Compiling Non-BEAM Code

It's almost guaranteed that you'll have some code in your project that won't be written in Elixir, Erlang, or another BEAM language. Nerves provides multiple ways of integrating this code and the one you choose depends on many things.

Here are rules of thumb:

1. Build large and complicated C and C++ projects using Buildroot by creating a Custom system
2. Build small C and C++ projects using elixir_make
3. Look for libraries like zigler for specific languages
4. When hope is lost, compile the programs outside of Nerves and include the binaries in a priv directory. Static linking is recommended.

In a perfect world, it would be easy to use whatever language you wanted and adding a program would be as simple as adding a reference to it to your mix deps. Sadly, that's not the case for embedded systems and sometimes an inferior library may be preferable just because it carries fewer dependencies or is easier to build.

Be aware of the following caveats with Nerves:

1. Nerves does not use the embedded Linux init systems like systemd or BusyBox init. Initialization is done in either an Application.start callback or in a GenServer so that it can be supervised.
2. D-Bus is not normally enabled on Nerves. It may be enabled in a custom system.
3. X Windows is not used. Again, it may be enabled, but it is far more common to have UI applications be fullscreen and not use a window manager.
4. Only a few commands are available to shell scripts. You're encouraged to use Elixir instead, but if that's not feasible, it's possible to add missing commands by enabling them in Busybox in a custom system.

Before you even start, experience has shown that searching the Erlang/OTP docs three times and skimming the Erlang source lead to all kinds of amazing discoveries that may not require you to port any code at all. If you do need to port code, keep in mind that while Nerves uses the Linux kernel, it highly favors Erlang/OTP ways of building systems and not embedded Linux ways. If you find yourself continually fighting Nerves and missing embedded Linux, your use case may be better met by installing Elixir on embedded Linux rather than trying to make Nerves look more like embedded Linux. Many embedded Elixir libraries work fine on both Nerves and embedded Linux.

Compilation environment variables

When compiling non-BEAM code, Nerves sets environment variables to guide compilation. These environment variables are available to mix, rebar3 and any code invoked from them. For example, these are frequently used in the Makefiles invoked by elixir_make.

Also see the elixir_make documentation for additional environment variables that may be useful.

Target CPU, ARCH, OS, and ABI

The TARGET_* variables are optionally set by the Nerves system. All official Nerves systems set them, but it is not mandatory for forks. These variables are useful for guiding compilation of LLVM-based tools.

The current way of deriving their values is to use zig and to select the combination that makes most sense for the target. To view the options, install zig and run:

zig targets | less

These variables are defined as custom environment variables in the Nerves system's mix.exs. For example, the following is the definition for the Raspberry Pi Zero:

  defp nerves_package do
    [
      type: :system,
      ...
      env: [
        {"TARGET_ARCH", "arm"},
        {"TARGET_CPU", "arm1176jzf_s"},
        {"TARGET_OS", "linux"},
        {"TARGET_ABI", "gnueabihf"}
      ]
      ...
    ]
  end

Library recommendations

In general, most Elixir and Erlang libraries that include NIFs and ports can be made to work with Nerves. Nerves is, however, less forgiving than normal compilation.

Three recommendations cannot be stressed enough:

First, always compile under _build. While it's much easier to compile in the source directory, this always leads to errors where an executable compiled for one architecture (the host) ends up being put on the target. Nerves will fail with an error when this happens, but it causes a lot of confusion.

Second, do not have a priv directory in your source tree. While Elixir provides a shortcut for copying files from a source priv directory to the build output priv directory, experience has been that this feature causes confusion when building native code. If you do have static assets that you want in the output priv directory, add a line to your Makefile or mix.exs to copy them.

Third, if you have the choice between using a NIF or a port to interface external code with Erlang VM, ports offer the benefit of safety since they run in an OS process. In other words, if the port crashes, Linux cleans up the mess. If a NIF crashes on Nerves, the BEAM crashes and Nerves reboots the device.

The Internet has many examples of how to write NIFs. For an example Makefile that works well with Nerves and embedded Linux, see the circuits_i2c Makefile. Also consider zigler for a safer alternative to C and C++ that works with Nerves.