GAR: Automating Entire OS Builds
I'm a member of the LNX-BBC Project. The LNX-BBC is a business card-sized CD-ROM containing a miniature distribution of GNU/Linux. In order to fit everything we wanted on the roughly 50MB volume, we had to build every single piece of software from scratch and weed out unnecessary files.
The first few LNX-BBC releases were done entirely by hand, pulling precompiled binaries from existing distributions and hand compiling some 200 software packages. While ultimately successful, it made the process of upgrading individual packages difficult. It was also impossible to keep our work in any sort of revision-control system.
User requests for source code revealed another problem with our development process. We had written some scripts to automate the creation of the compressed loopback filesystem and the El-Torito bootable ISO image, but the really difficult work was the compilation and installation of all the packages. We had no way of giving people a single tarball for building a BBC from scratch.
What we needed was a system for automating the compilation and installation of all of the third-party software. It needed to allow us to store our customizations in CVS and provide a simple mechanism for end users to build their own LNX-BBC ISO images.
These requirements had been noticed and met before. In 1994, Jordan Hubbard began work on the BSD Ports system. The FreeBSD operating system includes a great many programs and utilities, but it is not complete without a number of third-party programs. The BSD Ports system manages the compilation and installation of third-party software that has been ported to BSD.
Often when one asks for help with FreeBSD, an expert may answer by saying ``Just use ports!'' and listing the following commands:
cd /usr/ports/<category>/<package> make install
to suggest that the user needs to install a particular software package. This is similar to the way many Debian experts will tell people to apt-get install <package>. It's a simple way for a user to install software and related dependencies.
The Ports system is written entirely in pmake, the version of the make utility that comes with BSD. The choice to use make is both an obvious and a novel one. Make can be thought of as a language designed to automate software compilation and has many facilities for expressing build dependencies and rule-based build actions.
On the other hand, make has very limited flow control and lacks many features traditionally found in procedural programming languages. It can be rather unwieldy when used to build large projects.
As of the time of this writing, the core Ports runtime for FreeBSD has undergone 400 revisions since 1994. You can see all of the revisions and their changelog entries at http://www.freebsd.org/cgi/cvsweb.cgi/ports/Mk/bsd.port.mk. The collection of software currently contains nearly 4,000 packages.
I had made the mistake, in 1998, of making a fuss about the GNU system needing something like Ports. I made claims about how much more elegant Ports would be if written in GNU make and spent a lot of time reading the FSF's make book and the NetBSD Ports source code. It wasn't until an LNX-BBC meeting in 2001 that someone called my bluff, and I actually had to sit down and write the thing.
GAR ostensibly stands for the Gmake Autobuild Runtime because it's a library of Makefile rules that provide Ports-like functionality to individual packages. (It's actually just named GAR because that's my favorite interjection: ``Gar!'')
From the user's perspective, the GAR system may well be a tree of carefully maintained source code, ready to compile. The reality is that the system is just a tree of directories containing Makefiles, and the only thing that's stored in the GAR system itself is the information necessary to perform the steps a user would take in compiling and installing software.
The base of the GAR directory tree contains a number of directories. These directories are package categories, and within each category is a directory for each package. Inside a package directory is (among other things) a Makefile.
By way of the GAR system libraries, this Makefile provides seven basic targets for the package: fetch, checksum, extract, patch, configure, build and install.
Thus, to install Python using the BBC's GAR tree, one would cd to lang/python and run make install. To look at the source code to netcat, one would cd to net/netcat and run make extract.
Each of these seven targets runs all previous targets, though any of them may be undefined for a given package. If you run make patch, it's the same as running
make fetch checksum extract patch
fetch: this target downloads all files and patches needed to compile the package. Typically this is a single tarball, accompanied by the occasional patch file.
checksum: uses md5sum to ensure that the downloaded files match those with which the package maintainer worked.
extract: makes sure that all of the necessary source files are available in a working directory. In some cases (such as when downloading a single C source file), this will simply copy files over.
patch: if the package has to be patched (either via third-party patches or package maintainer patches), this target will perform that step.
configure: configures the package as specified in the Makefile. It will typically run the package's underlying configuration system (such as autoconf or Imake).
build: performs the actual step of compilation.
install: puts files in the proper locations and performs any necessary mop-up work.
These targets are named after their counterparts in the BSD Ports system and behave in the same manner.
One way to think of GAR is as a consistent interface for compiling a piece of software. No matter what mechanism the program uses to compile or install, a make install will install the program. The user need not care whether a program is a Python script or a C program that uses Imake.
In fact, the system will even satisfy dependencies. If the robotfindskitten program uses libncurses, the GAR system will pause the robotfindskitten build process right before the configure stage and then go install libncurses.
The GAR system is also centrally configurable, allowing the user to specify installation directories, CFLAGS and other useful environment variables. It is possible to configure default download locations, so that you can grab the code from a local cache or a nearby mirror of all the files.
GAR was designed with the notion that the build system and the target system may be different machines. It is also possible that the build and target system are different architectures. The following changes to gar.conf.mk show a possible way to configure GAR for a cross-compiling environment:
prefix = /usr/local/build/mips-build LDFLAGS += -L$(libdir) -L/usr/local/lib-mips CC = /usr/local/bin/gcc-mipsel LD_LIBRARY_PATH += :$(libdir):/usr/local/lib-mips export prefix LDFLAGS CC LD_LIBRARY_PATH
All of these settings will propagate down into each individual package.
The best way to understand the process of making a new package is to look at an example. The following is the relevant parts of the grep package:
GARNAME = grep GARVERSION = 2.4.2 CATEGORIES = utils MASTER_SITES = ftp://ftp.gnu.org/pub/gnu/grep/ DISTFILES = $(DISTNAME).tar.gz CONFIGURE_SCRIPTS = $(WORKSRC)/configure BUILD_SCRIPTS = $(WORKSRC)/Makefile INSTALL_SCRIPTS = $(WORKSRC)/Makefile CONFIGURE_ARGS = $(DIRPATHS) include ../../gar.mk
The first few lines set some basic bookkeeping information. The GARNAME and GARVERSION refer to the package name and version string as the GAR system will manipulate them. They are also used to create the convenience variable DISTNAME, which is $(GARNAME)-$(GARVERSION) by default, since most GNU or automake-using packages name their tarballs in that fashion.
The actual build of the system depends on the CONFIGURE_SCRIPTS, BUILD_SCRIPTS and INSTALL_SCRIPTS variables. These point to a space-separated list of files that are essential to the configure, build and install steps (respectively).
The GAR system will know what to do with most types of scripts. It knows to run CONFIGURE_SCRIPTs named ``configure'', and to run make or make install in the directory where a Makefile lives. It also knows that Imakefiles use xmkmf, and so forth. For most packages, no extra work needs to be done here.
We usually need to specify the CONFIGURE_ARGS to include the directory settings that we define in gar.conf.mk. This currently requires setting it to include $(DIRPATHS).
Finally, we include the gar.mk library. This takes all of the variable setting we've done and puts it to good use. This allows your short Makefile to provide the seven basic targets described above. It must come last (or, at least, after all of the variables have been set) in order to function properly.
The automatic satisfaction of dependencies is deceptively simple. In order to specify a library dependency, for example, simply set the LIBDEPS variable to be a space-separated list of paths from the base GAR directory to the package you depend on. For example, the GNU parted program has the following:
LIBDEPS = lib/e2fsprogs-libs
Before the configure step is run, GAR will cd into ../../lib/e2fsprogs-libs/ and run make install.
Quite often, the default behavior isn't what one would hope. You can pass parameters to a package's configure script with CONFIGURE_ARGS, but what if the GAR system doesn't know about your configuration script type? What if the configuration steps aren't enough? What if they're flat-out wrong? What if the same is true for the fetch, extract or install rules?
Fortunately, the system provides for this in a number of ways. What follows are a few of the mechanisms that a package maintainer can use to override or enhance the default behaviors.
For each of the seven basic targets, there exists slots for per-package pre- and post- rules. That means that the package maintainer can specify work to be done immediately before or after a rule.
To define a rule in make, simply place it at the beginning of a line, followed by a colon. The shell commands it should execute follow, preceded by a tab on each line.
As an example, let's consider the util-linux package. It doesn't use a standard autoconf-style configure script, but it can be configured by setting a variable at the top of the MCONFIG file. Thus, the end of our utils/util-linux/Makefile looks like the following:
CONFIGURE_SCRIPTS = $(WORKSRC)/configure BUILD_SCRIPTS = $(WORKSRC)/Makefile INSTALL_SCRIPTS = $(WORKSRC)/Makefile CONFIGURE_ARGS = $(DIRPATHS) pre-configure: echo "DESTDIR=$(prefix)" > $(WORKSRC)/MCONFIG.NEW cat $(WORKSRC)/MCONFIG << $(WORKSRC)/MCONFIG.NEW mv $(WORKSRC)/MCONFIG.NEW $(WORKSRC)/MCONFIG $(MAKECOOKIE) include ../../gar.mk
So before the configure script is run, the package-defined preconfigure rule adds code setting DESTDIR to the $(prefix) variable in MCONFIG.
The $(MAKECOOKIE) variable is a macro that creates a cookie file signifying the completion of the preconfigure step; $(MAKECOOKIE) performs some housekeeping that ensures rules are run only once, so it's important to end each rule with $(MAKECOOKIE).
As another example, the Bourne Again SHell (bash) can be linked to with the name ``sh'' in order to cause it to behave (somewhat) like a POSIX Bourne Shell. To do this, the end of our shells/bash/Makefile looks like:
CONFIGURE_SCRIPTS = $(WORKSRC)/configure BUILD_SCRIPTS = $(WORKSRC)/Makefile INSTALL_SCRIPTS = $(WORKSRC)/Makefile CONFIGURE_ARGS = $(DIRPATHS) post-install: (cd $(bindir); ln -sf bash sh) $(MAKECOOKIE) include ../../gar.mk
Thus creating the symbolic link to sh.
Suppose you want to perform a configure step, but there is no actual program or script associated with that step. For example, you could conceivably want to run a series of shell commands to create a configuration file. Simply pick a file associated with the configure step (or ``none'' if there isn't one) and set your CONFIGURE_SCRIPTS to that. Then write a configure- rule to handle that file (or ``configure-none'' if you used ``none'').
For example, look at the way the traceroute package (Gavron's Traceroute) performs its build step. There was no Makefile, so the maintainer had to provide a method for the compilation commands to be run semi-manually. So the BUILD_SCRIPTS is set to the .c file, and a build- rule for that file is created:
BUILD_SCRIPTS = $(WORKSRC)/traceroute.c build-$(WORKSRC)/traceroute.c: mkdir -p $(COOKIEDIR)/build-$(WORKSRC) $(CC) $(CFLAGS) -o $(WORKSRC)/traceroute \ $(WORKSRC)/traceroute.c -lresolv -lm $(MAKECOOKIE) include ../../gar.mk
This is in fact similar to the way that the default rules are defined in the GAR libraries. For example, here is how the library handles Makefiles as BUILD_SCRIPTS:
build-%/Makefile: mkdir -p $(COOKIEDIR)/build-$* $(BUILD_ENV) $(MAKE) -C $* $(BUILD_ARGS)In make, the % is a wildcard (matching) character when used in the name of a rule, and the $* variable holds whatever it matched. This means that if we told the system to build using work/robotfindskitten-1.0/Makefile, then $* would be set to work/robotfindskitten-1.0 when this rule is run.
The GAR library takes advantage of many of the advanced features of GNU make and is written with the strengths of make in mind. This has been advantageous for a number of reasons. Namely, the size of the GAR library files is a few hundred lines. The BSD Ports libraries are written more like shell scripts and weigh in at several thousand lines.
In addition, the pattern-and-rule handling of package-specific build features fits the realm of make nicely. Also, make has facilities that help avoid redundant behavior. If a package has been installed, running make install will cause it to quickly run through the seven steps, verifying that it had already performed them. To force a package to rebuild from scratch, the make clean target is provided.
GAR provides many features that are not present in most binary package systems, such as dpkg or RPM. This makes user customization of compiles much easier and allows you to rebuild your whole system with whatever optimization flags you choose. The software GAR builds can all be put in /usr/local, for example. The software can also be installed to a scratch directory or secondary volume for later packaging into a complete OS image.
GAR has largely been a product of the LNX-BBC Project, but has also been adopted by many GNOME developers as a means for building the latest CVS build of GNOME from scratch. For more information on the LNX-BBC Project and the GAR system, visit http://lnx-bbc.org/. If you have any specific questions about GAR, feel free to ask on the lnx-bbc-devel list at http://zork.net/mailman/listinfo/lnx-bbc-devel/.
Nick Moffitt is a free software enthusiast in San Francisco, California where he maintains a multiuser community shell server. He is a member of the LNX-BBC Project and maintains GAR, nwall and the popular game of robotfindskitten. When he's not building software projects in make or m4, he prefers the more conventional Scheme and Python languages.