Cross-compile just enough to get a native compiler for the new environment, and then emulate the new environment with QEMU to build the final system natively.
The intermediate system is built and run using only the following eight packages:
What we want to do is build a minimal intermediate system with just enough packages to be able to compile stuff, chroot into that, and build the final system from there. This isolates the host from the target, which means you should be able to build under a wide variety of distributions. It also means the final system is built with a known set of tools, so you get a consistent result.
A minimal build environment consists of a C library, a compiler, and BusyBox. So in theory you just need three packages:
Unfortunately, that doesn't work yet.
Environmental dependencies are things that need to be installed before you can build or run a given package. Lots of packages depend on things like zlib, SDL, texinfo, and all sorts of other strange things. (The GnuCash project stalled years ago after it released a version with so many environmental dependencies it was impossible to build or install. Environmental dependencies have a complexity cost, and are thus something to be minimized.)
A good build system will scan its environment to figure out what it has available, and disable functionality that depends on stuff that isn't available. (This is generally done with autoconf, which is disgusting but suffers from a lack of alternatives.) That way, the complexity cost is optional: you can build a minimal version of the package if that's all you need.
A really good build system can be told that the environment it's building in and the environment the result will run in are different, so just because it finds zlib on the build system doesn't mean that the target system will have zlib installed on it. (And even if it does, it may not be the same version. This is one of the big things that makes cross-compiling such a pain. One big reason for statically linking programs is to eliminate this kind of environmental dependency.)
The Firmware Linux build process is structured the way it is to eliminate environmental dependencies. Some are unavoidable (such as C libraries needing kernel headers or gcc needing binutils), but the intermediate system is the minimal fully functional Linux development environment I currently know how to build, and then we chroot into that and work our way back up from there by building more packages in the new environment.
To build uClibc you need kernel headers identifying the syscalls and such it can make to the OS. Way back when you could use the kernel headers straight out of the Linux kernel 2.4 tarball and they'd work fine, but sometime during 2.5 the kernel developers decided that exporting a sane API to userspace wasn't the kernel's job, and stopped doing it.
The 0.8x series of Firmware Linux used kernel headers manually cleaned up by Mariusz Mazur, but after the 2.6.12 kernel he had an attack of real life and fell too far behind to catch up again.
The current practice is to use the 2.6.18 kernel's "make headers_install" target, created by David Woodhouse. This runs various scripts against the kernel headers to sanitize them for use by userspace. This was merged in 2.6.18-rc1, so as of 2.6.18 we can use the Linux Kernel tarball as a source of headers again.
Another problem is that the busybox shell situation is a mess with four implementations that share little or no code (depending on how they're configured). The first question when trying to fix them is "which of the four do you fix?", and I'm just not going there. So until bbsh goes in we substitute bash.
Finally, most packages expect gcc. The tcc project isn't a drop-in gcc replacement yet, and doesn't include a "make" program. Most importantly, tcc development appears stalled because Fabrice Bellard's other major project (qemu) is taking up all his time these days. In 2004 Fabrice built a modified Linux kernel with tcc, and listed what needed to be upgraded in TCC to build an unmodified kernel, but since then he hardly seems to have touched tcc. Hopefully, someday he'll get back to it and put out a 1.0 release of tcc that's a drop-in gcc replacment. (And if he does, I'll add a make implementation to BusyBox so we don't need to use any of the gnu toolchain). But in the meantime the only open source compiler that can build a complete Linux system is still the gnu compiler.
The gnu compiler actually consists of three packages (binutils, gcc, and make), which is why it's generally called the gnu "toolchain". (The split between binutils and gcc is for purely historical reasons, and you have to match the right versions with each other or things break.)
This means that to compile a minimal build environment, you need seven packages, and to actually run the result we use an eighth package (QEMU).
This can actually be made to work. The next question is how?
The first problem is that we're cross-compiling. We can't help it. You're cross-compiling any time you create target binaries that won't run on the host system. Even when both the host and target are on the same processor, if they're sufficiently different that one can't run the other's binaries, then you're cross-compiling. In our case, the host is usually running both a different C library and an older kernel version than the target, even when it's the same processor.
The second problem is that we want to avoid requiring root access to build Firmware Linux. If the build can run as a normal user, it's a lot more portable and a lot less likely to muck up the host system if something goes wrong. This means we can't modify the host's / directory (making anything that requires absolute paths problematic). We also can't mknod, chown, chgrp, mount (for --bind, loopback, tmpfs)...
In addition, the gnu toolchain (gcc/binutils) is chock-full of hardwired assumptions, such as what C library it's linking binaries against, where to look for #included headers, where to look for libraries, the absolute path the compiler is installed at... Silliest of all, it assumes that if the host and target use the same processor, you're not cross-compiling (even if they have a different C library and a different kernel, and even if you ./configure it for cross-compiling it switches that back off because it knows better than you do). This makes it very brittle, and it also tends to leak its assumptions into the programs it builds. New versions may someday fix this, but for now we have to hit it on the head repeatedly with a metal bar to get anything remotely useful out of it, and run it in a separate filesystem (chroot environment) so it can't reach out and grab the wrong headers or wrong libraries despite everything we've told it.
The absolute paths problem affects target binaries because all dynamically linked apps expect their shared library loader to live at an absolute path (in this case /lib/ld-uClibc.so.0). This directory is only writeable by root, and even if we could install it there polluting the host like that is just ugly.
The Firmware Linux build has to assume it's cross-compiling because the host is generally running glibc, and the target is running uClibc, so the libraries the target binaries need aren't installed on the host. Even if they're statically linked (which also mitigates the absolute paths problem somewhat), the target often has a newer kernel than the host, so the set of syscalls uClibc makes (thinking it's talking to the new kernel, since that's what the ABI the kernel headers it was built against describe) may not be entirely understood by the old kernel, leading to segfaults. (One of the reasons glibc is larger than uClibc is it checks the kernel to see if it supports things like long filenames or 32-bit device nodes before trying to use them. uClibc should always work on a newer kernel than the one it was built to expect, but not necessarily an older one.)
Cross compiling is a pain. There are a lot of ways to get it to sort of kinda work for certain versions of certain packages built on certain versions of certain distributions. But making it reliable or generally applicable is hard to do.
I wrote an introduction to cross-compiling which explains the terminology, plusses and minuses, and why you might want to do it. Keep in mind that I wrote that for a company that specializes in cross-compiling. Personally, I consider cross-compiling a necessary evil to be minimized, and that's how Firmware Linux is designed. We cross-compile just enough stuff to get a working native compiler for the new platform, which we then run under emulation.
The emulator Firmware Linux 0.8x used was User Mode Linux (here's a UML mini-howto I wrote while getting this to work). Since we already need the linux-kernel source tarball anyway, building User Mode Linux from it was convenient and minimized the number of packages we needed to build the minimal system.
The first stage of the build compiled a UML kernel and ran the rest of the build under that, using UML's hostfs to mount the parent's root filesystem as the root filesystem for the new UML kernel. This solved both the kernel version and the root access problems. The UML kernel was the new version, and supported all the new syscalls and ioctls and such that the uClibc was built to expect, translating them to calls to the host system's C library as necessary. Processes running under User Mode Linux had root access (at least as far as UML was concerned), and although they couldn't write to the hostfs mounted root partition, they could create an ext2 image file, loopback mount it, --bind mount in directories from the hostfs partition to get the apps they needed, and chroot into it. Which is what the build did.
Current Firmware Linux has switched to a different emulator, QEMU, because as long as we're we're cross-compiling anyway we might as well have the ability to cross-compile for non-x86 targets. We still build a new kernel to run the uClibc binaries with the new kernel ABI, we just build a bootable kernel and run it under QEMU.
The main difference with QEMU is a sharper dividing line between the host system and the emulated target. Under UML we could switch to the emulated system early and still run host binaries (via the hostfs mount). This meant we could be much more relaxed about cross compiling, because we had one environment that ran both types of binaries. But this doesn't work if we're building an ARM, PPC, or x86-64 system on an x86 host.
Instead, we need to sequence more carefully. We build a cross-compiler, use that to cross-compile a minimal intermediate system from the seven packages listed earlier, and build a kernel and QEMU. Then we run the kernel under QEMU with the new intermediate system, and have it build the rest natively.
It's possible to use other emulators instead of QEMU, and I have a todo item to look at armulator. (I looked at another nommu system simulator at Ottawa Linux Symposium, but after resolving the third unnecessary environmental dependency and still not being able to get it to finish compiling yet, I gave up. Armulator may be a patch against an obsolete version of gdb, but I could at least get it to build.)
The main downsides of emulation are that is it's slow, can use a lot of memory, and can be tricky to debug if something goes wrong in the emulated environment. Cross compiling is sufficiently harder than native compiling that I consider it a good trade-off, but there are alternatives.
Some other build systems (such as uClibc's Buildroot) use a package called fakeroot, which is sort of a halfway emulator. It creates an environment where binaries run as if they had root access, but without being able to do anything that actually requires root access. This is nice if you want to create tarballs with device nodes and different ownership in them, but not so useful if you want to actually use one of those device nodes, or twiddle mount points. Firmware Linux doesn't use fakeroot (we use a real emulator instead), but it's an option.
In theory, we could work around the "host hasn't got uClibc" problem by statically linking our apps for the intermediate system, and work around the "host kernel older than the kernel headers we're using" problem by either building the intermediate version of uClibc with the host's kernel headers or just linking against glibc instead of uClibc.
This has a number of downsides: harvesting the host's kernel headers is distribution-specific, and could easily leak bits of the host into the final system. Linking the host tools against glibc (or a temporary version of uClibc built with different kernel headers) doesn't give us as much evidence that the resulting system will be able to rebuild itself under itself, and statically linking against glibc wastes a regrettable amount of space. None of this works with real cross-compiling between different processors (such as building an ARM system from x86).
We'd still have to solve the other problems (such as gcc wanting absolute paths) anyway, there just wouldn't be a switchover point where we could run the binaries we were building and start native compiling. Instead we'd have to keep cross-compiling all the way to the final system, and if anything's wrong with it we wouldn't find out until we tried to run it. With the native build, we've given the tools a bit of a workout during the build, so if the build completes then the finished system shouldn't have anything too fundamentally wrong with it.
(Note: QEMU can export a host directory to the target through the emulated network card as an smb filesystem, but you don't want to run your root filesystem on smb.)
Firmware Linux's directory hierarchy is a bit idiosyncratic: some redundant directories have been merged, with symlinks from the standard positions pointing to their new positions. On the bright side, this makes it easy to make the root partition read-only.
The set "bin->usr/bin, sbin->usr/sbin, lib->usr/lib" all serve to consolidate all the executables under /usr. This has a bunch of nice effects: making a a read-only run-from-CD filesystem easier to do, allowing du /usr to show the whole system size, allowing everything outside of there to be mounted noexec, and of course having just one place to look for everything. (Normal executables are in /usr/bin. Root only executables are in /usr/sbin. Libraries are in /usr/lib.)
For those of you wondering why /bin and /usr/sbin were split in the first place, the answer is it's because Ken Thompson and Dennis Ritchie ran out of space on the original 2.5 megabyte RK-05 disk pack their root partition lived on in 1971, and leaked the OS into their second RK-05 disk pack where the user home directories lived. When they got more disk space, they created a new direct (/home) and moved all the user home directories there.
The real reason we kept it is tradition. The execuse is that the root partition contains early boot stuff and /usr may get mounted later, but these days we use initial ramdisks (initrd and initramfs) to handle that sort of thing. The version skew issues of actually trying to mix and match different versions of /lib/libc.so.* living on a local hard drive with a /usr/bin/* from the network mount are not pretty.
I.E. The seperation is just a historical relic, and I've consolidated it in the name of simplicity.
The one bit where this can cause a problem is merging /lib with /usr/lib, which means that the same library can show up in the search path twice, and when that happens binutils gets confused and bloats the resulting executables. (They become as big as statically linked, but still refuse to run without opening the shared libraries.) This is really a bug in either binutils or collect2, and has probably been fixed since I first noticed it. In any case, the proper fix is to take /lib out of the binutils search path, which we do. The symlink is left there in case somebody's using dlopen, and for "standards compliance".
On a related note, there's no reason for "/opt". After the original Unix leaked into /usr, Unix shipped out into the world in semi-standardized forms (Version 7, System III, the Berkeley Software Distribution...) and sites that installed these wanted places to add their own packages to the system without mixing their additions in with the base system. So they created "/usr/local" and created a third instance of bin/sbin/lib and so on under there. Then Linux distributors wanted a place to install optional packages, and they had /bin, /usr/bin, and /usr/local/bin to choose from, but the problem with each of those is that they were already in use and thus might be cluttered by who knows what. So a new directory was created, /opt, for "optional" packages like firefox or open office.
It's only a matter of time before somebody suggests /opt/local, and I'm not humoring this. Executables for everybody go in /usr/bin, ones usable only by root go in /usr/sbin. There's no /usr/local or /opt. /bin and /sbin are symlinks to the corresponding /usr directories, but there's no reason to put them in the $PATH.
All the editable stuff has been moved under "var", starting with symlinking tmp->var/tmp. Although /tmp is much less useful these days than it used to be, some things (like X) still love to stick things like named pipes in there. Long ago in the days of little hard drive space and even less ram, people made extensive use of temporary files and they threw them in /tmp because ~home had an ironclad quota. These days, putting anything in /tmp with a predictable filename is a security issue (symlink attacks, you can be made to overwrite any arbitrary file you have access to). Most temporary files for things like the printer or email migrated to /var/spool (where there are persistent subdirectories with known ownership and permissions) or in the user's home directory under something like "~/.kde".
The theoretical difference between /tmp and /var/tmp is that the contents of /var/tmp should definitely be deleted by the system init scripts on every reboot, but the contents of /tmp may be preserved across reboots. Except deleting everyting out of /tmp during a reboot is a good idea anyway, and any program that actually depends on the contents of /tmp being preserved across a reboot is obviously broken, so there's no reason not to symlink them together.
(I case it hasn't become apparent yet, there's 30 years of accumulated cruft in the standards, convering a lot of cases that don't apply outside of supercomputing centers where 500 people share accounts on a mainframe that has a dedicated support staff. They serve no purpose on a laptop, let alone an embedded system.)
The corner case is /etc, which can be writeable (we symlink it to var/etc) or a read-only part of the / partition. It's really a question of whether you want to update configuration information and user accounts in a running system, or whether that stuff should be fixed before deploying. We're doing some cleanup, but leaving /etc writeable (as a symlink to /var/etc). Firmware Linux symlinks /etc/mtab->/proc/mounts, which is required by modern stuff like shared subtrees. If you want a read-only /etc, use "find /etc -type f | xargs ls -lt" to see what gets updated on the live system. Some specific cases are that /etc/adjtime was moved to /var by LSB and /etc/resolv.conf should be a symlink somewhere writeable.
The result of all this is that a running system can have / be mounted read only (with /usr living under that), /var can be ramfs or tmpfs with a tarball extracted to initialize it on boot, /dev can be ramfs/tmpfs managed by udev or mdev (with /dev/pts as devpts under that: note that /dev/shm naturally inherits /dev's tmpfs and some things like User Mode Linux get upset if /dev/shm is mounted noexec), /proc can be procfs, /sys can bs sysfs. Optionally, /home can be be an actual writeable filesystem on a hard drive or the network.
Remember to put root's home directory somewhere writeable (I.E. /root should move to either /var/root or /home/root, change the passwd entry to do this), and life is good.