities all over the world clamored for the chance to try out Unix themselves. Under a 1958 consent decree in settlement of an antitrust case, AT&T (the parent organization of Bell Labs) had been forbidden from entering the computer business. Unix could not, therefore, be turned into a product; indeed, under the terms of the consent decree, Bell Labs was required to license its nontelephone technology to anyone who asked. Ken Thompson quietly began answering requests by shipping out tapes and disk packs — each, according to legend, with a note signed “love, ken”.

This was years before personal computers. Not only was the hardware needed to run Unix too expensive to be within an individual's reach, but nobody imagined that would change in the foreseeable future. So Unix machines were only available by the grace of big organizations with big budgets: corporations, universities, government agencies. But use of these minicomputers was less regulated than the even-bigger mainframes, and Unix development rapidly took on a countercultural air. It was the early 1970s; the pioneering Unix programmers were shaggy hippies and hippie-wannabes. They delighted in playing with an operating system that not only offered them fascinating challenges at the leading edge of computer science, but also subverted all the technical assumptions and business practices that went with Big Computing. Card punches, COBOL, business suits, and batch IBM mainframes were the despised old wave; Unix hackers reveled in the sense that they were simultaneously building the future and flipping a finger at the system.

The excitement of those days is captured in this quote from Douglas Comer: “Many universities contributed to UNIX. At the University of Toronto, the department acquired a 200-dot-per-inch printer/plotter and built software that used the printer to simulate a phototypesetter. At Yale University, students and computer scientists modified the UNIX shell. At Purdue University, the Electrical Engineering Department made major improvements in performance, producing a version of UNIX that supported a larger number of users. Purdue also developed one of the first UNIX computer networks. At the University of California at Berkeley, students developed a new shell and dozens of smaller utilities. By the late 1970s, when Bell Labs released Version 7 UNIX, it was clear that the system solved the computing problems of many departments, and that it incorporated many of the ideas that had arisen in universities. The end result was a strengthened system. A tide of ideas had started a new cycle, flowing from academia to an industrial laboratory, back to academia, and finally moving on to a growing number of commercial sites” [Comer].

The first Unix of which it can be said that essentially all of it would be recognizable to a modern Unix programmer was the Version 7 release in 1979.[15] The first Unix user group had formed the previous year. By this time Unix was in use for operations support all through the Bell System[Hauben], and had spread to universities as far away as Australia, where John Lions's 1976 notes [Lions] on the Version 6 source code became the first serious documentation of the Unix kernel internals. Many senior Unix hackers still treasure a copy.

The Lions book was a samizdat publishing sensation. Because of copyright infringement or some such it couldn't be published in the U.S., so copies of copies seeped everywhere. I still have my copy, which was at least 6th generation. Back then you couldn't be a kernel hacker without a Lions.

– Ken Arnold

The beginnings of a Unix industry were coalescing as well. The first Unix company (the Santa Cruz Operation, SCO) began operations in 1978, and the first commercial C compiler (Whitesmiths) sold that same year. By 1980 an obscure software company in Seattle was also getting into the Unix game, shipping a port of the AT&T version for microcomputers called XENIX. But Microsoft's affection for Unix as a product was not to last very long (though Unix would continue to be used for most internal development work at the company until after 1990).

TCP/IP and the Unix Wars: 1980-1990

The Berkeley campus of the University of California emerged early as the single most important academic hot-spot in Unix development. Unix research had begun there in 1974, and was given a substantial impetus when Ken Thompson taught at the University during a 1975-76 sabbatical. The first BSD release had been in 1977 from a lab run by a then-unknown grad student named Bill Joy. By 1980 Berkeley was the hub of a sub-network of universities actively contributing to their variant of Unix. Ideas and code from Berkeley Unix (including the vi(1) editor) were feeding back from Berkeley to Bell Labs.

Then, in 1980, the Defense Advanced Research Projects Agency needed a team to implement its brand-new TCP/IP protocol stack on the VAX under Unix. The PDP-10s that powered the ARPANET at that time were aging, and indications that DEC might be forced to cancel the 10 in order to support the VAX were already in the air. DARPA considered contracting DEC to implement TCP/IP, but rejected that idea because they were concerned that DEC might not be responsive to requests for changes in their proprietary VAX/VMS operating system [Libes-Ressler]. Instead, DARPA chose Berkeley Unix as a platform — explicitly because its source code was available and unencumbered [Leonard].

Berkeley's Computer Science Research Group was in the right place at the right time with the strongest development tools; the result became arguably the most critical turning point in Unix's history since its invention.

Until the TCP/IP implementation was released with Berkeley 4.2 in 1983, Unix had had only the weakest networking support. Early experiments with Ethernet were unsatisfactory. An ugly but serviceable facility called UUCP (Unix to Unix Copy Program) had been developed at Bell Labs for distributing software over conventional telephone lines via modem.[16] UUCP could forward Unix mail between widely separated machines, and (after Usenet was invented in 1981) supported Usenet, a distributed bulletin-board facility that allowed users to broadcast text messages to anywhere that had phone lines and Unix systems.

Still, the few Unix users aware of the bright lights of the ARPANET felt like they were stuck in a backwater. No FTP, no telnet, only the most restricted remote job execution, and painfully slow links. Before TCP/IP, the Internet and Unix cultures did not mix. Dennis Ritchie's vision of computers as a way to “encourage close communication” was one of collegial communities clustered around individual timesharing machines or in the same computing center; it didn't extend to the continent-wide distributed ‘network nation’ that ARPA users had started to form in the mid-1970s. Early ARPANETters, for their part, considered Unix a crude makeshift limping along on risibly weak hardware.

After TCP/IP, everything changed. The ARPANET and Unix cultures began to merge at the edges, a development that would eventually save both from destruction. But there would be hell to pay first as the result of two unrelated disasters; the rise of Microsoft and the AT&T divestiture.

In 1981, Microsoft made its historic deal with IBM over the new IBM PC. Bill Gates bought QDOS (Quick and Dirty Operating System), a clone of CP/M that its programmer Tim Paterson had thrown together in six weeks, from Paterson's employer Seattle Computer Products. Gates, concealing the IBM deal from Paterson and SCP, bought the rights for $50,000. He then talked IBM into allowing Microsoft to market MS-DOS separately from the PC hardware. Over the next decade, leveraging code he didn't write made Bill Gates a multibillionaire, and business tactics even sharper than the original deal gained Microsoft a monopoly lock on desktop computing. XENIX as a product was rapidly deep-sixed, and eventually sold to SCO.

It was not apparent at the time how successful (or how destructive) Microsoft was going to be. Since the IBM PC-1 didn't have the hardware capacity to run Unix, Unix people barely noticed it at all (though, ironically enough, DOS 2.0 eclipsed CP/M largely because Microsoft's co-founder Paul Allen merged in Unix features including subdirectories and pipes). There were things that seemed much more interesting going on — like the 1982 launching of Sun Microsystems.

Sun Microsystems founders Bill Joy, Andreas Bechtolsheim, and Vinod Khosla set out to build a dream Unix machine with built-in networking capability. They combined hardware designed at Stanford with the Unix developed at Berkeley to produce a smashing success, and founded the workstation industry. At the time, nobody much minded watching source-code access to one branch of the Unix tree gradually dry up as Sun began to behave less like a freewheeling startup and more like a conventional firm. Berkeley was still distributing BSD with source code. Officially, System III source licenses cost $40,000 each; but Bell Labs was turning a blind eye to the number of bootleg Bell Labs Unix tapes in circulation, the universities were still swapping code with Bell Labs, and it looked like Sun's commercialization of Unix might just be the best thing to happen to it yet.

1982 was also the year that C first showed signs of establishing itself outside the Unix world as the systems-programming language of choice. It would only take about five years for C to drive machine assemblers almost completely out of use. By the early 1990s C and C++ would dominate not only systems but application programming; by the late 1990s all other conventional compiled languages would be effectively obsolete.

When DEC canceled development on the PDP-10's successor machine (Jupiter) in 1983, VAXes running Unix began to take over as the dominant Internet machines, a position they would hold until being displaced by Sun workstations. By 1985, about 25% of all VAXes would be running Unix despite DEC's stiff opposition. But the longest-term effect of the Jupiter cancellation was a less obvious one; the death of the MIT AI Lab's PDP-10-centered hacker culture motivated a programmer named Richard Stallman to begin writing GNU, a complete free clone of Unix.

By 1983 there were no fewer than six Unix-workalike operating systems for the IBM-PC: uNETix, Venix, Coherent, QNX, Idris, and the port hosted on the Sritek PC daughtercard. There was still no port of Unix in either the System V or BSD versions; both groups considered the 8086 microprocessor woefully underpowered and wouldn't go near it. None of the Unix-workalikes were significant as commercial successes, but they indicated a significant demand for Unix on cheap hardware that the major vendors were not supplying. No individual could afford to meet it, either, not with the $40,000 price-tag on a source-code license.

Sun was already a success (with imitators!) when, in 1983, the U.S. Department of Justice won its second antitrust case against AT&T and broke up the Bell System. This relieved AT&T from the 1958 consent decree that had prevented them from turning Unix into a product. AT&T promptly rushed to commercialize Unix System V—a move that nearly killed Unix.

So true. But their marketing did spread Unix internationally.

– Ken Thompson

Most Unix boosters thought that the divestiture was great news. We thought we saw in the post-divestiture AT&T, Sun Microsystems, and Sun's smaller imitators the nucleus of a healthy Unix industry — one that, using inexpensive 68000-based workstations, would challenge and eventually break the oppressive monopoly that then loomed over the computer industry — IBM's.

What none of us realized at the time was that the productization of Unix would destroy the free exchanges of source code that had nurtured so much of the system's early vitality. Knowing no other model than secrecy for collecting profits from software and no other model than centralized control for developing a commercial product, AT&T clamped down hard on source-code distribution. Bootleg Unix tapes became far less interesting in the knowledge that the threat of lawsuit might come with them. Contributions from universities began to dry up.

To make matters worse, the big new players in the Unix market promptly committed major strategic blunders. One was to seek advantage by product differentiation — a tactic which resulted in the interfaces of different Unixes diverging. This threw away cross-platform compatibility and fragmented the Unix market.

The other, subtler error was to behave as if personal computers and Microsoft were irrelevant to Unix's prospects. Sun Microsystems failed to see that commoditized PCs would inevitably become an attack on its workstation market from below. AT&T, fixated on minicomputers and mainframes, tried several different strategies to become a major player in computers, and badly botched all of them. A dozen small companies formed to support Unix on PCs; all were underfunded, focused on selling to developers and engineers, and never aimed at the business and home market that Microsoft was targeting.

In fact, for years after divestiture the Unix community was preoccupied with the first phase of the Unix wars — an internal dispute, the rivalry between System V Unix and BSD Unix. The dispute had several levels, some technical (sockets vs. streams, BSD tty vs. System V termio) and some cultural. The divide was roughly between longhairs and shorthairs; programmers and technical people tended to line up with Berkeley and BSD, more business-oriented types with AT&T and System V. The longhairs, repeating a theme from Unix's early days ten years before, liked to see themselves as rebels against a corporate empire; one of the small companies put out a poster showing an X-wing-like space fighter marked “BSD” speeding away from a huge AT&T ‘death star’ logo left broken and in flames. Thus we fiddled while Rome burned.

But something else happened in the year of the AT&T divestiture that would have more long-term importance for Unix. A programmer/linguist named Larry Wall quietly invented the patch(1) utility. The patch program, a simple tool that applies changebars generated by diff(1) to a base file, meant that Unix developers could cooperate by passing around patch sets — incremental changes to code — rather than entire code files. This was important not only because patches are less bulky than full files, but because patches would often apply cleanly even if much of the base file had changed since the patch-sender fetched his copy. With this tool, streams of development on a common source-code base could diverge, run in parallel, and re-converge. The patch program did more than any other single tool to enable collaborative development over the Internet — a method that would revitalize Unix after 1990.

In 1985 Intel shipped the first 386 chip, capable of addressing 4 gigabytes of memory with a flat address space. The clumsy segment addressing of the 8086 and 286 became immediately obsolete. This was big news, because it meant that for the first time, a microprocessor in the dominant Intel family had the capability to run Unix without painful compromises. The handwriting was on the wall for Sun and the other workstation makers. They failed to see it.

1985 was also the year that Richard Stallman issued the GNU manifesto [Stallman] and launched the Free Software Foundation. Very few people took him or his GNU project seriously, a judgment that turned out to be seriously mistaken. In an unrelated development of the same year, the originators of the X window system released it as source code without royalties, restrictions, or license code. As a direct result of this decision, it became a safe neutral area for collaboration between Unix vendors, and defeated proprietary contenders to become Unix's graphics engine.

Serious standardization efforts aimed at reconciling the System V and Berkeley APIs also began in 1983 with the /usr/group standard. This was followed in 1985 by the POSIX standards, an effort backed by the IEEE. These described the intersection set of the BSD and SVR3 (System V Release 3) calls, with the superior Berkeley signal handling and job control but with SVR3 terminal control. All later Unix standards would incorporate POSIX at their core, and later Unixes would adhere to it closely. The only major addition to the modern Unix kernel API to come afterwards was BSD sockets.

In 1986 Larry Wall, previously the inventor of patch(1), began work on Perl, which would become the first and most widely used of the open-source scripting languages. In early 1987 the first version of the GNU C compiler appeared, and by the end of 1987 the core of the GNU toolset was falling into place : editor, compiler, debugger, and other basic development tools. Meanwhile, the X windowing system was beginning to show up on relatively inexpensive workstations. Together, these would provide the armature for the open-source Unix developments of the 1990s.

1986 was also the year that PC technology broke free of IBM's grip. IBM, still trying to preserve a price-vs.-power curve across its product line that would favor its high-margin mainframe business, rejected the 386 for most of its new line of PS/2 computers in favor of the weaker 286. The PS/2 series, designed around a proprietary bus architecture to lock out clonemakers, became a colossally expensive failure.[17] Compaq, the most aggressive of the clonemakers, trumped IBM's move by releasing the first 386 machine. Even with a clock speed of a mere 16 MHz, the 386 made a tolerable Unix machine. It was the first PC of which that could be said.

It was beginning to be possible to imagine that Stallman's GNU project might mate with 386 machines to produce Unix workstations almost an order of magnitude less costly than anyone was offering. Curiously, no one seems to have actually got this far in their thinking. Most Unix programmers, coming from the minicomputer and workstation worlds, continued to disdain cheap 80x86 machines in favor of more elegant 68000-based designs. And, though a lot of programmers contributed to the GNU project, among Unix people it tended to be considered a quixotic gesture that was unlikely to have near-term practical consequences.

The Unix community had never lost its rebel streak. But in retrospect, we were nearly as blind to the future bearing down on us as IBM or AT&T. Not even Richard Stallman, who had declared a moral crusade against proprietary software a few years before, really understood how badly the productization of Unix had damaged the community around it; his concerns were with more abstract and long-term issues. The rest of us kept hoping that some clever variation on the corporate formula would solve the problems of fragmentation, wretched marketing, and strategic drift, and redeem Unix's pre-divestiture promise. But worse was still to come.

1988 was the year Ken Olsen (CEO of DEC) famously described Unix as “snake oil”. DEC had been shipping its own variant of Unix on PDP-11s since 1982, but really wanted the business to go to its proprietary VMS operating system. DEC and the minicomputer industry were in deep trouble, swamped by waves of powerful low-cost machines coming out of Sun Microsystems and the rest of the workstation vendors. Most of those workstations ran Unix.

But the Unix industry's own problems were growing more severe. In 1988 AT&T took a 20% stake in Sun Microsystems. These two companies, the leaders in the Unix market, were beginning to wake up to the threat posed by PCs, IBM, and Microsoft, and to realize that the preceding five years of bloodletting had gained them little. The AT&T/Sun alliance and the development of technical standards around POSIX eventually healed the breach between the System V and BSD Unix lines. But the second phase of the Unix wars began when the second-tier vendors (IBM, DEC, Hewlett-Packard, and others) formed the Open Software Foundation and lined up against the AT&T/Sun axis (represented by Unix International). More rounds of Unix fighting Unix ensued.

Meanwhile, Microsoft was making billions in the home and small-business markets that the warring Unix factions had never found the will to address. The 1990 release of Windows 3.0 — the first successful graphical operating system from Redmond — cemented Microsoft's dominance, and created the conditions that would allow them to flatten and monopolize the market for desktop applications in the 1990s.

The years from 1989 to 1993 were the darkest in Unix's history. It appeared then that all the Unix community's dreams had failed. Internecine warfare had reduced the proprietary Unix industry to a squabbling shambles that never summoned either the determination or the capability to challenge Microsoft. The elegant Motorola chips favored by most Unix programmers had lost out to Intel's ugly but inexpensive processors. The GNU project failed to produce the free Unix kernel it had been promising since 1985, and after years of excuses its credibility was beginning to wear thin. PC technology was being relentlessly corporatized. The pioneering Unix hackers of the 1970s were hitting middle age and slowing down. Hardware was getting cheaper, but Unix was still too expensive. We were belatedly becoming aware that the old monopoly of IBM had yielded to a newer monopoly of Microsoft, and Microsoft's mal-engineered software was rising around us like a tide of sewage.

Blows against the Empire: 1991-1995

The first glimmer of light in the darkness was the 1990 effort by William Jolitz to port BSD onto a 386 box, publicized by a series of magazine articles beginning in 1991. The 386BSD port was possible because, partly influenced by Stallman, Berkeley hacker Keith Bostic had begun an effort to clean AT&T proprietary code out of the BSD sources in 1988. But the 386BSD project took a severe blow when, near the end of 1991, Jolitz walked away from it and destroyed his own work. There are conflicting explanations, but a common thread in all is that Jolitz wanted his code to be released as unencumbered source and was upset when the corporate sponsors of the project opted for a more proprietary licensing model.

In August 1991 Linus Torvalds, then an unknown university student from Finland, announced the Linux project. Torvalds is on record that one of his main motivations was the high cost of Sun's Unix at his university. Torvalds has also said that he would have joined the BSD effort had he known of it, rather than founding his own. But 386BSD was not shipped until early 1992, some months after the first Linux release.

The importance of both these projects became clear only in retrospect. At the time, they attracted little notice even within the Internet hacker culture — let alone in the wider Unix community, which was still fixated on more capable machines than PCs, and on trying to reconcile the special properties of Unix with the conventional proprietary model of a software business.

It would take another two years and the great Internet explosion of 1993–1994 before the true importance of Linux and the open-source BSD distributions became evident to the rest of the Unix world. Unfortunately for the BSDers, an AT&T lawsuit against BSDI (the startup company that had backed the Jolitz port) consumed much of that time and motivated some key Berkeley developers to switch to Linux.

Code copying and theft of trade secrets was alleged. The actual infringing code was not identified for nearly two years. The lawsuit could have dragged on for much longer but for the fact that Novell bought USL from AT&T and sought a settlement. In the end, three files were removed from the 18,000 that made up the distribution, and a number of minor changes were made to other files. In addition, the University agreed to add USL copyrights to about 70 files, with the stipulation that those files continued to be freely redistributed.

– Marshall Kirk McKusick

The settlement set an important precedent by freeing an entire working Unix from proprietary control, but its effects on BSD itself were dire. Matters were not helped when, in 1992–1994, the Computer Science Research Group at Berkeley shut down; afterwards, factional warfare within the BSD community split it into three competing development efforts. As a result, the BSD lineage lagged behind Linux at a crucial time and lost to it the lead position in the Unix community.

The Linux and BSD development efforts were native to the Internet in a way previous Unixes had not been. They relied on distributed development and Larry Wall's patch(1) tool, and recruited developers via email and through Usenet newsgroups. Accordingly, they got a tremendous boost when Internet Service Provider businesses began to proliferate in 1993, enabled by changes in telecomm technology and the privatization of the Internet backbone that are outside the scope of this history. The demand for cheap Internet was created by something else: the 1991 invention of the World Wide Web. The Web was the “killer app” of the Internet, the graphical user interface technology that made it irresistible to a huge population of nontechnical end users.

The mass-marketing of the Internet both increased the pool of potential developers and lowered the transaction costs of distributed development. The results were reflected in efforts like XFree86, which used the Internet-centric model to build a more effective development organization than that of the official X Consortium. The first XFree86 in 1992 gave Linux and the BSDs the graphical-user-interface engine they had been missing. Over the next decade XFree86 would lead in X development, and an increasing portion of the X Consortium's activity would come to consist of funneling innovations originated in the XFree86 community back to the Consortium's industrial sponsors.

By late 1993, Linux had both Internet capability and X. The entire GNU toolkit had been hosted on it from the beginning, providing high-quality development tools. Beyond GNU tools, Linux acted as a basin of attraction, collecting and concentrating twenty years of open-source software that had previously been scattered across a dozen different proprietary Unix platforms. Though the Linux kernel was still officially in beta (at 0.99 level), it was remarkably crash-free. The breadth and quality of the software in Linux distributions was already that of a production-ready operating system.

A few of the more flexible-minded among old-school Unix developers began to notice that the long-awaited dream of a cheap Unix system for everybody had snuck up on them from an unexpected direction. It didn't come from AT&T or Sun or any of the traditional vendors. Nor did it rise out of an organized effort in academia. It was a bricolage that bubbled up out of the Internet by what seemed like spontaneous generation, appropriating and recombining elements of the Unix tradition in surprising ways.

Elsewhere, corporate maneuvering continued. AT&T divested its interest in Sun in 1992; then sold its Unix Systems Laboratories to Novell in 1993; Novell handed off the Unix trademark to the X/Open standards group in 1994; AT&T and Novell joined OSF in 1994, finally ending the Unix wars. In 1995 SCO bought UnixWare (and the rights to the original Unix sources) from Novell. In 1996, X/Open and OSF merged, creating one big Unix standards group.

But the conventional Unix vendors and the wreckage of their wars came to seem steadily less and less relevant. The action and energy in the Unix community were shifting to Linux and BSD and the open-source developers. By the time IBM, Intel, and SCO announced the Monterey project in 1998 — a last-gasp attempt to merge One Big System out of all the proprietary Unixes left standing — developers and the trade press reacted with amusement, and the project was abruptly canceled in 2001 after three years of going nowhere.

The industry transition could not be said to have completed until 2000, when SCO sold UnixWare and the original Unix source-code base to Caldera — a Linux distributor. But after 1995, the story of Unix became the story of the open-source movement. There's another side to that story; to tell it, we'll need to return to 1961 and the origins of the Internet hacker culture.


[13] Ken Thompson reminded me that today's cellphones have more RAM than the PDP-7 had RAM and disk storage combined; a large disk, in those days, was less than a megabyte of storage.

[14] There is a Web FAQ on the PDP computers that explains the otherwise extremely obscure PDP-7's place in history.

[15] The version 7 manuals can be browsed on-line at

[16] UUCP was hot stuff when a fast modem was 300 baud.

[17] The PS/2 did, however, leave one mark on later PCs — they made the mouse a standard peripheral, which is why the mouse connector on the back of your chassis is called a “PS/2 port”.

Origins and History of the Hackers, 1961-1995

The Unix tradition is an implicit culture that has always carried with it more than just a bag of technical tricks. It transmits a set of values about beauty and good design; it has legends and folk heroes. Intertwined with the history of the Unix tradition is another implicit culture that is more difficult to label neatly. It has its own values and legends and folk heroes, partly overlapping with those of the Unix tradition and partly derived from other sources. It has most often been called the “hacker culture”, and since 1998 has largely coincided with what the computer trade press calls “the open source movement”.

The relationships between the Unix tradition, the hacker culture, and the open-source movement are subtle and complex. They are not simplified by the fact that all three implicit cultures have frequently been expressed in the behaviors of the same human beings. But since 1990 the story of Unix is largely the story of how the open-source hackers changed the rules and seized the initiative from the old-line proprietary Unix vendors. Therefore, the other half of the history behind today's Unix is the history of the hackers.

At Play in the Groves of Academe: 1961-1980

The roots of the hacker culture can be traced back to 1961, the year MIT took delivery of its first PDP-1 minicomputer. The PDP-1 was one of the earliest interactive computers, and (unlike other machines) of the day was inexpensive enough that time on it did not have to be rigidly scheduled. It attracted a group of curious students from the Tech Model Railroad Club who experimented with it in a spirit of fun. Hackers: Heroes of the Computer Revolution[Levy] entertainingly describes the early days of the club. Their most famous achievement was SPACEWAR, a game of dueling rocketships loosely inspired by the Lensman space operas of E.E. “Doc” Smith.[18]

Several of the TMRC experimenters later went on to become core members of the MIT Artificial Intelligence Lab, which in the 1960s and 1970s became one of the world centers of cutting-edge computer science. They took some of TMRC's slang and in-jokes with them, including a tradition of elaborate (but harmless) pranks called “hacks”. The AI Lab programmers appear to have been the first to describe themselves as “hackers”.

After 1969 the MIT AI Lab was connected, via the early ARPANET, to other leading computer science research laboratories at Stanford, Bolt Beranek & Newman, Carnegie-Mellon University and elsewhere. Researchers and students got the first foretaste of the way fast network access abolishes geography, often making it easier to collaborate and form friendships with distant people on the net than it would be to do likewise with colleagues closer-by but less connected.

Software, ideas, slang, and a good deal of humor flowed over the experimental ARPANET links. Something like a shared culture began to form. One of its earliest and most enduring artifacts was the Jargon File, a list of shared slang terms that originated at Stanford in 1973 and went through several revisions at MIT after 1976. Along the way it accumulated slang from CMU, Yale, and other ARPANET sites.

Technically, the early hacker culture was largely hosted on PDP-10 minicomputers. They used a variety of operating systems that have since passed into history: TOPS-10, TOPS-20, Multics, ITS, SAIL. They programmed in assembler and dialects of Lisp. PDP-10 hackers took over running the ARPANET itself because nobody else wanted the job. Later, they became the founding cadre of the Internet Engineering Task Force (IETF) and originated the tradition of standardization through Requests For Comment (RFCs).

Socially, they were young, exceptionally bright, almost entirely male, dedicated to programming to the point of addiction, and tended to have streaks of stubborn nonconformism — what years later would be called ‘geeks’. They, too, tended to be shaggy hippies and hippie-wannabes. They, too, had a vision of computers as community-building devices. They read Robert Heinlein and J. R. R. Tolkien, played in the Society for Creative Anachronism, and tended to have a weakness for puns. Despite their quirks (or perhaps because of them!) many of them were among the brightest programmers in the world.

They were not Unix programmers. The early Unix community was drawn largely from the same pool of geeks in academia and government or commercial research laboratories, but the two cultures differed in important ways. One that we've already touched on is the weak networking of early Unix. There was effectively no Unix-based ARPANET access until after 1980, and it was uncommon for any individual to have a foot in both camps.

Collaborative development and the sharing of source code was a valued tactic for Unix programmers. To the early ARPANET hackers, on the other hand, it was more than a tactic: it was something rather closer to a shared religion, partly arising from the academic “publish or perish” imperative and (in its more extreme versions) developing into an almost Chardinist idealism about networked communities of minds. The most famous of these hackers, Richard M. Stallman, became the ascetic saint of that religion.

Internet Fusion and the Free Software Movement: 1981-1991

After 1983 and the BSD port of TCP/IP, the Unix and ARPANET cultures began to fuse together. This was a natural development once the communication links were in place, since both cultures were composed of the same kind of people (indeed, in a few but significant cases the same people). ARPANET hackers learned C and began to speak the jargon of pipes, filters, and shells; Unix programmers learned TCP/IP and started to call each other “hackers”. The process of fusion was accelerated after the Project Jupiter cancellation in 1983 killed the PDP-10's future. By 1987 the two cultures had merged so completely that most hackers programmed in C and casually used slang terms that went back to the Tech Model Railroad Club of twenty-five years earlier.

(In 1979 I was unusual in having strong ties to both the Unix and ARPANET cultures. In 1985 that was no longer unusual. By the time I expanded the old ARPANET Jargon File into the New Hacker's Dictionary[Raymond96] in 1991, the two cultures had effectively fused. The Jargon File, born on the ARPANET but revised on Usenet, aptly symbolized the merger.)

But TCP/IP networking and slang were not the only things the post-1980 hacker culture inherited from its ARPANET roots. It also got Richard Stallman, and Stallman's moral crusade.

Richard M. Stallman (generally known by his login name, RMS) had already proved by the late 1970s that he was one of the most able programmers alive. Among his many inventions was the Emacs editor. For RMS, the Jupiter cancellation in 1983 only finished off a disintegration of the MIT AI Lab culture that had begun a few years earlier as many of its best went off to help run competing Lisp-machine companies. RMS felt ejected from a hacker Eden, and decided that proprietary software was to blame.

In 1983 Stallman founded the GNU project, aimed at writing an entire free operating system. Though Stallman was not and had never been a Unix programmer, under post-1980 conditions implementing a Unix-like operating system became the obvious strategy to pursue. Most of RMS's early contributors were old-time ARPANET hackers newly decanted into Unix-land, in whom the ethos of code-sharing ran rather stronger than it did among those with a more Unix-centered background.

In 1985, RMS published the GNU Manifesto. In it he consciously created an ideology out of the values of the pre-1980 ARPANET hackers — complete with a novel ethico-political claim, a self-contained and characteristic discourse, and an activist plan for change. RMS aimed to knit the diffuse post-1980 community of hackers into a coherent social machine for achieving a single revolutionary purpose. His behavior and rhetoric half-consciously echoed Karl Marx's attempts to mobilize the industrial proletariat against the alienation of their work.

RMS's manifesto ignited a debate that is still live in the hacker culture today. His program went way beyond maintaining a codebase, and essentially implied the abolition of intellectual-property rights in software. In pursuit of this goal, RMS popularized the term “free software”, which was the first attempt to label the product of the entire hacker culture. He wrote the General Public License (GPL), which was to become both a rallying point and a focus of great controversy, for reasons we will examine in Chapter 16. You can learn more about RMS's position and the Free Software Foundation at the GNU website.

The term “free software” was partly a description and partly an attempt to define a cultural identity for hackers. On one level, it was quite successful. Before RMS, people in the hacker culture recognized each other as fellow-travelers and used the same slang, but nobody bothered arguing about what a ‘hacker’ is or should be. After him, the hacker culture became much more self-conscious; value disputes (often framed in RMS's language even by those who opposed his conclusions) became a normal feature of debate. RMS, a charismatic and polarizing figure, himself became so much a culture hero that by the year 2000 he could hardly be distinguished from his legend. Free as in Freedom[Williams] gives us an excellent portrait.

RMS's arguments influenced the behavior even of many hackers who remained skeptical of his theories. In 1987, he persuaded the caretakers of BSD Unix that cleaning out AT&T's proprietary code so they could release an unencumbered version would be a good idea. However, despite his determined efforts over more than fifteen years, the post-1980 hacker culture never unified around his ideological vision.

Other hackers were rediscovering open, collaborative development without secrets for more pragmatic, less ideological reasons. A few buildings away from Richard Stallman's 9th-floor office at MIT, the X development team thrived during the late 1980s. It was funded by Unix vendors who had argued each other to a draw over the control and intellectual-property-rights issues surrounding the X windowing system, and saw no better alternative than to leave it free to everyone. In 1987–1988 the X development prefigured the really huge distributed communities that would redefine the leading edge of Unix five years later.

X was one of the first large-scale open-source projects to be developed by a disparate team of individuals working for different organizations spread across the globe. E-mail allowed ideas to move rapidly among the group so that issues could be resolved as quickly as necessary, and each individual could contribute in whatever capacity suited them best. Software updates could be distributed in a matter of hours, enabling every site to act in a concerted manner during development. The net changed the way software could be developed.

– Keith Packard

The X developers were no partisans of the GNU master plan, but they weren't actively opposed to it, either. Before 1995 the most serious opposition to the GNU plan came from the BSD developers. The BSD people, who remembered that they had been writing freely redistributable and modifiable software years before RMS's manifesto, rejected GNU's claim to historical and ideological primacy. They specifically objected to the infectious or “viral” property of the GPL, holding out the BSD license as being “more free” because it placed fewer restrictions on the reuse of code.

It did not help RMS's case that, although his Free Software Foundation had produced most of the rest of a full software toolkit, it failed to deliver the central piece. Ten years after the founding of the GNU project, there was still no GNU kernel. While individual tools like Emacs and GCC proved tremendously useful, GNU without a kernel neither threatened the hegemony of proprietary Unixes nor offered an effective counter to the rising problem of the Microsoft monopoly.

After 1995 the debate over RMS's ideology took a somewhat different turn. Opposition to it became closely associated with both Linus Torvalds and the author of this book.

Linux and the Pragmatist Reaction: 1991-1998

Even as the HURD (the GNU kernel) effort was stalling, new possibilities were opening up. In the early 1990s the combination of cheap, powerful PCs with easy Internet access proved a powerful lure for a new generation of young programmers looking for challenges to test their mettle. The user-space toolkit written by the Free Software Foundation suggested a way forward that was free of the high cost of proprietary software development tools. Ideology followed economics rather than leading the charge; some of the newbies signed up with RMS's crusade and adopted the GPL as their banner, and others identified more with the Unix tradition as a whole and joined the anti-GPL camp, but most dismissed the whole dispute as a distraction and just wrote code.

Linus Torvalds neatly straddled the GPL/anti-GPL divide by using the GNU toolkit to surround the Linux kernel he had invented and the GPL's infectious properties to protect it, but rejecting the ideological program that went with RMS's license. Torvalds affirmed that he thought free software better in general but occasionally used proprietary programs. His refusal to be a zealot even in his own cause made him tremendously attractive to the majority of hackers who had been uncomfortable with RMS's rhetoric, but had lacked any focus or convincing spokesperson for their skepticism.

Torvalds's cheerful pragmatism and adept but low-key style catalyzed an astonishing string of victories for the hacker culture in the years 1993–1997, including not merely technical successes but the solid beginnings of a distribution, service, and support industry around the Linux operating system. As a result his prestige and influence skyrocketed. Torvalds became a hero on Internet time; by 1995, he had achieved in just four years the kind of culture-wide eminence that RMS had required fifteen years to earn — and far exceeded Stallman's record at selling “free software” to the outside world. By contrast with Torvalds, RMS's rhetoric began to seem both strident and unsuccessful.

Between 1991 and 1995 Linux went from a proof-of-concept surrounding an 0.1 prototype kernel to an operating system that could compete on features and performance with proprietary Unixes, and beat most of them on important statistics like continuous uptime. In 1995, Linux found its killer app: Apache, the open-source webserver. Like Linux, Apache proved remarkably stable and efficient. Linux machines running Apache quickly became the platform of choice for ISPs worldwide; Apache captured about 60% of websites,[19] handily beating out both of its major proprietary competitors.

The one thing Torvalds did not offer was a new ideology — a new rationale or generative myth of hacking, and a positive discourse to replace RMS's hostility to intellectual property with a program more attractive to people both within and outside the hacker culture. I inadvertently supplied this lack in 1997 as a result of trying to understand why Linux's development had not collapsed in confusion years before. The technical conclusions of my published papers [Raymond01] will be summarized in Chapter 19. For this historical sketch, it will be sufficient to note the impact of the first one's central formula: “Given a sufficiently large number of eyeballs, all bugs are shallow”.

This observation implied something nobody in the hacker culture had dared to really believe in the preceding quarter-century: that its methods could reliably produce software that was not just more elegant but more reliable and better than our proprietary competitors' code. This consequence, quite unexpectedly, turned out to present exactly the direct challenge to the discourse of “free software” that Torvalds himself had never been interested in mounting. For most hackers and almost all nonhackers, “Free software because it works better” easily trumped “Free software because all software should be free”.

The paper's contrast between ‘cathedral’ (centralized, closed, controlled, secretive) and ‘bazaar’ (decentralized, open, peer-review-intensive) modes of development became a central metaphor in the new thinking. In an important sense this was merely a return to Unix's pre-divestiture roots — it is continuous with McIlroy's 1991 observations about the positive effects of peer pressure on Unix development in the early 1970s and Dennis Ritchie's 1979 reflections on fellowship, cross-fertilized with the early ARPANET's academic tradition of peer review and with its idealism about distributed communities of mind.

In early 1998, the new thinking helped motivate Netscape Communications to release the source code of its Mozilla browser. The press attention surrounding that event took Linux to Wall Street, helped drive the technology-stock boom of 1999–2001, and proved to be a turning point in both the history of the hacker culture and of Unix.


[18] SPACEWAR was not related to Ken Thompson's Space Travel game, other than by the fact that both appealed to science-fiction fans.

[19] Current and historical webserver share figures are available at the monthly Netcraft Web Server Survey.

The Open-Source Movement: 1998 and Onward

By the time of the Mozilla release in 1998, the hacker community could best be analyzed as a loose collection of factions or tribes that included Richard Stallman's Free Software Movement, the Linux community, the Perl community, the Apache community, the BSD community, the X developers, the Internet Engineering Task Force (IETF), and at least a dozen others. These factions overlap, and an individual developer would be quite likely to be affiliated with two or more.

A tribe might be grouped around a particular codebase that they maintain, or around one or more charismatic influence leaders, or around a language or development tool, or around a particular software license, or around a technical standard, or around a caretaker organization for some part of the infrastructure. Prestige tends to correlate with longevity and historical contribution as well as more obvious drivers like current market-share and mind-share; thus, perhaps the most universally respected of the tribes is the IETF, which can claim continuity back to the beginnings of the ARPANET in 1969. The BSD community, with continuous traditions back to the late 1970s, commands considerable prestige despite having a much lower installation count than Linux. Stallman's Free Software Movement, dating back to the early 1980s, ranks among the senior tribes both on historical contribution and as the maintainer of several of the software tools in heaviest day-to-day use.

After 1995 Linux acquired a special role as both the unifying platform for most of the community's other software and the hackers' most publicly recognizable brand name. The Linux community showed a corresponding tendency to absorb other sub-tribes — and, for that matter, to co-opt and absorb the hacker factions associated with proprietary Unixes. The hacker culture as a whole began to draw together around a common mission: push Linux and the bazaar development model as far as it could go.

Because the post-1980 hacker culture had become so deeply rooted in Unix, the new mission was implicitly a brief for the triumph of the Unix tradition. Many of the hacker community's senior leaders were also Unix old-timers, still bearing scars from the post-divestiture civil wars of the 1980s and getting behind Linux as the last, best hope to fulfill the rebel dreams of the early Unix days.

The Mozilla release helped further concentrate opinions. In March of 1998 an unprecedented summit meeting of community influence leaders representing almost all of the major tribes convened to consider common goals and tactics. That meeting adopted a new label for the common development method of all the factions: open source.

Within six months almost all the tribes in the hacker community would accept “open source” as its new banner. Older groups like IETF and the BSD developers would begin to apply it retrospectively to what they had been doing all along. In fact, by 2000 the rhetoric of open source would not just unify the hacker culture's present practice and plans for the future, but re-color its view of its own past.

The galvanizing effect of the Netscape announcement, and of the new prominence of Linux, reached well beyond the Unix community and the hacker culture. Beginning in 1995, developers from various platforms in the path of Microsoft's Windows juggernaut (MacOS; Amiga; OS/2; DOS; CP/M; the weaker proprietary Unixes; various mainframe, minicomputer, and obsolete microcomputer operating systems) had banded together around Sun Microsystems's Java language. Many disgruntled Windows developers joined them in hopes of maintaining at least some nominal independence from Microsoft. But Sun's handling of Java was (as we discuss in Chapter 14) clumsy and alienating on several levels. Many Java developers liked what they saw in the nascent open-source movement, and followed Netscape's lead into Linux and open source just as they had previously followed Netscape into Java.

Open-source activists welcomed the surge of immigrants from everywhere. The old Unix hands began to share the new immigrants' dreams of not merely passively out-enduring the Microsoft monopoly, but actually reclaiming key markets from it. The open-source community as a whole prepared a major push for mainstream respectability, and began to welcome alliances with major corporations that increasingly feared losing control of their own businesses as Microsoft's lock-in tactics grew ever bolder.

There was one exception: Richard Stallman and the Free Software Movement. “Open source” was explicitly intended to replace Stallman's preferred “free software” with a public label that was ideologically neutral, acceptable both to historically opposed groups like the BSD hackers and those who did not wish to take a position in the GPL/anti-GPL debate. Stallman flirted with adopting the term, then rejected it on the grounds that it failed to represent the moral position that was central to his thinking. The Free Software Movement has since insisted on its separateness from “open source”, creating perhaps the most significant political fissure in the hacker culture of 2003.

The other (and more important) intention behind “open source” was to present the hacker community's methods to the rest of the world (especially the business mainstream) in a more market-friendly, less confrontational way. In this role, fortunately, it proved an unqualified success — and led to a revival of interest in the Unix tradition from which it sprang.

The Lessons of Unix History

The largest-scale pattern in the history of Unix is this: when and where Unix has adhered most closely to open-source practices, it has prospered. Attempts to proprietarize it have invariably resulted in stagnation and decline.

In retrospect, this should probably have become obvious much sooner than it did. We lost ten years after 1984 learning our lesson, and it would probably serve us very ill to ever again forget it.

Being smarter than anyone else about important but narrow issues of software design didn't prevent us from being almost completely blind about the consequences of interactions between technology and economics that were happening right under our noses. Even the most perceptive and forward-looking thinkers in the Unix community were at best half-sighted. The lesson for the future is that over-committing to any one technology or business model would be a mistake — and maintaining the adaptive flexibility of our software and the design tradition that goes with it is correspondingly imperative.

Another lesson is this: Never bet against the cheap plastic solution. Or, equivalently, the low-end/high-volume hardware technology almost always ends up climbing the power curve and winning. The economist Clayton Christensen calls this disruptive technology and showed in The Innovator's Dilemma[Christensen] how this happened with disk drives, steam shovels, and motorcycles. We saw it happen as minicomputers displaced mainframes, workstations and servers replaced minis, and commodity Intel machines replaced workstations and servers. The open-source movement is winning by commoditizing software. To prosper, Unix needs to maintain the knack of co-opting the cheap plastic solution rather than trying to fight it.

Finally, the old-school Unix community failed in its efforts to be “professional” by welcoming in all the command machinery of conventional corporate organization, finance, and marketing. We had to be rescued from our folly by a rebel alliance of obsessive geeks and creative misfits —who then proceeded to show us that professionalism and dedication really meant what we had been doing before we succumbed to the mundane persuasions of “sound business practices”.

The application of these lessons with respect to software technologies other than Unix is left as an easy exercise for the reader.

Chapter 3. Contrasts

If you have any trouble sounding condescending, find a Unix user to show you how it's done.

– Scott Adams Dilbert newsletter 3.0, 1994

The design of operating systems conditions the style of software development under them in many ways both obvious and subtle. Much of this book traces connections between the design of the Unix operating system and the philosophy of program design that has evolved around it. For contrast, it will therefore be instructive to compare the classic Unix way with the styles of design and programming native to other major operating systems.

The Elements of Operating-System Style

Before we can start discussing specific operating systems, we'll need an organizing framework for the ways that operating-system design can affect programming style for good or ill.

Overall, the design and programming styles associated with different operating systems seem to derive from three different sources: (a) the intentions of the operating-system designers, (b) uniformities forced on designs by costs and limitations in the programming environment, and (c) random cultural drift, early practices becoming traditional simply because they were there first.

Even if we take it as given that there is some random cultural drift in every operating-system community, considering the intentions of the designers and the costs and limitations of the results does reveal some interesting patterns that can help us understand the Unix style better by contrast. We can make the patterns explicit by analyzing some of the most important ways that operating systems differ.

What Is the Operating System's Unifying Idea?

Unix has a couple of unifying ideas or metaphors that shape its APIs and the development style that proceeds from them. The most important of these are probably the “everything is a file” model and the pipe metaphor[20] built on top of it. In general, development style under any given operating system is strongly conditioned by the unifying ideas baked into the system by its designers — they percolate upwards into applications programming from the models provided by system tools and APIs.

Accordingly, the most basic question to ask in contrasting Unix with another operating system is: Does it have unifying ideas that shape its development, and if so how do they differ from Unix's?

To design the perfect anti-Unix, have no unifying idea at all, just an incoherent pile of ad-hoc features.

Multitasking Capability

One of the most basic ways operating systems can differ is in the extent to which they can support multiple concurrent processes. At the lowest end (such as DOS or CP/M) the operating system is basically a sequential program loader with no capacity to multitask at all. Operating systems of this kind are no longer competitive on general-purpose computers.

At the next level up, an operating system may have cooperative multitasking. Such systems can support multiple processes, but a process has to voluntarily give up its hold on the processor before the next one can run (thus, simple programming errors can readily freeze the machine). This style of operating system was a transient adaptation to hardware that was powerful enough for concurrency but lacked either a periodic clock interrupt[21] or a memory-management unit or both; it, too, is obsolete and no longer competitive.

Unix has preemptive multitasking, in which timeslices are allocated by a scheduler which routinely interrupts or pre-empts the running process in order to hand control to the next one. Almost all modern operating systems support preemption.

Note that “multitasking” is not the same as “multiuser”. An operating system can be multitasking but single-user, in which case the facility is used to support a single console and multiple background processes. True multiuser support requires multiple user privilege domains, a feature we'll cover in the discussion of internal boundaries a bit further on.

To design the perfect anti-Unix, don't support multitasking at all — or, support multitasking but cripple it by surrounding process management with a lot of restrictions, limitations, and special cases that mean it's quite difficult to get any actual use out of multitasking.

Cooperating Processes

In the Unix experience, inexpensive process-spawning and easy inter-process communication (IPC) makes a whole ecology of small tools, pipes, and filters possible. We'll explore this ecology in Chapter 7; here, we need to point out some consequences of expensive process-spawning and IPC.

The pipe was technically trivial, but profound in its effect. However, it would not have been trivial without the fundamental unifying notion of the process as an autonomous unit of computation, with process control being programmable. As in Multics, a shell was just another process; process control did not come from God inscribed in JCL.

– Doug McIlroy

If an operating system makes spawning new processes expensive and/or process control is difficult and inflexible, you'll usually see all of the following consequences:

These are examples of common stylistic traits (even in applications programming) being driven by a limitation in the OS environment.

A subtle but important property of pipes and the other classic Unix IPC methods is that they require communication between programs to be held down to a level of simplicity that encourages separation of function. Conversely, the result of having no equivalent of the pipe is that programs can only be designed to cooperate by building in full knowledge of each others' internals.

In operating systems without flexible IPC and a strong tradition of using it, programs communicate by sharing elaborate data structures. Because the communication problem has to be solved anew for all programs every time another is added to the set, the complexity of this solution rises as the square of the number of cooperating programs. Worse than that, any change in one of the exposed data structures can induce subtle bugs in an arbitrarily large number of other programs.

Word and Excel and PowerPoint and other Microsoft programs have intimate — one might say promiscuous — knowledge of each others' internals. In Unix, one tries to design programs to operate not specifically with each other, but with programs as yet unthought of.

– Doug McIlroy

We'll return to this theme in Chapter 7.

To design the perfect anti-Unix, make process-spawning very expensive, make process control difficult and inflexible, and leave IPC as an unsupported or half-supported afterthought.

Internal Boundaries

Unix has wired into it an assumption that the programmer knows best. It doesn't stop you or request confirmation when you do dangerous things with your own data, like issuing rm -rf *. On the other hand, Unix is rather careful about not letting you step on other people's data. In fact, Unix encourages you to have multiple accounts, each with its own attached and possibly differing privileges, to help you protect yourself from misbehaving programs.[22] System programs often have their own pseudo-user accounts to confer access to special system files without requiring unlimited (or superuser) access.

Unix has at least three levels of internal boundaries that guard against malicious users or buggy programs. One is memory management; Unix uses its hardware's memory management unit (MMU) to ensure that separate processes are prevented from intruding on the others' memory-address spaces. A second is the presence of true privilege groups for multiple users — an ordinary (nonroot) user's processes cannot alter or read another user's files without permission. A third is the confinement of security-critical functions to the smallest possible pieces of trusted code. Under Unix, even the shell (the system command interpreter) is not a privileged program.

The strength of an operating system's internal boundaries is not merely an abstract issue of design: It has important practical consequences for the security of the system.

To design the perfect anti-Unix, discard or bypass memory management so that a runaway process can crash, subvert, or corrupt any running program. Have weak or nonexistent privilege groups, so users can readily alter each others' files and the system's critical data (e.g., a macro virus, having seized control of your word processor, can format your hard drive). And trust large volumes of code, like the entire shell and GUI, so that any bug or successful attack on that code becomes a threat to the entire system.

File Attributes and Record Structures

Unix files have neither record structure nor attributes. In some operating systems, files have an associated record structure; the operating system (or its service libraries) knows about files with a fixed record length, or about text line termination and whether CR/LF is to be read as a single logical character.

In other operating systems, files and directories can have name/attribute pairs associated with them — out-of-band data used (for example) to associate a document file with an application that understands it. (The classic Unix way to handle these associations is to have applications recognize ‘magic numbers’, or other type data within the file itself.)

OS-level record structures are generally an optimization hack, and do little more than complicate APIs and programmers' lives. They encourage the use of opaque record-oriented file formats that generic tools like text editors cannot read properly.

File attributes can be useful, but (as we will see in Chapter 20) can raise some awkward semantic issues in a world of byte-stream-oriented tools and pipes. When file attributes are supported at the operating-system level, they predispose programmers to use opaque formats and lean on the file attributes to tie them to the specific applications that interpret them.

To design the perfect anti-Unix, have a cumbersome set of record structures that make it a hit-or-miss proposition whether any given tool will be able to even read a file as the writer intended it. Add file attributes and have the system depend on them heavily, so that the semantics of a file will not be determinable by looking at the data within it.

Binary File Formats

If your operating system uses binary formats for critical data (such as user-account records) it is likely that no tradition of readable textual formats for applications will develop. We explain in more detail why this is a problem in Chapter 5. For now it's sufficient to note the following consequences:

To design the perfect anti-Unix, make all file formats binary and opaque, and require heavyweight tools to read and edit them.

Preferred User Interface Style

In Chapter 11 we will develop in some detail the consequences of the differences between command-line interfaces (CLIs) and graphical user interfaces (GUIs). Which kind an operating system's designers choose as its normal mode of presentation will affect many aspects of the design, from process scheduling and memory management on up to the application programming interfaces (APIs) presented for applications to use.

It has been enough years since the first Macintosh that very few people need to be convinced that weak GUI facilities in an operating system are a problem. The Unix lesson is the opposite: that weak CLI facilities are a less obvious but equally severe deficit.

If the CLI facilities of an operating system are weak or nonexistent, you'll also see the following consequences:

To design the perfect anti-Unix, have no CLI and no capability to script programs — or, important facilities that the CLI cannot drive.

Intended Audience

The design of operating systems varies in response to the expected audience for the system. Some operating systems are intended for back rooms, some for desktops. Some are designed for technical users, others for end users. Some are intended to work standalone in real-time control applications, others for an environment of timesharing and pervasive networking.

One important distinction is client vs. server. ‘Client’ translates as: being lightweight, suppporting only a single user, able to run on small machines, designed to be switched on when needed and off when the user is done, lacking pre-emptive multitasking, optimized for low latency, and putting a lot of its resources into fancy user interfaces. ‘Server’ translates as: being heavyweight, capable of running continuously, optimized for throughput, fully pre-emptively multitasking to handle multiple sessions. In origin all operating systems were server operating systems; the concept of a client operating system only emerged in the late 1970s with inexpensive but underpowered PC hardware. Client operating systems are more focused on a visually attractive user experience than on 24/7 uptime.

All these variables have an effect on development style. One of the most obvious is the level of interface complexity the target audience will tolerate, and how it tends to weight perceived complexity against other variables like cost and capability. Unix is often said to have been written by programmers for programmers — an audience that is notoriously tolerant of interface complexity.

This is a consequence rather than a goal. I abhor a system designed for the “user”, if that word is a coded pejorative meaning “stupid and unsophisticated”.

– Ken Thompson

To design the perfect anti-Unix, write an operating system that thinks it knows what you're doing better than you do. And then adds injury to insult by getting it wrong.

Entry Barriers to Development

Another important dimension along which operating systems differ is the height of the barrier that separates mere users from becoming developers. There are two important cost drivers here. One is the monetary cost of development tools, the other is the time cost of gaining proficiency as a developer. Some development cultures evolve social barriers to entry, but these are usually an effect of the underlying technology costs, not a primary cause.

Expensive development tools and complex, opaque APIs produce small and elitist programming cultures. In those cultures, programming projects are large, serious endeavors — they have to be in order to offer a payoff that justifies the cost of both hard and soft (human) capital invested. Large, serious projects tend to produce large, serious programs (and, far too often, large expensive failures).

Inexpensive tools and simple interfaces support casual programming, hobbyist cultures, and exploration. Programming projects can be small (often, formal project structure is plain unnecessary), and failure is not a catastrophe. This changes the style in which people develop code; among other things, they show less tendency to over-commit to failed approaches.

Casual programming tends to produce lots of small programs and a self-reinforcing, expanding community of knowledge. In a world of cheap hardware, the presence or absence of such a community is an increasingly important factor in whether an operating system is long-term viable at all.

Unix pioneered casual programming. One of the things Unix was first at doing was shipping with a compiler and scripting tools as part of the default installation available to all users, supporting a hobbyist software-development culture that spanned multiple installations. Many people who write code under Unix do not think of it as writing code — they think of it as writing scripts to automate common tasks, or as customizing their environment.

To design the perfect anti-Unix, make casual programming impossible.


[20] For readers without Unix experience, a pipe is a way of connecting the output of one program to the input of another. We'll explore the ways this idea can be used to help programs cooperate in Chapter 7.

[21] A periodic clock interrupt from the hardware is useful as a sort of heartbeat for a timesharing system; each time it fires, it tells the system that it may be time to switch to another task, defining the size of the unit timeslice. In 2003 Unixes usually set the heartbeat to either 60 or 100 times a second.

[22] The modern buzzword for this is role-based security.

[23] This problem was considered quite serious by Microsoft itself during their rebuild of Hotmail. See [BrooksD].

Operating-System Comparisons

The logic of Unix's design choice stands out more clearly when we contrast it with other operating systems. Here we will attempt only a design overview; for detailed discussion of the technical features of different operating systems.[24]

Figure 3.1. Schematic history of timesharing.

Schematic history of timesharing.

Figure 3.1 indicates the genetic relationships among the timesharing operating systems we'll survey. A few other operating systems (marked in gray, and not necessarily timesharing) are included for context. Sytems in solid boxes are still live. The ‘birth’ are dates of first shipment;[25] the ‘death’ dates are generally when the system was end-of-lifed by its vendor.

Solid arrows indicate a genetic relationship or very strong design influence (e.g., a later system with an API deliberately reverse-engineered to match an earlier one). Dashed lines indicate significant design influence. Dotted lines indicate weak design influence. Not all the genetic relationships are acknowledged by the developers; indeed, some have been officially denied for legal or corporate-strategy reasons but are open secrets in the industry.

The ‘Unix’ box includes all proprietary Unixes, including both AT&T and early Berkeley versions. The ‘Linux’ box includes the open-source Unixes, all of which launched in 1991. They have genetic inheritance from early Unix through code that was freed from AT&T proprietary control by the settlement of a 1993 lawsuit.[26]


VMS is the proprietary operating system originally developed for the VAX minicomputer from Digital Equipment Corporation. It was first released in 1978, was an important production operating system in the 1980s and early 1990s, and continued to be maintained when DEC was acquired by Compaq and Compaq was acquired by Hewlett-Packard. It is still sold and supported in mid-2003, though little new development goes on in it today.[27] VMS is surveyed here to show the contrast between Unix and other CLI-oriented operating systems from the minicomputer era.

VMS has full preemptive multitasking, but makes process-spawning very expensive. The VMS file system has an elaborate notion of record types (though not attributes). These traits have all the consequences we outlined earlier on, especially (in VMS's case) the tendency for programs to be huge, clunky monoliths.

VMS features long, readable COBOL-like system commands and command options. It has very comprehensive on-line help (not for APIs, but for the executable programs and command-line syntax). In fact, the VMS CLI and its help system are the organizing metaphor of VMS. Though X windows has been retrofitted onto the system, the verbose CLI remains the most important stylistic influence on program design. This has the following major implications:

VMS has a respectable system of internal boundaries. It was designed for true multiuser operation and fully employs the hardware MMU to protect processes from each other. The system command interpreter is privileged, but the encapsulation of critical functions is otherwise reasonably good. Security cracks against VMS have been rare.

VMS tools were initially expensive, and its interfaces are complex. Enormous volumes of VMS programmer documentation are only available in paper form, so looking up anything is a time-consuming, high-overhead operation. This has tended to discourage exploratory programming and learning a large toolkit. Only since being nearly abandoned by its vendor has VMS developed casual programming and a hobbyist culture, and that culture is not particularly strong.

Like Unix, VMS predated the client/server distinction. It was successful in its day as a general-purpose timesharing operating system. The intended audience was primarily technical users and software-intensive businesses, implying a moderate tolerance for complexity.


The Macintosh operating system was designed at Apple in the early 1980s, inspired by pioneering work on GUIs done earlier at Xerox's Palo Alto Research Center. It saw its debut with the Macintosh in 1984. MacOS has gone through two significant design transitions since, and is undergoing a third. The first transition was the shift from supporting only a single application at a time to being able to cooperatively multitask multiple applications (MultiFinder); the second was the shift from 68000 to PowerPC processors, which both preserved backward binary compatibility with 68K applications and brought in an advanced shared library management system for PowerPC applications, replacing the original 68K trap instruction-based code-sharing system. The third was the merger of MacOS design ideas with a Unix-derived infrastructure in MacOS X. Except where specifically noted, the discussion here applies to pre-OS-X versions.

MacOS has a very strong unifying idea that is very different from Unix's: the Mac Interface Guidelines. These specify in great detail what an application GUI should look like and how it should behave. The consistency of the Guidelines influenced the culture of Mac users in significant ways. Not infrequently, simple-minded ports of DOS or Unix programs that did not follow the Guidelines have been summarily rejected by the Mac user base and failed in the marketplace.

One key idea of the Guidelines is that things stay where you put them. Documents, directories, and other objects have persistent locations on the desktop that the system doesn't mess with, and the desktop context persists through reboots.

The Macintosh's unifying idea is so strong that most of the other design choices we discussed above are either forced by it or invisible. All programs have GUIs. There is no CLI at all. Scripting facilities are present but much less commonly used than under Unix; many Mac programmers never learn them. MacOS's captive-interface GUI metaphor (organized around a single main event loop) leads to a weak scheduler without preemption. The weak scheduler, and the fact that all MultiFinder applications run in a single large address space, implies that it is not practical to use separated processes or even threads rather than polling.

MacOS applications are not, however, invariably monster monoliths. The system's GUI support code, which is partly implemented in a ROM shipped with the hardware and partly implemented in shared libraries, communicates with MacOS programs through an event interface that has been quite stable since its beginnings. Thus, the design of the operating system encourages a relatively clean separation between application engine and GUI interface.

MacOS also has strong support for isolating application metadata like menu structures from the engine code. MacOS files have both a ‘data fork’ (a Unix-style bag of bytes that contains a document or program code) and a ‘resource fork’ (a set of user-definable file attributes). Mac applications tend to be designed so that (for example) the images and sound used in them are stored in the resource fork and can be modified separately from the application code.

The MacOS system of internal boundaries is very weak. There is a wired-in assumption that there is but a single user, so there are no per-user privilege groups. Multitasking is cooperative, not pre-emptive. All MultiFinder applications run in the same address space, so bad code in any application can corrupt anything outside the operating system's low-level kernel. Security cracks against MacOS machines are very easy to write; the OS has been spared an epidemic mainly because very few people are motivated to crack it.

Mac programmers tend to design in the opposite direction from Unix programmers; that is, they work from the interface inward, rather than from the engine outward (we'll discuss some of the implications of this choice in Chapter 20). Everything in the design of the MacOS conspires to encourage this.

The intended role for the Macintosh was as a client operating system for nontechnical end users, implying a very low tolerance for interface complexity. Developers in the Macintosh culture became very, very good at designing simple interfaces.

The incremental cost of becoming a developer, assuming you have a Macintosh already, has never been high. Thus, despite rather complex interfaces, the Mac developed a strong hobbyist culture early on. There is a vigorous tradition of small tools, shareware, and user-supported software.

Classic MacOS has been end-of-lifed. Most of its facilities have been imported into MacOS X, which mates them to a Unix infrastructure derived from the Berkeley tradition.[28] At the same time, leading-edge Unixes such as Linux are beginning to borrow ideas like file attributes (a generalization of the resource fork) from MacOS.


OS/2 began life as an IBM development project called ADOS (‘Advanced DOS’), one of three competitors to become DOS 4. At that time, IBM and Microsoft were formally collaborating to develop a next-generation operating system for the PC. OS/2 1.0 was first released in 1987 for the 286, but was unsuccessful. The 2.0 version for the 386 came out in 1992, but by that time the IBM/Microsoft alliance had already fractured. Microsoft went in a different (and more lucrative) direction with Windows 3.0. OS/2 attracted a loyal minority following, but never attracted a critical mass of developers and users. It remained third in the desktop market, behind the Macintosh, until being subsumed into IBM's Java initiative after 1996. The last released version was 4.0 in 1996. Early versions found their way into embedded systems and still, as of mid-2003, run inside many of the world's automated teller machines.

Like Unix, OS/2 was built to be preemptively multitasking and would not run on a machine without an MMU (early versions simulated an MMU using the 286's memory segmentation). Unlike Unix, OS/2 was never built to be a multiuser system. Process-spawning was relatively cheap, but IPC was difficult and brittle. Networking was initially focused on LAN protocols, but a TCP/IP stack was added in later versions. There were no programs analogous to Unix service daemons, so OS/2 never handled multi-function networking very well.

OS/2 had both a CLI and GUI. Most of the positive legendry around OS/2 was about the Workplace Shell (WPS), the OS/2 desktop. Some of this technology was licensed from the developers of the AmigaOS Workbench,[29] a pioneering GUI desktop that still as of 2003 has a loyal fan base in Europe. This is the one area of the design in which OS/2 achieved a level of capability which Unix arguably has not yet matched. The WPS was a clean, powerful, object-oriented design with understandable behavior and good extensibility. Years later it would become a model for Linux's GNOME project.

The class-hierarchy design of WPS was one of OS/2's unifying ideas. The other was multithreading. OS/2 programmers used threading heavily as a partial substitute for IPC between peer processes. No tradition of cooperating program toolkits developed.

OS/2 had the internal boundaries one would expect in a single-user OS. Running processes were protected from each other, and kernel space was protected from user space, but there were no per-user privilege groups. This meant the file system had no protection against malicious code. Another consequence was that there was no analog of a home directory; application data tended to be scattered all over the system.

A further consequence of the lack of multiuser capability was that there could be no privilege distinctions in userspace. Thus, developers tended to trust only kernel code. Many system tasks that in Unix would be handled by user-space daemons were jammed into the kernel or the WPS. Both bloated as a result.

OS/2 had a text vs. binary mode (that is, a mode in which CR/LF was read as a single end-of-line, versus one in which no such interpretation was performed), but no other file record structure. It supported file attributes, which were used for desktop persistence after the manner of the Macintosh. System databases were mostly in binary formats.

The preferred UI style was through the WPS. User interfaces tended to be ergonomically better than Windows, though not up to Macintosh standards (OS/2's most active period was relatively early in the history of MacOS Classic). Like Unix and Windows, OS/2's user interface was themed around multiple, independent per-task groups of windows, rather than capturing the desktop for the running application.

The intended audience for OS/2 was business and nontechnical end users, implying a low tolerance for interface complexity. It was used both as a client operating system and as a file and print server.

In the early 1990s, developers in the OS/2 community began to migrate to a Unix-inspired environment called EMX that emulated POSIX interfaces. Ports of Unix software started routinely showing up under OS/2 in the latter half of the 1990s.

Anyone could download EMX, which included the GNU Compiler Collection and other open-source development tools. IBM intermittently gave away copies of the system documentation in the OS/2 developer's toolkit, which was posted on many BBSs and FTP sites. Because of this, the “Hobbes” FTP archive of user-developed OS/2 software had already grown to over a gigabyte in size by 1995. A very vigorous tradition of small tools, exploratory programming, and shareware developed and retained a loyal following for some years after OS/2 itself was clearly headed for the dustbin of history.

After the release of Windows 95 the OS/2 community, feeling beleaguered by Microsoft and encouraged by IBM, became increasingly interested in Java. After the Netscape source code release in early 1998, the direction of migration changed (rather suddenly), toward Linux.

OS/2 is interesting as a case study in how far a multitasking but single-user operating-system design can be pushed. Most of the observations in this case study would apply well to other operating systems of the same general type, notably AmigaOS[30] and GEM.[31] A wealth of OS/2 material is still available on the Web in 2003, including some good histories.[32]

Windows NT

Windows NT (New Technology) is Microsoft's operating system for high-end personal and server use; it is shipped in several variants that can all be considered the same for our purposes. All of Microsoft's operating systems since the demise of Windows ME in 2000 have been NT-based; Windows 2000 was NT 5, and Windows XP (current in 2003) is NT 5.1. NT is genetically descended from VMS, with which it shares some important characteristics.

NT has grown by accretion, and lacks a unifying metaphor corresponding to Unix's “everything is a file” or the MacOS desktop.[33] Because core technologies are not anchored in a small set of persistent central metaphors, they become obsolete every few years. Each of the technology generations — DOS (1981), Windows 3.1 (1992), Windows 95 (1995), Windows NT 4 (1996), Windows 2000 (2000), Windows XP (2002), and Windows Server 2003 (2003) — has required that developers relearn fundamental things in a different way, with the old way declared obsolete and no longer well supported.

There are other consequences as well:

NT has file attributes in some of its file system types. They are used in a restricted way, to implement access-control lists on some file systems, and don't affect development style very much. It also has a record-type distinction, between text and binary files, that produces occasional annoyances (both NT and OS/2 inherited this misfeature from DOS).

Though pre-emptive multitasking is supported, process-spawning is expensive — not as expensive as in VMS, but (at about 0.1 seconds per spawn) up to an order of magnitude more so than on a modern Unix. Scripting facilities are weak, and the OS makes extensive use of binary file formats. In addition to the expected consequences we outlined earlier are these:

System and user configuration data are centralized in a central properties registry rather than being scattered through numerous dotfiles and system data files as in Unix. This also has consequences throughout the design:

NT systems on the Internet are notoriously vulnerable to worms, viruses, defacements, and cracks of all kinds. There are many reasons for this, some more fundamental than others. The most fundamental is that NT's internal boundaries are extremely porous.

NT has access control lists that can be used to implement per-user privilege groups, but a great deal of legacy code ignores them, and the operating system permits this in order not to break backward compatibility. There are no security controls on message traffic between GUI clients, either,[34] and adding them would also break backward compatibility.

While NT will use an MMU, NT versions after 3.5 have the system GUI wired into the same address space as the privileged kernel for performance reasons. Recent versions even wire the webserver into kernel space in an attempt to match the speed of Unix-based webservers.

These holes in the boundaries have the synergistic effect of making actual security on NT systems effectively impossible.[35] If an intruder can get code run as any user at all (e.g., through the Outlook email-macro feature), that code can forge messages through the window system to any other running application. And any buffer overrun or crack in the GUI or webserver can be exploited to take control of the entire system.

Because Windows does not handle library versioning properly, it suffers from a chronic configuration problem called “DLL hell”, in which installing new programs can randomly upgrade (or even downgrade!) the libraries on which existing programs depend. This applies to the vendor-supplied system libraries as well as to application-specific ones: it is not uncommon for an application to ship with specific versions of system libraries, and break silently when it does not have them.[36]

On the bright side, NT provides sufficient facilities to host Cygwin, which is a compatibility layer implementing Unix at both the utilities and the API level, with remarkably few compromises.[37] Cygwin permits C programs to make use of both the Unix and the native APIs, and is the first thing many Unix hackers install on such Windows systems as they are compelled by circumstances to make use of.

The intended audience for the NT operating systems is primarily nontechnical end users, implying a very low tolerance for interface complexity. It is used in both client and server roles.

Early in its history Microsoft relied on third-party development to supply applications. They originally published full documentation for the Windows APIs, and kept the price of development tools low. But over time, and as competitors collapsed, Microsoft's strategy shifted to favor in-house development, they began hiding APIs from the outside world, and development tools grew more expensive. As early as Windows 95, Microsoft was requiring nondisclosure agreements as a condition for purchasing professional-quality development tools.

The hobbyist and casual-developer culture that had grown up around DOS and earlier Windows versions was large enough to be self-sustaining even in the face of increasing efforts by Microsoft to lock them out (including such measures as certification programs designed to delegitimize amateurs). Shareware never went away, and Microsoft's policy began to reverse somewhat after 2000 under market pressure from open-source operating systems and Java. However, Windows interfaces for ‘professional’ programming continued to grow more complex over time, presenting an increasing barrier to casual (or serious!) coding.

The result of this history is a sharp dichotomy between the design styles practiced by amateur and professional NT developers — the two groups barely communicate. While the hobbyist culture of small tools and shareware is very much alive, professional NT projects tend to produce monster monoliths even bulkier than those characteristic of ‘elitist’ operating systems like VMS.

Unix-like shell facilities, command sets, and library APIs are available under Windows through third-party libraries including UWIN, Interix, and the open-source Cygwin.


Be, Inc. was founded in 1989 as a hardware vendor, building pioneering multiprocessing machines around the PowerPC chip. BeOS was Be's attempt to add value to the hardware by inventing a new, network-ready operating system model incorporating the lessons of both Unix and the MacOS family, without being either. The result was a tasteful, clean, and exciting design with excellent performance in its chosen role as a multimedia platform.

BeOS's unifying ideas were ‘pervasive threading’, multimedia flows, and the file system as database. BeOS was designed to minimize latency in the kernel, making it well-suited for processing large volumes of data such as audio and video streams in real time. BeOS ‘threads’ were actually lightweight processes in Unix terminology, since they supported thread-local storage and therefore did not necessarily share all address spaces. IPC via shared memory was fast and efficient.

BeOS followed the Unix model in having no file structure above the byte level. Like the MacOS, it supported and used file attributes. In fact, the BeOS file system was actually a database that could be indexed by any attribute.

One of the things BeOS took from Unix was intelligent design of internal boundaries. It made full use of an MMU, and sealed running processes off from each other effectively. While it presented as a single-user operating system (no login), it supported Unix-like privilege groups in the file system and elsewhere in the OS internals. These were used to protect system-critical files from being touched by untrusted code; in Unix terms, the user was logged in as an anonymous guest at boot time, with the only other ‘user’ being root. Full multiuser operation would have been a small change to the upper levels of the system, and there was in fact a BeLogin utility.

BeOS tended to use binary file formats and the native database built into the file system, rather than Unix-like textual formats.

The preferred UI style of BeOS was GUI, and it leaned heavily on MacOS experience in interface design. CLI and scripting were, however, also fully supported. The command-line shell of BeOS was a port of bash(1), the dominant open-source Unix shell, running through a POSIX compatibility library. Porting of Unix CLI software was, by design, trivially easy. Infrastructure to support the full panoply of scripting, filters, and service daemons that goes with the Unix model was in place.

BeOS's intended role was as a client operating system specialized for near-real-time multimedia processing (especially sound and video manipulation). Its intended audience included technical and business end users, implying a moderate tolerance for interface complexity.

Entry barriers to BeOS development were low; though the operating system was proprietary, development tools were inexpensive and full documentation was readily available. The BeOS effort began as part of one of the efforts to unseat Intel's hardware with RISC technology, and was continued as a software-only effort after the Internet explosion. Its strategists were paying attention during Linux's formative period in the early 1990s, and were fully aware of the value of a large casual-developer base. In fact they succeeded in attracting an intensely loyal following; as of 2003 no fewer than five separate projects are attempting to resurrect BeOS in open source.

Unfortunately, the business strategy surrounding BeOS was not as astute as the technical design. The BeOS software was originally bundled with dedicated hardware, and marketed with only vague hints about intended applications. Later (1998) BeOS was ported to generic PCs and more closely focused on multimedia applications, but never attracted a critical mass of applications or users. BeOS finally succumbed in 2001 to a combination of anticompetitive maneuvering by Microsoft (lawsuit in progress as of 2003) and competition from variants of Linux that had been adapted for multimedia handling.


MVS (Multiple Virtual Storage) is IBM's flagship operating system for its mainframe computers. Its roots stretch back to OS/360, which began life in the mid-1960s as the operating system IBM wanted its customers to use on the then-new System/360 computer systems. Descendants of this code remain at the heart of today's IBM mainframe operating systems. Though the code has been almost entirely rewritten, the basic design is largely untouched; backward compatibility has been religiously maintained, to the point that applications written for OS/360 run unmodified on the MVS of 64-bit z/Series mainframe computers three architectural generations later.

Of all the operating systems surveyed here, MVS is the only one that could be considered older than Unix (the ambiguity stems from the degree to which it has evolved over time). It is also the least influenced by Unix concepts and technology, and represents the strongest design contrast with Unix. The unifying idea of MVS is that all work is batch; the system is designed to make the most efficient possible use of the machine for batch processing of huge amounts of data, with minimal concessions to interaction with human users.

Native MVS terminals (the 3270 series) operate only in block mode. The user is presented with a screen that he fills in, modifying local storage in the terminal. No interrupt is presented to the mainframe until the user presses the send key. Character-level interaction, in the manner of Unix's raw mode, is impossible.

TSO, the closest equivalent to the Unix interactive environment, is limited in native capabilities. Each TSO user is represented to the rest of the system as a simulated batch job. The facility is expensive — so much so that its use is typically limited to programmers and support staff. Ordinary users who need to merely run applications from a terminal almost never use TSO. Instead, they work through transaction monitors, a kind of multiuser application server that does cooperative multitasking and supports asynchronous I/O. In effect, each kind of transaction monitor is a specialized timesharing plugin (almost, but not entirely unlike a webserver running CGI).

Another consequence of the batch-oriented architecture is that process spawning is a slow operation. The I/O system deliberately trades high setup cost (and associated latency) for better throughput. These choices are a good match for batch operation, but deadly to interactive response. A predictable result is that TSO users nowadays spend almost all their time inside a dialog-driven interactive environment, ISPF. It is rare for a programmer to do anything inside native TSO except start up an instance of ISPF. This does away with process-spawn overhead, at the cost of introducing a very large program that does everything but start the machine room coffeepot.

MVS uses the machine MMU; processes have separate address spaces. Interprocess communication is supported only through shared memory. There are facilities for threading (which MVS calls “subtasking”), but they are lightly used, mainly because the facility is only easily accessible from programs written in assembler. Instead, the typical batch application is a short series of heavyweight program invocations glued together by JCL (Job Control Language) which provides scripting, though in a notoriously difficult and inflexible way. Programs in a job communicate through temporary files; filters and the like are nearly impossible to do in a usable manner.

Every file has a record format, sometimes implied (inline input files in JCL are implied to have an 80-byte fixed-length record format inherited from punched cards, for example), but more often explicitly specified. Many system configuration files are in text format, but application files are usually in binary formats specific to the application. Some general tools for examining files have evolved out of sheer necessity, but it is still not an easy problem to solve.

File system security was an afterthought in the original design. However, when security was found to be necessary, IBM added it in an inspired fashion: They defined a generic security API, then made all file access requests pass by that interface before being processed. As a result, there are at least three competing security packages with differing design philosophies — and all of them are quite good, with no known cracks against them between 1980 and mid-2003. This variety allows an installation to select the package that best suits local security policy.

Networking facilities are another afterthought. There is no concept of one interface for both network connections and local files; their programming interfaces are separate and quite different. This did allow TCP/IP to supplant IBM's native SNA (Systems Network Architecture) as the network protocol of choice fairly seamlessly. It is still common in 2003 to see both in use at a given installation, but SNA is dying out.

Casual programming for MVS is almost nonexistent except within the community of large enterprises that run MVS. This is not due so much to the cost of the tools themselves as it is to the cost of the environment — when one must spend several million dollars on the computer system, a few hundred dollars a month for a compiler is almost incidental. Within that community, however, there is a thriving culture of freely available software, mainly programming and system-administration tools. The first computer user's group, SHARE, was founded in 1955 by IBM users, and is still going strong today.

Considering the vast architectural differences, it is a remarkable fact that MVS was the first non-System-V operating system to meet the Single Unix Specification (there is less to this than meets the eye, however, as ports of Unix software from elsewhere have a strong tendency to founder on ASCII-vs.-EBCDIC character-set issues). It's possible to start a Unix shell from TSO; Unix file systems are specially formatted MVS data sets. The MVS Unix character set is a special EBCDIC codepage with newline and linefeed swapped (so that what appears as linefeed to Unix appears like newline to MVS), but the system calls are real system calls implemented in the MVS kernel.

As the cost of the environment drops into the hobbyist range, there is a small but growing group of users of the last public-domain version of MVS (3.8, dating from 1979). This system, as well as development tools and the emulator to run them, are all available for the cost of a CD.[38]

The intended role of MVS has always been in the back office. Like VMS and Unix itself, MVS predates the server/client distinction. Interface complexity for back-office users is not only tolerated but expected, in the name of making the computer spend fewer expensive resources on interfaces and more on the work it's there to get done.


VM/CMS is IBM's other mainframe operating system. Historically speaking, it is Unix's uncle: the common ancestor is the CTSS system, developed at MIT around 1963 and running on the IBM 7094 mainframe. The group that developed CTSS then went on to write Multics, the immediate ancestor of Unix. IBM established a group in Cambridge to write a timesharing system for the IBM 360/40, a modified 360 with (for the first time on an IBM system) a paging MMU.[39] The MIT and IBM programmers continued to interact for many years thereafter, and the new system got a user interface that was very CTSS-like, complete with a shell named EXEC and a large supply of utilities analogous to those used on Multics and later on Unix.

In another sense, VM/CMS and Unix are funhouse mirror images of one another. The unifying idea of the system, provided by the VM component, is virtual machines, each of which looks exactly like the underlying physical machine. They are preemptively multitasked, and run either the single-user operating system CMS or a complete multitasking operating system (typically MVS, Linux, or another instance of VM itself). Virtual machines correspond to Unix processes, daemons, and emulators, and communication between them is accomplished by connecting the virtual card punch of one machine to the virtual card reader of another. In addition, a layered tools environment called CMS Pipelines is provided within CMS, directly modeled on Unix's pipes but architecturally extended to support multiple inputs and outputs.

When communication between them has not been explicitly set up, virtual machines are completely isolated from each other. The operating system has the same high reliability, scalability, and security as MVS, and has far greater flexibility and is much easier to use. In addition, the kernel-like portions of CMS do not need to be trusted by the VM component, which is maintained completely separately.

Although CMS is record-oriented, the records are essentially equivalent to the lines that Unix textual tools use. Databases are much better integrated into CMS Pipelines than is typically the case on Unix, where most databases are quite separate from the operating system. In recent years, CMS has been augmented to fully support the Single Unix Specification.

The UI style of CMS is interactive and conversational, very unlike MVS but like VMS and Unix. A full-screen editor called XEDIT is heavily used.

VM/CMS predates the client/server distinction, and is nowadays used almost entirely as a server operating system with emulated IBM terminals. Before Windows came to dominate the desktop so completely, VM/CMS provided word-processing services and email both internally to IBM and between mainframe customer sites — indeed, many VM systems were installed exclusively to run those applications because of VM's ready scalability to tens of thousands of users.

A scripting language called Rexx supports programming in a style not unlike shell, awk, Perl or Python. Consequently, casual programming (especially by system administrators) is very important on VM/CMS. Free cycles permitting, admins often prefer to run production MVS in a virtual machine rather than directly on the bare iron, so that CMS is also available and its flexibility can be taken advantage of. (There are CMS tools that permit access to MVS file systems.)

There are even striking parallels between the history of VM/CMS within IBM and Unix within Digital Equipment Corporation (which made the hardware that Unix first ran on). It took IBM years to understand the strategic importance of its unofficial timesharing system, and during that time a community of VM/CMS programmers arose that was closely analogous in behavior to the early Unix community. They shared ideas, shared discoveries about the system, and above all shared source code for utilities. No matter how often IBM tried to declare VM/CMS dead, the community — which included IBM's own MVS system developers! — insisted on keeping it alive. VM/CMS even went through the same cycle of de facto open source to closed source back to open source, though not as thoroughly as Unix did.

What VM/CMS lacks, however, is any real analog to C. Both VM and CMS were written in assembler and have remained so implemented. The nearest equivalent to C was various cut-down versions of PL/I that IBM used for systems programming, but did not share with its customers. Therefore, the operating system remains trapped on its original architectural line, though it has grown and expanded as the 360 architecture became the 370 series, the XA series, and finally the current z/Series.

Since the year 2000, IBM has been promoting VM/CMS on mainframes to an unprecedented degree — as ways to host thousands of virtual Linux machines at once.


Linux, originated by Linus Torvalds in 1991, leads the pack of new-school open-source Unixes that have emerged since 1990 (also including FreeBSD, NetBSD, OpenBSD, and Darwin), and is representative of the design direction being taken by the group as a whole. The trends in it can be taken as typical for this entire group.

Linux does not include any code from the original Unix source tree, but it was designed from Unix standards to behave like a Unix. In the rest of this book, we emphasize the continuity between Unix and Linux. That continuity is extremely strong, both in terms of technology and key developers — but here we emphasize some directions Linux is taking that mark a departure from ‘classical’ Unix tradition.

Many developers and activists in the Linux community have ambitions to win a substantial share of end-user desktops. This makes Linux's intended audience quite a bit broader than was ever the case for the old-school Unixes, which have primarily aimed at the server and technical-workstation markets. This has implications for the way Linux hackers design software.

The most obvious change is a shift in preferred interface styles. Unix was originally designed for use on teletypes and slow printing terminals. Through much of its lifetime it was strongly associated with character-cell video-display terminals lacking either graphics or color capabilities. Most Unix programmers stayed firmly wedded to the command line long after large end-user applications had migrated to X-based GUIs, and the design of both Unix operating systems and their applications have continued to reflect this fact.

Linux users and developers, on the other hand, have been adapting themselves to address the nontechnical user's fear of CLIs. They have moved to building GUIs and GUI tools much more intensively than was the case in old-school Unix, or even in contemporary proprietary Unixes. To a lesser but significant extent, this is true of the other open-source Unixes as well.

The desire to reach end users has also made Linux developers much more concerned with smoothness of installation and software distribution issues than is typically the case under proprietary Unix systems. One consequence is that Linux features binary-package systems far more sophisticated than any analogs in proprietary Unixes, with interfaces designed (as of 2003, with only mixed success) to be palatable to nontechnical end users.

The Linux community wants, more than the old-school Unixes ever did, to turn their software into a sort of universal pipefitting for connecting together other environments. Thus, Linux features support for reading and (often) writing the file system formats and networking methods native to other operating systems. It also supports multiple-booting with them on the same hardware, and simulating them in software inside Linux itself. The long-term goal is subsumption; Linux emulates so it can absorb.[40]

The goal of subsuming the competition, combined with the drive to reach the end-user, has motivated Linux developers to adopt design ideas from non-Unix operating systems to a degree that makes traditional Unixes look rather insular. Linux applications using Windows .INI format files for configuration is a minor example we'll cover in Chapter 10; Linux 2.5's incorporation of extended file attributes, which among other things can be used to emulate the semantics of the Macintosh resource fork, is a recent major one at time of writing.

But the day Linux gives the Mac diagnostic that you can't open a file because you don't have the application is the day Linux becomes non-Unix.

– Doug McIlroy

The remaining proprietary Unixes (such as Solaris, HP-UX, AIX, etc.) are designed to be big products for big IT budgets. Their economic niche encourages designs optimized for maximum power on high-end, leading-edge hardware. Because Linux has part of its roots among PC hobbyists, it emphasizes doing more with less. Where proprietary Unixes are tuned for multiprocessor and server-cluster operation at the expense of performance on low-end hardware, core Linux developers have explicitly chosen not to accept more complexity and overhead on low-end machines for marginal performance gains on high-end hardware.

Indeed, a substantial fraction of the Linux user community is understood to be wringing usefulness out of hardware as technically obsolete today as Ken Thompson's PDP-7 was in 1969. As a consequence, Linux applications are under pressure to stay lean and mean that their counterparts under proprietary Unix do not experience.

These trends have implications for the future of Unix as a whole, a topic we'll return to in Chapter 20.


[24] See the OSData website.

[25] Except for Multics which exerted most of its influence between the time its specifications were published in 1965 and when it actually shipped in 1969.

[26] For details on the lawsuit, see Marshall Kirk McKusick's paper in [OpenSources].


[27] More information is available at the site.

[28] MacOS X actually consists of two proprietary layers (ports of the OpenStep and Classic Mac GUIs) layered over an open-source Unix core (Darwin).


[29] In return for some Amiga technology, IBM gave Commodore a license for its REXX scripting language. The deal is described at

[30] AmigaOS Portal.

[31] The GEM Operating System.

[32] See, for example, the OS Voice and OS/2 BBS.COM sites.

[33] Perhaps. It has been argued that the unifying metaphor of all Microsoft operating systems is “the customer must be locked in”.


[35] Microsoft actually admitted publicly that NT security is impossible in March 2003. See

[36] The DLL hell problem is somewhat mitigated by the .NET development framework, which handles library versioning — but as of 2003 .NET only ships on the highest-end server versions of NT.

[37] Cygwin is largely compliant with the Single Unix Specification, but programs requiring direct hardware access run into limitations in the Windows kernel that hosts it. Ethernet cards are notoriously problematic.


[39] The development machine and initial target was a 40 with customized microcode, but it proved insufficiently powerful; production deployment was on the 360/67.

[40] The results of Linux's emulate-and-subsume strategy differ noticeably from the embrace-and-extend practiced by some of its competitors. For starters, Linux does not break compatibility with what it is emulating in order to lock customers into the “extended” version.

What Goes Around, Comes Around

We attempted to select for comparison timesharing systems that either are now or have in the past been competitive with Unix. The field of plausible candidates is not wide. Most (Multics, ITS, DTSS, TOPS-10, TOPS-20, MTS, GCOS, MPE and perhaps a dozen others) are so long dead that they are fading from the collective memory of the computing field. Of those we surveyed, VMS and OS/2 are moribund, and MacOS has been subsumed by a Unix derivative. MVS and VM/CMS were limited to a single proprietary mainframe line. Only Microsoft Windows remains as a viable competitor independent of the Unix tradition.

We pointed out Unix's strengths in Chapter 1, and they are certainly part of the explanation. But it's actually more instructive to look at the obverse of that answer and ask which weaknesses in Unix's competitors did them in.

The most obvious shared problem is nonportability. Most of Unix's pre-1980 competitors were tied to a single hardware platform, and died with that platform. One reason VMS survived long enough to merit inclusion here as a case study is that it was successfully ported from its original VAX hardware to the Alpha processor (and in 2003 is being ported from Alpha to Itanium). MacOS successfully made the jump from the Motorola 68000 to PowerPC chips in the late 1980s. Microsoft Windows escaped this problem by being in the right place when commoditization flattened the market for general-purpose computers into a PC monoculture.

From 1980 on, another particular weakness continually reemerges as a theme in different systems that Unix either steamrollered or outlasted: an inability to support networking gracefully.

In a world of pervasive networking, even an operating system designed for single-user use needs multiuser capability (multiple privilege groups) — because without that, any network transaction that can trick a user into running malicious code will subvert the entire system (Windows macro viruses are only the tip of this iceberg). Without strong multitasking, the ability of an operating system to handle network traffic and run user programs at the same time will be impaired. The operating system also needs efficient IPC so that its network programs can communicate with each other and with the user's foreground applications.

Windows gets away with having severe deficiencies in these areas only by virtue of having developed a monopoly position before networking became really important, and by having a user population that has been conditioned to accept a shocking frequency of crashes and security breaches as normal. This is not a stable situation, and it is one that partisans of Linux have successfully (in 2003) exploited to make major inroads in the server-operating-system market.

Around 1980, during the early heyday of personal computers, operating-system designers dismissed Unix and traditional timesharing as heavyweight, cumbersome, and inappropriate for the brave new world of single-user personal machines — despite the fact that GUI interfaces tended to demand the reinvention of multitasking to cope with threads of execution bound to different windows and widgets. The trend toward client operating systems was so intense that server operating systems were at times dismissed as steam-powered relics of a bygone age.

But as the designers of BeOS noticed, the requirements of pervasive networking cannot be met without implementing something very close to general-purpose timesharing. Single-user client operating systems cannot thrive in an Internetted world.

This problem drove the reconvergence of client and server operating systems. The first, pre-Internet attempts at peer-to-peer networking over LANs, in the late 1980s, began to expose the inadequacy of the client-OS design model. Data on a network has to have rendezvous points in order to be shared; thus, we can't do without servers. At the same time, experience with the Macintosh and Windows client operating systems raised the bar on the minimum quality of user experience customers would tolerate.

With non-Unix models for timesharing effectively dead by 1990, there were not many possible responses client operating-system designers could mount to this challenge. They could co-opt Unix (as MacOS X has done), re-invent roughly equivalent features a patch at a time (as Windows has done), or attempt to reinvent the entire world (as BeOS tried and failed to do). But meanwhile, open-source Unixes were growing client-like capabilities to use GUIs and run on inexpensive personal machines.

These pressures turned out, however, not to be as symmetrically balanced as the above description might imply. Retrofitting server-operating-system features like multiple privilege classes and full multitasking onto a client operating system is very difficult, quite likely to break compatibility with older versions of the client, and generally produces a fragile and unsatisfactory result rife with stability and security problems. Retrofitting a GUI onto a server operating system, on the other hand, raises problems that can largely be finessed by a combination of cleverness and throwing ever-more-inexpensive hardware resources at the problem. As with buildings, it's easier to repair superstructure on top of a solid foundation than it is to replace the foundations without trashing the superstructure.

Besides having the native architectural strengths of a server operating system, Unix was always agnostic about its intended audience. Its designers and implementers never assumed they knew all potential uses the system would be put to.

Thus, the Unix design proved more capable of reinventing itself as a client than any of its client-operating-system competitors were of reinventing themselves as servers. While many other factors of technology and economics contributed to the Unix resurgence of the 1990s, this is one that really foregrounds itself in any discussion of operating-system design style.


Chapter 4. Modularity

There are two ways of constructing a software design. One is to make it so simple that there are obviously no deficiencies; the other is to make it so complicated that there are no obvious deficiencies. The first method is far more difficult.

– C. A. R. Hoare The Emperor's Old Clothes, CACM February 1981

There is a natural hierarchy of code-partitioning methods that has evolved as programmers have had to manage ever-increasing levels of complexity. In the beginning, everything was one big lump of machine code. The earliest procedural languages brought in the notion of partition by subroutine. Then we invented service libraries to share common utility functions among multiple programs. Next, we invented separated address spaces and communicating processes. Today we routinely distribute program systems across multiple hosts separated by thousands of miles of network cable.

The early developers of Unix were among the pioneers in software modularity. Before them, the Rule of Modularity was computer-science theory but not engineering practice. In Design Rules[Baldwin-Clark], a path-breaking study of the economics of modularity in engineering design, the authors use the development of the computer industry as a case study and argue that the Unix community was in fact the first to systematically apply modular decomposition to production software, as opposed to hardware. Modularity of hardware has of course been one of the foundations of engineering since the adoption of standard screw threads in the late 1800s.

The Rule of Modularity bears amplification here: The only way to write complex software that won't fall on its face is to build it out of simple modules connected by well-defined interfaces, so that most problems are local and you can have some hope of fixing or optimizing a part without breaking the whole.

The tradition of being careful about modularity and of paying close attention to issues like orthogonality and compactness are still much deeper in the bone among Unix programmers than elsewhere.

Early Unix programmers became good at modularity because they had to be. An OS is one of the most complicated pieces of code around. If it is not well structured, it will fall apart. There were a couple of early failures at building Unix that were scrapped. One can blame the early (structureless) C for this, but basically it was because the OS was too complicated to write. We needed both refinements in tools (like C structures) and good practice in using them (like Rob Pike's rules for programming) before we could tame that complexity.

– Ken Thompson

Early Unix hackers struggled with this in many ways. In the languages of 1970 function calls were expensive, either because call semantics were complicated (PL/1. Algol) or because the compiler was optimizing for other things like fast inner loops at the expense of call time. Thus, code tended to be written in big lumps. Ken and several of the other early Unix developers knew modularity was a good idea, but they remembered PL/1 and were reluctant to write small functions lest performance go to hell.

Dennis Ritchie encouraged modularity by telling all and sundry that function calls were really, really cheap in C. Everybody started writing small functions and modularizing. Years later we found out that function calls were still expensive on the PDP-11, and VAX code was often spending 50% of its time in the CALLS instruction. Dennis had lied to us! But it was too late; we were all hooked...

– Steve Johnson

All programmers today, Unix natives or not, are taught to modularize at the subroutine level within programs. Some learn the art of doing this at the module or abstract-data-type level and call that ‘good design’. The design-patterns movement is making a noble effort to push up a level from there and discover successful design abstractions that can be applied to organize the large-scale structure of programs.

Getting better at all these kinds of problem partitioning is a worthy goal, and many excellent treatments of them are available elsewhere. We shall not attempt to cover all the issues relating to modularity within programs in too much detail: first, because that is a subject for an entire volume (or several volumes) in itself; and second, because this is a book about the art of Unix programming.

What we will do here is examine more specifically what the Unix tradition teaches us about how to follow the Rule of Modularity. In this chapter, our examples will live within process units. Later, in Chapter 7, we'll examine the circumstances under which partitioning programs into multiple cooperating processes is a good idea, and discuss more specific techniques for doing that partitioning.

Encapsulation and Optimal Module Size

The first and most important quality of modular code is encapsulation. Well-encapsulated modules don't expose their internals to each other. They don't call into the middle of each others' implementations, and they don't promiscuously share global data. They communicate using application programming interfaces (APIs) — narrow, well-defined sets of procedure calls and data structures. This is what the Rule of Modularity is about.

The APIs between modules have a dual role. On the implementation level, they function as choke points between the modules, preventing the internals of each from leaking into its neighbors. On the design level, it is the APIs (not the bits of implementation between them) that really define your architecture.

One good test for whether an API is well designed is this one: if you try to write a description of it in purely human language (with no source-code extracts allowed), does it make sense? It is a very good idea to get into the habit of writing informal descriptions of your APIs before you code them. Indeed, some of the most able developers start by defining their interfaces, writing brief comments to describe them, and then writing the code — since the process of writing the comment clarifies what the code must do. Such descriptions help you organize your thoughts, they make useful module comments, and eventually you might want to turn them into a roadmap document for future readers of the code.

As you push module decomposition harder, the pieces get smaller and the definition of the APIs gets more important. Global complexity, and consequent vulnerability to bugs, decreases. It has been received wisdom in computer science since the 1970s (exemplified in papers such as [Parnas]) that you ought to design your software systems as hierarchies of nested modules, with the grain size of the modules at each level held to a minimum.

It is possible, however, to push this kind of decomposition too hard and make your modules too small. There is evidence [Hatton97] that when one plots defect density versus module size, the curve is U-shaped and concave upwards (see Figure 4.1). Very small and very large modules are associated with more bugs than those of intermediate size. A different way of viewing the same data is to plot lines of code per module versus total bugs. The curve looks roughly logarithmic up to a ‘sweet spot’ where it flattens (corresponding to the minimum in the defect density curve), after which it goes up as the square of the number of the lines of code (which is what one might intuitively expect for the whole curve, following Brooks's Law[41]).

Figure 4.1. Qualitative plot of defect count and density vs. module size.

Qualitative plot of defect count and density vs. module size.

This unexpectedly increasing incidence of bugs at small module sizes holds across a wide variety of systems implemented in different languages. Hatton has proposed a model relating this nonlinearity to the chunk size of human short-term memory.[42] Another way to interpret the nonlinearity is that at small module grain sizes, the increasing complexity of the interfaces becomes the dominating term; it's difficult to read the code because you have to understand everything before you can understand anything. In Chapter 7 we'll examine more advanced forms of program partitioning; there, too, the complexity of interface protocols comes to dominate the total complexity of the system as the component processes get smaller.

In nonmathematical terms, Hatton's empirical results imply a sweet spot between 200 and 400 logical lines of code that minimizes probable defect density, all other factors (such as programmer skill) being equal. This size is independent of the language being used — an observation which strongly reinforces the advice given elsewhere in this book to program with the most powerful languages and tools you can. Beware of taking these numbers too literally however. Methods for counting lines of code vary considerably according to what the analyst considers a logical line, and other biases (such as whether comments are stripped). Hatton himself suggests as a rule of thumb a 2x conversion between logical and physical lines, suggesting an optimal range of 400–800 physical lines.


[41] Brooks's Law predicts that adding programmers to a late project makes it later. More generally, it predicts that costs and error rates rise as the square of the number of programmers on a project.

[42] In Hatton's model, small differences in the maximum chunk size a programmer can hold in short-term memory have a large multiplicative effect on the programmer's efficiency. This might be a major contributor to the order-of-magnitude (or larger) variations in effectiveness observed by Fred Brooks and others.

Compactness and Orthogonality

Code is not the only sort of thing with an optimal chunk size. Languages and APIs (such as sets of library or system calls) run up against the same sorts of human cognitive constraints that produce Hatton's U-curve.

Accordingly, Unix programmers have learned to think very hard about two other properties when designing APIs, command sets, protocols, and other ways to make computers do tricks: compactness and orthogonality.


Compactness is the property that a design can fit inside a human being's head. A good practical test for compactness is this: Does an experienced user normally need a manual? If not, then the design (or at least the subset of it that covers normal use) is compact.

Compact software tools have all the virtues of physical tools that fit well in the hand. They feel pleasant to use, they don't obtrude themselves between your mind and your work, they make you more productive — and they are much less likely than unwieldy tools to turn in your hand and injure you.

Compact is not equivalent to ‘weak’. A design can have a great deal of power and flexibility and still be compact if it is built on abstractions that are easy to think about and fit together well. Nor is compact equivalent to ‘easily learned’; some compact designs are quite difficult to understand until you have mastered an underlying conceptual model that is tricky, at which point your view of the world changes and compact becomes simple. For a lot of people, the Lisp language is a classic example of this.

Nor does compact mean ‘small’. If a well-designed system is predictable and ‘obvious’ to the experienced user, it might have quite a few pieces.

– Ken Arnold

Very few software designs are compact in an absolute sense, but many are compact in a slightly looser sense of the term. They have a compact working set, a subset of capabilities that suffices for 80% or more of what expert users normally do with them. Practically speaking, such designs normally need a reference card or cheat sheet but not a manual. We'll call such designs semi-compact, as opposed to strictly compact.

The concept is perhaps best illustrated by examples. The Unix system call API is semi-compact, but the standard C library is not compact in any sense. While Unix programmers easily keep a subset of the system calls sufficient for most applications programming (file system operations, signals, and process control) in their heads, the C library on modern Unixes includes many hundreds of entry points, e.g., mathematical functions, that won't all fit inside a single programmer's cranium.

The Magical Number Seven, Plus or Minus Two: Some Limits on Our Capacity for Processing Information[Miller] is one of the foundation papers in cognitive psychology (and, incidentally, the specific reason that U.S. local telephone numbers have seven digits). It showed that the number of discrete items of information human beings can hold in short-term memory is seven, plus or minus two. This gives us a good rule of thumb for evaluating the compactness of APIs: Does a programmer have to remember more than seven entry points? Anything larger than this is unlikely to be strictly compact.

Among Unix tools, make(1) is compact; autoconf(1) and automake(1) are not. Among markup languages, HTML is semi-compact, but DocBook (a documentation markup language we shall discuss in Chapter 18) is not. The man(7) macros are compact, but troff(1) markup is not.

Among general-purpose programming languages, C and Python are semi-compact; Perl, Java, Emacs Lisp, and shell are not (especially since serious shell programming requires you to know half-a-dozen other tools like sed(1) and awk(1)). C++ is anti-compact — the language's designer has admitted that he doesn't expect any one programmer to ever understand it all.

Some designs that are not compact have enough internal redundancy of features that individual programmers end up carving out compact dialects sufficient for that 80% of common tasks by choosing a working subset of the language. Perl has this kind of pseudo-compactness, for example. Such designs have a built-in trap; when two programmers try to communicate about a project, they may find that differences in their working subsets are a significant barrier to understanding and modifying the code.

Noncompact designs are not automatically doomed or bad, however. Some problem domains are simply too complex for a compact design to span them. Sometimes it's necessary to trade away compactness for some other virtue, like raw power and range. Troff markup is a good example of this. So is the BSD sockets API. The purpose of emphasizing compactness as a virtue is not to condition you to treat compactness as an absolute requirement, but to teach you to do what Unix programmers do: value compactness properly, design for it whenever possible, and not throw it away casually.


Orthogonality is one of the most important properties that can help make even complex designs compact. In a purely orthogonal design, operations do not have side effects; each action (whether it's an API call, a macro invocation, or a language operation) changes just one thing without affecting others. There is one and only one way to change each property of whatever system you are controlling.

Your monitor has orthogonal controls. You can change the brightness independently of the contrast level, and (if the monitor has one) the color balance control will be independent of both. Imagine how much more difficult it would be to adjust a monitor on which the brightness knob affected the color balance: you'd have to compensate by tweaking the color balance every time after you changed the brightness. Worse, imagine if the contrast control also affected the color balance; then, you'd have to adjust both knobs simultaneously in exactly the right way to change either contrast or color balance alone while holding the other constant.

Far too many software designs are non-orthogonal. One common class of design mistake, for example, occurs in code that reads and parses data from one (source) format to another (target) format. A designer who thinks of the source format as always being stored in a disk file may write the conversion function to open and read from a named file. Usually the input could just as well have been any file handle. If the conversion routine were designed orthogonally, e.g., without the side effect of opening a file, it could save work later when the conversion has to be done on a data stream supplied from standard input, a network socket, or any other source.

Doug McIlroy's advice to “Do one thing well” is usually interpreted as being about simplicity. But it's also, implicitly and at least as importantly, about orthogonality.

It's not a problem for a program to do one thing well and other things as side effects, provided supporting those other things doesn't raise the complexity of the program and its vulnerability to bugs. In Chapter 9 we'll examine a program called ascii that prints synonyms for the names of ASCII characters, including hex, octal, and binary values; as a side effect, it can serve as a quick base converter for numbers in the range 0–255. This second use is not an orthogonality violation because the features that support it are all necessary to the primary function; they do not make the program more difficult to document or maintain.

The problems with non-orthogonality arise when side effects complicate a programmer's or user's mental model, and beg to be forgotten, with results ranging from inconvenient to dire. Even when you do not forget the side effects, you're often forced to do extra work to suppress them or work around them.

There is an excellent discussion of orthogonality and how to achieve it in The Pragmatic Programmer [Hunt-Thomas]. As they point out, orthogonality reduces test and development time, because it's easier to verify code that neither causes side effects nor depends on side effects from other code — there are fewer combinations to test. If it breaks, orthogonal code is more easily replaced without disturbance to the rest of the system. Finally, orthogonal code is easier to document and reuse.

The concept of refactoring, which first emerged as an explicit idea from the ‘Extreme Programming’ school, is closely related to orthogonality. To refactor code is to change its structure and organization without changing its observable behavior. Software engineers have been doing this since the birth of the field, of course, but naming the practice and identifying a stock set of refactoring techniques has helped concentrate peoples' thinking in useful ways. Because these fit so well with the central concerns of the Unix design tradition, Unix developers have quickly coopted the terminology and ideas of refactoring.[43]

The basic Unix APIs were designed for orthogonality with imperfect but considerable success. We take for granted being able to open a file for write access without exclusive-locking it for write, for example; not all operating systems are so graceful. Old-style (System III) signals were non-orthogonal, because signal receipt had the side-effect of resetting the signal handler to the default die-on-receipt. There are large non-orthogonal patches like the BSD sockets API and very large ones like the X windowing system's drawing libraries.

But on the whole the Unix API is a good example: Otherwise it not only would not but could not be so widely imitated by C libraries on other operating systems. This is also a reason that the Unix API repays study even if you are not a Unix programmer; it has lessons about orthogonality to teach.

The SPOT Rule

The Pragmatic Programmer articulates a rule for one particular kind of orthogonality that is especially important. Their “Don't Repeat Yourself” rule is: every piece of knowledge must have a single, unambiguous, authoritative representation within a system. In this book we prefer, following a suggestion by Brian Kernighan, to call this the Single Point Of Truth or SPOT rule.

Repetition leads to inconsistency and code that is subtly broken, because you changed only some repetitions when you needed to change all of them. Often, it also means that you haven't properly thought through the organization of your code.

Constants, tables, and metadata should be declared and initialized once and imported elsewhere. Any time you see duplicate code, that's a danger sign. Complexity is a cost; don't pay it twice.

Often it's possible to remove code duplication by refactoring; that is, changing the organization of your code without changing the core algorithms. Data duplication sometimes appears to be forced on you. But when you see it, here are some valuable questions to ask:

There is an analog of the SPOT rule for data structures: “No junk, no confusion”. “No junk” says that the data structure (the model) should be minimal, e.g., not made so general that it can represent situations which cannot exist. “No confusion” says that states which must be kept distinct in the real-world problem must be kept distinct in the model. In short, the SPOT rule advocates seeking a data structure whose states have a one-to-one correspondence with the states of the real-world system to be modeled.

From deeper within the Unix tradition, we can add some of our own corollaries of the SPOT rule:

The reader should begin to see a pattern emerging here.

In the Unix world, the SPOT Rule as a unifying idea has seldom been explicit — but heavy use of code generators to implement particular kinds of SPOT are very much part of the tradition. We'll survey these techniques in Chapter 9.

Compactness and the Strong Single Center

One subtle but powerful way to promote compactness in a design is to organize it around a strong core algorithm addressing a clear formal definition of the problem, avoiding heuristics and fudging.

Formalization often clarifies a task spectacularly. It is not enough for a programmer to recognize that bits of his task fall within standard computer-science categories — a little depth-first search here and a quicksort there. The best results occur when the nub of the task can be formalized, and a clear model of the job at hand can be constructed. It is not necessary that ultimate users comprehend the model. The very existence of a unifying core will provide a comfortable feel, unencumbered with the why-in-hell-did-they-do-that moments that are so prevalent in using Swiss-army-knife programs.

– Doug McIlroy

This is an often-overlooked strength of the Unix tradition. Many of its most effective tools are thin wrappers around a direct translation of some single powerful algorithm.

Perhaps the clearest example of this is diff(1), the Unix tool for reporting differences between related files. This tool and its dual, patch(1), have become central to the network-distributed development style of modern Unix. A valuable property of diff is that it seldom surprises anyone. It doesn't have special cases or painful edge conditions, because it uses a simple, mathematically sound method of sequence comparison. This has consequences:

By virtue of a mathematical model and a solid algorithm, Unix diff contrasts markedly with its imitators. First, the central engine is solid, small, and has never needed one line of maintenance. Second, the results are clear and consistent, unmarred by surprises where heuristics fail.

– Doug McIlroy

Thus, people who use diff can develop an intuitive feel for what it will do in any given situation without necessarily understanding the central algorithm perfectly. Other well-known examples of this special kind of clarity achieved through a strong central algorithm abound in Unix:

All three of these programs are so bug-free that their correct functioning is taken utterly for granted, and compact enough to fit easily in a programmer's hand. Only a part of these good qualities are due to the polishing that comes with a long service life and frequent use; most of it is that, having been constructed around a strong and provably correct algorithmic core, they never needed much polishing in the first place.

The opposite of a formal approach is using heuristics —rules of thumb leading toward a solution that is probabilistically, but not certainly, correct. Sometimes we use heuristics because a deterministically correct solution is impossible. Think of spam filtering, for example; an algorithmically perfect spam filter would need a full solution to the problem of understanding natural language as a module. Other times, we use heuristics because known formally correct methods are impossibly expensive. Virtual-memory management is an example of this; there are near-perfect solutions, but they require so much runtime instrumentation that their overhead would swamp any theoretical gain over heuristics.

The trouble with heuristics is that they proliferate special cases and edge cases. If nothing else, you usually have to backstop a heuristic with some sort of recovery mechanism when it fails. All the usual problems with escalating complexity follow. To manage the resulting tradeoffs, you have to start by being aware of them. Always ask if a heuristic actually pays off in performance what it costs in code complexity — and don't guess at the performance difference, actually measure it before making a decision.

The Value of Detachment

We began this book with a reference to Zen: “a special transmission, outside the scriptures”. This was not mere exoticism for stylistic effect; the core concepts of Unix have always had a spare, Zen-like simplicity that continues to shine through the layers of historical accidents that have accreted around them. This quality is reflected in the cornerstone documents of Unix, like The C Programming Language [Kernighan-Ritchie] and the 1974 CACM paper that introduced Unix to the world; one of the famous quotes from that paper observes “...constraint has encouraged not only economy, but also a certain elegance of design”. That simplicity came from trying to think not about how much a language or operating system could do, but of how little it could do — not by carrying assumptions but by starting from zero (what in Zen is called “beginner's mind” or “empty mind”).

To design for compactness and orthogonality, start from zero. Zen teaches that attachment leads to suffering; experience with software design teaches that attachment to unnoticed assumptions leads to non-orthogonality, noncompact designs, and projects that fail or become maintenance nightmares.

To achieve enlightenment and surcease from suffering, Zen teaches detachment. The Unix tradition teaches the value of detachment from the particular, accidental conditions under which a design problem was posed. Abstract. Simplify. Generalize. Because we write software to solve problems, we cannot completely detach from the problems — but it is well worth the mental effort to see how many preconceptions you can throw away, and whether the design becomes more compact and orthogonal as you do that. Possibilities for code reuse often result.

Jokes about the relationship between Unix and Zen are a live part of the Unix tradition as well.[45] This is not an accident.


[43] In the foundation text on this topic, Refactoring [Fowler], the author comes very close to stating that the principal goal of refactoring is to improve orthogonality. But lacking the concept, he can only approximate this idea from several different directions: eliminating code duplication and various other “bad smells” many of which are some sort of orthogonality violation.

[44] An archetypal example of bad caching is the rehash directive in csh(1); type man 1 csh for details. See the section called “Caching Operation Results” for another example.

[45] For a recent example of Unix/Zen crossover, see Appendix D.

Software Is a Many-Layered Thing

Broadly speaking, there are two directions one can go in designing a hierarchy of functions or objects. Which direction you choose, and when, has a profound effect on the layering of your code.

Top-Down versus Bottom-Up

One direction is bottom-up, from concrete to abstract — working up from the specific operations in the problem domain that you know you will need to perform. For example, if one is designing firmware for a disk drive, some of the bottom-level primitives might be ‘seek head to physical block’, ‘read physical block’, ‘write physical block’, ‘toggle drive LED’, etc.

The other direction is top-down, abstract to concrete — from the highest-level specification describing the project as a whole, or the application logic, downwards to individual operations. Thus, if one is designing software for a mass-storage controller that might drive several different sorts of media, one might start with abstract operations like ‘seek logical block’, ‘read logical block’, ‘write logical block’, ‘toggle activity indication’. These would differ from the similarly named hardware-level operations above in that they're intended to be generic across different kinds of physical devices.

These two examples could be two ways of approaching design for the same collection of hardware. Your choice, in cases like this, is one of these: either abstract the hardware (so the objects encapsulate the real things out there and the program is merely a list of manipulations on those things), or organize around some behavioral model (and then embed the actual hardware manipulations that carry it out in the flow of the behavioral logic).

An analogous choice shows up in a lot of different contexts. Suppose you're writing MIDI sequencer software. You could organize that code around its top level (sequencing tracks) or around its bottom level (switching patches or samples and driving wave generators).

A very concrete way to think about this difference is to ask whether the design is organized around its main event loop (which tends to have the high-level application logic close to it) or around a service library of all the operations that the main loop can invoke. A designer working from the top down will start by thinking about the program's main event loop, and plug in specific events later. A designer working from the bottom up will start by thinking about encapsulating specific tasks and glue them together into some kind of coherent order later on.

For a larger example, consider the design of a Web browser. The top-level design of a Web browser is a specification of the expected behavior of the browser: what types of URL (like http: or ftp: or file:) it interprets, what kinds of images it is expected to be able to render, whether and with what limitations it will accept Java or JavaScript, etc. The layer of the implementation that corresponds to this top-level view is its main event loop; each time around, the loop waits for, collects, and dispatches on a user action (such as clicking a Web link or typing a character into a field).

But the Web browser has to call a large set of domain primitives to do its job. One group of these is concerned with establishing network connections, sending data over them, and receiving responses. Another set is the operations of the GUI toolkit the browser will use. Yet a third set might be concerned with the mechanics of parsing retrieved HTML from text into a document object tree.

Which end of the stack you start with matters a lot, because the layer at the other end is quite likely to be constrained by your initial choices. In particular, if you program purely from the top down, you may find yourself in the uncomfortable position that the domain primitives your application logic wants don't match the ones you can actually implement. On the other hand, if you program purely from the bottom up, you may find yourself doing a lot of work that is irrelevant to the application logic — or merely designing a pile of bricks when you were trying to build a house.

Ever since the structured-programming controversies of the 1960s, novice programmers have generally been taught that the correct approach is the top-down one: stepwise refinement, where you specify what your program is to do at an abstract level and gradually fill in the blanks of implementation until you have concrete working code. Top-down tends to be good practice when three preconditions are true: (a) you can specify in advance precisely what the program is to do, (b) the specification is unlikely to change significantly during implementation, and (c) you have a lot of freedom in choosing, at a low level, how the program is to get that job done.

These conditions tend to be fulfilled most often in programs relatively close to the user and high in the software stack — applications programming. But even there those preconditions often fail. You can't count on knowing what the ‘right’ way for a word processor or a drawing program to behave is until the user interface has had end-user testing. Purely top-down programming often has the effect of overinvesting effort in code that has to be scrapped and rebuilt because the interface doesn't pass a reality check.

In self-defense against this, programmers try to do both things — express the abstract specification as top-down application logic, and capture a lot of low-level domain primitives in functions or libraries, so they can be reused when the high-level design changes.

Unix programmers inherit a tradition that is centered in systems programming, where the low-level primitives are hardware-level operations that are fixed in character and extremely important. They therefore lean, by learned instinct, more toward bottom-up programming.

Whether you're a systems programmer or not, bottom-up can also look more attractive when you are programming in an exploratory way, trying to get a grasp on hardware or software or real-world phenomena you don't yet completely understand. Bottom-up programming gives you time and room to refine a vague specification. Bottom-up also appeals to programmers' natural human laziness — when you have to scrap and rebuild code, you tend to have to throw away larger pieces if you're working top-down than you do if you're working bottom-up.

Real code, therefore tends to be programmed both top-down and bottom-up. Often, top-down and bottom-up code will be part of the same project. That's where ‘glue’ enters the picture.

Glue Layers

When the top-down and bottom-up drives collide, the result is often a mess. The top layer of application logic and the bottom layer of domain primitives have to be impedance-matched by a layer of glue logic.

One of the lessons Unix programmers have learned over decades is that glue is nasty stuff and that it is vitally important to keep glue layers as thin as possible. Glue should stick things together, but should not be used to hide cracks and unevenness in the layers.

In the Web-browser example, the glue would include the rendering code that maps a document object parsed from incoming HTML into a flattened visual representation as a bitmap in a display buffer, using GUI domain primitives to do the painting. This rendering code is notoriously the most bug-prone part of a browser. It attracts into itself kluges to address problems that originate both in the HTML parsing (because there is a lot of ill-formed markup out there) and the GUI toolkit (which may not have quite the primitives that are really needed).

A Web browser's glue layer has to mediate not merely between specification and domain primitives, but between several different external specifications: the network behavior standardized in HTTP, HTML document structure, and various graphics and multimedia formats as well as the users' behavioral expectations from the GUI.

And one single bug-prone glue layer is not the worst fate that can befall a design. A designer who is aware that the glue layer exists, and tries to organize it into a middle layer around its own set of data structures or objects, can end up with two layers of glue — one above the midlayer and one below. Programmers who are bright but unseasoned are particularly apt to fall into this trap; they'll get each fundamental set of classes (application logic, midlayer, and domain primitives) right and make them look like the textbook examples, only to flounder as the multiple layers of glue needed to integrate all that pretty code get thicker and thicker.

The thin-glue principle can be viewed as a refinement of the Rule of Separation. Policy (the application logic) should be cleanly separated from mechanism (the domain primitives), but if there is a lot of code that is neither policy nor mechanism, chances are that it is accomplishing very little besides adding global complexity to the system.

Case Study: C Considered as Thin Glue

The C language itself is a good example of the effectiveness of thin glue.

In the late 1990s, Gerrit Blaauw and Fred Brooks observed in Computer Architecture: Concepts and Evolution [BlaauwBrooks] that the architectures in every generation of computers, from early mainframes through minicomputers through workstations through PCs, had tended to converge. The later a design was in its technology generation, the more closely it approximated what Blaauw & Brooks called the “classical architecture”: binary representation, flat address space, a distinction between memory and working store (registers), general-purpose registers, address resolution to fixed-length bytes, two-address instructions, big-endianness,[46] and data types a consistent set with sizes a multiple of either 4 or 6 bits (the 6-bit families are now extinct).

Thompson and Ritchie designed C to be a sort of structured assembler for an idealized processor and memory architecture that they expected could be efficiently modeled on most conventional computers. By happy accident, their model for the idealized processor was the PDP-11, a particularly mature and elegant minicomputer design that closely approximated Blaauw & Brooks's classical architecture. By good judgment, Thompson and Ritchie declined to wire into their language most of the few traits (such as little-endian byte order) where the PDP-11 didn't match it.[47]

The PDP-11 became an important model for the following generations of microprocessor architectures. The basic abstractions of C turned out to capture the classical architecture rather neatly. Thus, C started out as a good fit for microprocessors and, rather than becoming irrelevant as its assumptions fell out of date, actually became a better fit as hardware converged more closely on the classical architecture. One notable example of this convergence was when Intel's 386, with its large flat memory-address space, replaced the 286's awkward segmented-memory addressing after 1985; pure C was actually a better fit for the 386 than it had been for the 286.

It is not a coincidence that the experimental era in computer architectures ended in the mid-1980s at the same time that C (and its close descendant C++) were sweeping all before them as general-purpose programming languages. C, designed as a thin but flexible layer over the classical architecture, looks with two decades' additional perspective like almost the best possible design for the structured-assembler niche it was intended to fill. In addition to compactness, orthogonality, and detachment (from the machine architecture on which it was originally designed), it also has the important quality of transparency that we will discuss in Chapter 6. The few language designs since that are arguably better have needed to make large changes (like introducing garbage collection) in order to get enough functional distance from C not to be swamped by it.

This history is worth recalling and understanding because C shows us how powerful a clean, minimalist design can be. If Thompson and Ritchie had been less wise, they would have designed a language that did much more, relied on stronger assumptions, never ported satisfactorily off its original hardware platform, and withered away as the world changed out from under it. Instead, C has flourished — and the example Thompson and Ritchie set has influenced the style of Unix development ever since. As the writer, adventurer, artist, and aeronautical engineer Antoine de Saint-Exupéry once put it, writing about the design of airplanes: «La perfection est atteinte non quand il ne reste rien à ajouter, mais quand il ne reste rien à enlever». (“Perfection is attained not when there is nothing more to add, but when there is nothing more to remove”.)

Ritchie and Thompson lived by this maxim. Long after the resource constraints on early Unix software had eased, they worked at keeping C as thin a layer over the hardware as possible.

Dennis used to say to me, when I would ask for some particularly extravagant feature in C, “If you want PL/1, you know where to get it”. He didn't have to deal with some marketer saying “But we need a check in the box on the sales viewgraph!”

– Mike Lesk

The history of C is also a lesson in the value of having a working reference implementation before you standardize. We'll return to this point in Chapter 17 when we discuss the evolution of C and Unix standards.


[46] The terms big-endian and little-endian refer to architectural choices about the order in which bits are interpreted within a machine word. Though it has no canonical location, a Web search for On Holy Wars and a Plea for Peace should turn up a classic and entertaining paper on this subject.

[47] The widespread belief that the autoincrement and autodecrement features entered C because they represented PDP-11 machine instructions is a myth. According to Dennis Ritchie, these operations were present in the ancestral B language before the PDP-11 existed.


One consequence of the emphasis that the Unix programming style put on modularity and well-defined APIs is a strong tendency to factor programs into bits of glue connecting collections of libraries, especially shared libraries (the equivalents of what are called dynamically-linked libraries or DLLs under Windows and other operating systems).

If you are careful and clever about design, it is often possible to partition a program so that it consists of a user-interface-handling main section (policy) and a collection of service routines (mechanism) with effectively no glue at all. This approach is especially appropriate when the program has to do a lot of very specific manipulations of data structures like graphic images, network-protocol packets, or control blocks for a hardware interface. Some good general architectural advice from within the Unix tradition, particularly applicable to the resource-management challenges of this sort of library is collected in The Discipline and Method Architecture for Reusable Libraries[Vo].

Under Unix, it is normal practice to make this layering explicit, with the service routines collected in a library that is separately documented. In such programs, the front end gets to specialize in user-interface considerations and high-level protocol. With a little more care in design, it may be possible to detach the original front end and replace it with others adapted for different purposes. Some other advantages should become evident from our case study.

There is a flip side to this. In the Unix world, libraries which are delivered as libraries should come with exerciser programs.

APIs should come with programs, and vice versa. An API that you must write C code to use, which cannot be invoked easily from the command line, is harder to learn and use. And contrariwise, it's a royal pain to have interfaces whose only open, documented form is a program, so you cannot invoke them easily from a C program — for example, route(1) in older Linuxes.

– Henry Spencer

Besides easing the learning curve, library exercisers often make excellent test frameworks. Experienced Unix programmers therefore see them not just as a form of thoughtfulness to the library's users but as an indication that the code has probably been well tested.

An important form of library layering is the plugin, a library with a set of known entry points that is dynamically loaded after startup time to perform a specialized task. For plugins to work, the calling program has to be organized largely as a documented service library that the plugin can call back into.

Case Study: GIMP Plugins

The GIMP (GNU Image Manipulation program) is a graphics editor designed to be driven through an interactive GUI. But GIMP is built as a library of image-manipulation and housekeeping routines called by a relatively thin layer of control code. The driver code knows about the GUI, but not directly about image formats; the library routines reverse this by knowing about image formats and operations but not about the GUI.

The library layer is documented (and, in fact shipped as “libgimp” for use by other programs). This means that C programs called “plugins” can be dynamically loaded by GIMP and call the library to do image manipulation, effectively taking over control at the same level as the GUI (see Figure 4.2).

Figure 4.2. Caller/callee relationships in GIMP with a plugin loaded.

Caller/callee relationships in GIMP with a plugin loaded.

Plugins are used to perform lots of special-purpose transformations such as colormap hacking, blurring and despeckling; also for reading and writing file formats not native to the GIMP core; for extensions like editing animations and window manager themes; and for lots of other sorts of image-hacking that can be automated by scripting the image-hacking logic in the GIMP core. A registry of GIMP plugins is available on the World Wide Web.

Though most GIMP plugins are small, simple C programs, it is also possible to write a plugin that exposes the library API to a scripting language; we'll discuss this possibility in Chapter 11 when we examine the ‘polyvalent program’ pattern.

Unix and Object-Oriented Languages

Since the mid-1980s most new language designs have included native support for object-oriented programming (OO). Recall that in object-oriented programming, the functions that act on a particular data structure are encapsulated with the data in an object that can be treated as a unit. By contrast, modules in non-OO languages make the association between data and the functions that act on it rather accidental, and modules frequently leak data or bits of their internals into each other.

The OO design concept initially proved valuable in the design of graphics systems, graphical user interfaces, and certain kinds of simulation. To the surprise and gradual disillusionment of many, it has proven difficult to demonstrate significant benefits of OO outside those areas. It's worth trying to understand why.

There is some tension and conflict between the Unix tradition of modularity and the usage patterns that have developed around OO languages. Unix programmers have always tended to be a bit more skeptical about OO than their counterparts elsewhere. Part of this is because of the Rule of Diversity; OO has far too often been promoted as the One True Solution to the software-complexity problem. But there is something else behind it as well, an issue which is worth exploring as background before we evaluate specific OO (object-oriented) languages in Chapter 14. It will also help throw some characteristics of the Unix style of non-OO programming into sharper relief.

We observed above that the Unix tradition of modularity is one of thin glue, a minimalist approach with few layers of abstraction between the hardware and the top-level objects of a program. Part of this is the influence of C. It takes serious effort to simulate true objects in C. Because that's so, piling up abstraction layers is an exhausting thing to do. Thus, object hierarchies in C tend to be relatively flat and transparent. Even when Unix programmers use other languages, they tend to want to carry over the thin-glue/shallow-layering style that Unix models have taught them.

OO languages make abstraction easy — perhaps too easy. They encourage architectures with thick glue and elaborate layers. This can be good when the problem domain is truly complex and demands a lot of abstraction, but it can backfire badly if coders end up doing simple things in complex ways just because they can.

All OO languages show some tendency to suck programmers into the trap of excessive layering. Object frameworks and object browsers are not a substitute for good design or documentation, but they often get treated as one. Too many layers destroy transparency : It becomes too difficult to see down through them and mentally model what the code is actually doing. The Rules of Simplicity, Clarity, and Transparency get violated wholesale, and the result is code full of obscure bugs and continuing maintenance problems.

This tendency is probably exacerbated because a lot of programming courses teach thick layering as a way to satisfy the Rule of Representation. In this view, having lots of classes is equated with embedding knowledge in your data. The problem with this is that too often, the ‘smart data’ in the glue layers is not actually about any natural entity in whatever the program is manipulating — it's just about being glue. (One sure sign of this is a proliferation of abstract subclasses or ‘mixins’.)

Another side effect of OO abstraction is that opportunities for optimization tend to disappear. For example, a + a + a + a can become a * 4 and even a << 2 if a is an integer. But if one creates a class with operators, there is nothing to indicate if they are commutative, distributive, or associative. Since one isn't supposed to look inside the object, it's not possible to know which of two equivalent expressions is more efficient. This isn't in itself a good reason to avoid using OO techniques on new projects; that would be premature optimization. But it is reason to think twice before transforming non-OO code into a class hierarchy.

Unix programmers tend to share an instinctive sense of these problems. This tendency appears to be one of the reasons that, under Unix, OO languages have failed to displace non-OO workhorses like C, Perl (which actually has OO facilities, but they're not heavily used), and shell. There is more vocal criticism of OO in the Unix world than orthodoxy permits elsewhere; Unix programmers know when not to use OO; and when they do use OO languages, they spend more effort on trying to keep their object designs uncluttered. As the author of The Elements of Networking Style once observed in a slightly different context [Padlipsky]: “If you know what you're doing, three layers is enough; if you don't, even seventeen levels won't help”.

One reason that OO has succeeded most where it has (GUIs, simulation, graphics) may be because it's relatively difficult to get the ontology of types wrong in those domains. In GUIs and graphics, for example, there is generally a rather natural mapping between manipulable visual objects and classes. If you find yourself proliferating classes that have no obvious mapping to what goes on in the display, it is correspondingly easy to notice that the glue has gotten too thick.

One of the central challenges of design in the Unix style is how to combine the virtue of detachment (simplifying and generalizing problems from their original context) with the virtue of thin glue and shallow, flat, transparent hierarchies of code and design.

We'll return to some of these points and apply them when we discuss object-oriented languages in Chapter 14.

Coding for Modularity

Modularity is expressed in good code, but it primarily comes from good design. Here are some questions to ask about any code you work on that might help you improve its modularity:

You might find it instructive to compare these with our checklist of questions about transparency, and discoverability in Chapter 6.


[48] Globals also mean your code cannot be reentrant; that is, multiple instances in the same process are likely to step on each other.

[49] Many years ago, I learned from Kernighan & Plauger's The Elements of Programming Style a useful rule. Write that one-line comment immediately after the prototype of your function. For every function, without exception.

[50] A cheap way to collect this information is to analyze the tags files generated by a utility like etags(1) or ctags(1).

Chapter 5. Textuality

It's a well-known fact that computing devices such as the abacus were invented thousands of years ago. But it's not well known that the first use of a common computer protocol occurred in the Old Testament. This, of course, was when Moses aborted the Egyptians' process with a control-sea.

– Tom Galloway rec.arts.comics, February 1992

In this chapter, we'll look at what the Unix tradition has to tell us about two different kinds of design that are closely related: the design of file formats for retaining application data in permanent storage, and the design of application protocols for passing data and commands between cooperating programs, possibly over a network.

What unifies these two kinds of design is that they both involve the serialization of in-memory data structures. For the internal operation of computer programs, the most convenient representation of a complex data structure is one in which all fields have the machine's native data format (e.g. two's-complement binary for integers) and all pointers are actual memory addresses (as opposed, say, to being named references). But these representations are not well suited to storage and transmission; memory addresses in the data structure lose their meaning outside memory, and emitting raw native data formats causes interoperability problems passing data between machines with different conventions (big- vs. little-endian, say, or 32-bit vs. 64-bit).

For transmission and storage, the traversable, quasi-spatial layout of data structures like linked lists needs to be flattened or serialized into a byte-stream representation from which the structure can later be recovered. The serialization (save) operation is sometimes called marshaling and its inverse (load) operation unmarshaling. These terms are usually applied with respect to objects in an OO language like C++ or Python or Java, but could be used with equal justice of operations like loading a graphics file into the internal storage of a graphics editor and saving it out after modifications.

A significant percentage of what C and C++ programmers maintain is ad-hoc code for marshaling and unmarshaling operations — even when the serialized representation chosen is as simple as a binary structure dump (a common technique under non-Unix environments). Modern languages like Python and Java tend to have built-in unmarshal and marshal functions that can be applied to any object or byte-stream representing an object, and that reduce this labor substantially.

But these naïve methods are often unsatisfactory for various reasons, including both the machine-interoperability problems we mentioned above and the negative trait of being opaque to other tools. When the application is a network protocol, economy may demand that an internal data structure (such as, say, a message with source and destination addresses) be serialized not into a single blob of data but into a series of attempted transactions or messages which the receiving machine may reject (so that, for example, a large message can be rejected if the destination address is invalid).

Interoperability, transparency, extensibility, and storage or transaction economy: these are the important themes in designing file formats and application protocols. Interoperability and transparency demand that we focus such designs on clean data representations, rather than putting convenience of implementation or highest possible performance first. Extensibility also favors textual protocols, since binary ones are often harder to extend or subset cleanly. Transaction economy sometimes pushes in the opposite direction — but we shall see that putting that criterion first is a form of premature optimization that it is often wise to resist.

Finally, we must note a difference between data file formats and the run-control files that are often used to set the startup options of Unix programs. The most basic difference is that (with sporadic exceptions like GNU Emacs's configuration interface) programs don't normally modify their own run-control files — the information flow is one-way, from file read at startup time to application settings. Data-file formats, on the other hand, associate properties with named resources and are both read and written by their applications. Configuration files are generally hand-edited and small, whereas data files are program-generated and can become arbitrarily large.

Historically, Unix has related but different sets of conventions for these two kinds of representation. The conventions for run control files are surveyed in Chapter 10; only conventions for data files are examined in this chapter.

The Importance of Being Textual

Pipes and sockets will pass binary data as well as text. But there are good reasons the examples we'll see in Chapter 7 are textual: reasons that hark back to Doug McIlroy's advice quoted in Chapter 1. Text streams are a valuable universal format because they're easy for human beings to read, write, and edit without specialized tools. These formats are (or can be designed to be) transparent.

Also, the very limitations of text streams help enforce encapsulation. By discouraging elaborate representations with rich, densely encoded structure, text streams also discourage programs from being promiscuous with each other about their internal states and help enforce encapsulation. We'll return to this point at the end of Chapter 7 when we discuss RPC.

When you feel the urge to design a complex binary file format, or a complex binary application protocol, it is generally wise to lie down until the feeling passes. If performance is what you're worried about, implementing compression on the text protocol stream either at some level below or above the application protocol will give you a cleaner and perhaps better-performing design than a binary protocol (text compresses well, and quickly).

A bad example of binary formats in Unix history was the way device-independent troff read a binary file containing device information, supposedly for speed. The initial implementation generated that binary file from a text description in a somewhat unportable way. Faced with a need to port ditroff quickly to a new machine, rather than reinvent the binary goo, I ripped it out and just had ditroff read the text file. With carefully crafted file-reading code, the speed penalty was negligible.

– Henry Spencer

Designing a textual protocol tends to future-proof your system. One specific reason is that ranges on numeric fields aren't implied by the format itself. Binary formats usually specify the number of bits allocated to a given value, and extending them is difficult. For example, IPv4 only allows 32 bits for an address. To extend address size to 128 bits (as done by IPv6) requires a major revamping.[51] In contrast, if you need a larger value in a text format, just write it. It may be that a given program can't receive values in that range, but it's usually easier to modify the program than to modify all the data stored in that format.

The only good justification for a binary protocol is if you're going to be manipulating large enough data sets that you're genuinely worried about getting the most bit-density out of your media, or if you're very concerned about the time or instruction budget required to interpret the data into an in-core structure. Formats for large images and multimedia are sometimes an example of the former, and network protocols with hard latency requirements sometimes an example of the latter.

The reciprocal problem with SMTP or HTTP-like text protocols is that they tend to be expensive in bandwidth and slow to parse. The smallest X request is 4 bytes: the smallest HTTP request is about 100 bytes. X requests, including amortized overhead of transport, can be executed in the order of 100 instructions; at one point, an Apache [web server] developer proudly indicated they were down to 7000 instructions. For graphics, bandwidth becomes everything on output; hardware is designed such that these days the graphics-card bus is the bottleneck for small operations, so any protocol had better be very tight if it is not to be a worse bottleneck. This is the extreme case.

– Jim Gettys

These concerns are valid in other extreme cases as well as in X — for example, in the design of graphics file formats intended to hold very large images. But they are usually just another case of premature-optimization fever. Textual formats don't necessarily have much lower bit density than binary ones; they do after all use seven out of eight bits per byte. And what you gain by not having to parse text, you generally lose the first time you need to generate a test load, or to eyeball a program-generated example of your format and figure out what's in there.

In addition, the kind of thinking that goes into designing tight binary formats tends to fall down on making them cleanly extensible. The X designers experienced this:

Against the current X framework is the fact we didn't design enough of a structure to make it easier to ignore trivial extensions to the protocol; we can do this some of the time, but a bit better framework would have been good.

– Jim Gettys

When you think you have an extreme case that justifies a binary file format or protocol, you need to think very carefully about extensibility and leaving room in the design for growth.

Case Study: Unix Password File Format

On many operating systems, the per-user data required to validate logins and start a user's session is an opaque binary database. Under Unix, by contrast, it's a text file with records one per line and colon-separated fields.

Example 5.1 consists of some randomly-chosen example lines:

Example 5.1. Password file example.

ftp:*:14:50:FTP User:/home/ftp:
esr:0SmFuPnH5JlNs:23:23:Eric S. Raymond:/home/esr:

Without even knowing anything about the semantics of the fields, we can notice that it would be hard to pack the data much tighter in a binary format. The colon sentinel characters would have to have functional equivalents taking at least as much space (usually either count bytes or NULs). The per-user records would either have to have terminators (which could hardly be shorter than a single newline) or else be wastefully padded out to a fixed length.

Actually the prospects for saving space through binary encoding pretty much vanish if you know the actual semantics of the data. The numeric user ID (3rd) and group ID (4th) fields are integers, thus on most machines a binary representation would be at least 4 bytes, and longer than the text for values up to 999. But let's agree to ignore this for now and suppose the best case that the numeric fields have a 0-255 range.

We could tighten up the numeric fields (3rd and 4th) by collapsing the numerics to single bytes, and the password strings (2nd) to an 8-bit encoding. On this example, that would give about an 8% size decrease.

That 8% of putative inefficiency buys us a lot. It avoids putting an arbitrary limit on the range of the numeric fields. It gives us the ability to modify the password file with any old text editor of our choice, rather than having to build a specialized tool to edit a binary format (though in the case of the password file itself, we have to be extra careful about concurrent edits). And it gives us the ability to do ad-hoc searches and filters and reports on the user account information with text-stream tools such as grep(1).

We do have to be a bit careful about not embedding a colon in any of the textual fields. Good practice is to tell the file write code to precede embedded colons with an escape character, and then to tell the file read code to interpret it. Unix tradition favors backslash for this use.

The fact that structural information is conveyed by field position rather than an explicit tag makes this format faster to read and write, but a bit rigid. If the set of properties associated with a key is expected to change with any frequency, one of the tagged formats described below might be a better choice.

Economy is not a major issue with password files to begin with, as they're normally read seldom[52] and infrequently modified. Interoperability is not an issue, since various data in the file (notably user and group numbers) are not portable off the originating machine. For password files, it's therefore quite clear that going where the transparency criterion leads was the right thing.

Case Study: .newsrc Format

Usenet news is a worldwide distributed bulletin-board system that anticipated today's P2P networking by two decades. It uses a message format very similar to that of RFC 822 electronic-mail messages, except that instead of being directed to personal recipients messages are sent to topic groups. Articles posted at any participating site are broadcast to each site that it has registered as a neighbor, and eventually flood-fill to all news sites.

Almost all Usenet news readers understand the .newsrc file, which records which Usenet messages have been seen by the calling user. Though it is named like a run-control file, it is not only read at startup but typically updated at the end of the newsreader run. The .newsrc format has been fixed since the first newsreaders around 1980. Example 5.2 is a representative section from a .newsrc file.

Example 5.2. A .newsrc example.

rec.arts.sf.misc! 1-14774,14786,14789! 1-2534
rec.arts.sf.written: 1-876513
news.answers! 1-199359,213516,215735
news.announce.newusers! 1-4399
news.newusers.questions! 1-645661
news.groups.questions! 1-32676! 1-95504,137265,137274,140059,140091,140117
alt.test! 1-1441498

Each line sets properties for the newsgroup named in the first field. The name is immediately followed by a character that indicates whether the owning user is currently subscribed to the group or not; a colon indicates subscription, and an exclamation mark indicates nonsubscription. The remainder of the line is a sequence of comma-separated article numbers or ranges of article numbers, indicating which articles the user has seen.

Non-Unix programmers might have automatically tried to design a fast binary format in which each newsgroup status was described by either a long but fixed-length binary record, or a sequence of self-describing binary packets with internal length fields. The main point of such a binary representation would be to express ranges with binary data in paired word-length fields, in order to avoid the overhead of parsing all the range expressions at startup.

Such a layout could be read and written faster than a textual format, but it would have other problems. A naïve implementation in fixed-length records would have placed artificial length limits on newsgroup names and (more seriously) on the maximum number of ranges of seen-article numbers. A more sophisticated binary-packet format would avoid the length limits, but could not be edited with the user's eyeballs and fingers — a capability that can be quite useful when you want to reset just some of the read bits in an individual newsgroup. Also, it would not necessarily be portable to different machine types.

The designers of the original newsreader chose transparency and interoperability over economy. The case for going in the other direction was not completely ridiculous; .newsrc files can get very large, and one modern reader (GNOME's Pan) uses a speed-optimized private format to avoid startup lag. But to other implementers, textual representation looked like a good tradeoff in 1980, and has looked better as machines increased in speed and storage dropped in price.

Case Study: The PNG Graphics File Format

PNG (Portable Network Graphics) is a file format for bitmap graphics. It is like GIF, and unlike JPEG, in that it uses lossless compression and is optimized for applications such as line art and icons rather than photographic images. Documentation and open-source reference libraries of high quality are available at the Portable Network Graphics website.

PNG is an excellent example of a thoughtfully designed binary format. A binary format is appropriate since graphics files may contain very large amounts of data, such that storage size and Internet download time would go up significantly if the pixel data were stored textually. Transaction economy was the prime consideration, with transparency sacrificed.[53] The designers were, however, careful about interoperability; PNG specifies byte orders, integer word lengths, endianness, and (lack of) padding between fields.

A PNG file consists of a sequence of chunks, each in a self-describing format beginning with the chunk type name and the chunk length. Because of this organization, PNG does not need a release number. New chunk types can be added at any time; the case of the first letter in the chunk type name informs PNG-using software whether or not each chunk can be safely ignored.

The PNG file header also repays study. It has been cleverly designed to make various common kinds of file corruption (e.g., by 7-bit transmission links, or mangling of CR and LF characters) easy to detect.

The PNG standard is precise, comprehensive, and well written. It could serve as a model for how to write file format standards.


[51] There is a legend that some early airline reservation systems allocated exactly one byte for a plane's passenger count. Supposedly they became very confused by the arrival of the Boeing 747, the first plane that could carry more than 255 passengers.

[52] Password files are normally read once per user session at login time, and after that occasionally by file-system utilities like ls(1) that must map from numeric user and group IDs to names.

[53] Confusingly, PNG supports a different kind of transparency — transparent pixels in the PNG image.

Data File Metaformats

A data file metaformat is a set of syntactic and lexical conventions that is either formally standardized or sufficiently well established by practice that there are standard service libraries to handle marshaling and unmarshaling it.

Unix has evolved or adopted metaformats suitable for a wide range of applications. It is good practice to use one of these (rather than an idiosyncratic custom format) wherever possible. The benefits begin with the amount of custom parsing and generation code that you may be able to avoid writing by using a service library. But the most important benefit is that developers and even many users will instantly recognize these formats and feel comfortable with them, which reduces the friction costs of learning new programs.

In the following discussion, when we refer to “traditional Unix tools” we are intending the combination of grep(1), sed(1), awk(1), tr(1), and cut(1) for doing text searches and transformations. Perl and other scripting languages tend to have good native support for parsing the line-oriented formats that these tools encourage.

Here, then, are the standard formats that can serve you as models.

DSV Style

DSV stands for Delimiter-Separated Values. Our first case study in textual metaformats was the /etc/passwd file, which is a DSV format with colon as the value separator. Under Unix, colon is the default separator for DSV formats in which the field values may contain whitespace.

/etc/passwd format (one record per line, colon-separated fields) is very traditional under Unix and frequently used for tabular data. Other classic examples include the /etc/group file describing security groups and the /etc/inittab file used to control startup and shutdown of Unix service programs at different run levels of the operating system.

Data files in this style are expected to support inclusion of colons in the data fields by backslash escaping. More generally, code that reads them is expected to support record continuation by ignoring backslash-escaped newlines, and to allow embedding nonprintable character data by C-style backslash escapes.

This format is most appropriate when the data is tabular, keyed by a name (in the first field), and records are typically short (less than 80 characters long). It works well with traditional Unix tools.

One occasionally sees field separators other than the colon, such as the pipe character | or even an ASCII NUL. Old-school Unix practice used to favor tabs, a preference reflected in the defaults for cut(1) and paste(1); but this has gradually changed as format designers became aware of the many small irritations that ensue from the fact that tabs and spaces are not visually distinguishable.

This format is to Unix what CSV (comma-separated value) format is under Microsoft Windows and elsewhere outside the Unix world. CSV (fields separated by commas, double quotes used to escape commas, no continuation lines) is rarely found under Unix.

In fact, the Microsoft version of CSV is a textbook example of how not to design a textual file format. Its problems begin with the case in which the separator character (in this case, a comma) is found inside a field. The Unix way would be to simply escape the separator with a backslash, and have a double escape represent a literal backslash. This design gives us a single special case (the escape character) to check for when parsing the file, and only a single action when the escape is found (treat the following character as a literal). The latter conveniently not only handles the separator character, but gives us a way to handle the escape character and newlines for free. CSV, on the other hand, encloses the entire field in double quotes if it contains the separator. If the field contains double quotes, it must also be enclosed in double quotes, and the individual double quotes in the field must themselves be repeated twice to indicate that they don't end the field.

The bad results of proliferating special cases are twofold. First, the complexity of the parser (and its vulnerability to bugs) is increased. Second, because the format rules are complex and underspecified, different implementations diverge in their handling of edge cases. Sometimes continuation lines are supported, by starting the last field of the line with an unterminated double quote — but only in some products! Microsoft has incompatible versions of CSV files between its own applications, and in some cases between different versions of the same application (Excel being the obvious example here).

RFC 822 Format

The RFC 822 metaformat derives from the textual format of Internet electronic mail messages; RFC 822 is the principal Internet RFC describing this format (since superseded by RFC 2822). MIME (Multipurpose Internet Media Extension) provides a way to embed typed binary data within RFC-822-format messages. (Web searches on either of these names will turn up the relevant standards.)

In this metaformat, record attributes are stored one per line, named by tokens resembling mail header-field names and terminated with a colon followed by whitespace. Field names do not contain whitespace; conventionally a dash is substituted instead. The attribute value is the entire remainder of the line, exclusive of trailing whitespace and newline. A physical line that begins with tab or whitespace is interpreted as a continuation of the current logical line. A blank line may be interpreted either as a record terminator or as an indication that unstructured text follows.

Under Unix, this is the traditional and preferred textual metaformat for attributed messages or anything that can be closely analogized to electronic mail. More generally, it's appropriate for records with a varying set of fields in which the hierarchy of data is flat (no recursion or tree structure).

Usenet news uses it; so do the HTTP 1.1 (and later) formats used by the World Wide Web. It is very convenient for editing by humans. Traditional Unix search tools are still good for attribute searches, though finding record boundaries will be a little more work than in a record-per-line format.

One weakness of RFC 822 format is that when more than one RFC 822 message or record is put in a file, the record boundaries may not be obvious — how is a poor literal-minded computer to know where the unstructured text body of a message ends and the next header begins? Historically, there have been several different conventions for delimiting messages in mailboxes. The oldest and most widely supported, leading each message with a line that begins with the string "From " and sender information, is not appropriate for other kinds of records; it also requires that lines in message text beginning with "From " be escaped (typically with >) — a practice which not infrequently leads to confusion.

Some mail systems use delimiter lines consisting of control characters unlikely to appear in messages, such as several ASCII 01 (control-A) characters in succession. The MIME standard gets around the problem by including an explicit message length in the header, but this is a fragile solution which is very likely to break if messages are ever manually edited. For a somewhat better solution, see the record-jar style described later in this chapter.

For examples of RFC 822 format, look in your mailbox.

Cookie-Jar Format

Cookie-jar format is used by the fortune(1) program for its database of random quotes. It is appropriate for records that are just bags of unstructured text. It simply uses newline followed by %% (or sometimes newline followed by %) as a record separator. Example 5.3 is an example section from a file of email signature quotes:

Example 5.3. A fortune file example.

"Among the many misdeeds of British rule in India, history will look
upon the Act depriving a whole nation of arms as the blackest."
        -- Mohandas Gandhi, "An Autobiography", pg 446
The people of the various provinces are strictly forbidden to have 
in their possession any swords, short swords, bows, spears, firearms,
or other types of arms. The possession of unnecessary implements 
makes difficult the collection of taxes and dues and tends to foment 
        -- Toyotomi Hideyoshi, dictator of Japan, August 1588
"One of the ordinary modes, by which tyrants accomplish their 
purposes without resistance, is, by disarming the people, and making 
it an offense to keep arms."
        -- Supreme Court Justice Joseph Story, 1840

It is good practice to accept whitespace after % when looking for record delimiters. This helps cope with human editing mistakes. It's even better practice to use %%, and ignore all text from %% to end-of-line.

The cookie-jar separator was originally %%\n. I wanted something a bit more visible than % would have been. In fact, any stuff after the %% is treated as a comment (or at least that's how I wrote it).

– Ken Arnold

Simple cookie-jar format is appropriate for pieces of text that have no natural ordering, distinguishable structure above word level, or search keys other than their text context.

Record-Jar Format

Cookie-jar record separators combine well with the RFC 822 metaformat for records, yielding a format we'll call ‘record-jar’. If you need a textual format that will support multiple records with a variable repertoire of explicit fieldnames, one of the least surprising and human-friendliest ways to do it would look like Example 5.4.

Example 5.4. Basic data for three planets in a record-jar format.

Planet: Mercury
Orbital-Radius: 57,910,000 km
Diameter: 4,880 km
Mass: 3.30e23 kg
Planet: Venus
Orbital-Radius: 108,200,000 km
Diameter: 12,103.6 km
Mass: 4.869e24 kg
Planet: Earth
Orbital-Radius: 149,600,000 km
Diameter: 12,756.3 km
Mass: 5.972e24 kg
Moons: Luna

Of course, the record delimiter could be a blank line, but a line consisting of " %%\n " is more explicit and less likely to be introduced by accident during editing (two printable characters are better than one because it can't be generated by a single-character typo). In a format like this it is good practice to simply ignore blank lines.

If your records have an unstructured text part, your record-jar format is closely approaching a mailbox format. In this case, it's important that you have a well-defined way to escape the record delimiter so it can appear in text; otherwise, your record reader is going to choke on an ill-formed text part someday. Some technique analogous to byte-stuffing (described later in this chapter) is indicated.

Record-jar format is appropriate for sets of field-attribute associations that are like DSV files, but have a variable repertoire of fields, and possibly unstructured text associated with them.


XML is a very simple syntax resembling HTML — angle-bracketed tags and ampersand-led literal sequences. It is about as simple as a plain-text markup can be and yet express recursively nested data structures. XML is just a low-level syntax; it requires a document type definition (such as XHTML) and associated application logic to give it semantics.

XML is well suited for complex data formats (the sort of things for which the old-school Unix tradition would use an RFC-822-like stanza format) though overkill for simpler ones. It is especially appropriate for formats that have a complex nested or recursive structure of the sort that the RFC 822 metaformat does not handle well. For a good introduction to the format, see XML in a Nutshell[Harold-Means].

Among the hardest things to get right in designing any text file format are issues of quoting, whitespace and other low-level syntax details. Custom file formats often suffer from slightly broken syntax that doesn't quite match other similar formats. Using a standard format such as XML, which is verifiable and parsed by a standard library, eliminates most of these issues.

– Keith Packard

Example 5.5 is a simple example of an XML-based configuration file. It is part of the kdeprint tool shipped with the open-source KDE office suite hosted under Linux. It describes options for an image-to-PostScript filtering operation, and how to map them into arguments for a filter command. For another instructive example, see the discussion of Glade in Chapter 8.

Example 5.5. An XML example.

<?xml version="1.0"?>
<kprintfilter name="imagetops">
 data="imagetops %filterargs %filterinput %filteroutput" />
        <filterarg name="center" 
 description="Image centering" 
 format="-nocenter" type="bool" default="true">
            <value name="true" description="Yes" />
            <value name="false" description="No" />
        <filterarg name="turn" 
 description="Image rotation" 
 format="-%value" type="list" default="auto">
            <value name="auto" description="Automatic" />
            <value name="noturn" description="None" />
            <value name="turn" description="90 deg" />
        <filterarg name="scale" 
 description="Image scale" 
 format="-scale %value" 
 min="0.0" max="1.0" default="1.000" />
        <filterarg name="dpi" 
 description="Image resolution" 
 format="-dpi %value" 
 type="int" min="72" max="1200" default="300" />
        <filterarg name="file" format="%in" />
        <filterarg name="pipe" format="" />
        <filterarg name="file" format="> %out" />
        <filterarg name="pipe" format="" />

One advantage of XML is that it is often possible to detect ill-formed, corrupted, or incorrectly generated data through a syntax check, without knowing the semantics of the data.

The most serious problem with XML is that it doesn't play well with traditional Unix tools. Software that wants to read an XML format needs an XML parser; this means bulky, complicated programs. Also, XML is itself rather bulky; it can be difficult to see the data amidst all the markup.

One application area in which XML is clearly winning is in markup formats for document files (we'll have more to say about this in Chapter 18). Tagging in such documents tends to be relatively sparse among large blocks of plain text; thus, traditional Unix tools still work fairly well for simple text searches and transformations.

One interesting bridge between these worlds is PYX format — a line-oriented translation of XML that can be hacked with traditional line-oriented Unix text tools and then losslessly translated back to XML. A Web search for “Pyxie” will turn up resources. The xmltk toolkit takes the opposite tack, providing stream-oriented tools analogous to grep(1) and sort(1) for filtering XML documents; Web search for “xmltk” to find it.

XML can be a simplifying choice or a complicating one. There is a lot of hype surrounding it, but don't become a fashion victim by either adopting or rejecting it uncritically. Choose carefully and bear the KISS principle in mind.

Windows INI Format

Many Microsoft Windows programs use a textual data format that looks like Example 5.6. This example associates optional resources named account, directory, numeric_id, and developer with named projects python, sng, fetchmail, and py-howto. The DEFAULT entry supplies values that will be used when a named entry fails to supply them.

Example 5.6. A .INI file example.

account = esr

directory = /home/esr/cvs/python/
developer = 1

directory = /home/esr/WWW/sng/
numeric_id = 1012
developer = 1

numeric_id = 18364

account = eric
directory = /home/esr/cvs/py-howto/
developer = 1

This style of data-file format is not native to Unix, but some Linux programs (notably Samba, the suite of tools for accessing Windows file shares from Linux) support it under Windows's influence. This format is readable and not badly designed, but like XML it doesn't play well with grep(1) or conventional Unix scripting tools.

The .INI format is appropriate if your data naturally falls into its two-level organization of name-attribute pairs clustered under named records or sections. It's not good for data with a fully recursive treelike structure (XML is more appropriate for that), and it would be overkill for a simple list of name-value associations (use DSV format for that).

Unix Textual File Format Conventions

There are long-standing Unix traditions about how textual data formats ought to look. Most of these derive from one or more of the standard Unix metaformats we've just described. It is wise to follow these conventions unless you have strong and specific reasons to do otherwise.

In Chapter 10 we will discuss a different set of conventions used for program run-control files, but you should notice that it will share some of these same rules (especially about the lexical level, the rules by which characters are assembled into tokens).

The Pros and Cons of File Compression

Many modern Unix projects, such as and AbiWord, now use XML compressed with zip(1) or gzip(1) as a data file format. Compressed XML combines space economy with some of the advantages of a textual format — notably, it avoids the problem that binary formats must often allocate space for information that may not be used in particular cases (e.g., for unusual options or large ranges). But there is some dispute about this, dispute which turns on some of the central tradeoffs discussed in this chapter.

On the one hand, experiments have shown that documents in a compressed XML file are usually significantly smaller than the Microsoft Word's native file format, a binary format that one might imagine would take less space. The reason relates to a fundamental of the Unix philosophy: Do one thing well. Creating a single tool to do the compression job well is more effective than ad-hoc compression on parts of the file, because the tool can look across all the data and exploit all repetition in the information.

Also, by separating the representation design from the particular compression method used, you leave open the possibility of using different compression methods in the future with no more than minimal changes to the actual file parsing — perhaps, with no changes at all.

On the other hand, compression does some damage to transparency. While a human being can estimate from context whether uncompressing the file is likely to show him anything useful, tools such as file(1) cannot as of mid-2003 see through the wrapping.

Some would advocate a less structured compression format — straight gzip(1)-compressed XML data, say, without the internal structure and self-identifying header chunk provided by zip(1). While using a format similar to that of zip(1) solves the identification problem, it means that decoding such files will be tricky for programs written in the simpler scripting languages.

Any of these solutions (straight text, straight binary, or compressed text) may be optimal depending on the relative weight you give to storage economy, discoverability, or making browsing tools as simple as possible to write. The point of the preceding discussion is not to advocate any one of these approaches over the others, but rather to suggest how you can think about the options and design tradeoffs clearly.

This having been said, the truly Unixy solution would probably be to fix file(1) to see file prefixes through the compression — and, failing that, to write a shellscript wrapper around file(1) that would interpret compression as a direction to apply gunzip(1) and take a second look.

Application Protocol Design

In Chapter 7, we'll discuss the advantages of breaking complicated applications up into cooperating processes speaking an application-specific command set or protocol with each other. All the good reasons for data file formats to be textual apply to these application-specific protocols as well.

When your application protocol is textual and easily parsed by eyeball, many good things become easier. Transaction dumps become much easier to interpret. Test loads become easier to write.

Server processes are often invoked by harness programs such as inetd(8) in such a way that the server sees commands on standard input and ships responses to standard output. We describe this “CLI server” pattern in more detail in Chapter 11.

A CLI server with a command set that is designed for simplicity has the valuable property that a human tester will be able to type commands direct to the server process to probe the software's behavior.

Another issue to bear in mind is the end-to-end design principle. Every protocol designer should read the classic End-to-End Arguments in System Design [Saltzer]. There are often serious questions about which level of the protocol stack should handle features like security and authentication; this paper provides some good conceptual tools for thinking about them. Yet a third issue is designing application protocols for good performance. We'll cover that issue in more detail in Chapter 12.

The traditions of Internet application protocol design evolved separately from Unix before 1980.[54] But since the 1980s these traditions have become thoroughly naturalized into Unix practice.

We'll illustrate the Internet style by looking at three application protocols that are both among the most heavily used, and are widely regarded among Internet hackers as paradigmatic: SMTP, POP3, and IMAP. All three address different aspects of mail transport (one of the net's two most important applications, along with the World Wide Web), but the problems they address (passing messages, setting remote state, indicating error conditions) are generic to non-email application protocols as well and are normally addressed using similar techniques.

Case Study: SMTP, the Simple Mail Transfer Protocol

Example 5.7 is an example transaction in SMTP (Simple Mail Transfer Protocol), which is described by RFC 2821. In the example, C: lines are sent by a mail transport agent (MTA) sending mail, and S: lines are returned by the MTA receiving it. Text emphasized like this is comments, not part of the actual transaction.

Example 5.7. An SMTP session example.

C: <client connects to service port 25>
C: HELO sending host identifies self
S: 250 OK Hello snark, glad to meet you receiver acknowledges
C: MAIL FROM: <>         identify sending user
S: 250 <>... Sender ok receiver acknowledges
C: RCPT TO: identify target user
S: 250 root... Recipient ok receiver acknowledges
S: 354 Enter mail, end with "." on a line by itself
C: Scratch called. He wants to share
C: a room with us at Balticon.
C:.                                    end of multiline send
S: 250 WAA01865 Message accepted for delivery
C: QUIT sender signs off
S: 221 closing connection receiver disconnects
C: <client hangs up>

This is how mail is passed among Internet machines. Note the following features: command-argument format of the requests, responses consisting of a status code followed by an informational message, the fact that the payload of the DATA command is terminated by a line consisting of a single dot.

SMTP is one of the two or three oldest application protocols still in use on the Internet. It is simple, effective, and has withstood the test of time. The traits we have called out here are tropes that recur frequently in other Internet protocols. If there is any single archetype of what a well-designed Internet application protocol looks like, SMTP is it.

Case Study: POP3, the Post Office Protocol

Another one of the classic Internet protocols is POP3, the Post Office Protocol. It is also used for mail transport, but where SMTP is a ‘push’ protocol with transactions initiated by the mail sender, POP3 is a ‘pull’ protocol with transactions initiated by the mail receiver. Internet users with intermittent access (like dial-up connections) can let their mail pile up on a mail-drop machine, then use a POP3 connection to pull mail up the wire to their personal machines.

Example 5.8 is an example POP3 session. In the example, C: lines are sent by the client, and S: lines by the mail server. Observe the many similarities with SMTP. This protocol is also textual and line-oriented, sends payload message sections terminated by a line consisting of a single dot followed by line terminator, and even uses the same exit command, QUIT. Like SMTP, each client operation is acknowledged by a reply line that begins with a status code and includes an informational message meant for human eyes.

Example 5.8. A POP3 example session.

C: <client connects to service port 110>
S: +OK POP3 server ready <>
C: USER bob
S: +OK bob
C: PASS redqueen
S: +OK bob's maildrop has 2 messages (320 octets)
S: +OK 2 320
S: +OK 2 messages (320 octets)
S: 1 120
S: 2 200
S: +OK 120 octets
S: <the POP3 server sends the text of message 1>
S: +OK message 1 deleted
S: +OK 200 octets
S: <the POP3 server sends the text of message 2>
S: +OK message 2 deleted
S: +OK dewey POP3 server signing off (maildrop empty)
C: <client hangs up>

There are a few differences. The most obvious one is that POP3 uses status tokens rather than SMTP's 3-digit status codes. Of course the requests have different semantics. But the family resemblance (one we'll have more to say about when we discuss the generic Internet metaprotocol later in this chapter) is clear.

Case Study: IMAP, the Internet Message Access Protocol

To complete our triptych of Internet application protocol examples, we'll look at IMAP, another post office protocol designed in a slightly different style. See Example 5.9; as before, C: lines are sent by the client, and S: lines by the mail server. Text emphasized like this is comments, not part of the actual transaction.

Example 5.9. An IMAP session example.

C: <client connects to service port 143>
S: * OK IMAP4rev1 v12.264 server ready
C: A0001 USER "frobozz" "xyzzy"
S: * OK User frobozz authenticated
S: * FLAGS (\Answered \Flagged \Deleted \Draft \Seen)
S: * OK [UNSEEN 1] first unseen message in /var/spool/mail/esr
S: A0002 OK [READ-WRITE] SELECT completed
C: A0003 FETCH 1 RFC822.SIZE Get message sizes
S: * 1 FETCH (RFC822.SIZE 2545)
S: A0003 OK FETCH completed
C: A0004 FETCH 1 BODY[HEADER]                   Get first message header
S: * 1 FETCH (RFC822.HEADER {1425}
<server sends 1425 octets of message payload>
S: A0004 OK FETCH completed
C: A0005 FETCH 1 BODY[TEXT]                     Get first message body
S: * 1 FETCH (BODY[TEXT] {1120}
<server sends 1120 octets of message payload>
S: * 1 FETCH (FLAGS (\Recent \Seen))
S: A0005 OK FETCH completed
S: * BYE IMAP4rev1 server terminating connection
S: A0006 OK LOGOUT completed
C: <client hangs up>

IMAP delimits payloads in a slightly different way. Instead of ending the payload with a dot, the payload length is sent just before it. This increases the burden on the server a little bit (messages have to be composed ahead of time, they can't just be streamed up after the send initiation) but makes life easier for the client, which can tell in advance how much storage it will need to allocate to buffer the message for processing as a whole.

Also, notice that each response is tagged with a sequence label supplied by the request; in this example they have the form A000n, but the client could have generated any token into that slot. This feature makes it possible for IMAP commands to be streamed to the server without waiting for the responses; a state machine in the client can then simply interpret the responses and payloads as they come back. This technique cuts down on latency.

IMAP (which was designed to replace POP3) is an excellent example of a mature and powerful Internet application protocol design, one well worth study and emulation.


[54] One relic of this pre-Unix history is that Internet protocols normally use CR-LF as a line terminator rather than Unix's bare LF.

Application Protocol Metaformats

Just as data file metaformats have evolved to simplify serialization for storage, application protocol metaformats have evolved to simplify serialization for transactions across networks. The tradeoffs are a little different in this case; because network bandwidth is more expensive than storage, there is more of a premium on transaction economy. Still, the transparency and interoperability benefits of textual formats are sufficiently strong that most designers have resisted the temptation to optimize for performance at the cost of readability.

The Classical Internet Application Metaprotocol

Marshall Rose's RFC 3117, On the Design of Application Protocols,[55] provides an excellent overview of the design issues in Internet application protocols. It makes explicit several of the tropes in classical Internet application protocols that we observed in our examination of SMTP, POP, and IMAP, and provides an instructive taxonomy of such protocols. It is recommended reading.

The classical Internet metaprotocol is textual. It uses single-line requests and responses, except for payloads which may be multiline. Payloads are shipped either with a preceding length in octets or with a terminator that is the line ".\r\n". In the latter case the payload is byte-stuffed; all lines that start with a period get another period prepended, and the receiver side is responsible for both recognizing the termination and stripping away the stuffing. Response lines consist of a status code followed by a human-readable message.

One final advantage of this classical style is that it is readily extensible. The parsing and state-machine framework doesn't need to change much to accommodate new requests, and it is easy to code implementations so that they can parse unknown requests and return an error or simply ignore them. SMTP, POP3, and IMAP have all been extended in minor ways fairly often during their lifetimes, with minimal interoperability problems. Naïvely designed binary protocols are, by contrast, notoriously brittle.

HTTP as a Universal Application Protocol

Ever since the World Wide Web reached critical mass around 1993, application protocol designers have shown an increasing tendency to layer their special-purpose protocols on top of HTTP, using web servers as generic service platforms.

This is a viable option because, at the transaction layer, HTTP is very simple and general. An HTTP request is a message in an RFC-822/MIME-like format; typically, the headers contain identification and authentication information, and the first line is a method call on some resource specified by a Universal Resource Indicator (URI). The most important methods are GET (fetch the resource), PUT (modify the resource) and POST (ship data to a form or back-end process). The most important form of URI is a URL or Uniform Resource Locator, which identifies the resource by service type, host name, and a location on the host. An HTTP response is simply an RFC-822/MIME message and can contain arbitrary content to be interpreted by the client.

Web servers handle the transport and request-multiplexing layers of HTTP, as well as standard service types like http and ftp. It is relatively easy to write web server plugins that will handle custom service types, and to dispatch on other elements of the URI format.

Besides avoiding a lot of lower-level details, this method means the application protocol will tunnel through the standard HTTP service port and not need a TCP/IP service port of its own. This can be a distinct advantage; most firewalls leave port 80 open, but trying to punch another hole through can be fraught with both technical and political difficulties.

With this advantage comes a risk. It means that your web server and its plugins grow more complex, and cracks in any of that code can have large security implications. It may become more difficult to isolate and shut down problem services. The usual tradeoffs between security and convenience apply.

RFC 3205, On the Use of HTTP As a Substrate,[56] has good design advice for anyone considering using HTTP as the underlayer of an application protocol, including a summary of the tradeoffs and problems involved.

Case Study: The CDDB/ Database

Audio CDs consist of a sequence of music tracks in a digital format called CDDA-WAV. They were designed to be played by very simple consumer-electronics devices a few years before general-purpose computers developed enough raw speed and sound capability to decode them on the fly. Because of this, there is no provision in the format for even simple metainformation such as the album and track titles. But modern computer-hosted CD players want this information so the user can assemble and edit play lists.

Enter the Internet. There are (at least two) repositories that provide a mapping between a hash code computed from the track-length table on a CD and artist/album-title/track-title records. The original was, but another site called which is probably now more complete and widely used. Both sites rely on their users for the enormous task of keeping the database current as new CDs come out; arose from a developer revolt after CDDB elected to take all that user-contributed information proprietary .

Queries to these services could have been implemented as a custom application protocol on top of TCP/IP, but that would have required steps such as getting a new TCP/IP port number assigned and fighting to get a hole for it punched through thousands of firewalls. Instead, the service is implemented over HTTP as a simple CGI query (as if the CD's hash code had been supplied by a user filling in a Web form).

This choice makes all the existing infrastructure of HTTP and Web-access libraries in various programming languages available to support programs for querying and updating this database. As a result, adding such support to a software CD player is nearly trivial, and effectively every software CD player knows how to use them.

Case Study: Internet Printing Protocol

Internet Printing Protocol (IPP) is a successful, widely implemented standard for the control of network-accessible printers. Pointers to RFCs, implementations, and much other related material are available at the IETF's Printer Working Group site.

IPP uses HTTP 1.1 as a transport layer. All IPP requests are passed via an HTTP POST method call; responses are ordinary HTTP responses. (Section 4.2 of RFC 2568, Rationale for the Structure of the Model and Protocol for the Internet Printing Protocol, does an excellent job of explaining this choice; it repays study by anyone considering writing a new application protocol.)

From the software side, HTTP 1.1 is widely deployed. It already solves many of the transport-level problems that would otherwise distract protocol developers and implementers from concentrating on the domain semantics of printing. It is cleanly extensible, so there is room for IPP to grow. The CGI programming model for handling the POST requests is well understood and development tools are widely available.

Most network-aware printers already embed a web server, because that's the natural way to make the status of the printer remotely queryable by human beings. Thus, the incremental cost of adding IPP service to the printer firmware is not large. (This is an argument that could be applied to a remarkably wide range of other network-aware hardware, including vending machines and coffee makers[57] and hot tubs!)

About the only serious drawback of layering IPP over HTTP is that the protocol is completely driven by client requests. Thus there is no space in the model for printers to ship asynchronous alert messages back to clients. (However, smarter clients could run a trivial HTTP server to receive such alerts formatted as HTTP requests from the printer.)

BEEP: Blocks Extensible Exchange Protocol

BEEP (formerly BXXP) is a generic protocol machine that competes with HTTP for the role of universal underlayer for application protocols. There is a niche open because there is not as yet any other more established metaprotocol that is appropriate for truly peer-to-peer applications, as opposed to the client-server applications that HTTP handles well. A project website provides access to standards and open-source implementations in several languages.

BEEP has features to support both client-server and peer-to-peer modes. The authors designed the BEEP protocol and support library so that picking the right options abstracts away messy issues like data encoding, flow control, congestion-handling, support of end-to-end encryption, and assembling a large response composed of multiple transmissions,

Internally, BEEP peers exchange sequences of self-describing binary packets not unlike chunk types in PNG. The design is tuned more for economy and less for transparency than the classical Internet protocols or HTTP, and might be a better choice when data volumes are large. BEEP also avoids the HTTP problem that all requests have to be client-initiated; it would be better in situations in which a server needs to send asynchronous status messages back to the client.

BEEP is still new technology in mid-2003, and has only a few demonstration projects. But the BEEP papers are good analytical surveys of best practice in protocol design; even if BEEP itself fails to gain widespread adoption, the papers will retain considerable tutorial value.

XML-RPC, SOAP, and Jabber

There is a developing trend in application protocol design toward using XML within MIME to structure requests and payloads. BEEP peers use this format for channel negotiations. Three major protocols are going the XML route throughout: XML-RPC and SOAP (Simple Object Access Protocol) for remote procedure calls, and Jabber for instant messaging and presence. All three are XML document types.

XML-RPC is very much in the Unix spirit (its author observes that he learned how to program in the 1970s by reading the original source code for Unix). It's deliberately minimalist but nevertheless quite powerful, offering a way for the vast majority of RPC applications that can get by on passing around scalar boolean/integer/float/string datatypes to do their thing in a way that is lightweight and easy to understand and monitor. XML-RPC's type ontology is richer than that of a text stream, but still simple and portable enough to act as a valuable check on interface complexity. Open-source implementations are available. An excellent XML-RPC home page points to specifications and multiple open-source implementations.

SOAP is a more heavyweight RPC protocol with a richer type ontology that includes arrays and C-like structs. It was inspired by XML-RPC, but has been plausibly accused of being an overdesigned victim of the second-system effect. As of mid-2003 the SOAP standard is still a work in progress, but a trial implementation in Apache is tracking the drafts. Open-source client modules in Perl, Python, Tcl, and Java are readily discoverable by a Web search. The W3C draft specification is available on the Web.

XML-RPC and SOAP, considered as remote procedure call methods, have some associated risks that we discuss at the end of Chapter 7.

Jabber is a peer-to-peer protocol designed to support instant messaging and presence. What makes it interesting as an application protocol is that it supports passing around XML forms and live documents. Specifications, documentation, and open-source implementations are available at the Jabber Software Foundation site.


[55] See RFC 3117.

[56] See RFC 3205.

[57] See RFC 2324 and RFC 2325.

Chapter 6. Transparency

Beauty is more important in computing than anywhere else in technology because software is so complicated. Beauty is the ultimate defense against complexity.

– David Gelernter Machine Beauty: Elegance and the Heart of Technology (1998)

In the previous chapter we discussed the importance of textual data formats and application protocols, representations that are easy for human beings to examine and interact with. These promote qualities in design that are much valued in the Unix tradition but seldom if ever talked about explicitly: transparency and discoverability.

Software systems are transparent when they don't have murky corners or hidden depths. Transparency is a passive quality. A program is transparent when it is possible to form a simple mental model of its behavior that is actually predictive for all or most cases, because you can see through the machinery to what is actually going on.

Software systems are discoverable when they include features that are designed to help you build in your mind a correct mental model of what they do and how they work. So, for example, good documentation helps discoverability to a user. Good choice of variable and function names helps discoverability to a programmer. Discoverability is an active quality. To achieve it in your software you cannot merely fail to be obscure, you have to go out of your way to be helpful.[58]

Transparency and discoverability are important for both users and software developers. But they're important in different ways. Users like these properties in a UI because they mean an easier learning curve. UI transparency and discoverability are a large part of what people mean when they say a UI is ‘intuitive’; most of the rest is the Rule of Least Surprise. We'll examine the properties that make user interfaces pleasant and effective in more depth in Chapter 11.

Software developers like these qualities in the code itself (the part users don't see) because they so often need to understand it well enough to modify and debug it. Also, a program designed so that its internal data flows are readily comprehensible is more likely to be one that does not fail because of bad interactions that the designer didn't notice, and more likely to be able to evolve forward gracefully (including accommodating change when new maintainers pick up the baton).

Transparency is a major component of what David Gelernter refers to as “beauty” in this chapter's epigraph. Unix programmers, borrowing from mathematicians, often use the more specific term “elegance” for the quality Gelernter speaks of. Elegance is a combination of power and simplicity. Elegant code does much with little. Elegant code is not only correct but visibly, transparently correct. It does not merely communicate an algorithm to a computer, but also conveys insight and assurance to the mind of a human that reads it. By seeking elegance in our code, we build better code. Learning to write transparent code is a first, long step toward learning how to write elegant code — and taking care to make code discoverable helps us learn how to make it transparent. Elegant code is both transparent and discoverable.

It may be easier to appreciate the difference between transparency and discoverability with a pair of extreme examples. The Linux kernel source is remarkably transparent (given the intrinsic complexity of what it does) but not at all discoverable — acquiring the minimum knowledge needed to live in the code and understand the idiom of the developers is difficult, but once you do the whole makes sense.[59] On the other hand, the Emacs Lisp libraries are discoverable but not transparent. It's easy to acquire enough knowledge to tweak just one thing, but quite difficult to comprehend the whole system.

In this chapter, we'll examine features of Unix designs that promote transparency and discoverability not just in UIs but in the parts users don't normally see. We'll develop some useful rules you can apply to your coding and development practice. Later on, in Chapter 19 we'll see how good release-engineering practices (like having a README file with appropriate content) can make your source code as discoverable as your design.

If you need a practical reminder why these qualities are important, remember that the sanity you save by writing transparent, discoverable systems may well be that of your own future self.


[58] An economically-minded friend comments: “Discoverability is about reducing barriers to entry; transparency is about reducing the cost of living in the code”.

[59] The Linux kernel makes a number of attempts at discoverability, including the Documentation subdirectory in the Linux kernel source tarball and quite a number of tutorial websites and books. These attempts are frustrated by the speed at which the kernel changes; the documentation has a chronic tendency to fall behind.

Studying Cases

Normal practice in this book has been to intersperse case studies with philosophy. But in this chapter we'll begin by looking at several Unix designs that exhibit transparency and discoverability, and attempt to draw lessons from them only after all have been presented. Each major point of the analysis in the latter half of this chapter draws on several of these, and the arrangement avoids forward references to case studies the reader hasn't seen yet.

Case Study: audacity

First, we'll look at an example of transparency in UI design. It is audacity, an open-source editor for sound files that runs on Unix systems, Mac OS X, and Windows. Sources, downloadable binaries, documentation, and screen shots are available at the project site.

This program supports cutting, pasting, and editing of audio samples. It supports multitrack editing and mixing. The UI is superbly simple; the sound waveforms are shown in the audacity window. The image of the waveform can be cut and pasted; operations on that image are directly reflected in the audio sample as soon as they are performed.

Figure 6.1. Screen shot of audacity.

Screen shot of audacity.

Multitrack editing is supported in the simplest possible way; the screen splits into multiple per-track displays in a spatial relationship that conveys their concurrency and makes it easy to match features by inspection. Tracks can be dragged right or left with the mouse to change their relative timing.

Several features of this UI are subtly excellent and worthy of emulation: the large, easily visible and clickable operation buttons with distinguishing colors, the presence of an undo command that removes most of the risk from experimentation, the volume slider that makes softness/loudness visually obvious in its shape.

But these are details. The central virtue of this program is that it has a superbly transparent and natural user interface, one that erects as few barriers between the user and the sound file as possible.

Case Study: fetchmail's -v option

fetchmail is a network gateway program. Its main purpose is to translate between POP3 or IMAP remote-mail protocols and the Internet's native SMTP protocol for email exchange. It is in extremely widespread use on Unix machines that use intermittent SLIP or PPP connections to Internet service providers, and as such probably touches an appreciable fraction of the Internet's mail traffic.

fetchmail has no fewer than 60 command-line options (which, as we'll establish later in this book, is probably too many), and a number of other options that are settable from the run-control file but not from the command line. Of all these, the most important — by far — is -v, the verbose option.

When -v is on, fetchmail dumps each one of its POP, IMAP, and SMTP transactions to standard output as they happen. A developer can actually see the code doing protocol with remote mailservers and the mail transport program it forwards to, in real time. Users can send session transcripts with their bug reports. Example 6.1 shows a representative session transcript.

Example 6.1. An example fetchmail -v transcript.

fetchmail: 6.1.0 querying (protocol IMAP) 
 at Mon, 09 Dec 2002 08:41:37 -0500 (EST): poll started
fetchmail: running ssh %h /usr/sbin/imapd 
              (host service imap)
fetchmail: IMAP< * PREAUTH [] IMAP4rev1 v12.264 server ready
fetchmail: IMAP> A0001 CAPABILITY
fetchmail: IMAP< A0001 OK CAPABILITY completed
fetchmail: IMAP> A0002 SELECT "INBOX"
fetchmail: IMAP< * 2 EXISTS
fetchmail: IMAP< * 1 RECENT
fetchmail: IMAP< * OK [UIDVALIDITY 1039260713] UID validity status
fetchmail: IMAP< * OK [UIDNEXT 23982] Predicted next UID
fetchmail: IMAP< * FLAGS (\Answered \Flagged \Deleted \Draft \Seen)
               (\* \Answered \Flagged \Deleted \Draft \Seen)] 
 Permanent flags
fetchmail: IMAP< * OK [UNSEEN 2] first unseen in /var/spool/mail/esr
fetchmail: IMAP< A0002 OK [READ-WRITE] SELECT completed
fetchmail: IMAP> A0003 EXPUNGE
fetchmail: IMAP< A0003 OK Mailbox checkpointed, no messages expunged
fetchmail: IMAP> A0004 SEARCH UNSEEN
fetchmail: IMAP< * SEARCH 2
fetchmail: IMAP< A0004 OK SEARCH completed
2 messages (1 seen) for esr at
fetchmail: IMAP> A0005 FETCH 1:2 RFC822.SIZE
fetchmail: IMAP< * 1 FETCH (RFC822.SIZE 2545)
fetchmail: IMAP< * 2 FETCH (RFC822.SIZE 8328)
fetchmail: IMAP< A0005 OK FETCH completed
skipping message (2545 octets) not flushed
fetchmail: IMAP> A0006 FETCH 2 RFC822.HEADER
fetchmail: IMAP< * 2 FETCH (RFC822.HEADER {1586}
reading message of 2 (1586 header octets)
fetchmail: SMTP< 220 ESMTP Sendmail 8.12.5/8.12.5; 
 Mon, 9 Dec
2002 08:41:41 -0500
fetchmail: SMTP> EHLO localhost
fetchmail: SMTP< 
 Hello localhost [], pleased to meet you
fetchmail: SMTP< 250-8BITMIME
fetchmail: SMTP< 250-SIZE
fetchmail: SMTP> MAIL FROM:<> SIZE=8328
fetchmail: SMTP< 250 2.1.0 <>... Sender ok
fetchmail: SMTP> RCPT TO:<esr@localhost>
fetchmail: SMTP< 250 2.1.5 <esr@localhost>... Recipient ok
fetchmail: SMTP> DATA
fetchmail: SMTP< 354 Enter mail, end with "." on a line by itself
fetchmail: IMAP< )
fetchmail: IMAP< A0006 OK FETCH completed
fetchmail: IMAP> A0007 FETCH 2 BODY.PEEK[TEXT]
fetchmail: IMAP< * 2 FETCH (BODY[TEXT] {6742}
 (6742 body octets) *********************.**************************.
fetchmail: IMAP< )
fetchmail: IMAP< A0007 OK FETCH completed
fetchmail: SMTP>. (EOM)
fetchmail: SMTP< 250 2.0.0 gB9ffWo08245 Message accepted for delivery
fetchmail: IMAP> A0008 STORE 2 +FLAGS (\Seen \Deleted)
fetchmail: IMAP< * 2 FETCH (FLAGS (\Recent \Seen \Deleted))
fetchmail: IMAP< A0008 OK STORE completed
fetchmail: IMAP> A0009 EXPUNGE
fetchmail: IMAP< * 2 EXPUNGE
fetchmail: IMAP< * 1 EXISTS
fetchmail: IMAP< * 0 RECENT
fetchmail: IMAP< A0009 OK Expunged 1 messages
fetchmail: IMAP> A0010 LOGOUT
fetchmail: IMAP< * BYE hurkle IMAP4rev1 server terminating connection
fetchmail: IMAP< A0010 OK LOGOUT completed
fetchmail: 6.1.0 querying (protocol IMAP) 
 at Mon, 09 Dec 2002 08:41:42 -0500: poll completed
fetchmail: SMTP> QUIT
fetchmail: SMTP< 221 2.0.0 closing connection
fetchmail: normal termination, status 0

The -v option makes what fetchmail is doing discoverable (by letting you see the protocol exchanges). This is immensely useful. I considered it so important that I wrote special code to mask account passwords out of -v transaction dumps so that they could be passed around and posted without anyone having to remember to edit sensitive information out of them.

This turned out to be a good call. At least eight out of ten problems reported get diagnosed within seconds of a knowledgeable person's eyes seeing a session transcript. There are several knowledgeable people on the fetchmail mailing list — in fact, because most bugs are easy to diagnose, I seldom have to handle them myself.

Over the years, fetchmail has acquired a reputation as a rather bulletproof program. It can be misconfigured, but it very seldom outright breaks. Betting that this has nothing to do with the fact that the exact circumstances of eight out of ten bugs are rapidly discoverable would not be smart.

We can learn from this example. The lesson is this: Don't let your debugging tools be mere afterthoughts or treat them as throwaways. They are your windows into the code; don't just knock crude holes in the walls, finish and glaze them. If you plan to keep the code maintained, you're always going to need to let light into it.

Case Study: GCC

GCC, the GNU C compiler used on most modern Unixes, is perhaps an even better example of engineering for transparency. GCC is organized as a sequence of processing stages knit together by a driver program. The stages are: preprocessor, parser, code generator, assembler, and linker.

Each of the first three stages takes in a readable textual format and emits a readable textual format (the assembler has to emit and the linker to accept binary formats, pretty much by definition). With various command-line options of the gcc(1) driver, you can see not just the results after C preprocessing, after assembly generation, and after object code generation — but you can also monitor the results of many intermediate steps in parsing and code generation.

This is exactly the structure of cc, the first (PDP-11) C compiler.

– Ken Thompson

There are many benefits of this organization. One that is particularly important for GCC is regression testing.[60] Because most of the various intermediate formats are textual, deviations from expected results in a regression test are easily spotted and analyzed using simple textual diff operations on the intermediate results; there is no need for specialist dump-analysis tools that may well harbor their own bugs, and in any case would represent an additional maintenance burden.

The design pattern to extract from this example is that the driver program has monitoring switches that merely (but sufficiently) expose the textual data flows among the components. As with fetchmail's -v option, these options are not afterthoughts; they are designed in for discoverability.

Case Study: kmail

kmail is the GUI mailreader distributed with the KDE environment. The kmail UI is tastefully and well designed, with many good features including automatic display of enclosed images in a MIME multipart and support for PGP key encryption/decryption. It is friendly to end-users — my beloved but nontechie wife uses and enjoys it.

Many mail user agents make one gesture in the direction of discoverability by having a command that toggles display of all the mail headers, as opposed to a select few like From and Subject. The UI of kmail takes this a long step further.

A running kmail displays status notifications in a one-line subwindow at the bottom of its window, in small type over a steel-gray background clearly modeled on the Netscape/Mozilla status bar. When you open a mailbox, for example, the status bar displays counts of total and unread messages. The visual presentation is unobtrusive; it is easy to ignore the notifications, but also easy to focus on them if you want to.

Figure 6.2. Screen shot of kmail.

Screen shot of kmail.

The kmail GUI is good user-interface design. It's informative, but not distracting; it gets around the reason we adduce in Chapter 11 that the best policy for Unix tools operating normally is usually silence. The authors showed excellent taste in borrowing the look and feel of the browser status bar.

But the extent of the kmail developers' tastefulness will not become clear until you have to troubleshoot an installation that is having trouble sending mail. If you watch closely during the send, you will observe that each line of the SMTP transaction with the remote mail transport is echoed into the kmail status bar as it happens.

The kmail developers neatly avoid a trap that often makes GUI programs like kmail a terrible pain in a troubleshooter's fundament. Most design teams with kmail's objectives would have suppressed those messages entirely, fearing that they would give Aunt Tillie a touch of the vapors that would drive her back to the meretricious pseudo-simplicity of a Windows box.

Instead, they designed for transparency — they made the transaction messages show, but also made them visually easy to ignore. By getting the presentation right, they managed to please both Aunt Tillie and her geeky nephew Melvin who fixes her computer problems. This was brilliant; it's a technique other GUI interfaces could and should emulate.

Ultimately, of course, the visibility of those messages is good for Aunt Tillie, because they mean Melvin is far less likely to throw up his hands in frustration while trying to solve her email problems.

The lesson here is clear. Dumbing down your UI is only the half-smart thing to do. The really smart thing is to find a way to leave the details accessible, but make them unobtrusive.

Case Study: SNG

The program sng translates between PNG format and an all-text representation of it (SNG or Scriptable Network Graphics format) that can be examined and modified with an ordinary text editor. Run on a PNG file, it produces an SNG file; run on an SNG file, it recovers the equivalent PNG. The transformation is 100% faithful and lossless in both directions.

In syntactic style, SNG resembles CSS (Cascading Style Sheets), another language for controlling presentation of graphics; this makes at least a gesture in the direction of the Rule of Least Surprise. Here is a test example:

Example 6.2. An SNG Example.

#SNG: This is a synthetic SNG test file

# Our first test is a paletted (type 3) image.
 width: 16;
 height: 19;
 bitdepth: 8;
 using color: palette; 
 with interlace;

# Sample bit depth chunk
sBIT: {
 red: 8;
 green: 8;
 blue: 8;

# An example palette: three colors, one of which 
# we will render transparent 
   (0, 0, 255)
   (255, 0, 0)
   "dark slate gray",
# Suggested palette
sPLT {
 name: "A random suggested palette";
 depth: 8;
   (0, 0, 255), 255, 7;
   (255, 0, 0), 255, 5;
   ( 70, 70, 70), 255, 3;

# The viewer will actually use this...
 pixels base64

tEXt: {                                # Ordinary text chunk
 keyword: "Title";
 text: "Sample SNG script";

# Test file ends here

The point of this tool is to enable users to edit various obscure PNG chunk types that are not necessarily supported by conventional graphics editors. Rather than writing special-purpose code to grovel through the PNG binary format, the user can simply flip an image into an all-text representation, edit that, and massage it back. Another potential application is in making images amenable to version control; under most version-control systems, text files are much easier to manage than binary blobs, and diff operations on SNG representations actually have some possibility of yielding useful information.

The gains here go beyond the time not spent writing special-purpose code for manipulating binary PNGs, however. The code of the sng program itself is not especially transparent, but it promotes transparency in larger systems of programs by making the entire contents of PNGs discoverable.

Case Study: The Terminfo Database

The terminfo database is a collection of descriptions of video-display terminals. Each entry describes the escape sequences that perform various manipulations on the terminal screen, such as inserting or deleting lines, erasing from the cursor position to end of line or screen, or beginning and ending screen highlights such as reverse video, underline, or blink.

The terminfo database is primarily used by the curses(3) libraries. These underlie the “roguelike” interface style we discuss in Chapter 11, and some very widely used programs such as mutt(1), lynx(1), and slrn(1). Though the terminal emulators such as xterm(1) that run on today's bitmapped displays all have capabilities that are minor variations on those of the ANSI X3.64 standard and the venerable VT100 terminal, there is still enough variation that hardwiring ANSI capabilities into applications would be a bad idea. Terminfo is also worth studying because problems that are logically similar to the one it addressed arise constantly in managing other kinds of peripheral hardware that doesn't have a standard way to report their own capabilities.

The design of terminfo benefits from experience with an earlier capability format called termcap. The database of termcap descriptions lived in a textual format in one big file, /etc/termcap; though this format is now obsolete, your Unix system almost certainly includes a copy.

Normally, the key used to look up your terminal type entry is the environment variable TERM, which for purposes of this case study is set by magic.[61] Applications that use terminfo (or termcap) pay a small penalty in startup lag; when the curses(3) library initializes itself, it has to look up the entry corresponding to TERM and load the entry into memory.

Experience with termcap showed that the startup penalty was dominated by the time required to parse the textual representation of capabilities. Accordingly, terminfo entries are binary structure dumps that can be marshaled and unmarshaled more quickly. There is a master textual format for the entire database, the terminfo capability file. That file (or individual entries) can be compiled to binary form with the terminfo compiler tic(1); binary entries can be decompiled to the editable text format by infocmp(1).

The design superficially contradicts the advice we gave in Chapter 5 against binary caches, but this is actually the extreme case in which that's a good tactic. Edits to the text masters are very rare — in fact, Unixes normally ship with the terminfo database precompiled and the text master serving primarily as documentation. Thus, the synchronization and inconsistency problems that would normally militate against this approach almost never arise.

The designers of terminfo could have optimized for speed in a second way. The entire database of binary entries could have been put in some kind of big opaque database file. What they actually did instead was more clever and more in the Unix spirit. Terminfo entries live in a directory hierarchy, usually on modern Unixes under /usr/share/terminfo. Consult the terminfo(5) man page to find the location on your system.

If you look in the terminfo directory, you'll see subdirectories named by single printable characters. Under each of these are the entries for each terminal type that has a name beginning with that letter. The goal of this organization was to avoid having to do a linear search of a very large directory; under more modern Unix file systems, which represent directories with B-trees or other structures optimized for fast lookup, the subdirectories won't be necessary.

I found that even on a fairly modern Unix, splitting a big directory up into subdirectories can improve performance substantially. It was tens of thousands of files, an authorized-user database for a big educational institution, on a late-model DEC Alpha running DEC's Unix. (Subdirectories named by first and last letter of name — e.g., "johnson" would be in directory "j_n" — worked best of the schemes we tested. Using the first two letters wasn't nearly as good, because there were a lot of systematically-generated names which differed only toward the end.) This may just say that sophisticated directory indexing is still not as common as it should be... but even so, that makes an organization which works well without it more portable than one which requires it.

– Henry Spencer

Thus, the cost of opening a terminfo entry is two file system lookups and a file open. But since mining the same entry from one big database would have required a lookup and open for the database, the incremental cost for terminfo's organization is at most one file system lookup. Actually, it's less than that; it's the cost difference between one file system lookup and whatever retrieval method the one big database would have used. This is probably marginal, and quite tolerable once per application at startup time.

Terminfo uses the file system itself as a simple hierarchical database. This is a superb bit of constructive laziness, obeying the Rule of Economy and the Rule of Transparency. It means that all the ordinary tools for navigating, examining and modifying the file system can be used to navigate, examine, and modify the terminfo database; no special ones (other than tic(1) and infocmp(1) for packing and unpacking the individual records) need to be written and debugged. It also means that work on speeding up database access would be work on speeding up the file system itself, tuning that would benefit many more applications than just users of curses(3).

There is one additional advantage of this organization that doesn't come up in the terminfo case; you get to use Unix's permissions mechanism rather than having to invent your own access-control layer with its own bugs. This falls out as a consequence of adopting the “everything is a file” philosophy of Unix rather than trying to fight it.

The terminfo directory layout is rather space-inefficient on most Unix file systems. The entries are usually between 400 and 1400 bytes long, but file systems normally allocate a minimum of 4K for every nonempty disk file. The designers accepted this cost for the same reason they chose a packed binary format, to cut the startup latency of terminfo-using programs to a minimum. Disk capacity for constant price has exploded over a thousandfold since, tending to vindicate that decision.

The contrast with the formats used by the Microsoft Windows registry files is instructive. Registries are property databases used by both Windows itself and applications. Each registry lives in one big file. Registries contain a mix of text and binary data that requires specialized editing tools. The one-big-file approach leads, among other things, to the notorious ‘registry creep’ phenomenon; average access time rises without bound as new entries are added. Because there is no standard API for editing the registry provided by the system, applications use ad-hoc code to edit it themselves, making it notoriously subject to corruption that can lock up the entire system.

Using the Unix file system as a database is a tactic other applications with simple database requirements might do well to emulate. Good reasons not to do it are more likely to have to do with the database keys not naturally looking like filenames than they are with any performance problems. In any case, it's the sort of good fast hack that can be very useful in prototyping.

Case Study: Freeciv Data Files

Freeciv is an open-source strategy game inspired by Sid Meier's classic Civilization II. In it, each player begins with a wandering band of neolithic nomads and builds a civilization. Player civilizations may explore and colonize the world, fight wars, engage in trade, and research technological advances. Some players may actually be artificial intelligences; solitaire play against these can be challenging. One wins either by conquering the world or by being the first player to reach a technology level sufficient to get a starship to Alpha Centauri. Sources and documentation are available at the project site.

Figure 6.3. Main window of a Freeciv game.

Main window of a Freeciv game.

In Chapter 7 we'll exhibit the Freeciv strategy game as an example of client-server partitioning, with the server maintaining shared state and the client concentrating on GUI presentation. But this game has another notable architectural feature; much of the game's fixed data, rather than being wired into the server code, is expressed in a property registry read in by the game server at startup time.

The game's registry files are written in a textual data-file format that assembles text strings (with associated text and numeric properties) into various internal lists of important data (such as nations and unit types) in the game server. The minilanguage has an include directive, so game data can be broken up into semantic units (different files) that are each separately editable. This design choice has been carried through to such an extent that it's possible to define new nations and new unit types simply by creating new declarations in the data files, without touching the server code at all.

The Freeciv server's startup parsing has an interesting feature that creates something of a conflict between two of Unix's design rules, and is therefore worth closer examination. The server ignores property names it doesn't know how to use. This makes it possible to declare properties that the server doesn't yet use without breaking the startup parsing. It means that development of the game data (policy) and the server engine (mechanism) can be cleanly separated. On the other hand, it also means startup parsing won't catch simple misspellings of attribute names. This quiet failure seems to violate the Rule of Repair.

To resolve this conflict, notice that it's the server's job to use the registry data, but the task of carefully error-checking that data could be handed off to another program to be run by human editors each time the registry is modified. One Unix solution would be a separate auditing program that analyzes either a machine-readable specification of the ruleset format or the source of the server code to determine the set of properties it uses, parses the Freeciv registry to determine the set of properties it provides, and prepares a difference report.[62]

The aggregate of all Freeciv data files is functionally similar to a Windows registry, and even uses a syntax resembling the textual portions of registries. But the creep and corruption problems we noted with the Windows registry don't crop up here because no program (either within or outside the Freeciv suite) writes to these files. It's a read-only registry edited only by the game's maintainers.

The performance impact of data-file parsing is minimized because for each file the operation is performed only once, at either client or server startup time.


[60] Regression testing is a method for detecting bugs introduced as software is modified. It consists of periodically checking the output of the changing software for some fixed test input against a snapshot of output captured at an earlier stage of the process and known (or assumed) to be correct.

[61] Actually, TERM is set by the system at login time. For actual terminals on serial lines, the mapping from tty lines to TERM values is set from a system configuration file at boot time; the details vary among Unixes. Terminal emulators like xterm(1) set this variable themselves.

[62] The ur-ancestor of such validator programs under Unix was lint, a validator for C code separate from the C compiler. Though GCC has absorbed its functions, old Unix hands are still apt to refer to the process of running a validator as ‘linting’, and the name survives in utilities such as xmllint.

Designing for Transparency and Discoverability

To design for transparency and discoverability, you need to apply every tactic for keeping your code simple, and also concentrate on the ways in which your code is a communication to other human beings. The first questions to ask, after “Will this design work?” are “Will it be readable to other people? Is it elegant?” We hope it is clear by now that these questions are not fluff and that elegance is not a luxury. These qualities in the human reaction to software are essential for reducing its bugginess and increasing its long-term maintainability.

The Zen of Transparency

One pattern that emerges from the examples we've examined so far in this chapter is this: If you want transparent code, the most effective route is simply not to layer too much abstraction over what you are manipulating with the code.

In Chapter 4 's section on the value of detachment, our advice was to abstract and simplify and generalize, to try and detach from the particular, accidental conditions under which a design problem was posed. The advice to abstract does not actually contradict the advice against excessive abstractions we're developing here, because there is a difference between getting free of assumptions and forgetting the problem you're trying to solve. This is part of what we were driving at when we developed the idea that glue layers need to be kept thin.

One of the main lessons of Zen is that we ordinarily see the world through a haze of preconceptions and fixed ideas that proceed from our desires. To achieve enlightenment, we must follow the Zen teaching not merely to let go of desire and attachment, but to experience reality exactly as it is — without the preconceptions and the fixed ideas getting in the way.

This is excellent pragmatic advice for software designers. It's part of what's implicit in the classic Unix advice to be minimalist. Software designers are clever people who form ideas (abstractions) about the application domains they deal with. They organize the software they write around those ideas. Then, when debugging, they often find they have great trouble seeing through those ideas to what is actually going on.

Any Zen master would recognize this problem instantly, yell “Three pounds of flax!”, and probably clout the student a good one.[63] Consciously designing for transparency is a slightly less mystical way of addressing it.

In Chapter 4 we criticized object-oriented programming in terms likely to prove a bit shocking to programmers who were raised on the 1990s gospel of OO. Object-oriented design doesn't have to be over-complicated design, but we've observed that too often it is. Too many OO designs are spaghetti-like tangles of is-a and has-a relationships, or feature thick layers of glue in which many of the objects seem to exist simply to hold places in a steep-sided pyramid of abstractions. Such designs are the opposite of transparent; they are (notoriously) opaque and difficult to debug.

As we've previously noted, Unix programmers are the original zealots about modularity, but tend to go about it in a quieter way. Keeping glue layers thin is part of it; more generally, our tradition teaches us to build lower, hugging the ground with algorithms and structures that are designed to be simple and transparent.

As with Zen art, the simplicity of good Unix code depends on exacting self-discipline and a high level of craft, neither of which are necessarily apparent on casual inspection. Transparency is hard work, but worth the effort for more than merely artistic reasons. Unlike Zen art, software requires debugging — and usually needs continuing maintenance, forward-porting, and adaptation throughout its lifetime. Transparency is therefore more than an esthetic triumph; it is a victory that will be reflected in lower costs throughout the software's life cycle.

Coding for Transparency and Discoverability

Transparency and discoverability, like modularity, are primarily properties of designs, not code. It is not sufficient to get right the low-level elements of style, such as indenting code in a clear and consistent way or having good variable-naming conventions. These qualities have much more to do with code properties that are less obvious to inspection. Here are a few to think about:

It's best for code to be simple. But if it answers these sorts of questions well, it can be very complex without putting an impossible cognitive burden on a human maintainer.

The reader might find it instructive to compare these with our checklist questions about modularity in Chapter 4.

Transparency and Avoiding Overprotectiveness

Close kin to the programmer tendency to build overelaborate castles of abstractions is a tendency to overprotect others from the low-level details. While it's not bad practice to hide those details in the program's normal mode of operation ( fetchmail 's -v switch is off by default), they should be discoverable. There's an important difference between hiding them and making them inaccessible.

Programs that cannot reveal what they are doing make troubleshooting far more difficult. Thus, experienced Unix users actually take the presence of debugging and instrumentation switches as a good sign, and their absence as possibly a bad one. Absence suggests an inexperienced or careless developer; presence suggests one with enough wisdom to follow the Rule of Transparency.

The temptation to overprotect is especially strong in GUI applications targeted for end users, like mail readers. One reason Unix developers have been cool toward GUI interfaces is that, in their designers' haste to make them ‘user-friendly’ each one often becomes frustratingly opaque to anyone who has to solve user problems — or, indeed, interact with it anywhere outside the narrow range predicted by the user-interface designer.

Worse, programs that are opaque about what they are doing tend to have a lot of assumptions baked into them, and to be frustrating or brittle or both in any use case not anticipated by the designer. Tools that look glossy but shatter under stress are not good long-term value.

Unix tradition pushes for programs that are flexible for a broader range of uses and troubleshooting situations, including the ability to present as much state and activity information to the user as the user indicates he is willing to handle. This is good for troubleshooting; it is also good for growing smarter, more self-reliant users.

Transparency and Editable Representations

Another theme that emerges from these examples is the value of programs that flip a problem out of a domain in which transparency is hard into one in which it is easy. Audacity, sng(1) and the tic(1)/infocmp(1) pair all have this property. The objects they manipulate are not readily conformable to the hand and eye; audio files are not visual objects, and although images expressed in PNG format are visual, the complexities of PNG annotation chunks are not. All three applications turn manipulation of their binary file formats into a problem to which human beings can more readily apply intuition and competences gained from everyday experience.

A rule all these examples follow is that they degrade the representation as little as possible — in fact, they translate it reversibly and losslessly. This property is very important, and worth implementing even if there is no obvious application demand for that kind of 100% fidelity. It gives potential users confidence that they can experiment without degrading their data.

All the advantages of textual data-file formats that we discussed in Chapter 5 also apply to the textual formats that sng(1), infocmp(1) and their kin generate. One important application for sng(1) is robotic generation of PNG image annotations by scripts — because sng(1) exists, such scripts are easier to write.

Whenever you face a design problem that involves editing some kind of complex binary object, the Unix tradition encourages asking first off whether you can write a tool analogous to sng(1) or the tic(1)/infocmp(1) pair that can do a lossless mapping to an editable textual format and back. There is no established term for programs of this kind, but we'll call them textualizers.

If the binary object is dynamically generated or very large, then it may not be practical or possible to capture all the state with a textualizer. In that case, the equivalent task is to write a browser. The paradigm example is fsdb(1), the file-system debugger supported under various Unixes; there is a Linux equivalent called debugfs(1). The psql(1) used to browse PostgreSQL databases, and the smbclient(1) program that can be used to query Windows file shares on a SAMBA-equipped Linux machine, are two more. All five are simple CLI programs that could be driven by scripts and test harnesses.

Writing a textualizer or browser is a valuable exercise for at least four reasons:

After you've done this, you may well discover that it's possible to apply the “separated engine and interface” pattern (see Chapter 11) using your textualizer/debugger as the engine. All the usual benefits of this pattern will apply.

It is desirable, although often difficult, for a textualizer to be able to read and write even a damaged binary object. For one thing, it lets you generate damaged test cases to stress-test software; for another, it can make emergency repairs a whole lot easier. It may be hard to handle cases in which the structure of the object is messed up, but at least you should handle cases in which the content of the structure is nonsense, e.g., by showing nonsense values in hex and converting the hex back to the values.

– Henry Spencer

Transparency, Fault Diagnosis, and Fault Recovery

Yet another benefit of transparency, related to ease of debugging, is that transparent systems are easier to perform recovery actions on after a bug bites — and, often, more resistant to damage from bugs in the first place.

In comparing the terminfo database with Windows registries we noted that registries are notoriously subject to being corrupted by buggy application code. This can make the entire system unusable. Even if it doesn't, recovery can be difficult if the corruption confuses the specialized registry-editing tools.

Our Unix case studies illustrate ways that designing for transparency can prevent this class of problem. Because the terminfo database is not one big file, botching one terminfo entry does not make the whole terminfo data set unusable. Fully textual one-big-file formats like termcap are usually parsed with methods which (unlike block reads of binary structure dumps) can recover from single-point errors. Syntax errors in an SNG file can be corrected by hand without requiring specialized editors that might refuse to load a damaged PNG image.

Going back to the kmail case study, that program makes fault diagnosis easier because it obeys the Rule of Repair: SMTP failures are noisy, usefully so. You don't have to decode a layer of obfuscatory messages generated by kmail itself to see what the interaction with the SMTP server looks like. All you have to do is look in the right place, because kmail is being transparent and not throwing away information about the error state. (It helps that SMTP itself is textual and includes human-readable status messages in its transactions.)

Discoverability tools like textualizers and browsers also make fault diagnosis easier. We've already touched on one reason: they make inspecting the state of the system easier. But there is another effect at work as well; textualized versions of data tend to have useful redundancies (such as using whitespace for visual separation as well as explicit delimiters for parsing). These are present to make them easier to read for humans, but also have the effect of making them more resistant to being irreparably trashed by point failures. A corrupted chunk in a PNG file is seldom recoverable, but the human capacity for pattern recognition and reasoning from context might be able to repair the equivalent SNG form.

Over and over again, the Rule of Robustness is clear. Simplicity plus transparency lowers costs, reduces everybody's stress, and frees people to concentrate on new problems rather than cleaning up after old mistakes.


[63] See the koan called Tozan's Three Pounds in the Gateless Gate [Mumon].

[64] An invariant is a property of a software design that is preserved by every operation in it. For example, in most databases it is an invariant that no two records may have the same key. In a C program that correctly manipulates strings, every string buffer must contain a terminating NUL byte on exit from each string function. In an inventory system, no parts count can hold a number less than zero.

Designing for Maintainability

Software is maintainable to the extent that people who are not its author can successfully understand and modify it. Maintainability demands more than code that works; it demands code that follows the Rule of Clarity and communicates successfully to human beings as well as the computer.

Unix programmers have a lot of implicit knowledge available to them about what makes for maintainable code, because Unix hosts source code that goes back decades. For reasons we'll discuss in Chapter 17, Unix programmers learn a tendency to scrap and rebuild rather than patching grubby code (see Rob Pike's meditation on this subject in Chapter 1). Thus, any sources that have survived more than a decade of evolutionary pressure have been selected for maintainability. These old, successful, well-established projects with maintainable code are the community's models for practice.

A question Unix programmers — and especially Unix programmers in the open-source world — learn to ask about tools they are evaluating for use is: “Is this code live, dormant, or dead?” Live code has an active developer community attached to it. Dormant code has often become dormant because the pain of maintaining it exceeded its utility to its originators. Dead code has been dormant for so long that it would be easier to reimplement an equivalent from scratch. If you want your code to live, investing effort to make it maintainable (and therefore attractive to future maintainers) will be one of the most effective ways you can spend your time.

Code that is designed to be both transparent and discoverable has gone a long way toward being maintainable. But there are other practices we can observe in the model projects in this chapter that are worth emulating.

One very important practice is an application of the Rule of Clarity: choosing simple algorithms. In Chapter 1 we quoted Ken Thompson: “When in doubt, use brute force”. Thompson understood the full cost of complicated algorithms — not just that they're more bug-prone when initially implemented, but that they're harder for maintainers down the line to understand.

Another important practice is the inclusion of hacker's guides. It has always been highly approved behavior for source code distributions to include guide documents informally describing the key data structures and algorithms in the code. In fact, Unix programmers have often been better about producing hacker's guides than they are about writing end-user documentation.

The open-source community has seized on and elaborated this custom. Besides being advice to future maintainers, hacker's guides for open-source projects are also designed to make it easy for casual contributors to add features or fix bugs. The Design Notes file shipped with fetchmail is representative. The Linux kernel sources include literally dozens of these.

In Chapter 19 we'll describe conventions that Unix developers have evolved for making source code distributions easy to examine and easy to build running code from. These practices, too, promote maintainability.

Chapter 7. Multiprogramming

If we believe in data structures, we must believe in independent (hence simultaneous) processing. For why else would we collect items within a structure? Why do we tolerate languages that give us the one without the other?

– Alan Perlis Epigrams in Programming, in ACM SIGPLAN (Vol 17 #9, 1982)

The most characteristic program-modularization technique of Unix is splitting large programs into multiple cooperating processes. This has usually been called ‘multiprocessing’ in the Unix world, but in this book we revive the older term ‘multiprogramming’ to avoid confusion with multiprocessor hardware implementations.

Multiprogramming is a particularly murky area of design, one in which there are few guidelines to good practice. Many programmers with excellent judgment about how to break up code into subroutines nevertheless wind up writing whole applications as monster single-process monoliths that founder on their own internal complexity.

The Unix style of design applies the do-one-thing-well approach at the level of cooperating programs as well as cooperating routines within a program, emphasizing small programs connected by well-defined interprocess communication or by shared files. Accordingly, the Unix operating system encourages us to break our programs into simpler subprocesses, and to concentrate on the interfaces between these subprocesses. It does this in at least three fundamental ways:

Inexpensive process-spawning and easy process control are critical enablers for the Unix style of programming. On an operating system such as VAX VMS, where starting processes is expensive and slow and requires special privileges, one must build monster monoliths because one has no choice. Fortunately the trend in the Unix family has been toward lower fork(2) overhead rather than higher. Linux, in particular, is famously efficient this way, with a process-spawn faster than thread-spawning on many other operating systems.[65]

Historically, many Unix programmers have been encouraged to think in terms of multiple cooperating processes by experience with shell programming. Shell makes it relatively easy to set up groups of multiple processes connected by pipes, running either in background or foreground or a mix of the two.

In the remainder of this chapter, we'll look at the implications of cheap process-spawning and discuss how and when to apply pipes, sockets, and other interprocess communication (IPC) methods to partition your design into cooperating processes. (In the next chapter, we'll apply the same separation-of-functions philosophy to interface design.)

While the benefit of breaking programs up into cooperating processes is a reduction in global complexity, the cost is that we have to pay more attention to the design of the protocols which are used to pass information and commands between processes. (In software systems of all kinds, bugs collect at interfaces.)

In Chapter 5 we looked at the lower level of this design problem — how to lay out application protocols that are transparent, flexible and extensible. But there is a second, higher level to the problem which we blithely ignored. That is the problem of designing state machines for each side of the communication.

It is not hard to apply good style to the syntax of application protocols, given models like SMTP or BEEP or XML-RPC. The real challenge is not protocol syntax but protocol logic —designing a protocol that is both sufficiently expressive and deadlock-free. Almost as importantly, the protocol has to be seen to be expressive and deadlock-free; human beings attempting to model the behavior of the communicating programs in their heads and verify its correctness must be able to do so.

In our discussion, therefore, we will focus on the kinds of protocol logic one naturally uses with each kind of interprocess communication.


[65] See, for example, the results quoted in Improving Context Switching Performance of Idle Tasks under Linux[Appleton].

Separating Complexity Control from Performance Tuning

First, though, we need to dispose of a few red herrings. Our discussion is not going to be about using concurrency to improve performance. Putting that concern before developing a clean architecture that minimizes global complexity is premature optimization, the root of all evil (see Chapter 12 for further discussion).

A closely related red herring is threads (that is, multiple concurrent processes sharing the same memory-address space). Threading is a performance hack. To avoid a long diversion here, we'll examine threads in more detail at the end of this chapter; the summary is that they do not reduce global complexity but rather increase it, and should therefore be avoided save under dire necessity.

Respecting the Rule of Modularity, on the other hand, is not a red herring; doing so can make your programs — and your life — simpler. All the reasons for process partitioning are continuous with the reasons for module partitioning that we developed in Chapter 4.

Another important reason for breaking up programs into cooperating processes is for better security. Under Unix, programs that must be run by ordinary users, but must have write access to security-critical system resources, get that access through a feature called the setuid bit.[66] Executable files are the smallest unit of code that can hold a setuid bit; thus, every line of code in a setuid executable must be trusted. (Well-written setuid programs, however, take all necessary privileged actions first and then drop their privileges back to user level for the remainder of their existence.)

Usually a setuid program only needs its privileges for one or a small handful of operations. It is often possible to break up such a program into cooperating processes, a smaller one that needs setuid and a larger one that does not. When we can do this, only the code in the smaller program has to be trusted. It is in significant part because this kind of partitioning and delegation is possible that Unix has a better security track record[67] than its competitors.


[66] A setuid program runs not with the privileges of the user calling it, but with the privileges of the owner of the executable. This feature can be used to give restricted, program-controlled access to things like the password file that nonadministrators should not be allowed to modify directly.

[67] That is, a better record measured in security breaches per total machine hours of Internet exposure.

Taxonomy of Unix IPC Methods

As in single-process program architectures, the simplest organization is the best. The remainder of this chapter will present IPC techniques roughly in order of escalating complexity of programming them. Before using a later, more complex technique, you should prove by demonstration — with prototypes and benchmark results — that no earlier and simpler technique will do. Often you will surprise yourself.

Handing off Tasks to Specialist Programs

In the simplest form of interprogram cooperation enabled by inexpensive process spawning, a program runs another to accomplish a specialized task. Because the called program is often specified as a Unix shell command through the system(3) call, this is often called shelling out to the called program. The called program inherits the user's keyboard and display and runs to completion. When it exits, the calling program resumes control of the keyboard and display and resumes execution.[68] Because the calling program does not communicate with the called program during the callee's execution, protocol design is not an issue in this kind of cooperation, except in the trivial sense that the caller may pass command-line arguments to the callee to change its behavior.

The classic Unix case of shelling out is calling an editor from within a mail or news program. In the Unix tradition one does not bundle purpose-built editors into programs that require general text-edited input. Instead, one allows the user to specify an editor of his or her choice to be called when editing needs to be done.

The specialist program usually communicates with its parent through the file system, by reading or modifying file(s) with specified location(s); this is how editor or mailer shellouts work.

In a common variant of this pattern, the specialist program may accept input on its standard input, and be called with the C library entry point popen(..., "w") or as part of a shellscript. Or it may send output to its standard output, and be called with popen(..., "r") or as part of a shellscript. (If it both reads from standard input and writes to standard output, it does so in a batch mode, completing all reads before doing any writes.) This kind of child process is not usually referred to as a shellout; there is no standard jargon for it, but it might well be called a ‘bolt-on’.

They key point about all these cases is that the specialist programs don't handshake with the parent while they are running. They have an associated protocol only in the trivial sense that whichever program (master or slave) is accepting input from the other has to be able to parse it.

Case Study: The mutt Mail User Agent

The mutt mail user agent is the modern representative of the most important design tradition in Unix email programs. It has a simple screen-oriented interface with single-keystroke commands for browsing and reading mail.

When you use mutt as a mail composer (either by calling it with an address as a command-line argument or by using one of the reply commands), it examines the process environment variable EDITOR, and then generates a temporary file name. The value of the EDITOR variable is called as a command with the tempfile name as an argument.[69] When that command terminates, mutt resumes on the assumption that the temporary file contains the desired mail text.

Almost all Unix mail- and netnews-composition programs observe the same convention. Because they do, composer implementers don't need to write a hundred inevitably diverging editors, and users don't need to learn a hundred divergent interfaces. Instead, users can carry their chosen editors with them.

An important variant of this strategy shells out to a small proxy program that passes the specialist job to an already-running instance of a big program, like an editor or a Web browser. Thus, developers who normally have an instance of emacs running on their X display can set EDITOR=emacsclient, and have a buffer pop open in their emacs when they request editing in mutt. The point of this is not really to save memory or other resources, it's to enable the user to unify all editing in a single emacs process (so that, for example, cut and paste among buffers can carry along internal emacs state information like font highlighting).

Pipes, Redirection, and Filters

After Ken Thompson and Dennis Ritchie, the single most important formative figure of early Unix was probably Doug McIlroy. His invention of the pipe construct reverberated through the design of Unix, encouraging its nascent do-one-thing-well philosophy and inspiring most of the later forms of IPC in the Unix design (in particular, the socket abstraction used for networking).

Pipes depend on the convention that every program has initially available to it (at least) two I/O data streams: standard input and standard output (numeric file descriptors 0 and 1 respectively). Many programs can be written as filters, which read sequentially from standard input and write only to standard output.

Normally these streams are connected to the user's keyboard and display, respectively. But Unix shells universally support redirection operations which connect these standard input and output streams to files. Thus, typing

ls >foo

sends the output of the directory lister ls(1) to a file named ‘foo’. On the other hand, typing:

wc <foo

causes the word-count utility wc(1) to take its standard input from the file ‘foo’, and deliver a character/word/line count to standard output.

The pipe operation connects the standard output of one program to the standard input of another. A chain of programs connected in this way is called a pipeline. If we write

ls | wc

we'll see a character/word/line count for the current directory listing. (In this case, only the line count is really likely to be useful.)

One favorite pipeline was “bc | speak”—a talking desk calculator. It knew number names up to a vigintillion.

– Doug McIlroy

It's important to note that all the stages in a pipeline run concurrently. Each stage waits for input on the output of the previous one, but no stage has to exit before the next can run. This property will be important later on when we look at interactive uses of pipelines, like sending the lengthy output of a command to more(1).

It's easy to underestimate the power of combining pipes and redirection. As an instructive example, The Unix Shell As a 4GL[Schaffer-Wolf] shows that with these facilities as a framework, a handful of simple utilities can be combined to support creating and manipulating relational databases expressed as simple textual tables.

The major weakness of pipes is that they are unidirectional. It's not possible for a pipeline component to pass control information back up the pipe other than by terminating (in which case the previous stage will get a SIGPIPE signal on the next write). Accordingly, the protocol for passing data is simply the receiver's input format.

So far, we have discussed anonymous pipes created by the shell. There is a variant called a named pipe which is a special kind of file. If two programs open the file, one for reading and the other for writing, a named pipe acts like a pipe-fitting between them. Named pipes are a bit of a historical relic; they have been largely displaced from use by named sockets, which we'll discuss below. (For more on the history of this relic, see the discussion of System V IPC below.)

Case Study: Piping to a Pager

Pipelines have many uses. For one example, Unix's process lister ps(1) lists processes to standard output without caring that a long listing might scroll off the top of the user's display too quickly for the user to see it. Unix has another program, more(1), which displays its standard input in screen-sized chunks, prompting for a user keystroke after displaying each screenful.

Thus, if the user types “ps | more ”, piping the output of ps(1) to the input of more(1), successive page-sized pieces of the list of processes will be displayed after each keystroke.

The ability to combine programs like this can be extremely useful. But the real win here is not cute combinations; it's that because both pipes and more(1) exist, other programs can be simpler. Pipes mean that programs like ls(1) (and other programs that write to standard out) don't have to grow their own pagers — and we're saved from a world of a thousand built-in pagers (each, naturally, with its own divergent look and feel). Code bloat is avoided and global complexity reduced.

As a bonus, if anyone needs to customize pager behavior, it can be done in one place, by changing one program. Indeed, multiple pagers can exist, and will all be useful with every application that writes to standard output.

In fact, this has actually happened. On modern Unixes, more(1) has been largely replaced by less(1), which adds the capability to scroll back in the displayed file rather than just forward.[70] Because less(1) is decoupled from the programs that use it, it's possible to simply alias ‘more’ to ‘less’ in your shell, set the environment variable PAGER to ‘less’ (see Chapter 10), and get all the benefits of a better pager with all properly-written Unix programs.

Case Study: Making Word Lists

A more interesting example is one in which pipelined programs cooperate to do some kind of data transformation for which, in less flexible environments, one would have to write custom code.

Consider the pipeline

tr -c '[:alnum:]' '[\n*]' | sort -iu | grep -v '^[0-9]*$'

The first command translates non-alphanumerics on standard input to newlines on standard output. The second sorts lines on standard input and writes the sorted data to standard output, discarding all but one copy of spans of adjacent identical lines. The third discards all lines consisting solely of digits. Together, these generate a sorted wordlist to standard output from text on standard input.

Case Study: pic2graph

Shell source code for the program pic2graph(1) ships with the groff suite of text-formatting tools from the Free Software Foundation. It translates diagrams written in the PIC language to bitmap images. Example 7.1 shows the pipeline at the heart of this code.

Example 7.1. The pic2graph pipeline.

(echo ".EQ"; echo $eqndelim; echo ".EN"; echo ".PS";cat;echo ".PE")|\
 groff -e -p $groffpic_opts -Tps >${tmp}.ps \
     && convert -crop 0x0 $convert_opts ${tmp}.ps ${tmp}.${format} \
     && cat ${tmp}.${format}

The pic2graph(1) implementation illustrates how much one pipeline can do purely by calling preexisting tools. It starts by massaging its input into an appropriate form, continues by feeding it through groff(1) to produce PostScript, and finishes by converting the PostScript to a bitmap. All these details are hidden from the user, who simply sees PIC source go in one end and a bitmap ready for inclusion in a Web page come out the other.

This is an interesting example because it illustrates how pipes and filtering can adapt programs to unexpected uses. The program that interprets PIC, pic(1), was originally designed only to be used for embedding diagrams in typeset documents. Most of the other programs in the toolchain it was part of are now semiobsolescent. But PIC remains handy for new uses, such as describing diagrams to be embedded in HTML. It gets a renewed lease on life because tools like pic2graph(1) can bundle together all the machinery needed to convert the output of pic(1) into a more modern format.

We'll examine pic(1) more closely, as a minilanguage design, in Chapter 8.

Case Study: bc(1) and dc(1)

Part of the classic Unix toolkit dating back to Version 7 is a pair of calculator programs. The dc(1) program is a simple calculator that accepts text lines consisting of reverse-Polish notation (RPN) on standard input and emits calculated answers to standard output. The bc(1) program accepts a more elaborate infix syntax resembling conventional algebraic notation; it includes as well the ability to set and read variables and define functions for elaborate formulas.

While the modern GNU implementation of bc(1) is standalone, the classic version passed commands to dc(1) over a pipe. In this division of labor, bc(1) does variable substitution and function expansion and translates infix notation into reverse-Polish — but doesn't actually do calculation itself, instead passing RPN translations of input expressions to dc(1) for evaluation.

There are clear advantages to this separation of function. It means that users get to choose their preferred notation, but the logic for arbitrary-precision numeric calculation (which is moderately tricky) does not have to be duplicated. Each of the pair of programs can be less complex than one calculator with a choice of notations would be. The two components can be debugged and mentally modeled independently of each other.

In Chapter 8 we will reexamine these programs from a slightly different example, as examples of domain-specific minilanguages.

Anti-Case Study: Why Isn't fetchmail a Pipeline?

In Unix terms, fetchmail is an uncomfortably large program that bristles with options. Thinking about the way mail transport works, one might think it would be possible to decompose it into a pipeline. Suppose for a moment it were broken up into several programs: a couple of fetch programs to get mail from POP3 and IMAP sites, and a local SMTP injector. The pipeline could pass Unix mailbox format. The present elaborate fetchmail configuration could be replaced by a shellscript containing command lines. One could even insert filters in the pipeline to block spam.

imap | spamblocker | smtp jrandom
imap | smtp jrandom
# pop | smtp jrandom

This would be very elegant and Unixy. Unfortunately, it can't work. We touched on the reason earlier; pipelines are unidirectional.

One of the things the fetcher program ( imap or pop) would have to do is decide whether to send a delete request for each message it fetches. In fetchmail 's present organization, it can delay sending that request to the POP or IMAP server until it knows that the local SMTP listener has accepted responsibility for the message. The pipelined, small-component version would lose that property.

Consider, for example, what would happen if the smtp injector fails because the SMTP listener reports a disk-full condition. If the fetcher has already deleted the mail, we lose. This means the fetcher cannot delete mail until it is notified to do so by the smtp injector. This in turn raises a host of questions. How would they communicate? What message, exactly, would the injector pass back? The global complexity of the resulting system, and its vulnerability to subtle bugs, would almost certainly be higher than that of a monolithic program.

Pipelines are a marvelous tool, but not a universal one.


The opposite of a shellout is a wrapper. A wrapper creates a new interface for a called program, or specializes it. Often, wrappers are used to hide the details of elaborate shell pipelines. We'll discuss interface wrappers in Chapter 11. Most specialization wrappers are quite simple, but nevertheless very useful.

As with shellouts, there is no associated protocol because the programs do not communicate during the execution of the callee; but the wrapper usually exists to specify arguments that modify the callee's behavior.

Case Study: Backup Scripts

Specialization wrappers are a classic use of the Unix shell and other scripting languages. One kind of specialization wrapper that is both common and representative is a backup script. It may be a one-liner as simple as this:

tar -czvf /dev/st0 "$@"

This is a wrapper for the tar(1) tape archiver utility which simply supplies one fixed argument (the tape device /dev/st0) and passes to tar all the other arguments supplied by the user (“ $@ ”).[71]

Security Wrappers and Bernstein Chaining

One common use of wrapper scripts is as security wrappers. A security script may call a gatekeeper program to check some sort of credential, then conditionally execute another based on the status value returned by the gatekeeper.

Bernstein chaining is a specialized security-wrapper technique first invented by Daniel J. Bernstein, who has employed it in a number of his packages. (A similar pattern appears in commands like nohup(1) and su(1), but the conditionality is absent.) Conceptually, a Bernstein chain is like a pipeline, but each successive stage replaces the previous one rather than running concurrently with it.

The usual application is to confine security-privileged applications to some sort of gatekeeper program, which can then hand state to a less privileged one. The technique pastes several programs together using execs, or possibly a combination of forks and execs. The programs are all named on one command line. Each program performs some function and (if successful) runs exec(2) on the rest of its command line.

Bernstein's rblsmtpd package is a prototypical example. It serves to look up a host in the antispam DNS zone of the Mail Abuse Prevention System. It does this by doing a DNS query on the IP address passed into it in the TCPREMOTEIP environment variable. If the query is successful, then rblsmtpd runs its own SMTP that discards the mail. Otherwise the remaining command-line arguments are presumed to constitute a mail transport agent that knows the SMTP protocol, and are handed to exec(2) to be run.

Another example can be found in Bernstein's qmail package. It contains a program called condredirect. The first parameter is an email address, and the remainder a gatekeeper program and arguments. condredirect forks and execs the gatekeeper with its arguments. If the gatekeeper exits successfully, condredirect forwards the email pending on stdin to the specified email address. In this case, opposite to that of rblsmtpd, the security decision is made by the child; this case is a bit more like a classical shellout.

A more elaborate example is the qmail POP3 server. It consists of three programs, qmail-popup, checkpassword, and qmail-pop3d. Checkpassword comes from a separate package cleverly called checkpassword, and unsurprisingly it checks the password. The POP3 protocol has an authentication phase and mailbox phase; once you enter the mailbox phase you cannot go back to the authentication phase. This is a perfect application for Bernstein chaining.

The first parameter of qmail-popup is the hostname to use in the POP3 prompts. The rest of its parameters are forked and passed to exec(2), after the POP3 username and password have been fetched. If the program returns failure, the password must be wrong, so qmail-popup reports that and waits for a different password. Otherwise, the program is presumed to have finished the POP3 conversation, so qmail-popup exits.

The program named on qmail-popup 's command line is expected to read three null-terminated strings from file descriptor 3.[72] These are the username, password, and response to a cryptographic challenge, if any. This time it's checkpassword which accepts as parameters the name of qmail-pop3d and its parameters. The checkpassword program exits with failure if the password does not match; otherwise it changes to the user's uid, gid, and home directory, and executes the rest of its command line on behalf of that user.

Bernstein chaining is useful for situations in which the application needs setuid or setgid privileges to initialize a connection, or to acquire some credential, and then drop those privileges so that following code does not have to be trusted. Following the exec, the child program cannot set its real user ID back to root. It's also more flexible than a single process, because you can modify the behavior of the system by inserting another program into the chain.

For example, rblsmtpd (mentioned above) can be inserted into a Bernstein chain, in between tcpserver (from the ucspi-tcp package) and the real SMTP server, typically qmail-smtpd. However, it works with inetd(8) and sendmail -bs as well.

Slave Processes

Occasionally, child programs both accept data from and return data to their callers through pipes connected to standard input and output, interactively. Unlike simple shellouts and what we have called ‘bolt-ons’ above, both master and slave processes need to have internal state machines to handle a protocol between them without deadlocking or racing. This is a drastically more complex and more difficult-to-debug organization than a simple shellout.

Unix's popen(3) call can set up either an input pipe or an output pipe for a shellout, but not both for a slave process — this seems intended to encourage simpler programming. And, in fact, interactive master-slave communication is tricky enough that it is normally only used when either (a) the implied protocol is utterly trivial, or (b) the slave process has been designed to speak an application protocol along the lines we discussed in Chapter 5. We'll return to this issue, and ways to cope with it, in Chapter 8.

When writing a master/slave pair, it is good practice for the master to support a command-line switch or environment variable that allows callers to set their own slave command. Among other things, this is useful for debugging; you will often find it handy during development to invoke the real slave process from within a harness that monitors and logs transactions between slave and master.

If you find that master/slave interactions in your program are becoming nontrivial, it may be time to think about going the rest of the way to a more peer-to-peer organization, using techniques like sockets or shared memory.

Case Study: scp and ssh

One common case in which the implied protocol really is trivial is progress meters. The scp(1) secure-copy command calls ssh(1) as a slave process, intercepting enough information from ssh's standard output to reformat the reports as an ASCII animation of a progress bar.[73]

Peer-to-Peer Inter-Process Communication

All the communication methods we've discussed so far have a sort of implicit hierarchy about them, with one program effectively controlling or driving another and zero or limited feedback passing in the opposite direction. In communications and networking we frequently need channels that are peer-to-peer, usually (but not necessarily) with data flowing freely in both directions. We'll survey peer-to-peer communications methods under Unix here, and develop some case studies in later chapters.


The use of tempfiles as communications drops between cooperating programs is the oldest IPC technique there is. Despite drawbacks, it's still useful in shellscripts, and in one-off programs where a more elaborate and coordinated method of communication would be overkill.

The most obvious problem with using tempfiles as an IPC technique is that it tends to leave garbage lying around if processing is interrupted before the tempfile can be deleted. A less obvious risk is that of collisions between multiple instances of a program using the same name for a tempfile. This is why it is conventional for shellscripts that make tempfiles to include $$ in their names; this shell variable expands to the process-ID of the enclosing shell and effectively guarantees that the filename will be unique (the same trick is supported in Perl).

Finally, if an attacker knows the location to which a tempfile will be written, it can overwrite on that name and possibly either read the producer's data or spoof the consumer process by inserting modified or spurious data into the file.[74] This is a security risk. If the processes involved have root privileges, this is a very serious risk. It can be mitigated by setting the permissions on the tempfile directory carefully, but such arrangements are notoriously likely to spring leaks.

All these problems aside, tempfiles still have a niche because they're easy to set up, they're flexible, and they're less vulnerable to deadlocks or race conditions than more elaborate methods. And sometimes, nothing else will do. The calling conventions of your child process may require that it be handed a file to operate on. Our first example of a shellout to an editor demonstrates this perfectly.


The simplest and crudest way for two processes on the same machine to communicate with each other is for one to send the other a signal. Unix signals are a form of soft interrupt; each one has a default effect on the receiving process (usually to kill it). A process can declare a signal handler that overrides the default action for the signal; the handler is a function that is executed asynchronously when the signal is received.

Signals were originally designed into Unix as a way for the operating system to notify programs of certain errors and critical events, not as an IPC facility. The SIGHUP signal, for example, is sent to every program started from a given terminal session when that session is terminated. The SIGINT signal is sent to whatever process is currently attached to the keyboard when the user enters the currently-defined interrupt character (often control-C). Nevertheless, signals can be useful for some IPC situations (and the POSIX-standard signal set includes two signals, SIGUSR1 and SIGUSR2, intended for this use). They are often employed as a control channel for daemons (programs that run constantly, invisibly, in background), a way for an operator or another program to tell a daemon that it needs to either reinitialize itself, wake up to do work, or write internal-state/debugging information to a known location.

I insisted SIGUSR1 and SIGUSR2 be invented for BSD. People were grabbing system signals to mean what they needed them to mean for IPC, so that (for example) some programs that segfaulted would not coredump because SIGSEGV had been hijacked.

This is a general principle — people will want to hijack any tools you build, so you have to design them to either be un-hijackable or to be hijacked cleanly. Those are your only choices. Except, of course, for being ignored—a highly reliable way to remain unsullied, but less satisfying than might at first appear.

– Ken Arnold

A technique often used with signal IPC is the so-called pidfile. Programs that will need to be signaled will write a small file to a known location (often in /var/run or the invoking user's home directory) containing their process ID or PID. Other programs can read that file to discover that PID. The pidfile may also function as an implicit lock file in cases where no more than one instance of the daemon should be running simultaneously.

There are actually two different flavors of signals. In the older implementations (notably V7, System III, and early System V), the handler for a given signal is reset to the default for that signal whenever the handler fires. The result of sending two of the same signal in quick succession is therefore usually to kill the process, no matter what handler was set.

The BSD 4.x versions of Unix changed to “reliable” signals, which do not reset unless the user explicitly requests it. They also introduced primitives to block or temporarily suspend processing of a given set of signals. Modern Unixes support both styles. You should use the BSD-style nonresetting entry points for new code, but program defensively in case your code is ever ported to an implementation that does not support them.

Receiving N signals does not necessarily invoke the signal handler N times. Under the older System V signal model, two or more signals spaced very closely together (that is, within a single timeslice of the target process) can result in various race conditions[75] or anomalies. Depending on what variant of signals semantics the system supports, the second and later instances may be ignored, may cause an unexpected process kill, or may have their delivery delayed until earlier instances have been processed (on modern Unixes the last is most likely).

The modern signals API is portable across all recent Unix versions, but not to Windows or classic (pre-OS X) MacOS.

System Daemons and Conventional Signals

Many well-known system daemons accept SIGHUP (originally the signal sent to programs on a serial-line drop, such as was produced by hanging up a modem connection) as a signal to reinitialize (that is, reload their configuration files); examples include Apache and the Linux implementations of bootpd(8), gated(8), inetd(8), mountd(8), named(8), nfsd(8), and ypbind(8). In a few cases, SIGHUP is accepted in its original sense of a session-shutdown signal (notably in Linux pppd(8)), but that role nowadays generally goes to SIGTERM.

SIGTERM (‘terminate’) is often accepted as a graceful-shutdown signal (this is as distinct from SIGKILL, which does an immediate process kill and cannot be blocked or handled). SIGTERM actions often involve cleaning up tempfiles, flushing final updates out to databases, and the like.

When writing daemons, follow the Rule of Least Surprise: use these conventions, and read the manual pages to look for existing models.

Case Study: fetchmail's Use of Signals

The fetchmail utility is normally set up to run as a daemon in background, periodically collecting mail from all remote sites defined in its run-control file and passing the mail to the local SMTP listener on port 25 without user intervention. fetchmail sleeps for a user-defined interval (defaulting to 15 minutes) between collection attempts, so as to avoid constantly loading the network.

When you invoke fetchmail with no arguments, it checks to see if you have a fetchmail daemon already running (it does this by looking for a pidfile). If no daemon is running, fetchmail starts up normally using whatever control information has been specified in its run-control file. If a daemon is running, on the other hand, the new fetchmail instance just signals the old one to wake up and collect mail immediately; then the new instance terminates. In addition, fetchmail -q sends a termination signal to any running fetchmail daemon.

Thus, typing fetchmail means, in effect, “poll now and leave a daemon running to poll later; don't bother me with the detail of whether a daemon was already running or not”. Observe that the detail of which particular signals are used for wakeup and termination is something the user doesn't have to know.


Sockets were developed in the BSD lineage of Unix as a way to encapsulate access to data networks. Two programs communicating over a socket typically see a bidirectional byte stream (there are other socket modes and transmission methods, but they are of only minor importance). The byte stream is both sequenced (that is, even single bytes will be received in the same order sent) and reliable (socket users are guaranteed that the underlying network will do error detection and retry to ensure delivery). Socket descriptors, once obtained, behave essentially like file descriptors.

Sockets differ from read/write in one important case. If the bytes you send arrive, but the receiving machine fails to ACK, the sending machine's TCP/IP stack will time out. So getting an error does not necessarily mean that the bytes didn't arrive; the receiver may be using them. This problem has profound consequences for the design of reliable protocols, because you have to be able to work properly when you don't know what was received in the past. Local I/O is ‘yes/no’. Socket I/O is ‘yes/no/maybe’. And nothing can ensure delivery — the remote machine might have been destroyed by a comet.

– Ken Arnold

At the time a socket is created, you specify a protocol family which tells the network layer how the name of the socket is interpreted. Sockets are usually thought of in connection with the Internet, as a way of passing data between programs running on different hosts; this is the AF_INET socket family, in which addresses are interpreted as host-address and service-number pairs. However, the AF_UNIX (aka AF_LOCAL) protocol family supports the same socket abstraction for communication between two processes on the same machine (names are interpreted as the locations of special files analogous to bidirectional named pipes). As an example, client programs and servers using the X windowing system typically use AF_LOCAL sockets to communicate.

All modern Unixes support BSD-style sockets, and as a matter of design they are usually the right thing to use for bidirectional IPC no matter where your cooperating processes are located. Performance pressure may push you to use shared memory or tempfiles or other techniques that make stronger locality assumptions, but under modern conditions it is best to assume that your code will need to be scaled up to distributed operation. More importantly, those locality assumptions may mean that portions of your system get chummier with each others' internals than ought to be the case in a good design. The separation of address spaces that sockets enforce is a feature, not a bug.

To use sockets gracefully, in the Unix tradition, start by designing an application protocol for use between them — a set of requests and responses which expresses the semantics of what your programs will be communicating about in a succinct way. We've already discussed the some major issues in the design of application protocols in Chapter 5.

Sockets are supported in all recent Unixes, under Windows, and under classic MacOS as well.

Case Study: PostgreSQL

PostgreSQL is an open-source database program. Had it been implemented as a monster monolith, it would be a single program with an interactive interface that manipulates database files on disk directly. Interface would be welded together with implementation, and two instances of the program attempting to manipulate the same database at the same time would have serious contention and locking issues.

Instead, the PostgreSQL suite includes a server called postmaster and at least three client applications. One postmaster server process per machine runs in background and has exclusive access to the database files. It accepts requests in the SQL query minilanguage through TCP/IP sockets, and returns answers in a textual format as well. When the user runs a PostgreSQL client, that client opens a session to postmaster and does SQL transactions with it. The server can handle several client sessions at once, and sequences requests so that they don't interfere with each other.

Because the front end and back end are separate, the server doesn't need to know anything except how to interpret SQL requests from a client and send SQL reports back to it. The clients, on the other hand, don't need to know anything about how the database is stored. Clients can be specialized for different needs and have different user interfaces.

This organization is quite typical for Unix databases — so much so that it is often possible to mix and match SQL clients and SQL servers. The interoperability issues are the SQL server's TCP/IP port number, and whether client and server support the same dialect of SQL.

Case Study: Freeciv

In Chapter 6, we introduced Freeciv as an example of transparent data formats. But more critical to the way it supports multiplayer gaming is the client/server partitioning of the code. This is a representative example of a program in which the application needs to be distributed over a wide-area network and handles communication through TCP/IP sockets.

The state of a running Freeciv game is maintained by a server process, the game engine. Players run GUI clients which exchange information and commands with the server through a packet protocol. All game logic is handled in the server. The details of GUI are handled in the client; different clients support different interface styles.

This is a very typical organization for a multiplayer online game. The packet protocol uses TCP/IP as a transport, so one server can handle clients running on different Internet hosts. Other games that are more like real-time simulations (notably first-person shooters) use raw Internet datagram protocol (UDP) and trade lower latency for some uncertainty about whether any given packet will be delivered. In such games, users tend to be issuing control actions continuously, so sporadic dropouts are tolerable, but lag is fatal.

Shared Memory

Whereas two processes using sockets to communicate may live on different machines (and, in fact, be separated by an Internet connection spanning half the globe), shared memory requires producers and consumers to be co-resident on the same hardware. But, if your communicating processes can get access to the same physical memory, shared memory will be the fastest way to pass information between them.

Shared memory may be disguised under different APIs, but on modern Unixes the implementation normally depends on the use of mmap(2) to map files into memory that can be shared between processes. POSIX defines a shm_open(3) facility with an API that supports using files as shared memory; this is mostly a hint to the operating system that it need not flush the pseudofile data to disk.

Because access to shared memory is not automatically serialized by a discipline resembling read and write calls, programs doing the sharing must handle contention and deadlock issues themselves, typically by using semaphore variables located in the shared segment. The issues here resemble those in multithreading (see the end of this chapter for discussion) but are more manageable because default is not to share memory. Thus, problems are better contained.

On systems where it is available and reliable, the Apache web server's scoreboard facility uses shared memory for communication between an Apache master process and the load-sharing pool of Apache images that it manages. Modern X implementations also use shared memory, to pass large images between client and server when they are resident on the same machine, to avoid the overhead of socket communication. Both uses are performance hacks justified by experience and testing, rather than being architectural choices.

The mmap(2) call is supported under all modern Unixes, including Linux and the open-source BSD versions; this is described in the Single Unix Specification. It will not normally be available under Windows, MacOS classic, and other operating systems.

Before purpose-built mmap(2) was available, a common way for two processes to communicate was for them to open the same file, and then delete that file. The file wouldn't go away until all open filehandles were closed, but some old Unixes took the link count falling to zero as a hint that they could stop updating the on-disk copy of the file. The downside was that your backing store was the file system rather than a swap device, the file system the deleted file lived on couldn't be unmounted until the programs using it closed, and attaching new processes to an existing shared memory segment faked up in this way was tricky at best.

After Version 7 and the split between the BSD and System V lineages, the evolution of Unix interprocess communication took two different directions. The BSD direction led to sockets. The AT&T lineage, on the other hand, developed named pipes (as previously discussed) and an IPC facility, specifically designed for passing binary data and based on shared-memory bidirectional message queues. This is called ‘System V IPC’—or, among old timers, ‘Indian Hill’ IPC after the AT&T facility where it was first written.

The upper, message-passing layer of System V IPC has largely fallen out of use. The lower layer, which consists of shared memory and semaphores, still has significant applications under circumstances in which one needs to do mutual-exclusion locking and some global data sharing among processes running on the same machine. These System V shared memory facilities evolved into the POSIX shared-memory API, supported under Linux, the BSDs, MacOS X and Windows, but not classic MacOS.

By using these shared-memory and semaphore facilities (shmget(2), semget(2), and friends) one can avoid the overhead of copying data through the network stack. Large commercial databases (including Oracle, DB2, Sybase, and Informix) use this technique heavily.


[68] A common error in programming shellouts is to forget to block signals in the parent while the subprocess runs. Without this precaution, an interrupt typed to the subprocess can have unwanted side effects on the parent process.

[69] Actually, the above is a slight oversimplification. See the discussion of EDITOR and VISUAL in Chapter 10 for the rest of the story.

[70] The less(1) man page explains the name by observing “Less is more”.

[71] A common error is to use $* rather than “ $@ ”. This does bad things when handed a filename with embedded spaces.

[72] qmail-popup 's standard input and standard output are the socket, and standard error (which will be file descriptor 2) goes to a log file. File descriptor 3 is guaranteed to be the next to be allocated. As an infamous kernel comment once observed: “You are not expected to understand this”.

[73] The friend who suggested this case study comments: “Yes, you can get away with this technique...if there are just a few easily-recognizable nuggets of information coming back from the slave process, and you have tongs and a radiation suit”.

[74] A particularly nasty variant of this attack is to drop in a named Unix-domain socket where the producer and consumer programs are expecting the tempfile to be.

[75] A ‘race condition’ is a class of problem in which correct behavior of the system relies on two independent events happening in the right order, but there is no mechanism for ensuring that they actually will. Race conditions produce intermittent, timing-dependent problems that can be devilishly difficult to debug.

Problems and Methods to Avoid

While BSD-style sockets over TCP/IP have become the dominant IPC method under Unix, there are still live controversies over the right way to partition by multiprogramming. Some obsolete methods have not yet completely died, and some techniques of questionable utility have been imported from other operating systems (often in association with graphics or GUI programming). We'll be touring some dangerous swamps here; beware the crocodiles.

Obsolescent Unix IPC Methods

Unix (born 1969) long predates TCP/IP (born 1980) and the ubiquitous networking of the 1990s and later. Anonymous pipes, redirection, and shellout have been in Unix since very early days, but the history of Unix is littered with the corpses of APIs tied to obsolescent IPC and networking models, beginning with the mx() facility that appeared in Version 6 (1976) and was dropped before Version 7 (1979).

Eventually BSD sockets won out as IPC was unified with networking. But this didn't happen until after fifteen years of experimentation that left a number of relics behind. It's useful to know about these because there are likely to be references to them in your Unix documentation that might give the misleading impression that they're still in use. These obsolete methods are described in more detail in Unix Network Programming [Stevens90].

The real explanation for all the dead IPC facilities in old AT&T Unixes was politics. The Unix Support Group was headed by a low-level manager, while some projects that used Unix were headed by vice presidents. They had ways to make irresistible requests, and would not brook the objection that most IPC mechanisms are interchangeable.

– Doug McIlroy
System V IPC

The System V IPC facilities are message-passing facilities based on the System V shared memory facility we described earlier.

Programs that cooperate using System V IPC usually define shared protocols based on exchanging short (up to 8K) binary messages. The relevant manual pages are msgctl(2) and friends. As this style has been largely superseded by text protocols passed between sockets, we do not give an example here.

The System V IPC facilities are present in Linux and other modern Unixes. However, as they are a legacy feature, they are not exercised very often. The Linux version is still known to have bugs as of mid-2003. Nobody seems to care enough to fix them.


Streams networking was invented for Unix Version 8 (1985) by Dennis Ritchie. A re-implementation called STREAMS (yes, it is all-capitals in the documentation) first became available in the 3.0 release of System V Unix (1986). The STREAMS facility provided a full-duplex interface (functionally not unlike a BSD socket, and like sockets, accessible through normal read(2) and write(2) operations after initial setup) between a user process and a specified device driver in the kernel. The device driver might be hardware such as a serial or network card, or it might be a software-only pseudodevice set up to pass data between user processes.

An interesting feature of both streams and STREAMS[76] is that it is possible to push protocol-translation modules into the kernel's processing path, so that the device the user process ‘sees’ through the full-duplex channel is actually filtered. This capability could be used, for example, to implement a line-editing protocol for a terminal device. Or one could implement protocols such as IP or TCP without wiring them directly into the kernel.

Streams originated as an attempt to clean up a messy feature of the kernel called ‘line disciplines’ — alternative modes of processing character streams coming from serial terminals and early local-area networks. But as serial terminals faded from view, Ethernet LANs became ubiquitous, and TCP/IP drove out other protocol stacks and migrated into Unix kernels, the extra flexibility provided by STREAMS had less and less utility. In 2003, System V Unix still supports STREAMS, as do some System V/BSD hybrids such as Digital Unix and Sun Microsystems' Solaris.

Linux and other open-source Unixes have effectively discarded STREAMS. Linux kernel modules and libraries are available from the LiS project, but (as of mid-2003) are not integrated into the stock Linux kernel. They will not be supported under non-Unix operating systems.

Remote Procedure Calls

Despite occasional exceptions such as NFS (Network File System) and the GNOME project, attempts to import CORBA, ASN.1, and other forms of remote-procedure-call interface have largely failed — these technologies have not been naturalized into the Unix culture.

There seem to be several underlying reasons for this. One is that RPC interfaces are not readily discoverable; that is, it is difficult to query these interfaces for their capabilities, and difficult to monitor them in action without building single-use tools as complex as the programs being monitored (we examined some of the reasons for this in Chapter 6). They have the same version skew problems as libraries, but those problems are harder to track because they're distributed and not generally obvious at link time.

As a related issue, interfaces that have richer type signatures also tend to be more complex, therefore more brittle. Over time, they tend to succumb to ontology creep as the inventory of types that get passed across interfaces grows steadily larger and the individual types more elaborate. Ontology creep is a problem because structs are more likely to mismatch than strings; if the ontologies of the programs on each side don't exactly match, it can be very hard to teach them to communicate at all, and fiendishly difficult to resolve bugs. The most successful RPC applications, such as the Network File System, are those in which the application domain naturally has only a few simple data types.

The usual argument for RPC is that it permits “richer” interfaces than methods like text streams — that is, interfaces with a more elaborate and application-specific ontology of data types. But the Rule of Simplicity applies! We observed in Chapter 4 that one of the functions of interfaces is as choke points that prevent the implementation details of modules from leaking into each other. Therefore, the main argument in favor of RPC is also an argument that it increases global complexity rather than minimizing it.

With classical RPC, it's too easy to do things in a complicated and obscure way instead of keeping them simple. RPC seems to encourage the production of large, baroque, over-engineered systems with obfuscated interfaces, high global complexity, and serious version-skew and reliability problems — a perfect example of thick glue layers run amok.

Windows COM and DCOM are perhaps the archetypal examples of how bad this can get, but there are plenty of others. Apple abandoned OpenDoc, and both CORBA and the once wildly hyped Java RMI have receded from view in the Unix world as people have gained field experience with them. This may well be because these methods don't actually solve more problems than they cause.

Andrew S. Tanenbaum and Robbert van Renesse have given us a detailed analysis of the general problem in A Critique of the Remote Procedure Call Paradigm[Tanenbaum-VanRenesse], a paper which should serve as a strong cautionary note to anyone considering an architecture based on RPC.

All these problems may predict long-term difficulties for the relatively few Unix projects that use RPC. Of these projects, perhaps the best known is the GNOME desktop effort.[77] These problems also contribute to the notorious security vulnerabilities of exposing NFS servers.

Unix tradition, on the other hand, strongly favors transparent and discoverable interfaces. This is one of the forces behind the Unix culture's continuing attachment to IPC through textual protocols. It is often argued that the parsing overhead of textual protocols is a performance problem relative to binary RPCs — but RPC interfaces tend to have latency problems that are far worse, because (a) you can't readily anticipate how much data marshaling and unmarshaling a given call will involve, and (b) the RPC model tends to encourage programmers to treat network transactions as cost-free. Adding even one additional round trip to a transaction interface tends to add enough network latency to swamp any overhead from parsing or marshaling.

Even if text streams were less efficient than RPC, the performance loss would be marginal and linear, the kind better addressed by upgrading your hardware than by expending development time or adding architectural complexity. Anything you might lose in performance by using text streams, you gain back in the ability to design systems that are simpler — easier to monitor, to model, and to understand.

Today, RPC and the Unix attachment to text streams are converging in an interesting way, through protocols like XML-RPC and SOAP. These, being textual and transparent, are more palatable to Unix programmers than the ugly and heavyweight binary serialization formats they replace. While they don't solve all the more general problems pointed out by Tanenbaum and van Renesse, they do in some ways combine the advantages of both text-stream and RPC worlds.

Threads — Threat or Menace?

Though Unix developers have long been comfortable with computation by multiple cooperating processes, they do not have a native tradition of using threads (processes that share their entire address spaces). These are a recent import from elsewhere, and the fact that Unix programmers generally dislike them is not merely accident or historical contingency.

From a complexity-control point of view, threads are a bad substitute for lightweight processes with their own address spaces; the idea of threads is native to operating systems with expensive process-spawning and weak IPC facilities.

By definition, though daughter threads of a process typically have separate local-variable stacks, they share the same global memory. The task of managing contentions and critical regions in this shared address space is quite difficult and a fertile source of global complexity and bugs. It can be done, but as the complexity of one's locking regime rises, the chance of races and deadlocks due to unanticipated interactions rises correspondingly.

Threads are a fertile source of bugs because they can too easily know too much about each others' internal states. There is no automatic encapsulation, as there would be between processes with separate address spaces that must do explicit IPC to communicate. Thus, threaded programs suffer from not just ordinary contention problems, but from entire new categories of timing-dependent bugs that are excruciatingly difficult to even reproduce, let alone fix.

Thread developers have been waking up to this problem. Recent thread implementations and standards show an increasing concern with providing thread-local storage, which is intended to limit problems arising from the shared global address space. As threading APIs move in this direction, thread programming starts to look more and more like a controlled use of shared memory.

Threads often prevent abstraction. In order to prevent deadlock, you often need to know how and if the library you are using uses threads in order to avoid deadlock problems. Similarly, the use of threads in a library could be affected by the use of threads at the application layer.

– David Korn

To add insult to injury, threading has performance costs that erode its advantages over conventional process partitioning. While threading can get rid of some of the overhead of rapidly switching process contexts, locking shared data structures so threads won't step on each other can be just as expensive.

The X server, able to execute literally millions of ops/second, is not threaded; it uses a poll/select loop. Various efforts to make a multithreaded implementation have come to no good result. The costs of locking and unlocking get too high for something as performance-sensitive as graphics servers.

– Jim Gettys

This problem is fundamental, and has also been a continuing issue in the design of Unix kernels for symmetric multiprocessing. As your resource-locking gets finer-grained, latency due to locking overhead can increase fast enough to swamp the gains from locking less core memory.

One final difficulty with threads is that threading standards still tend to be weak and underspecified as of mid-2003. Theoretically conforming libraries for Unix standards such as POSIX threads (1003.1c) can nevertheless exhibit alarming differences in behavior across platforms, especially with respect to signals, interactions with other IPC methods, and resource cleanup times. Windows and classic MacOS have native threading models and interrupt facilities quite different from those of Unix and will often require considerable porting effort even for simple threading cases. The upshot is that you cannot count on threaded programs to be portable.

For more discussion and a lucid contrast with event-driven programming, see Why Threads Are a Bad Idea [Ousterhout96].


[76] STREAMS was much more complex. Dennis Ritchie is reputed to have said “Streams means something different when shouted”.

[77] GNOME's main competitor, KDE, started with CORBA but abandoned it in their 2.0 release. They have been on a quest for lighter-weight IPC methods ever since.

Process Partitioning at the Design Level

Now that we have all these methods, what should we do with them?

The first thing to notice is that tempfiles, the more interactive sort of master/slave process relationship, sockets, RPC, and all other methods of bidirectional IPC are at some level equivalent — they're all just ways for programs to exchange data during their lifetimes. Much of what we do in a sophisticated way using sockets or shared memory we could do in a primitive way using tempfiles as mailboxes and signals for notification. The differences are at the edges, in how programs establish communication, where and when one does the marshalling and unmarshalling of messages, in what sorts of buffering problems you may have, and atomicity guarantees you get on the messages (that is, to what extent you can know that the result of a single send action from one side will show up as a single receive event on the other).

We've seen from the PostgreSQL study that one effective way to hold down complexity is to break an application into a client/server pair. The PostgreSQL client and server communicate through an application protocol over sockets, but very little about the design pattern would change if they used any other bidirectional IPC method.

This kind of partitioning is particularly effective in situations where multiple instances of the application must manage access to resources that are shared among all. A single server process may manage all resource contention, or cooperating peers may each take responsibility for some critical resource.

Client-server partitioning can also help distribute cycle-hungry applications across multiple hosts. Or it may make them suitable for distributed computing across the Internet (as with Freeciv). We'll discuss the related CLI server pattern in Chapter 11.

Because all these peer-to-peer IPC techniques are alike at some level, we should evaluate them mainly on the amount of program-complexity overhead they incur, and how much opacity they introduce into our designs. This, ultimately, is why BSD sockets have won over other Unix IPC methods, and why RPC has generally failed to get much traction.

Threads are fundamentally different. Rather than supporting communication among different programs, they support a sort of timesharing within an instance of a single program. Rather than being a way to partition a big program into smaller ones with simpler behavior, threading is strictly a performance hack. It has all the problems normally associated with performance hacks, and a few special ones of its own.

Accordingly, while we should seek ways to break up large programs into simpler cooperating processes, the use of threads within processes should be a last resort rather than a first. Often, you may find you can avoid them. If you can use limited shared memory and semaphores, asynchronous I/O using SIGIO, or poll(2)/select(2) rather than threading, do it that way. Keep it simple; use techniques earlier on this list and lower on the complexity scale in preference to later ones.

The combination of threads, remote-procedure-call interfaces, and heavyweight object-oriented design is especially dangerous. Used sparingly and tastefully, any of these techniques can be valuable — but if you are ever invited onto a project that is supposed to feature all three, fleeing in terror might well be an appropriate reaction.

We have previously observed that programming in the real world is all about managing complexity. Tools to manage complexity are good things. But when the effect of those tools is to proliferate complexity rather than to control it, we would be better off throwing them away and starting from zero. An important part of the Unix wisdom is to never forget this.

Chapter 8. Minilanguages

A good notation has a subtlety and suggestiveness which at times makes it almost seem like a live teacher.

– Bertrand Russell The World of Mathematics (1956)

One of the most consistent results from large-scale studies of error patterns in software is that programmer error rates in defects per hundreds of lines are largely independent of the language in which the programmers are coding.[78] Higher-level languages, which allow you to get more done in fewer lines, mean fewer bugs as well.

Unix has a long tradition of hosting little languages specialized for a particular application domain, languages that can enable you to drastically reduce the line count of your programs. Domain-specific language examples include the numerous Unix typesetting languages ( troff, eqn, tbl, pic, grap), shell utilities ( awk, sed, dc, bc), and software development tools ( make, yacc, lex). There is a fuzzy boundary between domain-specific languages and the more flexible sort of application run-control file ( sendmail, BIND, X); another with data-file formats; and another with scripting languages (which we'll survey in Chapter 14).

Historically, domain-specific languages of this kind have been called ‘little languages’ or ‘minilanguages’ in the Unix world, because early examples were small and low in complexity relative to general-purpose languages (all three terms for the category are in common use). But if the application domain is complex (in that it has lots of different primitive operations or involves manipulation of intricate data structures), an application language for it may have to be rather more complex than some general-purpose languages. We'll keep the traditional term ‘minilanguage’ to emphasize that the wise course is usually to keep these designs as small and simple as possible.

The domain-specific little language is an extremely powerful design idea. It allows you to define your own higher-level language to specify the appropriate methods, rules, and algorithms for the task at hand, reducing global complexity relative to a design that uses hardwired lower-level code for the same ends. You can get to a minilanguage design in at least three ways, two of them good and one of them dangerous.

One right way to get there is to realize up front that you can use a minilanguage design to push a given specification of a programming problem up a level, into a notation that is more compact and expressive than you could support in a general-purpose language. As with code generation and data-driven programming, a minilanguage lets you take practical advantage of the fact that the defect rate in your software will be largely independent of the level of the language you are using; more expressive languages mean shorter programs and fewer bugs.

The second right way to get to a minilanguage design is to notice that one of your specification file formats is looking more and more like a minilanguage — that is, it is developing complex structures and implying actions in the application you are controlling. Is it trying to describe control flow as well as data layouts? If so, it may be time to promote that control flow from being implicit to being explicit in your specification language.

The wrong way to get to a minilanguage design is to extend your way to it, one patch and crufty added feature at a time. On this path, your specification file keeps sprouting more implied control flow and more tangled special-purpose structures until it has become an ad-hoc language without your noticing it. Some legendary nightmares have been spawned this way; the example every Unix guru will think of (and shudder over) is the configuration file associated with the sendmail mail transport.

Sadly, most people do their first minilanguage the wrong way, and only realize later what a mess it is. Then the question is: how to clean it up? Sometimes the answer implies rethinking the entire application design. Another notorious example of language-by-feature creep was the editor TECO, which grew first macros and then loops and conditionals as programmers wanted to use it to package increasingly complex editing routines. The resulting ugliness was eventually fixed by a redesign of the entire editor to be based on an intentional language; this is how Emacs Lisp (which we'll survey below) evolved.

All sufficiently complicated specification files aspire to the condition of minilanguages. Therefore, it will often be the case that your only defense against designing a bad minilanguage is knowing how to design a good one. This need not be a huge step or involve knowing a lot of formal language theory; with modern tools, learning a few relatively simple techniques and bearing good examples in mind as you design should be sufficient.

In this chapter we'll examine all the kinds of minilanguages normally supported under Unix, and try to identify the kinds of situation in which each of them represents an effective design solution. This chapter is not meant to be an exhaustive catalog of Unix languages, but rather to bring out the design principles involved in structuring an application around a minilanguage. We'll have much more to say about languages for general-purpose programming in Chapter 14.

We'll need to start by doing a little taxonomy, so we'll know what we're talking about later on.


[78] Les Hatton reports by email on the analysis in his book in preparation, Software Failure : “Provided you use executable line counts for the density measure, the injected defect densities vary less between languages than they do between engineers by about a factor of 10”.

Understanding the Taxonomy of Languages

All the languages in Figure 8.1 are described in case studies, either in this chapter or elsewhere in this book. For the general-purpose interpreters near the right-hand side, see Chapter 14.

Figure 8.1. Taxonomy of languages.