Skip to content

Welcome to armbian/build, the Armbian build system

Welcome to the Armbian build system (armbian/build). The build system has significantly changed since release 23.02, when the armbian-next effort was merged into our main branch. Please expect some rough edges and report them to us.

Errors, errors everywhere

Previously, when faced with an error, the build system would simply ignore them, hardly log them anywhere, and continue. This led to builds completing successfully, but with (sometimes very sneakily) broken packages and images. Get used to seeing errors stopping builds now. Inspect the whole log, specially the errors, the last few lines, and the stack trace. We try, as much as possible given Bash limitations, to show the source of the error, with a “stack trace” of sorts. The topmost elements of the stack trace are usually the most relevant. Each line is prefixed with the file and line number where the error occurred. Open the file and go to the line number to see the source of the error. Most times, this simple inspection is enough to understand the error.

(ANSI) Logging

Logging is a bit more structured now. Logs are output to output/logs, in a few different formats. We output ANSI color both to the screen and to the logs. Please add SHARE_LOG=yes to share your logs with us when reporting issues, that allows us to check the logs on a web browser and keep to correct formatting.

Command line syntax has changed

General CLI syntax: ./compile.sh PARAM=value OTHER_PARAM=other_value [<configfile> <configfile> ...] [<command>]

  • where command defaults to build if you don’t specify it; could also be kernel-config or u-boot etc
  • you can’t have a config file with the same name as a possible <command> (system will check & bomb if so)
  • also: there’s no more default config – you have to be explicit
  • also: there’s no more docker config – Docker is fully auto-managed now. The system will complain if you have one.
  • you can specify PARAM=value, <configfile> or <command> in any order

No more config-default.conf, you need to specify the config file in the command line

  • No “default” config is auto-loaded anymore. Default config lead to unreproducible failing builds and was a source of confusion.
  • The configs still go to the same place, userpatches/config-xyz.conf – but now you gotta tell the build system to use that config, like ./compile.sh BOARD=xxx xyz; otherwise works the same.

Artifacts, cache, what the …?

The armbian/build system is currently undergoing refactoring to improve its structure. Previously, the build system was a single, complex Bash script that mixed the building of .deb packages with the creation of images.

This was reworked into a 1-to-N image-to-artifact dependency tree; a certain image build will depend on N possible “artifacts”. Artifacts are either .deb packages, a .tar of multiple .deb packages, or a rootfs.tar.zstd. Each artifact can be individually built, and has a specific name and a version.

Each artifact is also now cached by default using OCI storage at ghcr.io (GitHub Container Registry). To achieve consistent caching, each artifact produces a version that includes hashes of its composing files, variables, patches, hooks, external git SHA1 references, etc. That way we can consistently check the remote OCI cache for previously-built artifacts, and possibly save image builders from having to build heavy packages just to produce an image.

TL-DR about artifacts and caching:

  • KERNEL_ONLY=yes and KERNEL_ONLY=no are no more. Use the kernel CLI command instead.
  • ARTIFACT_IGNORE_CACHE=yes can help if you’re getting false positives. Please also report the problem, with a complete logfile.

Automatic Docker/sudo launcher

  • compile.sh will prefer to use Docker if it detects Docker is installed and working. - This handles Docker Desktop and Rancher Desktop (in Docker emulation mode) under macOS/Darwin, including Apple M1/M2. - You don’t need and actually can’t have the old docker config file.
  • If Docker is not installed, it will try to use sudo to run the build as root.
  • If you run directly as root, it will warn you and ask you to run it without sudo.

Kernel Git Trees: shallow vs full

During the build, depending on which local or remote caches are hit, it might be needed to build the Linux Kernel.

The kernel’s git repo is huge. Most build systems resort to fetching “shallow” trees directly from upstream git servers, to save bandwidth. Unfortunately that creates immense extra CPU load on the git servers. To avoid this problem, Armbian produces daily automated git tree exports cached in ghcr.io OCI repositories, and only uses git fetch to update the relatively small new changes from the upstream git server.

There are two types of cached Kernel git trees:

  • full is a complete git tree, including all of Torvald’s master and all of the currently-supported stable branches. - full is very large download and requires a lot of disk space. - full is more useful over time and when building multiple different kernels on the same machine, like for CI servers or developer workstations.
  • shallow is a shallow tree for a specific stable branch - shallow is a much smaller download and requires less disk space - shallow is appropriate for restricted devices like SBCs which will build a single kernel

TL,DR: KERNEL_GIT=full or KERNEL_GIT=shallow or let the system decide for you.

Consider Forking before cloning the repo

Before cloning the repo, consider forking it first. This will allow you to make changes and submit pull requests. You will need a GitHub account to do this; see GitHub’s documentation for more information. If you fork, make sure to keep your fork up-to-date with the main repo, by rebasing your fork.

Some really confusing stuff still remains

This is (by far) not a complete list:

  • wifi/other kernel drivers are still using pre-armbian-next code, and are very hard to work with. it’s not only the contents are a mess, the way the whole thing works leads to more and more compounding work. To make it worse, family patches sometimes need to patch driver code, leading to a cycle of sadness for developers. We’re still coming up with a plan to completely replace this lest most of us go insane.
  • although “aggregation” has been rewritten in Python, it still mostly works using the legacy principle, by scanning directories and files in a very complex and error-prone way. This is a source of many bugs and confusion. We plan to replace this with pure extensions eventually.
  • “minimal” images are far from really minimal. It is common to see “extensions” that actually remove stuff put there by the core. Those should be refactored.
  • board-side scripts (armbian-config, firstrun, hwoptimization, etc) haven’t really changed with armbian-next, and are in dire need of a rewrite. They’re also a source of many bugs and confusion. They also need to be extensible.
  • we’ve mostly working Kernel headers (linux-headers pkg) for 5.10+ including some vendor kernels; no kernel-headers for 4.x kernels. sorry. We’ve no plans to support this.

Multiple u-boot’s for same board

Hmm. This is an interesting problem. We can build u-boot twice, using UBOOT_TARGET_MAP. Some example I did in https://github.com/armbian/build/blob/main/config/boards/odroidhc4.conf#L15-L20 may help.