U-Boot Design Principles

The 10 Golden Rules of U-Boot design

Keep it Small

U-Boot is a Boot Loader, i.e. its primary purpose in the shipping system is to load some operating system. That means that U-Boot is necessary to perform a certain task, but it’s nothing you want to throw any significant resources at. Typically U-Boot is stored in relatively small NOR flash memory, which is expensive compared to the much larger NAND devices often used to store the operating system and the application.

At the moment, U-Boot supports boards with just 128 KiB ROM or with 256 KiB NOR flash. We should not easily ignore such configurations - they may be the exception in among all the other supported boards, but if a design uses such a resource-constrained hardware setup it is usually because costs are critical, i. e. because the number of manufactured boards might be tens or hundreds of thousands or even millions…

A usable and useful configuration of U-Boot, including a basic interactive command interpreter, support for download over Ethernet and the capability to program the flash shall fit in no more than 128 KiB.

Keep it Fast

The end user is not interested in running U-Boot. In most embedded systems they are not even aware that U-Boot exists. The user wants to run some application code, and that as soon as possible after switching on their device.

It is therefore essential that U-Boot is as fast as possible, especially that it loads and boots the operating system as fast as possible.

To achieve this, the following design principles shall be followed:

  • Enable caches as soon and whenever possible

  • Initialize devices only when they are needed within U-Boot, i.e. don’t initialize the Ethernet interface(s) unless U-Boot performs a download over Ethernet; don’t initialize any IDE or USB devices unless U-Boot actually tries to load files from these, etc. (and don’t forget to shut down these devices after using them - otherwise nasty things may happen when you try to boot your OS).

Also, building of U-Boot shall be as fast as possible. This makes it easier to run a build for all supported configurations or at least for all configurations of a specific architecture, which is essential for quality assurance. If building is cumbersome and slow, most people will omit this important step.

Keep it Simple

U-Boot is a boot loader, but it is also a tool used for board bring-up, for production testing, and for other activities.

Keep it Portable

U-Boot is a boot loader, but it is also a tool used for board bring-up, for production testing, and for other activities that are very closely related to hardware development. So far, it has been ported to several hundreds of different boards on about 30 different processor families - please make sure that any code you add can be used on as many different platforms as possible.

Avoid assembly language whenever possible - only the reset code with basic CPU initialization, maybe a static DRAM initialization and the C stack setup should be in assembly. All further initializations should be done in C using assembly/C subroutines or inline macros. These functions represent some kind of HAL functionality and should be defined consistently on all architectures, e.g. basic MMU and cache control, stack pointer manipulation. Non-existing functions should expand into empty macros or error codes.

Don’t make assumptions about the environment where U-Boot is running. It may be communicating with a human operator on directly attached serial console, but it may be through a GSM modem as well, or driven by some automatic test or control system. So don’t output any fancy control character sequences or similar.

Keep it Configurable

Section “Keep it Small” already explains about the size restrictions for U-Boot on one side. On the other side, U-Boot is a powerful tool with many, many extremely useful features. The maintainer or user of each board will have to decide which features are important to them and what shall be included with their specific board configuration to meet their current requirements and restrictions.

Please make sure that it is easy to add or remove features from a board configuration, so everybody can make the best use of U-Boot on their system.

If a feature is not included, it should not have any residual code bloating the build.

Keep it Debuggable

Of course debuggable code is a big benefit for all of us contributing in one way or another to the development of the U-Boot project. But as already mentioned in section “Keep it Portable” above, U-Boot is not only a tool in itself, it is often also used for hardware bring-up, so debugging U-Boot often means that we don’t know if we are tracking down a problem in the U-Boot software or in the hardware we are running on. Code that is clean and easy to understand and to debug is all the more important to many of us.

  • One important feature of U-Boot is to enable output to the (usually serial) console as soon as possible in the boot process, even if this causes tradeoffs in other areas like memory footprint.

  • All initialization steps shall print some “begin doing this” message before they actually start, and some “done” message when they complete. For example, RAM initialization and size detection may print a “RAM: “ before they start, and “256 MB\n” when done. The purpose of this is that you can always see which initialization step was running if there should be any problem. This is important not only during software development, but also for the service people dealing with broken hardware in the field.

  • U-Boot should be debuggable with simple JTAG or BDM equipment. It shall use a simple, single-threaded execution model. Avoid any magic, which could prevent easy debugging even when only 1 or 2 hardware breakpoints are available.

Keep it Usable

Please always keep in mind that there are at least three different groups of users for U-Boot, with completely different expectations and requirements:

  • The end user of an embedded device just wants to run some application; they do not even want to know that U-Boot exists and only rarely interacts with it (for example to perform a reset to factory default settings etc.)

  • System designers and engineers working on the development of the application and/or the operating system want a powerful tool that can boot from any boot device they can imagine, they want it fast and scriptable and whatever - in short, they want as many features supported as possible. And some more.

  • The engineer who ports U-Boot to a new board and the board maintainer want U-Boot to be as simple as possible so porting it to and maintaining it on their hardware is easy for them.

  • Make it easy to test. Add debug code (but don’t re-invent the wheel - use existing macros like log_debug() or debug() depending on context).

Please always keep in mind that U-Boot tries to meet all these different requirements.

Keep it Maintainable

  • Avoid #ifdefs where possible

  • Use “weak” functions

  • Always follow the U-Boot Coding Style requirements.

Keep it Beautiful

  • Keep the source code clean: strictly follow the U-Boot Coding Style, keep lists (target names in the Makefiles, board names, etc.) alphabetically sorted, etc.

  • Keep U-Boot console output clean: output only really necessary information, be terse but precise, keep output vertically aligned, do not use control character sequences (e.g. backspaces or \r to do “spinning wheel” activity indicators), etc.

Keep it Open

Contribute your work back to the whole community. Submit your changes and extensions as patches to the U-Boot mailing list.

Lemmas from the golden rules

Generic Code is Good Code

New code shall be as generic as possible and added to the U-Boot abstraction hierarchy as high as possible. As few code as possible shall be added in board directories as people usually do not expect re-usable code there. Thus peripheral drivers should be put below “drivers” even if they start out supporting only one specific configuration. Note that it is not a requirement for such a first instance to be generic as genericity generally cannot be extrapolated from a single data point.