Linux Device Drivers (2nd ed)

Author: Alessandro Rubini & Jonathan Corbet



Publisher: O'Reilly

Reviewed by: Nicholas Clark

The second edition gives you 592 pages total, across 16 chapters, written by 2 authors covering Linux versions 2.4.x, 2.2.x and 2.0.x. This contrasts with a first edition of 450 pages, 17 chapters, 1 author, covering 2.0.x and 1.2.13. However, those statistics hide the enormous benefits to the rest of the book that the removal of that extra chapter brings. Chapter 17 in the first edition is entitled "Recent Developments" and gives 20 pages describing the major changes to Linux in the 2.1 development series up to 2.1.43. The first edition had the misfortune to overlap the 2.1 development series, so that by the time it was published in February 1998 it was already partially obsolete, as the 2.1 kernel had reached 2.1.84, and within a year it was out of date when the 2.2.0 kernel was released in January 1999. In contrast, the second edition was published in June 2001, only a few months into the stable 2.4 series, and 2.5.0 was not released until November 2001. Thus over a year later the book is still covering the current stable series, and looks safe for some time to come. The book is also available online under the GNU free documentation licence at I've not checked whether the online version has had updates since the hardcopy edition was committed to press.

Although the second edition deals with three distinct versions of the Linux kernel, it manages to almost completely eliminate any complexity that describing multiple versions could have brought. The second edition is much better than the first edition, which could give a feeling of disorientation as it jumped back and forth between the two kernel versions it covered. The second edition achieves its serenity by structuring all chapters so that the main body only deals with 2.4.x series kernels. Example code is presented which will compile on 2.4.x, rather than code cluttered with #ifdefs, and a note made of changes needed to give backwards compatibility to 2.2 and 2.0 series kernels. The detail of these changes is saved up to a Backwards Compatibility section at the end of each chapter, which keeps the main text flowing along nicely. Concepts or functionality only present in 2.4 (or 2.4 and 2.2) series kernels is clearly marked, as are kernel subsystems now principally maintained for backwards compatibility. This makes it easy to decide what features to use if you are only interested in writing drivers for 2.4 (or later) kernels, and what features to avoid if you want to run on earlier kernels. It also gives the reader a clear idea of how the kernel systems have evolved over time, something that would not be easy to infer with just the kernel sources.

The book is carefully structured to build up logically, starting with simpler concepts in the early chapters, and building up logically so that more complex subjects in the later chapters don't require the introduction of many new ideas simultaneously. Hence chapter 2 describes how to build modules load the into a running kernel, configure them, and safely unloaded them when done. This description is all safely concluded before the first mention of how to write device drivers, which begins in chapter 3. The preface says that the book is structured in two parts - part 1 (chapters 1 to 10) give everything you need to know about writing a character device, and move from software towards hardware, whereas part 2 (chapters 11 to 16) cover more advanced topics such as block devices and network devices. The ordering is identical to the first edition, and was dictated because prior to the 2.4 kernels block devices were implemented using many of the structures of char devices. The preface states that all the chapters cover a distinct problem, and generally they do, in a neat self contained fashion. One exception would be that the many small subsections on various features of char drivers collected in Chapters 3 and 5 ("Char Drivers" and "Enhanced Char Driver Operations") have to be ordered somewhat arbitrarily, but I cannot see a better way of smoothing the flow between the disjoint subtopics. Chapter 12 ("Loading Block Drivers") covers pretty much everything you ever wanted to know about block drivers (except mmap), request queues, partitioning, and then some. At 49 dense pages I feel that it could have benefited from breaking into 2 chapters (basic block devices, and "everything else") possibly with the mmap chapter put between as light relief. But apart from avoiding getting stuck in that tar pit the book is generally quite readable from start to finish without either your head exploding or drifting off to sleep. (And I confess that is what I have done, over the course of several weeks, without actually trying to program anything it has taught me).

All the example code in the book is supplied online at The book doesn't use code as padding - quite often code not necessary to illustrate the point is omitted from the printout in the book, with the reader referred to the downloadable archive. Likewise this is no "all the worlds a VAXC^Wx86". For the first edition all the code was tested on x86, alpha and sparc to ensure it was both 64 bit clean and endian neutral. For this edition the test battery is upped to x86, alpha, ARM, IA-64, M68K, PPC, Sparc, Sparc64 and VR41xx (MIPS) [Which is more than we're smoking perl on, as we have no M68K or IA-64 test systems.] In addition the code is SMP safe, and the book is explicit in how to avoid SMP pitfalls.

If I have a gripe about the book, it is the diagrams. There are beautifully drawn greyscale diagrams, credited to Robert Romano and Jessamyn Read, which look far more slick than say the austere simplicity of the black and white line drawings in the books of W. Richard Stevens. However, there are far fewer diagrams than in Stevens, and this text hardly refers to the diagrams. For example, figure 3.2 is captioned the arguments to read. The only mention in the text is the sentence "Figure 3.2 represents how a typical read implementation uses its arguments." No text talks through what the reader is supposed to learn from the diagram - in this case I believe it is illustrating how you need to be careful as a driver writer because some of the pointers you will be given will be to user space, others to kernel space, and that even though you need to treat them differently, there is no way for you (or the compiler) to infer which is which from their C types. Figure 5.2 is a complex diagram illustrating "the data structures of poll", with a 1 sentence reference from the body text. There is no caption explaining the significance of the contents, nor is their body text. In a Stevens book this figure would be split into 3 or even 4 parts, each a diagram in itself. Colouring or shading would be avoided unless it had a clear and explicitly stated contribution to the understanding of the concepts illustrated, and their would be one or more paragraphs explaining the important things to take in and remember from the diagrams.

At first I thought that this lack of integration of the diagrams was a side effect of the book also being available online. I was guessing that the illustrations were copyright O'Reilly and not online, and that there was a single text had to work with and without the diagrams. But this is not the case - the online version has the same diagrams as the printed version, so there is no excuse for this lack of integration. I feel that it means that the effort put into the diagrams is mostly wasted, as they add nothing to the text.

So does the book do the job? Yes, I think it does (but beware the caveat that I've not actually tried to write a device driver). The book seems to teach you things pertinent to the task of a device driver author that you couldn't pick up from a simple grep of the kernel sources, such as the hows and whys of interrupts, scheduling your tasks for later, allocating memory suitable for DMA, the kernel interfaces for working with various kinds of hardware buses. There are also introductions to related topics, but this book correctly does not get sidetracked into the details, which would be enough to fill a book in their own right. The isn't a general introduction to the Linux kernel (for example it does not mention filesystems anywhere), but it does not set out to be one. Likewise it avoids the specifics of hardware almost totally, basing most of its examples on either devices that implement ramdisks or devices that implement loopback network drivers. The book attempts to teach the reader best practice in writing drivers, emphasising robustness, portability and maintainability. It is also clear to teach how drivers fit into Linux, keeping the distinction between mechanism and policy clear, and emphasising the Unix philosophy that policy belongs in userspace.

In summary this is a good book for a specific audience - people who need to write device drivers for Linux. Not a kernel primer. Not a hardware primer. Covers stuff you need to know. Not stuff you don't.