CML2 Language and Tools Description

The Kernel Configuration Menu Language

Eric Steven Raymond

Thyrsus Enterprises

      esr@thyrsus.com
      

This is version 0.5.0

$Date: 2000/06/26 08:40:58 $

This paper describes CML2, the Configuration Menu Language designed to support Linux kernel configuration. It is written as a historical narrative because that provides a good frame for describing the design issues in the language.


Why CML2?

 

" Every program eventually becomes rococo, and then rubble. "

 Alan Perlis

The CML2 project was launched because in March 2000, in the process of building 2.3.51 kernels, I discovered that the original kbuild configure language had reached the rococo stage; it had become so brittle that any attempt to change or extend it would break everything.

This matters, because the kernel-configuration process has grown excessively complex. The configuration system's job is to turn the user's answers to configuration questions into a file of #define constructs used to condition features in or out of the C code. As Linux has grown more features and more driver support, the number of menus and prompts one must navigate to choose the appropriate subset of those features has become forbidding even to expert users, and outright impossible for the novice users Linux increasingly seeks to attract.

To properly manage the complexity we have created, we need a configuration interface that can support better chunking and progressive disclosure. Ideally, novices building kernels should see only the choices they need to make. As they gain competence and cue the configuration system that they are ready, they should see more options for tuning and exotic hardware. The full process should be accessible to expert kernel hackers, but not inflicted willy-nilly on everyone else as it now is.

With the brittle old configure language (which we'll call CML1) this kind of redesign would simply not have been possible at all. CML1 had started out life as a set of simple shellscripts, but evolved into a massive hairball featuring three different interpreters, C code customized to generate Tcl/Tk, and source files so difficult to read that (according to the unanimous report of CML1's own maintainers) latter-day CML1 users rarely try to program it by any means more sophisticated than a cut-paste-edit of the menus for existing features.

I had a second reason for getting involved besides wanting to see the configuration process simplified; I like designing little languages more than any other kind of programming. The interesting challenge I saw was to design a language that would (a) capture all the right domain-specific abstractions, but (b) remain flexible enough to allow configuration-system maintainers to experiment with different progressive-disclosure policies.