Where did "Aboriginal Linux" come from? Our story so far...
My name is Rob Landley, and I've been working on Aboriginal Linux on and off since the late 90's. It's what got me into BusyBox and uClibc, embedded development, compiler internals, and so on. Now it's where I put together everything else I'm doing (like toybox, tinycc, and the relocatable gcc wrapper) to see what actually works and give it a good stress-test. "Eating your own dogfood", and all that.
The following may not be interesting to anybody but me. (It's as much autobiography as technical history of the project. A big blog entry, really.) But just for the record:
Back in the late 90's, before linksys routers came out, I installed several masquerading gateways by putting Red Hat on old leftover 386 machines. This involved removing as many packages as possible from the base install, both to get the size down (to fit it on old hard drives) and to reduce the security exposure of all the daemons Red Hat ran by default (including a print server and an NFS server exposed to the world, for no readily apparent reason).
Back around Red Hat 6, the smallest base install was still hundreds of megabytes, and needed dozens of packages removed to get a reasonably stripped down system. (You couldn't choose _not_ to install things like ghostscript, or printer support, only remove them after the fact.) Package dependencies often forced me to delete stuff by hand: some packages' uninstall scripts outright failed, others had circular dependencies in long chains through dozens of packages, and there was no distinction between "this package provides optional functionality" and "it won't run without this", a dependency was a dependency as far as RPM was concerned.
Stripping down Linux installs was a time-consuming process that still left behind mountains of junk doing who knows what, which I didn't understand well enough to safely remove.
Stripping down a full distribution seemed like the long way around to get a minimal system. What I wanted was to build _up_ from an empty hard drive, adding only what I needed. I knew how to build packages from source to add them to a working system, but not how to _get_ a working system in the first place. When I went to the the third Atlanta Linux Showcase (in 1999), I pestered everyone I met to tell me how to build a complete Linux system from source code. Lots of people thought it was a great idea, but nobody could point me to the appropriate HOWTO. A few months later, one of the people I'd asked emailed me about the launch of the Linux From Scratch project, and from that I finally learned what I needed to know.
The LFS book linked to the Linux Bootdisk HOWTO, and for a while I got no further than that immensely educational resource. It explained exactly what you needed to copy from an existing root filesystem in order to run an arbitrary app from a bootable floppy. It explained the directory layout, which configuration files were actually necessary and what they did, the early boot process, and introduced me to the ldd command with which I could track down the shared libraries a given executable needed.
Around this time I also encountered tomsrtbt, which used the old "format a 1.44 megabyte floppy to to 1.7 megabytes" trick to fit an enormous amount of Linux system onto a single bootable floppy disk. (This was also my introduction to the BusyBox project, and later to the programming language LUA.)
The above approach of cherry-picking your own boot environment using prebuilt binaries didn't scale very well, and didn't let me mix and match components (such as substituting busybox for Red Hat's command line utilities), so when Linux From Scratch's 3.0 release came out I cleared a month to sit down and properly work through it, understanding what each step was doing. I turned their instructions into a bash script as part of the learning process, because I kept screwing up steps and having to start over, only to typo an earlier step as I repeated it by hand and have to start over _again_. I joined the Automated Linux From Scratch mailing list in hopes I could find (or help create) an official script to use, but they were all talk and no code. (Everybody had their own automation script, but the project wanted to create something big like Gentoo and seemed to think that publishing a simple script implementing the existing Linux From Scratch instructions was beneath them. So everybody had their own script, none of which were "official".)
My own script quickly evolved to remove packages like gettext and tcl/expect, things the masquerading servers I'd been assembling didn't actually need. I poked at adding X11 (something I'd installed myself by hand back under OS/2) and pondered running the system on my laptop someday, but the hundreds of packages I'd need to build and the constant maintenance of keeping it up to date kept that idea way down on my to-do list.
Version 0: The WebOffice version ("Yellowbox")
Towards the end of 2000 I met the founders of a local start-up through the Austin Linux Users Group (two Taiwanese guys named Lan and Lon), and at the end of the year joined their start-up company "WebOffice" as employee #4. The two founders were ex-AMD hardware guys who didn't really program, who had already hired a recently retired professional tennis player to do marketing for them. They had a prototype firewall product they'd demonstrated to get venture capital funding: a small yellow box running stock Red Hat 7. (When they first demonstrated it to me, I diagnosed and removed the "code red" virus from it.)
The money wasn't great, but the project was interesting, challenging, and full of learning opportunities. Full-time Linux positions were still somewhat rare back then, and to make up for the low salary (and the fact they weren't offering stock options; yes I asked, they were saving it for themselves and the VCs), I was promised that I could GPL most of the code I was working on as soon as it shipped. Back in 2000, that sounded like a pretty good deal.
For a the first few months I was their only programmer, doing everything from architecture to implementation (OS, applications, web interface, the works). I became lead developer and architect when they got a second round of VC funding they hired more developers.
Alas, mostly bad ones. The founders didn't know enough about programming to choose wisely, and I wasn't consulted on most hiring decisions because I wasn't "management". Only one actual IDIOT in the bunch, thank goodness, but they made me share an office with him. He was another Taiwanese guy (this one named "Luan") who the founders felt sorry for because his previous dot-com had gone under and he'd be deported if he didn't stay employed to maintain his H1B visa. (Yes, they admitted this to me when I complained about him.) Unfortunately, not only did he not know anything of use to the company, but he never showed any ability to learn, and after the third time "show me how to do this" turned into him handing in my example code verbatim as his work, our working relationship deteriorated somewhat. He literally could not work out how to write a "hello world" program by himself, and when I spent an hour explaining things to him rather than writing example code he could turn in he got frustrated and accused me of being obstructionist because I wouldn't do his job for him. (Of course he had an MCSE.)
And thus began my habit of taking my laptop to places other than my office, so I could get work done without interruption...
There are reasons this company didn't survive to the present day.
WebOffice's proposed product was an early multi-function embedded Linux device. It was a masquerading fireball which provided dhcp and DNS for its private subnet. It also provided a VPN bridging multiple such subnets (possibly from behind other existing firewalls, by bouncing connections off a public "star server"; an idea the founders of WebOffice tried to patent over my objections). It also provided network attached storage (samba) with a web-based user account management GUI. It also provided scheduled actions, such as automated backup. It also acted as a video server. And it did a dozen other things I don't even remember. (Their marketing material called the project the "iLand gateway 2000". I had no say in this.)
I called it "yellowbox" (because it was), and described it as a "swiss army server". The hardware was a standard PC motherboard and an 8 port ethernet switch in a small custom-designed bright yellow metal case with pretty decals on it. Inside was a 266mhz celeron, 128 megs of ram, a 2 gig hard drive, two network cards, and the aforementioned 8 port 100baseT ethernet switch removed from its case and screwed into a metal frame. The back of the box exposed a lot of ethernet ports (one "uplink" port and 8 switch ports, although only 7 of the switch's ports worked because the eighth was soldered to the second internal ethernet card; they labeled it "service" or some such because if the hole in the back of the case didn't let them expose it to the outside world it wouldn't fit right). The only other port was a place to plug in the power cable. The front had many blinky lights (one of which was a blue LED, which they were very proud of and considered a big selling point).
Most importantly, the motherboard's video/keyboard/mouse ports weren't exposed to the outside world: it was supposed to run as a headless box administered through the network via a web server with clever CGI. We could plug a keyboard and monitor into it during development, but only by taking the case off. Out in the field, it had to "just work", and would be a useless brick if it didn't boot all the way into the OS and run our applications.
This was my first exposure to embedded development. The hardware was standard PC/x86, it wasn't too badly underpowered for what it did (at least by the standards of the day), and it used wall current instead of battery power... But it was a headless self-administering box meant to function as an appliance to end users, which was new to me. It was also a challenge to strip down the whole OS into a small enough package that they could download entire new OS images using the internet speeds of 2001, and then update the new OS image without losing data or turning it into a brick.
WebOffice's original prototype device ran a stock Red Hat 7 install (the one that had the Code Red virus when they first demoed it to me after a LUG meeting). The whole OS image took up almost a gigabyte, and that's before they'd implemented any applications or web UI. I rebased the system on Linux From Scratch, using my LFS 3.0 script to build the base OS and creating a new script to build the additional packages (apache, postscript, ssh, and so on) the project used. I got the OS down under 100 megs (but not by much, it still used glibc and gnu coreutils and so on). I then spent the next year and a half learning how to properly strip down and secure an embedded system. I brushed against both busybox and uClibc during this period, but couldn't get either one to work in our project at the time. We needed more functionality than either provided back then.
I implemented all the web CGI stuff in Python; a part-time web designer would come in once a week to mock up pages using Dreamweaver, and I'd take the result and make my Python code spit heavily cleaned up versions, plus actual content and minus most of the and similar lunacy. Getting the stylesheets to work was interesting. (Working around the way Internet Explorer treated the end-form tag as a break tag and inserted extra vertical whitespace that didn't show up in Netscape or Konqueror was also fun, although it _didn't_ do this if your start form tag and end form tags were at different table levels. Yes, to make it display right I had to make tags cross, so IE didn't think it understood the data and thus get confused and do the wrong thing. I'm not proud of this, but it was IE.)
I learned how to configure and administer (and automate the administration of) apache, samba, postfix, ssh, bind, dhcpd... I created a scalable vpn (which freeswan _wasn't_, nor was the out-of-tree patch of the day remotely reliable) by combining iptables port forwarding with ssh and a wrapper daemon. (Again the founders tried to patent this; I objected strenuously that it was A) obvious, B) they'd said I could GPL it when it shipped. This went on for a while).
I also made an automated production process for WebOffice: my scripts built a CD-rom image which, when booted (with the case off there was a spare IDE port you could hook a cd-rom drive to), would partition and format /dev/hda and install the final OS image on it, eject the CD, play "charge" through the PC speaker, and power down the machine. (If something went wrong, it played "taps" instead.) Yes, these CDs were dangerous things to leave lying around, and I made sure to label 'em as such.
WebOffice wanted to be able to remotely upgrade the firmware, which meant sending a new OS image as a single file. The install had to be fairly atomic, if something went wrong during the upgrade (including a power failure, including the user switching it off because it was taking too long) the thing could easily become a brick. Obviously a traditional "extract tarball into partition" approach was unacceptable, even before "fsck" issues came up. (The only journaling filesystem in the stock kernel at the time was reiserfs, and that was way too fiddly and overcomplicated for me to trust my data to it. I moved the data partition to ext3 when that got merged, but wanted to make the base OS partition read-only for security reasons.)
I wound up creating a gpg-signed tarball with several files, one of which was the new kernel to boot, one of which was the initrd (remember: this was back before initramfs), and one of which was a filesystem image to read-only loopback mount as the new root filesystem. (For security reasons I wanted root mounted read only, which also suggested a compressed filesystem to save space. Squashfs didn't exist yet and the ext2 compression patches had already bit-rotted, so I used zisofs.) The tarball also contained a file with a version string, and a file with an sha1sum of the concatenation of the other four files.
Extracting a firmware tarball wrote these files into a new subdirectory (The tar invocation extracted those specific names, so an attacker couldn't write to arbitrary locations in the filesystem with a carefully crafted tarball; yes I was paranoid while learning about security), and made use of the "lilo -R" option to switch to the new firmware. That sets the LILO command line for the next boot only, so we left the default pointing to the old firmware but told LILO that on the next boot it should use the new firmware. If the new firmware came up and its self-diagnostic checked out, it would change the LILO default. If it didn't work, power cycle the box and the old firmware would come up. (This greatly reduced the chances of turning the headless box into a brick, and you couldn't do that with grub.)
At a technical level, there was a chicken and egg problem here: the root filesystem was a loopback mount, but the file to loopback mount has to live somewhere. So the system needed a writeable partition for logging and such anyway, so I made /dev/hda1 be ext3 and mounted it on /var, and put the firmware in that. So during the boot process the initrd needed to mount /dev/hda1 onto a /temp directory, loopback mount the /temp/image file onto a /sub directory, and before doing the pivot_root into /sub it needed to move the /temp mount into /sub/var. This turned out to be nontrivial.
Back under the 2.2 kernel you couldn't mount a partition in two places at once, so mounting the same /dev/hda1 on both /tmp and /sub/var wasn't an option. I had to use early (and buggy) 2.4 kernels to have any chance to make this work (and also to make the VPN work, which required the undocumented SO_ORIGINAL_DST getsockopt() existing in 2.4 but not 2.2).
The early 2.4 kernels sucked mightily. The memory management problems that resulted in the rik->andrea switch in 2.4.10 hit the yellowbox project kind of hard. I once drove the 2.4.7 kernel into a swap thrashing state, went to lunch (instead of rebooting, just to see if it would recover), and it was still swap thrashing and mostly paralyzed when I came back over an hour later. The disk cache (especially the dentry cache) could get unbalanced until it grew to evict all the anonymous pages and froze the system hard. (A big rsync would do that fairly reliably. Trying to avoid this I studied the md4 algorithm and the rsync description file and spent a week writing most of my own rsync implementation in python, but A) it maxed out at about 300k/second on the processor we were using, B) it also caused the hang because it was really a kernel issue and not an application issue.) It was frustrating, but we persevered.
Mounting a partition twice and leaking one of the mount points (the old /temp was inaccessible after the pivot_root) was kind of unclean anyway, the clean thing for the boot to do was actually move the /tmp mount to /sub/var after mounting /sub but before the pivot_root into /sub. But when I asked on linux-kernel how to do that, I was told that "mount --move" didn't exist yet. A couple releases later Al Viro added it, and I was one of the first happy users.
I also wanted to put the kernel, initrd, and loopback mountable root filesystem image together into a single file, so we didn't have to extract a tarball during a firmware upgrade but could actually _boot_ into the actual file we'd downloaded, after verifying its signature. (This avoided the problem of successfully downloading the tarball but not having enough space left to extract it. Since zisofs, zImage, and initrd were already gzipped, compressing the firmware image for transport wasn't a priority. Keep in mind: headless box, self-administering. Even little things like this could turn into a big problem in the field if you didn't handle them.)
You could already use "losetup -o" to loopback mount a file at an offset, and I made a "length" patch to LILO that let its config file tell it to boot only the _start_ of the kernel file you fed it. But dealing with the initrd in between was a pain, which is why I eventually became an early avid follower of initramfs, and wound up writing documentation for it when I couldn't find any and had to answer so many questions myself.
The end at WebOffice
The original promise that I could GPL the code I was working on (everything except the python CGI) once it shipped never came true. Partly the founders were ambivalent about this whole "open source" thing, wanting every competitive advantage they could get. (They kept trying to patent obvious things I did. Their patent lawyer was a really cool dude when he flew in from California.)
Another contributing factor was that the founders were from Taiwan and had no idea how to address the US market. Their marketer employee #3 hadn't stayed very long (not much endorsement value for a tennis player trying to sell servers), and they themselves only ever tried to sell the device overseas (which made demonstrating the thing somewhat difficult, and this also meant they were shipping a VPN with cryptographic checks on firmware upgrades to places like Turkey, back in the days of cryptographic export regulations).
But the biggest problem was unending feature creep: every time the founders saw or heard of a product that did something, we had to do that too. I had a shippable product ready a few months after I started, but they wouldn't ship it. I designed the firmware upgrade mechanism so we could ship what we had and add more later, but they felt that doing so would take focus away from developing more features. (For about a while there they were trying to turn it into a video server. I made a python CGI script for apache to cache large files, by downloading them from an upstream server and sending them out as they came in as if it had been a local file all along, while simultaneously writing them to the hard drive for other users. Of course, they tried to patent this too...)
The tendency towards feature creep left them vulnerable to their venture capitalist changing their business model. Another of the VC's start-ups was paying lots of money to license the RealVideo streaming server, so the VC convinced WebOffice to waste six months trying to reverse engineer it. (After all, our idea of offering mp4 files though Samba or Apache made us a video server, right? This was just another kind of video server...) I wasn't interested in this direction and left Austin for a while to spend time with my mother (who was suffering from cancer and New Jersey) while they got this out of their system. They hired over a half-dozen programmers to replace me during this period, but progress on the yellow box ground to a halt anyway (and even went backwards a bit with numerous regressions) until I came back. The quality of the new hires varied ("erratic", "mediocre", and "suck" were all represented).
WebOffice ballooned to a dozen employees (over half of whom reported to me when I came back, although I still had little say in hire/fire decisions). The company bought itself back from the first VC by mortgaging itself to a second VC, and refocused on the original do-everything "swiss army server" idea. But they still wouldn't just ship what they had as long as there were more features we could add, and ultimately they burned through their venture capital without ever sending more than a few prototypes to actual customers.
WebOffice ran out of money in 2002, and instituted a round of layoffs. I continued on half-time (at half-pay) for several more months, hoping that necessity would make them focus on shipping units and bringing in revenue, but it didn't happen. I left in November and spent the last couple months of that year in Florida watching my mother die of cancer, then driving around the country distributing her possessions to various relatives, and finally crashing on Eric Raymond's couch for a few months doing an "editing pass" on The Art of Unix Programming that expanded the book from 9 chapters to 20.
Version 1: Relaunch based on BusyBox and uClibc ("Firmware Linux")
The code discussed here is still online (if abandoned since 2005).
When I returned to Austin in August 2003, I bought a condo near the University of Texas (and near Metro, my favorite 24 hour coffee shop with wireless internet access), enrolled in grad school, and got back into poking at Linux From Scratch.
Linux From Scratch had reorganized itself. My old weboffice scripts had been based on LFS 3, which involved building enough of a system to chroot into and complete the build under that. The potential downside was that bits of the host system could leak into the final target system, such as copied headers or tools build by the host's compiler.
In 2002 LFS 4 introduced an intermediate set of statically linked tools in a "static" directory, which were deleted after the intermediate system was built. In November 2003 LFS 5 renamed this temporary directory to "tools". This new approach added the temporary directory to the end of the $PATH during the chroot, rebuilt itself using the temporary system, and then discarded the entire directory to eliminate leaks of host files. This was a big enough change that it was less work to start over from scratch than try to adapt my existing scripts.
Starting over also seemed like a good idea because I was unsure of the IP status of my old scripts. Although I'd been promised repeatedly I could GPL everything but the python CGI when the yellowbox shipped, actual shipping had never quite happened, and I didn't have that promise in writing. (I don't remember if I lost it or if I'd been without a contract all along. You could make an argument I owned all the code I'd done outright in the second case, certainly that's what the copyright notices on the individual files said, and I'd been working on early versions of those scripts before I brought them to weboffice in the first place and had never signed over those preexisting copyrights. But I just didn't want to go there.)
I also wanted to take the project in new directions, further into the embedded space. WebOffice had focused on adding more and more features to a bigger and bigger image, while I personally had focused on trimming it down and streamlining it (for example replacing the Postgresql database with a few flat text files to store configuration and user information, thus replacing 200 megabytes of disk usage with about 90k and speeding up the relevant code considerably).
For the new project I had two main goals: make the bootable single file idea work, and make the result much smaller and simpler. (I also wanted to clean up the build so it didn't require root access, package and document it all so anyone could use it, other similar tidying steps.)
The firmware tarball I'd implemented for WebOffice had always been a stopgap, something they could ship with quickly while I got a better solution ready. What I really wanted was a single bootable file containing kernel, initial ram disk, and root filesystem all in one. (Putting an entire large root filesystem into a ramdisk consumed too much memory, the root filesystem needed a backing store it could page files in from.)
The name Firmware Linux came from the goal of packaging an entire OS image in a single bootable file, which could run directly and be used to atomically upgrade embedded systems.
My other goal for Firmware Linux started with the desire to replace as much of the gnu tools as possible with something smaller and simpler. The old yellowbox images from WebOffice had weighed in at almost 100 megabytes, most of which was glibc, coreutils, diffutils, and so on. This was clearly crazy, my first hard drive back in 1990 was only 120 megabytes, and back under DOS that was enormous (and a huge step up from my friend Chip's system with a 32 megabyte hard drive, which I learned to program C on). When I looked at the gnu implementation of the "cat" command and found out its source file was 833 lines of C code (just to implement _cat_), I decided the FSF sucked at this whole "software" thing. (Ok, I discovered that reading the gcc source at Rutgers back in 1993, but at the time I thought only GCC was a horrible bloated mass of conflicting #ifdefs, not everything the FSF had ever touched. Back then I didn't know that the "Cathedral" in the original Cathedral and the Bazaar paper was specifically referring to the GNU project.)
Searching for alternatives, I went back to take a closer look at busybox and uClibc. I was familiar with both from Tom's Root Boot (tomsrtbt), a popular single floppy Linux system that packed an amazing amount of functionality into a single specially formatted (1.7 megabyte) 3.5" floppy disk. I'd been using tomsrtbt for years, I just hadn't tried to build anything like it myself. Compared to the tens of megabytes of gnu bloat the LFS project produced, busybox and uClibc seemed worth a look.
This old message was my first attempt at sniffing around at uClibc. I didn't get time to seriously play with it (or BusyBox) until much later.
It also occurred to me that if the newly introduced /tools directory was enough to build the final system, then all I needed for the system to be self-hosting was enough extra packages to rebuild /tools. If the prehistory stage had been about starting from a full distro and cutting it down, and the WebOffice version had been about starting from ground zero and piling up lots of functionality into a 100 megabyte tarball, this new stage was about starting from an empty directory and adding as little as possible to do what I wanted while staying small and simple.
So the real questions were:
I started by writing new scripts based on Linux From Scratch 4 (quickly switching to LFS 5) to build a stock LFS system. I wrote a script to build /tools, and another script run under a chroot to build a final LFS system within tools. The second script acted as a test that the /tools created by the first script was good enough. And once I had a known working system, I started doing a number of different things to it.
Stripping down LFS 5.0
The full list of Linux From Scratch 5.0 packages were: autoconf, automake, bash, binutils, bison, bzip2, coreutils, dejagnu, diffutils, e2fsprogs, ed, expect, file, findutils, flex, gawk, several fragments of gcc, gettext, glibc, grep, groff, grub, gzip, inetutils, kbd, less, libtool, the linux kernel, m4, make, MAKEDEV, man, man-pages, modutils, ncurses, net-tools, patch, perl, procinfo, procps, psmisc, sed, shadow, sysklogd, sysvinit, tar, tcl, texinfo, util-linux, vim, and zlib. There were also two LFS-specific packages, providing boot scripts, config files, and miscellaneous utilities.
I started by removing packages I didn't actually need. Tcl, expect, and dejagnu hadn't been in LFS 4, so obviously it was possible to do without them. (I was already starting to view newer versions of Linux From Scratch as "bloated" compared to old versions. I could always build and run test suites later, and rebuilding the system under itself to produce a working result was already a fairly extensive test.)
I could also eliminate ed (which patch can use for obsolete patch formats, but who cares?), gettext (only needed for internationalization, which is best done at the X11 level and not at the command line), libtool (which is a NOP on ELF Linux systems and always has been, blame the FSF for trying to get us to use it at all), and man (and man-pages, groff, and texinfo, which are used to build/display documentation).
A bunch of development tools (autoconf, automake, binutils, bison, flex, gcc, make, and m4) wouldn't be needed on a stripped down system (such as a router) that never needed to compile anything. (Perl might be in this group as well, since it was only included because glibc needed it to build. The linux kernel and glibc both supplied files used by the compiler, such as the headers in /usr/include, so this group depended on them even if they had other more direct uses.) Similarly, the e2fsprogs package was used to create a filesystem, but mkisofs and such could substitute for it.
The kernel and grub were basic infrastructure, not really part of the root filesystem and easy to build separately. (I was still using my modified LILO anyway.) The C library (glibc) was the next layer past that, every userspace program had to link against it either statically or dynamically. The boot scripts, MAKEDEV, sysvinit, and modutils were all similarly low-level infrastructure pieces to boot the system or talk to hardware. The shadow package provided login and /etc/passwd support. The ncurses and zlib packages were shared libraries I understood, but were both largely optional (and gzip/zlib seemed somehow redundant). Bash was a command shell, bzip2 and gzip were compression programs, tar an archiver, vim a text editor, and sysklogd a logging daemon that wrote stuff to /var/messages.
That left coreutils, diffutils, file, findutils, gawk, grep, inetutils, kbd, less, net-tools, patch, procinfo, procps, psmisc, sed, and util-linux as "other stuff in the $PATH" which were only really needed if some application (such as a package build) used them. After enough study, I felt comfortable I understood what they all did.
That's what chapter 6, which built the final Linux From Scratch system, contained. Chapter 5 had a much shorter list: binutils, gcc, linux (used just for headers), glibc, tcl, expect, dejagnu, gawk, coreutils, bzip2, gzip, diffutils, findutils, make, grep, sed, gettext, ncurses, patch, tar, texinfo, bash, util-linux, and perl. And chapter 5 _had_ to contain enough to build chapter 6, and thus rebuild the entire system from source.
Again, tcl, expect, dejagnu, gettext, and texinfo could be discarded. (Most of those weren't even present in the earlier versions of Linux From Scratch I'd used, they had to be optional.) That left just 19 packages. The compiler toolchain was just binutils, gcc, make, glibc, and the Linux headers (all that autoconf, automake, lex, and bison stuff was obviously optional and could be added later from within a working system). Perl was only used to build glibc, if that was replaced or fixed then the need for perl (at least at this stage) could go away. Busybox claimed to provide replacements for gawk, coreutils, bzip2, gzip, findutils, grep, sed, tar, bash, and util-linux. Since busybox didn't use ncurses, it should be possible to build that at the start of chapter 6. And what was diffutils doing here at all? It turns out that the perl configure stage uses "cmp" (which it provides), so if you didn't need perl you didn't need this.
Since Linux From Scratch's "chapter 6" started by rebuilding binutils and gcc (which were the big, complicated, tough packages), those obviously didn't need any more than was in chapter 5 to rebuild themselves.
All this analysis reduced Linux From Scratch's chapter 5 to four functional groups:
Replacing packages with BusyBox and uClibc
Once I ran out of obvious packages to remove, I experimented with package substitutions, swapping out the stock Linux From Scratch packages for other (smaller) implementations of the same functionality. The two obvious goals (again, pursued in parallel) were to swap glibc for uClibc, and to use busybox in place of as many other commands it could replace.
In theory, a self-hosting LFS chapter 5 root filesystem that could rebuild itself directly from source could be reduced to binutils, gcc, make, uClibc, linux-headers, and an _extensively_ upgraded busybox. (Of course such a modified chapter 5 should still be able to build the unmodified chapter 6. If it couldn't, there was something wrong with it, so that was a good test.)
Both BusyBox and uClibc were maintained by a guy named Erik Andersen, who had started them while working for a company called Lineo and continued them after he left (a little like the way I was continuing Firmware Linux). In both cases he'd found a long-stalled existing project to salvage and relaunch instead of starting from scratch, but in reality he'd taken dead projects, replaced all their existing code, and built a community around them.
Busybox was nice because I could introduce it piecemeal. I could replace commands one at a time, swap an existing /tools/bin binary with its busybox equivalent and run the build to see if it worked. If it didn't, I could compare the two versions of the build against each other to see what had changed, or try to replace a different (simpler) command.
The Linux From Scratch installation instructions also listed the files installed by each package, so I could look through the lists ( sed had just one, gzip installed a little over a dozen, http://archive.linuxfromscratch.org/lfs-museum/5.0/LFS-BOOK-5.0-HTML/chapter06/util-linux.html>util-linux installed over 60) to see what was actually needed ("sed" yes, "cal" not so much) and what busybox did and didn't provide already and what would need to be added or upgraded.
I focused on eliminating packages, which meant I started by tackling fairly complicated commands like "bunzip" and "sed", because getting those to work would let me drop an entire package. I quickly sent in so many bugfixes to sed I wound up maintaining the applet, and got distracted rewriting bunzip entirely (but my new implementation compiled to only 7k).
Eventually, I wound up getting busybox to replace bzip2, coreutils, e2fsprogs, file, findutils, gawk, grep, inetutils, less, modutils, net-tools, patch, procps, sed, shadow, sysklogd, sysvinit, tar, util-linux, and vim.
To do that, I wound up extensively upgrading (or rewriting from scratch) dozens of different busybox commands and adding several new ones from scratch.
One new command was switch_root, for initramfs support. I wrote an initramfs file for the kernel's Documentation directory because I investigated it for Firmware Linux. (And later gave my first OLS presentation on the topic, and write an article series about it when I worked at TimeSys.)
Another new command was mdev, which was a rewrite of a shell script I used to populate /dev, which Frank Sorenson ported to C and I extended (adding a config file based on irc conversations with Gentoo's Solar).
Replacing glibc with uClibc took some doing, but at the time the uClibc project was quite heavily developed and rapidly improving (coming out with 8 releases in 2002 and 8 more in 2003) so there was always something new to try. If something didn't work, they were happy to fix it.
uClibc version 0.9.26 (January 2004) was the breakthrough version that went from "here are the packages known to work, anything else probably won't" to "any package you try to build against this will most likely work, please let us know about anything that doesn't". When uClibc did finally work, it allowed me to remove perl from /tools (which was only need to build glibc, but not required by anything else in LFS).
I also experimented with dynamically linking /tools, as another way to get the size down. Linux From Scratch statically linked chapter 5 for simplicity sake, I tried to get the existing compiler to link against the C library I just built. This was quite a learning experience. Everything from changing the library loader path to making sure gcc could find crt0.o at a nonstandard location was all new, and fiddly, and cryptic, and didn't work. And thus began the long war between me and gcc's path logic. (Since I had static linking to fall back on, I could poke at this in parallel with my other work on the project, and didn't get it to actually _work_ for quite some time.)
At the time, programs were normally built against uClibc by using a wrapper around gcc that rewrote its command line arguments to link against a different C library. Of course I took the wrapper apart to see how it worked and how to make gcc link against uClibc without it. What I wanted was a compiler that naturally linked against uClibc, not an existing glibc compiler repurposed to do so.
Based on what the wrapper was doing and a lot of tweaking and questions on the mailing list (which Erik graciously answered), I got it working around the middle of 2003.
User Mode Linux
The Linux From Scratch build assumed you had root access, in order to mknod devices, chroot into the temporary system (chapter 5) directory to build the final system (the chapter 6 packages), and to loopback mount files to create system images.
Asking people to download random code and run it as root seemed kind of impolite at best, and actively dangerous at worst. (When buildroot first came out I ran "make uninstall" in it. The resulting host/target confusion it suffered deleted things like gzip off my host. Back then I was still using a Red Hat system which meant "pam", and when the security theatre modules suffered an auto-immune response to my attempts to patch the system back together with busybox, I had to reinstall the OS in order to be able to launch X11 again.)
Since the end result of my system builds was just a file (a tarball or a filesystem image), there was no real excuse for requiring root access. The packages built as a normal user already, in theory that was the hard part.
My solution was User Mode Linux. (Of course I wrote a quick User Mode Linux HOWTO containing everything I needed to know to do what I was doing with it.)
User Mode Linux was an early virtual machine, which could give me simulated root access (enough for my needs), but without extensive setup thanks to the "hostfs" (a bit like qemu's virtfs), and without requiring a new source package (I already had the Linux kernel sources, this was just another way of building them).
I first got UML working in a patched 2.6.9 kernel, and later integrated it into the build when I got unpatched 2.6.11 to build a usable User Mode Linux image (although I had to patch it a bit myself later). I could then use that to chroot (via UML's "hostfs") and loopback mount as a normal user, and then mknod within that loopback mount, and run the chapter 5 environment within UML to build chapter 6. Of course I wrote a quick User Mode Linux HOWTO.
Using UML was optional, and the scripts autodetected if you were running as root and would chroot directly instead of firing up what amounted to an emulator, but running as root was not recommended.
Why 2006 was a lost year
The rise of buildroot
Shortly after I figured out how the uClibc wrapper (and gcc in general) worked under the covers, the uClibc developers abandoned the wrapper in favor of a new project called "buildroot".
The buildroot project was the response to fresh gcc bloat: around gcc 3.0 a new shared library called libgcc_s.so showed up, more or less a dynamic version of the old libgcc.a. It contained various gcc internal functions (such as 64 bit division on 32 bit platforms), which most nontrivial programs wound up linking against. Unfortunately, shared libraries can link against other shared libraries, and libgcc_s.so linked against libc.so.6. So any program that linked against this library snuck in a reference to glibc and wouldn't load without it, even if it was otherwise linked against uClibc.
This meant the gcc wrapper was no longer sufficient, but since libgcc_s.so was part of gcc, the only way to get a new version of libgcc_s.so that linked against uClibc (instead of glibc) was to download the gcc source code and build gcc itself against uClibc.
And that's exactly what buildroot did: built uClibc and a new compiler out of the four source packages uClibc, binutils, gcc, and the linux kernel for the kernel headers, hooking the whole mess together with several ./configure options and source patches. Then it used the new uClibc compiler to build Erik's other project, BusyBox, as a test case that it was all working correctly.
In order to test both static and dynamic linking, buildroot created a new uClibc root filesystem directory containing the uClibc shared libraries, and a second instance of busybox dynamically linked against those, which you could chroot and test out the uClibc version of busybox. Since he already had the kernel source lying around, Eric even taught it to build a User Mode Linux binary that could do the chroot for you.
Problems with buildroot
Of course I took buildroot apart to see how it worked, wrote the first buildroot HOWTO (because I tend to document everything I didn't initially understand), made a number of design suggestions, and even offered patches.
But I didn't really like the design of buildroot (nested makefiles aren't the most straightforward approach to anything, its need to run as root meant that early versions ate my laptop (twice, contributing strongly to my conviction that builds should never run as root), and it couldn't rebuild itself under itself, and in general was a constantly broken moving target with no stable releases.
So I continued to work on my existing build project instead (which was several years old by that point). Buildroot was an instructive counterexample on many fronts: my project should not require root access to build, shell scripts were more readable than nested makefiles, releases were important, and it's vital to have boundaries so you can clearly state what your project DOESN'T do.
I also checked in with buildroot from time to time to see what it was doing (several patches to make various packages work with uClibc were only ever documented by being checked into the buildroot repository, and then the uClibc developers acted shocked I hadn't heard of them).
The biggest problem with buildroot was the impact it had on BusyBox and uClibc development. Although buildroot started out as just a test harness for uClibc and busybox, it quickly grew out of hand. Since the easy way to regression test that a package worked against uClibc was to add it to the test suite and build it as part of the test root filesystem, lots of packages got added fairly quickly. People packaged up and shipped the root filesystem created by buildroot when they wanted a simple uClibc+busybox root filesystem, and then complained when it didn't build some package they needed.
Within a few months, buildroot had exploded from a simple test harness into a half-assed Linux distribution. Erik tried to avoid this (he'd previously built a uClibc-native version of Debian Woody and knew perfectly well what a real distro looked like), but buildroot turned into a distro anyway because the project had no clear boundaries that allowed him to say "no, this new feature is not within the project's scope". He never drew a line in the sand that allowed him to say "no", and thus over time steady feature creep buried him.
As a distro, buildroot was a deeply flawed. It had no package management tools (such as rpm or deb or portage), nor did it have the tens of thousands of packages build descriptions in the large and carefully maintained repository of Red Hat, Ubuntu, Gentoo, or even Slackware. For the project's first five years, buildroot never even had a release, instead insisting users grab a random source control snapshot du jour and hope for the best. Despite this, buildroot was the focus of the development efforts of the BusyBox and uClibc communities, and became their standard repository of knowledge about how to build packages for all sorts of embedded environments.
Buildroot derails uClibc and BusyBox development
With no clear dividing line between "how to build" and "what to build", buildroot's scope and complexity exploded, and despite its limitations as a distro what buildroot could do was suck away an endless amount of development time from the BusyBox and uClibc developers. By 2005, both uClibc and BusyBox development were clearly suffering.
Erik started by abandoning busybox after the 1.0 release, both because 1.0 seemed like a good stopping point (since it was "done" now), and because he just didn't have time for it anymore. Other developers (including myself) still did new development, found bugs and fixed them, but there was no prospect of a new release.
Over on the uClibc side of things, Erik held on longer but the release frequency slowed, from seven releases in 2003 (0.9.17 through 0.9.24) to two in 2004, one in 2005, and none at all in the whole year 2006.
The uClibc 0.9.26 release (in January 2004) was the point at which Erik stopped maintaining a supported application list, because most things just worked now. That was the point at which uClibc became generally useful, and could have been the basis for a 1.0 release similar to BusyBox's. But due to buildroot diverting the development community, a uClibc 1.0 release became a more and more distant possibility as development lost focus: by the end of the decade, uClibc still had no clear plan to produce a 1.0 release.
BusyBox takes over my life
By 2005, Firmware Linux built a system that could rebuild itself, but it still used a lot of gnu packages. I continued to replace more of these executables with busybox, making BusyBox supply more of my development environment and submitting numerous patches, everything from minor bugfixes and complete ground-up applet rewrites. I also pestered other people (such as the awk maintainer) into repeatedly fixing their parts of the code when some package build failed and I narrowed it down to a reproducible test case.At this point, Firmware Linux wasn't held back by my build infrastructure (which was just a pile of shell scripts anyway), but by deficiencies in BusyBox. I became one of the most active BusyBox developers, using Firmware Linux as a test environment for my busybox changes, and gradually began spending more time working on BusyBox than the rest of Firmware Linux combined.
After the BusyBox 1.0 release gave Erik an excuse to step back, I continued to work intensely on the project, and the bugfixes I needed (and that other developers supplied) kept accumulating. I eventually collected enough together to make a bugfix-only release, which became the official 1.01 when Erik approved it. And it turns out "he who cuts releases is the maintainer". Erik officially handed over maintainership a few months later, when I cut my first new development release (busybox 1.1.0).
Tinycc, QEMU, and TimeSys.
I was introduced to Fabrice Bellard's "tinycc" by the October 27, 2004 slashdot story about tccboot, an ISO image that booted Linux _from_source_code_, booting into a compiler that compiled the linux kernel and a simple userspace in a matter of seconds, and then ran the resulting kernel.
Afterwards I kept track of tinycc with an eye towards replacing gcc and binutils, thus producing a gnu-free development environment. (And then I'd ask Richard Stallman if a system without a line of gnu code anywhere in it was still Gnu/Linux/Dammit and mock him when he said yes.)
Tinycc was a small and simple C compiler, fitting a complete combined compiler and linker into a single 100k executable. It compiled code so quickly that its most common use was turning C into a scripting language: by starting C source files with "#!/usr/bin/tinycc -run" and setting the executable bit on the source file, tcc could compile and launch it in a small fraction of a second.
When I encountered it, tinycc couldn't build an unmodified Linux kernel (the tccboot kernel was a hacked up subset of Linux 2.4), but was only the third compiler ever (after gcc and Intel's x86-only closed source icc) that had ever built a working Linux kernel, and the tinycc developers were working towards full C99 compliance and the ability to build Linux 2.6.
Alas Tinycc got derailed the same way uClibc did: it spawned a side project that sucked its developers away. In this case, the side project was the emulator QEMU. Inspired by the speed of tinycc, Fabrice came up with another compiler that took pages of foriegn binary code as its input, translating them on the fly a page at a time into equivalent binary code the current machine could run. As he explained in the QEMU 0.1 announcement, the purpose was to allow Wine to run on non-x86 machines, but QEMU quickly turned into a very fast general purpose emulator.
Towards the end of 2005 I started playing with QEMU, as a potential replacement for User Mode Linux. When I got it to work I decided that Firmware Linux should support every target QEMU did, meaning I had to learn cross compiling.
When I mentioned this on the #uClibc channel on Freenode, one of the regulars asked if I wanted to come work at his company, TimeSys.
My time at TimeSys
TimeSys did embedded development, via extensive cross compiling. They had a couple dozen extremely talented engineers and masses of accumulated expertise. It was such an interesting job and a great group of guys I was willing to move to Pittsburgh (with my fiancee). I started there January 15, 2006.
At TimeSys I worked on BusyBox (having become the project's official maintainer with the 1.1.0 release). I also poked at uClibc a bit and encouraged its development however I could. (There was another cake for the next release, that one made and delivered by Howard Tayler of Schlock Mercenary. I asked him on the theory that if he's truly mercenary, and lives in the same town as Erik, I just have to pay him enough. I note his price has probably gone up since then.)
But the reason I took the job was to learn stuff I didn't already know. There were a bunch of guys doing cross compiling, and I learned everything I could from them. What were the various hardware targets, and what was each good for? How do you make a cross compiler, and how do you use it? I spent the whole of 2006 learning stuff from a bunch of great guys. (I've never worked with a better team of engineers than TimeSys had in mid-2006.)
Unfortunately, it couldn't last. Shortly before I arrived at TimeSys they completed a cross compiling build system (based on "TSRPM", a giant Rube Goldberg set of Perl wrappers and LD_PRELOAD library intercepts that cross compiled RPM source packages), and used it to cross compile the whole of Fedora Core 2 to several different hardware platforms. This was a popular product that sold well, and the CEO of the company decided that engineering had completed its task and the company's emphasis would now shift to marketing. This decision destroyed the engineering department.
Engineering's task was NOT done. New Fedora releases came out every 6 months, and Red Hat only supported 3 releases back (18 months). The TSRPM build system was extremely brittle and only ran on a specific set of carefully crafted servers in our build room: machines running 32-bit Red Hat 9 (already obsolete even then). Timesys was selling other companies access to these machines ("TimeSys LinuxLink"), so their engineers could log into them through the network and run their builds on them. This didn't scale.
Nevertheless, the CEO froze the engineering budget transferred the resources to sales and marketing. Laurie, the head of marketing, took this as her cue to try to take over the company, diverting engineers to report to her (starting with the webmaster, who quit, and then the system administrator, who quit...)
We discovered an interesting Corrolary to Brooks' Law (adding more manpower to a late project makes it later, because your existing productive people get distracted training the new ones). The corollary is that shrinking teams get paralyzed by knowledge transfers: everybody spends all their time trying to learn what the departing engineers know, as you try to preserve a minimal level of necessary expertise within the team.
Through heroic effort the understaffed engineers managed to rebase LinuxLink to Fedora Core 5 just as Fedora Core 2 was end-of-lifed by Red Hat (since everything was cross compiled, this was an enormous undertaking requiring almost as much work as porting Fedora Core 2 had in the first place). And this burned them all out. A few people had quit during this: the ones transferred to report to Marketing, and a few instances of natural attrition as their friends launched start-ups they wanted in on. But those people weren't replaced. The engineers realized that senior management did not value them, and that they'd never get the resources to turn their prototype build farm into a real sustainable production system, not when those dollars could go to cold-call salesman selling shares in the prototype.
Timesys had been expanding engineering when I came onboard, but a few months later it had a reorg where the boss who'd hired me (Manas Saksena) was "promoted into a closet" with nobody reporting to him anymore. He quit and went to work at Red Hat, where he launched the Fedora for Arm project (thus eliminating a significant part of TimeSys' customer base for their ported Fedora versions). His successor (and the last new hire before the freeze, when Manas was attempting to expand engineering to keep up with the success of the LinuxLink subscriptions) was David Mandala. David struggled mightily to keep the department together and shield us from the insanity of the CEO and marketing head, but he eventually confronted the CEO with a "do this or the company is doomed" presentation that the CEO didn't even listen to the first page of. David resigned and went to work for Ubuntu, where he became the head of Ubuntu Mobile and Embedded (and later Ubuntu for Arm).
David's resignation was like a bomb going off in engineering, and everybody started sending out their resumes. Perhaps a fifth of engineering had already left (and not been replaced), but now the floodgates opened and the team of two dozen engineers I'd been hired into, with a sattelite office in california and telecommuters in New England and Germany, rapidly collapsed down to a dozen, then a half-dozen engineers.
We stopped doing knowledge transfers, and began speculating about a replacement build system given the now absolute necessity of throwing the old one out, since we could no longer reproduce it, let alone maintain it.
The design of Aboriginal LInux
I returned to my Firmware Linux
UNFINISHED AFTER THIS POINT. -------- Buildroot traffic slowly strangled uClibc development discussion on the uClibc list until I gave up and created a new list and politely kicked the buildroot traffic over there. Cake Me suggesting new buildroot list: http://lists.uclibc.org/pipermail/uclibc/2003-November/028342.html Instead buildroot set off to reinvent the wheel, maintaining their repository in Debian's repository contained over 45,000 packages
|Copyright 2002, 2011 Rob Landley <firstname.lastname@example.org>|