What is Digg?28 Comments
- daddymcguire, on 08/15/2008, -0/+16Or you could visit linuxfromscratch.org
- antibastard, on 08/15/2008, -1/+15i love ibm, the wealth of info on their website is great.
- orvtech, on 08/15/2008, -2/+9... ubuntu users ..
- klco, on 08/15/2008, -0/+7Because you CAN
- theaceoffire, on 08/15/2008, -0/+5Sorry about the length of this comment...
You can download this thing as a text file from here:
http://www.mediafire.com/?tgglroijipd
If you don't trust that, then here is the text directly:
~~~~~~~~~~~~~~~~~~~~~~~
~~~~~~~~~~~~~~~~~~~~~~~
Build an embedded Linux distro from scratch
https://www6.software.ibm.com/developerworks/educa ...
Level: Intermediate
Peter Seebach, Freelance author, Plethora.net
12 Aug 2008
Learn how to build a custom Linux® distribution to use in an embedded environment, in this case to drive a Technologic Systems TS-7800 single-board computer. In this tutorial, you learn about cross-compiling, the boot loader, file systems, the root file system, disk images, and the boot process, all with respect to the decisions you make as you're building the system and creating the distribution.
Before you start
#####################################
Objectives
#####################################
This tutorial shows you how to install Linux on a target system. Not a prebuilt Linux distribution, but your own, built from scratch. While the details of the procedure necessarily vary from one target to another, the same general principles apply.
The result of this tutorial (if you have a suitable target) is a functional Linux system you can get a shell prompt on.
#####################################
About this tutorial
#####################################
The tutorial begins with a discussion of cross-compilation issues, then discusses what the components of a Linux system are and how they are put together. Both the building and the installation and configuration of the target system are covered.
The specific target discussed, a Technologic Systems TS-7800, imposes its own default boot and bring-up behaviors; other systems will have other mechanics, and this tutorial does not go into great detail about every possible boot loader.
#####################################
Prerequisites and system requirements
#####################################
Developers who are interested in targeting embedded systems, or who just want to learn more about what Linux systems are like under the hood, will get the most out of this tutorial.
The host environment used is Ubuntu, but other systems work as well. Users are assumed to have basic familiarity with UNIX® or Linux system administration issues. The tutorial assumes root access to a host system.
This tutorial assumes that your shell is a Bourne shell derivative; if you use a C shell derivative, the prompt will probably look different, and you will need to use different commands to set environment variables.
For cross-compiling (which is useful when targeting embedded systems), I used crosstool-ng version 1.1.0, released in May of 2008. You may download it from the distribution site (see Resources). Details follow on installing and configuring it.
About the target and architecture
#####################################
Target
#####################################
The target I chose is a Technologic Systems TS-7800 (see Resources for more detail). This is a small embedded ARM system, with both built-in and removable flash storage, as well as a SATA controller. This tutorial walks you through configuring this system to boot to a login prompt, without relying on prebuilt binaries.
#####################################
Architecture
#####################################
I chose the ARM architecture to make it a little easier to check whether a given binary is host or target, and to make it easy to see when host pollution might be occurring. It is also nice having a machine that consumes a total of about 5W of power and runs completely silently.
Cross-compiling
#####################################
What is cross-compiling?
#####################################
Cross-compiling is using a compiler on one system to develop code to run on another. Cross-compilation is relatively rare among casual UNIX users, as the default is to have a compiler installed on the system for use on that system. However, cross-compilation becomes quite common (and relevant) when targeting embedded systems. Even when host and target are the same architecture, it is necessary to distinguish between their compilers; they may have different versions of libraries, or libraries built with different compiler options, so that something compiled with the host compiler could fail to run, or could behave unexpectedly, on the target.
#####################################
Obtaining cross-compilation tools
#####################################
It is, in theory, quite possible to build a cross-compiler yourself, but it is rarely practical. The series of bootstrap stages needed can be difficult and time-consuming, and it is often necessary to build a very minimal compiler, which is used to partially configure and build libraries, the headers of which are then used to rebuild the compiler so it can use them, and so on. A number of commercial sources for working cross-compilers for various architecture combinations are available, as well as several free cross-compilation toolkits.
#####################################
Introducing crosstool-ng
#####################################
Dan Kegel's crosstool (see Resources for details) collects a variety of expertise and a few specialized patches to automatically build toolchains for a number of systems. Crosstool has not been updated in a while, but the new crosstool-ng project builds on this work. For this tutorial, I used crosstool-ng version 1.1.0, released in May of 2008. Download it from the distribution site (see Resources).
#####################################
Installing crosstool-ng
#####################################
Crosstool-ng has a configure script. To configure it, just run the script using --prefix to set a location. For instance:
$ ./configure --prefix=$HOME/7800/ctng
Once you have configured it, build it using make and then make install. The build process creates a ctng directory in the 7800 working directory that holds the crosstool-ng build scripts. Add the ctng/bin subdirectory to your path:
$ PATH=$PATH:$HOME/7800/ctng/bin
#####################################
Configuring crosstool-ng
#####################################
Crosstool-ng uses a .config file similar to those used by the Linux kernel. You need to create a configuration file matching your target to use crosstool-ng. Make a working directory for a crosstool-ng build:
$ mkdir toolchain-build
$ cd toolchain-build
Now, copy in a default configuration. It's possible to manually configure crosstool-ng, but one of the sample configurations happens to fit the target perfectly:
$ cp ../ctng/lib/ct-ng-1.1.0/samples/arm-unknown-linux-uclibc/* .
Finally, rename the crosstool.config file:
$ mv crosstool.config .config
This copies in a configuration file that targets an armv5te processor, the model used on the TS-7800. It builds with uClibc, a libc variant intended for embedded systems. However, the configuration file does need one modification.
#####################################
Fixing the configuration path
#####################################
The default target directory for a crosstool-ng build is $HOME/x-tools/$TARGET. For instance, on this build, it would come out as x-tools/arm-unknown-linux-uclibc. This is very useful if you are building for a lot of targets, but not so useful if you are building for only one. Edit the .config file and change CT_PREFIX_DIR to ${HOME}/7800/toolchain.
#####################################
Building the toolchain
#####################################
To build the toolchain, run the ct-ng script with the build argument. To improve performance, especially on a multi-core system, you may want to run with multiple jobs, specified as build.#. For example, this command builds with four jobs:
$ ct-ng build.4
This may take quite a while, depending on your host system. When it's done, the toolchain is installed in $HOME/7800/toolchain. The directory and its contents are marked read-only; if you need to delete or move them, use chmod u+w The ct-ng script takes other arguments, such as help. Note that ct-ng is a script for the standard make utility, and as a result, the output from --help is just the standard make help; use ct-ng help to get the help for crosstool-ng.
If you haven't seen this trick before, it's a neat one. Modern UNIX systems interpret an executable file in which the first line starts with #! as a script, specifically, a script for the program named on the rest of the line. For instance, many shell scripts start with #!/bin/sh. The name of the file is passed to the program. For programs that treat their first argument as a script to run, this is sufficient. While make does not do that automatically, you can give it a file to run with using the -f flag. The first line of ct-ng is #!/usr/bin/make -rf. The -r flag suppresses the built-in default construction rules of make, and the -f flag tells it that the following word (which is the script's file name) is the name of a file to use instead of one named Makefile. The result is an executable script that uses make syntax instead of shell syntax.
#####################################
Using the toolchain
#####################################
For starters, add the directory containing the compiler to your path:
$ PATH=~/7800/toolchain/bin:$PATH
With that in your path, you can now compile programs:
$ arm-unknown-linux-uclibc-gcc -o hello hello.c $ file hello
hello: ELF 32-bit LSB executable, ARM, version 1 (SYSV), for
GNU/Linux 2.4.17, dynamically linked (uses shared libs), not stripped
#####################################
Where are the libraries?
#####################################
The libraries used by the toolchain to link binaries are stored in arm-unknown-linux-uclibc/sys-root, under the toolchain directory. This directory forms the basis of an eventual root file system, a topic I will come back to shortly, once the kernel is built.
#####################################
Kernel setup
#####################################
The kernel distribution tree provided by the vendor is already configured for cross compilation. In the simplest case (which this is), the only thing you have to do to cross compile a Linux kernel is to set the CROSS_COMPILE variable in the top-level Makefile. This is a prefix that is prepended to the names of the various programs (gcc, as, ld) used during the build. For instance, if you set CROSS_COMPILE to arm-, the compile will try to find a program named arm-gcc in your path. For this build, then, the correct value is arm-unknown-linux-uclibc. Or, if you don't want to rely on path settings, you can specify the whole path, as in this example:
CROSS_COMPILE ?= $(HOME)/7800/toolchain/bin/arm-unknown-linux-uclibc-
#####################################
Kernel configuration
#####################################
A complete discussion of kernel configuration is beyond the scope of this tutorial. In this case, the ts7800_defconfig target gave me a default usable configuration for the 7800, with one small hiccup: the CONFIG_DMA_ENGINE setting ended up off when it should have been on.
#####################################
Tweaking the kernel
#####################################
It is usually best to edit the kernel using make menuconfig, which offers a semi-graphical interface to kernel configuration. This interface is navigated using arrow keys to move the cursor, the Tab key to select options from the bottom of the screen, and the space or Enter keys to select options. For instance, to exit without changing anything, press Tab until the at the bottom of the screen is highlighted, then press Enter. Running make menuconfig again reopens the editor.
#####################################
Changing the default console
#####################################
The TS-7800 normally boots silently, because the default kernel configuration specifies a null console device to keep the display quiet. To change this, use the arrow keys to navigate down to "Boot options," and press Enter. The third line shows the default kernel options, which select the ramdisk, the startup script, and the console. Use the arrow keys to navigate down to this line, press Enter, and change console none to console ttyS0,115200. Then, press Tab to move the cursor to the at the bottom of the panel, and press Enter. Now press Tab to select and press Enter, bringing you back to the main menu.
For the goal of booting as fast as possible, the console device isn't useful, and indeed, even at a high baud rate, sending kernel messages can take a noticeable fraction of the time the system takes to boot. For debugging and playing around, though, you want the console.
#####################################
Enabling the DMA engine
#####################################
Navigate down to "Device drivers" and press Enter. This list is longer than the usual display, so you will have to scroll down to the very end to reach the option for "DMA Engines." Navigate to that with the arrow keys, and press Enter. There are two options at the top of this page that have square brackets indicating a boolean option. The second, "Support for DMA engines," was not enabled by default in the download I started with. Navigate to it with the arrow keys, and press space to toggle its state. Now use Tab and Enter to select from each screen to navigate out to the top level of the program, and then - mythicflux, on 08/15/2008, -0/+4I'd build my own distro because I have 6 old 512mb compact flash cards and a CF to IDE adapter. I've always wanted to build a simple distro for scratch just to see if I could.
It's about pursuing knowledge for your own interest rather than just using someelses work. - tnoy, on 08/16/2008, -0/+4There are a lot of distros, but not many for embedded systems. You did see that it said "embedded" in the title, didn't you? Picking a generic distribution for most embedded projects is a bad idea. Sure, you can use a generic install of Debian on a AILX board, but its going to use up far more CPU/RAM than you'd want it to.
I'd like to see you put Kubuntu on a ARM device with 16MB of RAM.
The board used in the tutorial is kinda fancy: gigabit ethernet, SATA, and a FPGA? Makes my AILX baord seem rather primative. - noolvidaremos, on 08/15/2008, -1/+4Nice, i have planned to release this weekend a detailed howto on this too about rolling your own firmware kernel and all in to the NSLU2 on http://www.linuxevolution.org/nslu2.html I hold it to next weekend and include some tools.
- John6000, on 08/15/2008, -0/+3use bugmenot - google it
- thcobbs, on 08/15/2008, -2/+5sure, if you have a couple days to figure out what they don't spell out.
- solarwind24, on 08/16/2008, -0/+3This is good, but the absolute best way to do it is to use buildroot (made by the uclibc dev team). - http://buildroot.uclibc.org/news.html
What build root does is automatically set up an entire development environment (cross compiler toolchain, target binaries, even target filesystem image with complete userland) just by using a simple menu configurator and issuing one make command. I'm using it this very moment to set up a system for my x86 embedded board. It cross compiles to any architecture supported by gcc/Linux.
It even has a large set of packages that you can chose to compile for your target - ranging from uclibc to busybox to a full blown X11 environment. Best embedded development environment I have ever used (besides QNX Momentics, that is). - noolvidaremos, on 08/15/2008, -0/+3Yeah thats why I said I was going to put it up this weekend. take a look at http://www.nslu2-linux.org/ if you want more information about that device. there are lots of projects there.
... or check next weekend ;)
BTW... that site is running on a NSLU2, 32MB of ram, processor of 133MHz - CATSCEO2, on 08/15/2008, -2/+5Uh, were is the tutorial?
- theaceoffire, on 08/15/2008, -0/+2This can be a custom program rather than a conventional UNIX-style init, but it is easier to work with the system if you provide a working environment complete with console shell and everything.
#####################################
Basic libraries and headers
#####################################
The crosstool-ng build created a default system root providing library code and headers. Make a copy of it:
$ cp -r toolchain/arm-unknown-linux-uclibc/sys-root rootfs
Note that no effort is being made to fix permissions right now; for a real distribution, you'd probably want to correct this, but it's harmless enough for demonstration purposes. This gets you the first basic parts of a root filesystem, containing only the libraries (and headers) future builds will use.
#####################################
Busybox
#####################################
When it comes to embedded root filesystems, busybox is probably the best starting point. Busybox offers a reasonably complete set of core system utilities all built into a single binary, which uses the name it gets invoked with to determine what function to perform. This, coupled with a lot of links, allows dozens of the main system programs to be bundled into small amount of space. This isn't always necessary, but it is convenient for our purposes.
#####################################
Downloading busybox
#####################################
Busybox can be downloaded from its downloads page (see Resources). The archive is just a standard tarball and can be unpacked in a build directory. I used version 1.10.2.
#####################################
Configuring busybox
#####################################
Busybox is built around similar configuration tools to those used for crosstool-ng and the Linux kernel. To build a default configuration, run make defconfig. This creates the default configuration. A following make menuconfig allows you to change settings; under "Busybox Settings," the "Build Options" menu allows you to specify a static build. Do that, because the default crosstool-ng build of uClibc requires it. You might sometimes prefer a dynamically linked binary, to reduce the size of the executable, but the size problem isn't such a big deal on a system where nearly every binary is just a symbolic link. As with the kernel, the top-level Makefile has a CROSS_COMPILE variable which holds the prefix to use on the names of compiler tools. Set it to the same value you used before.
#####################################
Building busybox
#####################################
To build busybox, just run make. Surprisingly, there was a compilation problem during the build, which required me to add to the main libbb.h header. With that out of the way, the compile completed quite quickly.
#####################################
Installing busybox
#####################################
To install busybox, you need to run make install, but specify the root filesystem path:
$ make CONFIG_PREFIX=$HOME/7800/rootfs install
This copies in the busybox binary and creates all of the necessary links. This root filesystem now has libraries and all the common UNIX utilities. What's missing?
#####################################
Device nodes
#####################################
You need device nodes. There are two ways to provide device nodes for a root file system. One is to statically create all the device nodes you plan to have, the other is to use something like udev, which provides an automatic current device node tree for the current kernel. While udev is extremely elegant, it can also be fairly slow, and embedded systems have the luxury of being able to predict their hardware components.
#####################################
Creating the needed device nodes
#####################################
Many programs depend on access to device nodes, such as /dev/console or /dev/null. Device nodes are typically made at runtime on ARM systems, using the mdev utility, part of busybox. In fact, this has already been done by the initrd filesystem; one of the changes above was to move the already populated /dev mount point onto the new root filesystem. You can manually create key device nodes, but it is usually not worth the trouble; mdev is smarter.
#####################################
Mount points
#####################################
Several mount points are required. The /sys and /proc mount points are used to move the sysfs and procfs virtual filesystems. The /dev directory is used as a mount point for a tmpfs filesystem, populated by mdev. These directories do not need any other contents:
$ cd $HOME/7800/rootfs
$ mkdir sys proc dev
#####################################
Etcetera
#####################################
The /etc directory, while not strictly necessary to get to a shell prompt, is very useful to get to a useful shell prompt. The most important part during initial booting is the init.d subdirectory, which init searches for system startup files. Create a trivial rcS script and give it execute permissions for initial system startup:
$ mkdir rootfs/etc/init.d
Create the rootfs/etc/init.d/rcS file with the following contents:
#!/bin/sh
echo "Hello, world!"
Now mark it executable:
$ chmod 755 rootfs/etc/init.d/rcS
Creating a disk image
#####################################
Make a large empty file
#####################################
You can create a file system image by copying blocks from /dev/zero into a file, or just by copying an existing image. Whatever size space you have available (remembering that 12MB are already reserved for the first three partitions), you can create a file of that size with dd. This creates a nice roomy 64MB root filesystem:
$ dd if=/dev/zero of=rootfs.local bs=1M count=64
#####################################
Format the file
#####################################
The various mkfs utilities can be run on files, not just on disks. To build an ext3 root filesystem (the kind the initrd looks for), run mkfs.ext3 on your disk image:
$ mkfs.ext3 rootfs.local
The mkfs utility prompts you as to whether or not to proceed even though the file is not a block special device:
rootfs.tmp is not a block special device.
Proceed anyway? (y,n) y
#####################################
Mount the new disk image
#####################################
To mount the disk image, use the -o loop option to the mount command again to mount the image somewhere.
$ mount -o loop rootfs.local /mnt
#####################################
Copy the rootfs files to the mounted disk
#####################################
The pax utility is particularly good at this.
$ cd 7800/rootfs
$ pax -r -w -p e . /mnt
This copies the directory tree you've just created to /mnt, preserving symlinks and special files (except for sockets, but you can live without the log socket).
#####################################
Unmount the image
#####################################
Once again, having finished work on the image, unmount it:
$ umount /mnt
Getting the images back on the card
#####################################
Copying files back
#####################################
You might wonder why the above procedure didn't just use the card as a filesystem. The answer is that, in general, it is better not to use flash filesystems very heavily, because flash devices have much shorter life expectancies than hard drive media under heavy load. Thus, the complicated edits and copies are done to disk images, then the disk images are copied back as a single write operation.
#####################################
Copy in the initrd
#####################################
Copy the initrd image in to the third partition of the SD card:
$ dd if=initrd.local of=/dev/sdd3 bs=16k
#####################################
Copy in the root filesystem
#####################################
Likewise, copy the filesystem image in to the fourth partition of the SD card.
$ dd if=rootfs.local of=/dev/sdd4 bs=16k
#####################################
Wait for device access
#####################################
Do not yank the SD card out of the reader right away. Sometimes it can take a few extra seconds for the system to finish writing. Wait until any activity lights have stabilized, then wait a few seconds longer just to be sure.
The boot process reviewed
#####################################
What does init do?
#####################################
The init program manages system startup, and then keeps the system running. For instance, when processes die, init is the program that collects their exit status values so the kernel can finish unloading them. The exact startup procedure varies from one version of Linux to another. In this case, it is the busybox version of init that will be booting our test system.
You may have wondered why the previous startup script uses exec /sbin/init rather than just invoking init. The reason is that the startup script was the previous init, and had the special process ID 1. The init program does not become the system's startup daemon unless its process ID is 1; exec causes it to take over the calling shell's process ID, rather than acquiring a new one.
#####################################
Initial startup
#####################################
The init program starts by running the first system startup script, /etc/init.d/rcS. (If there is no such file, it prints a warning message and continues without it.) After this, it runs according to the instructions in /etc/inittab, if it exists. In the absence of an /etc/inittab, busybox init runs a shell on the console, and deals gracefully with reboot and halt requests.
#####################################
Bringing it together
#####################################
With your own kernel, a slightly customized initrd, and your own root filesystem, you should now be able to boot the system to a shell prompt. If you created the rcS file, the system should greet you on boot. Now you have a working, if rather minimal, Linux system, built entirely from source, and assembled by hand.
What you do next is up to you. I suggest installing some video games, but you could always configure the system for a database server or a Web server. If you plan to do anything that will produce a lot of disk activity, consider using an external hard drive attached via USB.
Resources
Learn
* Check out this introduction to the Linux boot process (developerWorks, May 2006): http://www.ibm.com/developerworks/linux/library/l- ...
* Learn more about Dan Kegel's crosstool on its original site. http://www.kegel.com/crosstool/
* Read about Linux initial RAM disks and how to use them (developerWorks, July 2006). http://www.ibm.com/developerworks/linux/library/l- ...
* In the developerWorks Linux zone, find more resources for Linux developers, and scan our most popular articles and tutorials. http://www.ibm.com/developerworks/linux/?S_TACT=10 ... http://www.ibm.com/developerworks/linux/library/l- ...
* See all Linux tips and Linux tutorials on developerWorks. http://www.ibm.com/developerworks/views/linux/libr ...
http://www.ibm.com/developerworks/views/linux/libr ...
* Stay current with developerWorks technical events and Webcasts. http://www.ibm.com/developerworks/offers/techbrief ...
Get products and technologies
* Download the crosstool-ng sources and documentation to build your own cross compiler. http://ymorin.is-a-geek.org/dokuwiki/projects/cros ...
* Download the Busybox source code to get a complete Linux system in a tiny little package. http://busybox.net/downloads/
* Learn more about the Technologic Systems TS-7800 single-board computer. http://www.embeddedarm.com/products/board-detail.p ...
* Order the SEK for Linux, a two-DVD set containing the latest IBM trial software for Linux from DB2®, Lotus®, Rational®, Tivoli®, and WebSphere®. http://www.ibm.com/developerworks/offers/sek/?S_TA ...
* With IBM trial software, available for download directly from developerWorks, build your next development project on Linux. http://www.ibm.com/developerworks/downloads/?S_TAC ...
Discuss
* Get involved in the developerWorks community through blogs, forums, podcasts, and spaces. http://www.ibm.com/developerworks/community?S_TACT ... - theaceoffire, on 08/15/2008, -0/+2exit one more time to leave the program. When asked whether you wish to save your new kernel configuration, tab to and press Enter.
#####################################
Compiling the kernel
#####################################
Type make. Yes, it really is that simple. This builds a kernel, as well as a collection of modules. Once again, multi-core users might want multiple jobs; try make -j 5. For the purposes of this project, I'm going to ignore kernel modules, and favor compiling-in any needed features. The bootstrap ramdisk technique used to get needed drivers into the kernel early seems excessive, and building a root file system is already complicated enough. This, of course, brings up the question of how to get a kernel booting, the subject of the next section.
Boot loaders
#####################################
What is a boot loader?
#####################################
A boot loader (or bootstrap loader) is a program that loads another program. The boot loader is a small and specialized hunk of code, specific to a target system, that has just enough complexity to find the kernel and load it without being a full-featured kernel. Different systems use a variety of different boot loaders, from the huge and complicated BIOS programs common on desktop PCs, to very small and simple programs more common on embedded systems.
#####################################
A simple boot loader
#####################################
The TS-7800 uses an unusually simple boot loader, which simply picks up the kernel from a predetermined partition of the SD card or on-board flash. A jumper determines whether the board looks first at the on-board flash, or at the SD cards. There are no other configuration settings. More complicated boot loaders (such as grub, commonly used on desktop PCs) have options for configuring kernel options and so on at boot time. On this system, the kernel's default options must be compiled in, as described previously.
The decision to compile in kernel options is a typical example of a choice that makes some sense in an embedded system, but would be uncomfortably limiting on a desktop.
#####################################
Kernel formats
#####################################
There are a variety of formats in which kernels are stored. The initial Linux kernel binary, named vmlinux, is rarely the file that a boot loader will work with. On the TS-7800, the boot loader can use two files, either Image (uncompressed) or zImage (compressed). These files are created in the arch/arm/boot directory within the kernel tree.
People often describe the zImage kernel as a compressed kernel, and so expect the boot loader to need to provide decompression. In fact, it's rather more clever. The zImage kernel is an uncompressed executable, which contains a particularly large static data object which is a compressed kernel image. When the boot loader loads and runs the zImage executable, that executable then unpacks the kernel image and executes it. This way, you get most of the benefit of compression without imposing additional effort on the boot loader.
#####################################
Setting up the SD card
#####################################
The SD card needs an MBR table containing DOS-style partition information. A sample MBR table is available for download from Technologic's site; this is for a 512MB card, but it is easy to edit the fourth partition to a size suiting whatever size card you want to use. The first three partitions are 4MB each; the first is unused on smaller cards, and the second and third hold the kernel and initial ramdisk image respectively.
For my purposes, I used the sfdisk utility, which is not always recommended but has the desireable trait of trusting me when I ask it to create a partition that does not align on a partition boundary.
#####################################
Installing the initial kernel
#####################################
The kernel for the TS-7800 is dumped directly into the second partition of the SD card. The exact path to the card depends on how you are accessing it; typically, if you have the card on a USB card reader, it will be detected as a SCSI device. Note that there is no file system access involved; the raw kernel is just dumped into the partition. The dd command copies raw data from one source to another, as in this example:
$ dd if=zImage of=/dev/sdd2 bs=16k
93+1 records in
93+1 records out
1536688 bytes (1.5 MB) copied, 0.847047 s, 1.8 MB/s
This command dumps raw data from the zImage file to the second partition of /dev/sdd, using 16KB blocks.
#####################################
A bit about block sizes
#####################################
The output of this command is a little cryptic (as are its inputs, honestly). When dd runs, it copies data in "records," which are by default 512-byte blocks; this command specifies a block size of 16k (which dd understands to mean 16*1024).
The dd command reports the amount of data copied first in blocks; the number after the plus sign is the number of partial blocks copied. In this case, because 1,536,688 is not an exact multiple of the block size, the remaining bytes of the file are read (and written) separately as a partial block. This information is harmless for most modern devices but crucial to diagnosing problems with some older ones.
The ability to control block sizes (and reblock data when transferring it) was exceptionally useful for working with tape devices and other specialized media that required writes to be of particular fixed sizes, and also helps for performance and reliability reasons with flash devices.
While the kernel device representing flash media can often take writes of arbitrary sizes, it is common for the underlying device to work only in full blocks, often of somewhat larger sizes (4KB or larger). To do a partial write, the flash device must extract the current contents of the full block, modify them with the input data, and flash the whole block back. If a device uses 4KB blocks, and you write to it in 512-byte blocks, each device block gets rewritten eight times for a single copy. This is bad for the device's longevity, and also bad for performance. (A 512-byte write of the same file was half as fast on the flash card I used.)
#####################################
Booting the kernel
#####################################
Booting the kernel is simple enough. Set the jumper for an SD boot, put the card in the system, and power it up. If you started with a blank card, this produces a predictable result:
Kernel panic - not syncing: VFS: Unable to mount root fs on unknown-block(1,0)
This cryptic message indicates that the kernel has been unable to find its root filesystem. That means it's about time to create one.
Filesystems
#####################################
Root filesystems
#####################################
Normally, Linux has only a single root filesystem. However, during boot, it is common to use a temporary root filesystem to load and configure device drivers and the like, and then obtain the real root filesystem and switch to it. The TS-7800 uses this temporary filesystem (called an initial ramdisk, or initrd) for the SD driver and for a variety of configuration options. In fact, the ramdisk is used for many of the things (such as determining the final root filesystem to use) that might be handled by the boot loader on another system.
#####################################
Manipulating filesystems
#####################################
A general caveat applies: most filesystem operations require root privileges. There are some exceptions and some workarounds. The most notable is the fakeroot utility/library, which allows you to build applications with "fakeroot support." This allows the manipulation of a virtual root filesystem with control over ownership, permissions, and so on. However, the permissions and related material are maintained in a separate database, and no actual privileged operations are needed. This allows a non-root user to manipulate directory hierarchies with virtual root privileges, and ultimately create archives or filesystem images containing files that the user hasn't got the privileges to create. For this tutorial, however, I assume root privileges.
#####################################
Changing root
#####################################
The chroot command, while certainly part of the story, is not sufficient to change the root filesystem. The tool for this is pivot_root, which makes one named directory the new root directory, and "moves" (really, changes the mount point of) the old root directory to another name.
#####################################
A not-so-initial root
#####################################
For the purposes of this tutorial, I'm assuming that the default initrd image provided with the board is adequate. Linux kernel distributions have some support for generating suitable initrd images, and the details of this one are not crucial to an understanding of building an actual distribution. I'll just introduce the key things needed to get the system moving, then focus on what goes on the real root filesystem.
#####################################
Filesystem images
#####################################
Most Linux users are familiar with files containing images of ISO CD-ROM or DVD filesystems. Linux supports the creation and use of images of other types of filesystems. In fact, any file can be treated as though it were a disk partition, using the -o loop option to mount. For instance, while researching this, I made a copy of the initrd image used by the TS-7800:
$ dd if=/dev/sdd3 of=initrd.dist bs=16k
$ cp initrd.dist initrd.local
With this file available, it's possible to mount the filesystem to have a look at it. The copy named initrd.local is the local version to edit; the copy named initrd.dist is a safe pristine copy in case something gets broken later.
$ mount -o loop initrd.local /mnt
#####################################
What does the initrd do?
#####################################
The initrd provides a very basic initial filesystem used to bootstrap the device by finding and mounting the real root filesystem. The default kernel configuration is hard-wired to use an initrd (/dev/ram0) as root, and run the linuxrc script on startup. There are several provided variants of this script for different root file systems. The obvious choice is to look at the linuxrc-sdroot variant.
#####################################
Preparing initrd
#####################################
The linuxrc-sdroot program tries to configure the system to use an SD card as root. Assuming a correctly configured SD card with an ext3 filesystem on the fourth fdisk partition, the script mounts that filesystem and runs /sbin/init on it. This looks like a good strategy. So, a couple of small changes are needed to the initrd image. The first is to change the linuxrc symlink to point to linuxrc-sdroot. The second is to update the script to work with the uClibc build instead of the Debian build it was configured for.
#####################################
The SD driver
#####################################
The SD driver for the TS-7800 includes a proprietary module, so the tssdcard.ko file simply has to be manually copied in, and loaded at runtime. The file on the distribution initrd works fine with the new kernel, so no change is needed.
#####################################
Switching bootstrap routines
#####################################
The kernel simply runs the program called linuxrc. By default, this is a symbolic link to the program linuxrc-fastboot, which comes up quickly to the initrd root filesystem. Removing the link, and linking linuxrc instead to linuxrc-sdroot, causes the system to try to boot from the last partition on the SD once the drivers are loaded.
#####################################
Changing the bootstrap routine
#####################################
The script checks for errors, and if it finds any, mounts from the internal flash. If it does not, it mounts the filesystem from the SD card and then tries to change to it. The bootstrap routine actually has a small quirk that affects this build; it tries to use its own mount utility, rather than the rootfs one, for the last cleanup work after calling pivot_root. This works well with the provided Debian install, but fails with the minimalist uClibc install. The solution is to change the lines following the pivot_root command at the end of the file (after the else clause) as follows:
pivot_root . ./initrd
/bin/mount -n --move ./initrd/dev ./dev
/bin/mount -n --move ./initrd/sys ./sys
/bin/mount -n --move ./initrd/proc ./proc
exec /sbin/init < $CONSOLE > $CONSOLE 2>&1
This causes the script to run the new (uClibc, statically linked) executables that you are about to create, instead of trying to run dynamically linked executables from another system. (If you were doing this without a tutorial, you probably wouldn't know this until after you'd set everything else up and started getting cryptic error messages).
#####################################
Unmount the disk image
#####################################
Unmount the initrd disk image like so:
$ umount /mnt
Populating the root file system
#####################################
What goes into a root filesystem?
#####################################
The root filesystem needs an /sbin/init that will do whatever it is we want done. This can be a c - noolvidaremos, on 08/15/2008, -0/+2so .... how much are they charging ?
- orvtech, on 08/15/2008, -0/+2Nice!
- daddymcguire, on 08/16/2008, -0/+2I have built a couple linux systems from scratch, and while that's all well and good there is something to be said for scripts that automate the process. I would recommend lfscript to anyone looking for a good script that builds an entire linux system from scratch for you. The author has been updating it and adding additional features and it's really turning into something good.
While I do appreciate the automation part of it, I would still recommend first or second timers to actually build their system manually for the knowledge that would be gained. - theaceoffire, on 08/15/2008, -0/+2Here is the tutorial in txt format.
http://www.mediafire.com/?tgglroijipd
>.> I also posted it in this thread. Yeah, it will probably get buried, but hopefully it will still be around even when the tutorial gets moved and my text gets taken down. - CATSCEO2, on 08/15/2008, -2/+4***** IBM, you have to be registered to view it.
- ThantiK, on 08/17/2008, -0/+1Semi-advanced-but-not-quite-brave-enough-to-build-from-scratch linux users should definately consider also looking into Arch linux. It has an AWESOME walk-through and most of the build system is automated, it basically installs a scratch system and you use Arch's package manager "pacman" to install any tools you need (automagically resolves dependancies, etc). It's probably not any good for embedded systems that need to be extremely tiny, but it's amazing to build a VERY slimmed down linux for yourself without spending hours compiling things. Another good thing is that most possible applications you would need are kept in a rolling updated repository, keeping your system completely up to date is as simple as running "pacman -Syu" and it updates everything.
- noolvidaremos, on 08/15/2008, -1/+2 | s/ubuntu/kubuntu/g
he probably has vista installed on his desktop - ErikHK, on 08/15/2008, -0/+1Uhm, all the paragraphs on that website contain the same text...
- antibastard, on 08/15/2008, -0/+1way to miss the point dude!
- noolvidaremos, on 08/15/2008, -2/+2it is called Gentoo.
- latrosicarius, on 08/15/2008, -3/+2You should build a distro which specializes in helping you choose the right distro
- duckyinc, on 08/15/2008, -10/+2Why build another distro? There are already distros for everything now.
- 4DFX, on 08/15/2008, -14/+5...because there aren't enough distros available already....
Browsing Digg on your phone just got easier with our enhancements to the 
© Digg Inc. 2009
— Content created and posted by Digg users is