Linux-Based Firmware, How to Implement a Good Way to Update

Linux-Based Firmware, how to implement a good way to update?

I used the following approach. It was somewhat based on the paper "Building Murphy-compatible embedded Linux systems," available here. I used the versions.conf stuff described in that paper, not the cfgsh stuff.

Use a boot kernel whose job is to loop-back mount the "main" root file system. If you need a newer kernel, then kexec into that newer kernel right after you loop-back mount it. I chose to put the boot kernel's complete init in initramfs, along with busybox and kexec (both statically linked), and my init was a simple shell script that I wrote.
One or more "main OS" root file systems exist on an "OS image" file system as disk image files. The boot kernel chooses one of these based on a versions.conf file. I only maintain two main OS image files, the current and fall-back file. If the current one fails (more on failure detection later), then the boot kernel boots the fall-back. If both fail or there is no fall-back, the boot kernel provides a shell.
System config is on a separate partition. This normally isn't upgraded, but there's no reason it couldn't be.
There are four total partitions: boot, OS image, config, and data. The data partition is for user application stuff that is intended for frequent writing. boot is never mounted read/write. OS image is only (re-)mounted read/write during upgrades. config is only mounted read/write when config stuff needs to change (hopefully never). data is always mounted read/write.
The disk image files each contain a full Linux system, including a kernel, init scripts, user programs (e.g. busybox, product applications), and a default config that is copied to the config partition on the first boot. The files are whatever size is necessary to fit everything in them. As long I allowed enough room for growth so that the OS image partition is always big enough to fit three main OS image files (during an upgrade, I don't delete the old fall-back until the new one is extracted), I can allow for the main OS image to grow as needed. These image files are always (loop-back) mounted read-only. Using these files also takes out the pain of dealing with failures of upgrading individual files within a rootfs.
Upgrades are done by transferring a self-extracting tarball to a tmpfs. The beginning of this script remounts the OS image read/write, then extracts the new main OS image to the OS image file system, and then updates the versions.conf file (using the rename method described in the "murphy" paper). After this is done, I touch a stamp file indicating an upgrade has happened, then reboot.
The boot kernel looks for this stamp file. If it finds it, it moves it to another stamp file, then boots the new main OS image file. The main OS image file is expected to remove the stamp file when it starts successfully. If it doesn't, the watchdog will trigger a reboot, and then the boot kernel will see this and detect a failure.
You will note there are a few possible points of failure during an upgrade: syncing the versions.conf during the upgrade, and touching/removing the stamp files (three instances). I couldn't find a way to reduce these further and achieve everything I wanted. If anyone has a better suggestion, I'd love to hear it. File system errors or power failures while writing the OS image could also occur, but I'm hoping the ext3 file system will provide some chance of surviving in that case.

Implementing an update/upgrade system for embedded Linux devices

I believe you are looking wrong at the problem - any update which is non atomic (e.g. dd a file system image, replace files in a directory) is broken by design - if the power goes off in the middle of an update the system is a brick and for embedded system, power can go off in the middle of an upgrade.

I have written a white paper on how to correctly do upgrade/update on embedded Linux systems [1]. It was presented at OLS. You can find the paper here: https://www.kernel.org/doc/ols/2005/ols2005v1-pages-21-36.pdf

[1] Ben-Yossef, Gilad. "Building Murphy-compatible embedded Linux systems." Linux Symposium. 2005.

What are techniques for allowing safe software upgrades in embedded systems

It all depends on how critical the application is. The two basic approaches (backup and bootloader) are also combined sometimes.

Many systems have a read only bootloader (like redboot), and then two banks of flash memory (on the same chip, most often). The bootloader then has a flag to choose which bank to boot from. The flag will then change based on events like upgrades (failed or successful), and so on.

So, when upgrading, the running version copies the new load into the backup bank, checks the checksum, toggles the boot flag, and then reboots the device. The device reboots on the new bank, with the new load. After the reboot, the new load can copy itself into the backup bank.

Often there is also a watchdog timer with a hardware reset. This way, if the firmware goes insane, it fails to kick the watchdog, the hardware reset will reboot the device, and the bootloader will look for a sane load.

The Open Mesh project is a good example of this approach.

Linux-Based Firmware, How to Implement a Good Way to Update