Cross-platform Compilation
On any Linux platform, there are two ways to do cross-platform compilation. For example, to build an aarch64-linux
program on an x86_64-linux
host, you can use the following methods:
- Use the cross-compilation toolchain to compile the
aarch64
program.- The disadvantage is that you cannot use the NixOS binary cache, and you need to compile everything yourself (cross-compilation also has a cache, but there is basically nothing in it).
- The advantages are that you don't need to emulate the instruction set, and the performance is high.
- Use QEMU to emulate the
aarch64
architecture and then compile the program in the emulator.- The disadvantage is that the instruction set is emulated, and the performance is poor.
- The advantage is that you can use the NixOS binary cache, and you don't need to compile everything yourself.
If you use method one, you don't need to enable binfmt_misc
, but you need to execute the compilation through the cross-compilation toolchain.
If you use method two, you need to enable the binfmt_misc
of the aarch64
architecture in the NixOS configuration of the building machine.
Cross Compilation
nixpkgs
provides a set of predefined host platforms for cross-compilation called pkgsCross
. You can explore them in nix repl
.
› nix repl '<nixpkgs>'
warning: future versions of Nix will require using `--file` to load a file
Welcome to Nix 2.13.3. Type :? for help.
Loading installable ''...
Added 19273 variables.
nix-repl> pkgsCross.<TAB>
pkgsCross.aarch64-android pkgsCross.msp430
pkgsCross.aarch64-android-prebuilt pkgsCross.musl-power
pkgsCross.aarch64-darwin pkgsCross.musl32
pkgsCross.aarch64-embedded pkgsCross.musl64
pkgsCross.aarch64-multiplatform pkgsCross.muslpi
pkgsCross.aarch64-multiplatform-musl pkgsCross.or1k
pkgsCross.aarch64be-embedded pkgsCross.pogoplug4
pkgsCross.arm-embedded pkgsCross.powernv
pkgsCross.armhf-embedded pkgsCross.ppc-embedded
pkgsCross.armv7a-android-prebuilt pkgsCross.ppc64
pkgsCross.armv7l-hf-multiplatform pkgsCross.ppc64-musl
pkgsCross.avr pkgsCross.ppcle-embedded
pkgsCross.ben-nanonote pkgsCross.raspberryPi
pkgsCross.fuloongminipc pkgsCross.remarkable1
pkgsCross.ghcjs pkgsCross.remarkable2
pkgsCross.gnu32 pkgsCross.riscv32
pkgsCross.gnu64 pkgsCross.riscv32-embedded
pkgsCross.i686-embedded pkgsCross.riscv64
pkgsCross.iphone32 pkgsCross.riscv64-embedded
pkgsCross.iphone32-simulator pkgsCross.rx-embedded
pkgsCross.iphone64 pkgsCross.s390
pkgsCross.iphone64-simulator pkgsCross.s390x
pkgsCross.loongarch64-linux pkgsCross.sheevaplug
pkgsCross.m68k pkgsCross.vc4
pkgsCross.mingw32 pkgsCross.wasi32
pkgsCross.mingwW64 pkgsCross.x86_64-darwin
pkgsCross.mips-linux-gnu pkgsCross.x86_64-embedded
pkgsCross.mips64-linux-gnuabi64 pkgsCross.x86_64-freebsd
pkgsCross.mips64-linux-gnuabin32 pkgsCross.x86_64-netbsd
pkgsCross.mips64el-linux-gnuabi64 pkgsCross.x86_64-netbsd-llvm
pkgsCross.mips64el-linux-gnuabin32 pkgsCross.x86_64-unknown-redox
pkgsCross.mipsel-linux-gnu
pkgsCross.mmix
If you want to set pkgs
to a cross-compilation toolchain globally in a flake, you only need to add a Module in flake.nix
, as shown below:
{
description = "NixOS running on LicheePi 4A";
inputs = {
nixpkgs.url = "github:nixos/nixpkgs/nixos-23.05";
};
outputs = inputs@{ self, nixpkgs, ... }: {
nixosConfigurations.lp4a = nixpkgs.lib.nixosSystem {
# native platform
system = "x86_64-linux";
modules = [
# add this module, to enable cross-compilation.
{
nixpkgs.crossSystem = {
# target platform
system = "riscv64-linux";
};
}
# ...... other modules
];
};
};
}
The nixpkgs.crossSystem
option is used to set pkgs
to a cross-compilation toolchain, so that all the contents built will be riscv64-linux
architecture.
Compile through emulated system
The second method is to cross-compile through the emulated system. This method does not require a cross-compilation toolchain.
To use this method, first your building machine needs to enable the binfmt_misc module in the configuration. If your building machine is NixOS, add the following configuration to your NixOS Module to enable the simulated build system of aarch64-linux
and riscv64-linux
architectures:
{ ... }:
{
# ......
# Enable binfmt emulation.
boot.binfmt.emulatedSystems = [ "aarch64-linux" "riscv64-linux" ];
# ......
}
As for flake.nix
, its setting method is very simple, even simpler than the setting of cross-compilation, as shown below:
{
description = "NixOS running on LicheePi 4A";
inputs = {
nixpkgs.url = "github:nixos/nixpkgs/nixos-23.05";
};
outputs = inputs@{ self, nixpkgs, ... }: {
nixosConfigurations.lp4a = nixpkgs.lib.nixosSystem {
# native platform
system = "riscv64-linux";
modules = [
# ...... other modules
];
};
};
}
You do not need to add any additional modules, just specify system
as riscv64-linux
. Nix will automatically detect whether the current system is riscv64-linux
during the build. If not, it will automatically build through the emulated system(QEMU). For users, these underlying operations are completely transparent.
Linux binfmt_misc
The previous section only provided an introduction on how to use Nix's emulated system, but if you want to understand the underlying details, here's a brief introduction.
binfmt_misc
is a feature of the Linux kernel, which stands for Kernel Support for miscellaneous Binary Formats. It enables Linux to run programs for almost any CPU architecture, including X86_64, ARM64, RISCV64, and more.
To enable binfmt_misc
to run programs in various formats, two things are required: a specific identification method for the binary format and the location of the corresponding interpreter. Although binfmt_misc
sounds powerful, its implementation is surprisingly easy to understand. It works similarly to how the Bash interpreter determines the interpreter to use by reading the first line of a script file (e.g., #!/usr/bin/env python3
). binfmt_misc
defines a set of rules, such as reading the magic number at a specific location in the binary file or determining the executable file format based on the file extension (e.g., .exe, .py). It then invokes the corresponding interpreter to execute the program. The default executable file format in Linux is ELF, but binfmt_misc
expands the execution possibilities by allowing a wide range of binary files to be executed using their respective interpreters.
To register a binary program format, you need to write a line in the format :name:type:offset:magic:mask:interpreter:flags
to the /proc/sys/fs/binfmt_misc/register
file. The detailed explanation of the format is beyond the scope of this discussion.
Since manually writing the registration information for binfmt_misc
can be cumbersome, the community provides a container to assist with automatic registration. This container is called binfmt
and running it will install various binfmt_misc
emulators. Here's an example:
# Register all architectures
podman run --privileged --rm tonistiigi/binfmt:latest --install all
# Register only common arm/riscv architectures
docker run --privileged --rm tonistiigi/binfmt --install arm64,riscv64,arm
The binfmt_misc
module was introduced in Linux version 2.6.12-rc2 and has undergone several minor changes in functionality since then. In Linux 4.8, the "F" (fix binary) flag was added, allowing the interpreter to be invoked correctly in mount namespaces and chroot environments. To work properly in containers where multiple architectures need to be built, the "F" flag is necessary. Therefore, the kernel version needs to be 4.8 or above.
In summary, binfmt_misc
provides transparency compared to explicitly calling an interpreter to execute non-native architecture programs. With binfmt_misc
, users no longer need to worry about which interpreter to use when running a program. It allows programs of any architecture to be executed directly. The configurable "F" flag is an added benefit, as it loads the interpreter program into memory during installation and remains unaffected by subsequent environment changes.
Custom build toolchain
Sometimes we may need to use a custom toolchain for building, such as using our own gcc, or using our own musl libc, etc. This modification can be achieved through overlays.
For example, let's try to use a different version of gcc, and test it through nix repl
:
```shell
› nix repl -f '<nixpkgs>'
Welcome to Nix 2.13.3. Type :? for help.
Loading installable ''...
Added 17755 variables.
# replace gcc through overlays, this will create a new instance of nixpkgs
nix-repl> a = import <nixpkgs> { crossSystem = { config = "riscv64-unknown-linux-gnu"; }; overlays = [ (self: super: { gcc = self.gcc12; }) ]; }
# check the gcc version, it is indeed changed to 12.2
nix-repl> a.pkgsCross.riscv64.stdenv.cc
«derivation /nix/store/jjvvwnf3hzk71p65x1n8bah3hrs08bpf-riscv64-unknown-linux-gnu-stage-final-gcc-wrapper-12.2.0.drv»
# take a look at the default pkgs, it is still 11.3
nix-repl> pkgs.pkgsCross.riscv64.stdenv.cc
«derivation /nix/store/pq3g0wq3yfc4hqrikr03ixmhqxbh35q7-riscv64-unknown-linux-gnu-stage-final-gcc-wrapper-11.3.0.drv»
So how to use this method in Flakes? The example flake.nix
is as follows:
{
description = "NixOS running on LicheePi 4A";
inputs = {
nixpkgs.url = "github:nixos/nixpkgs/nixos-23.05-small";
};
outputs = { self, nixpkgs, ... }:
{
nixosConfigurations.lp4a = nixpkgs.lib.nixosSystem {
system = "x86_64-linux";
modules = [
{
nigpkgs.crossSystem = {
config = "riscv64-unknown-linux-gnu";
};
# replace gcc with gcc12 through overlays
nixpkgs.overlays = [ (self: super: { gcc = self.gcc12; }) ];
}
# other moduels ......
];
};
};
}
nixpkgs.overlays
is used to modify the pkgs
instance globally, and the modified pkgs
instance will take effect to the whole flake. It will likely cause a large number of cache missing, and thus require building a large number of Nix packages locally.
To avoid this problem, a better way is to create a new pkgs
instance, and only use this instance when building the packages we want to modify. The example flake.nix
is as follows:
{
description = "NixOS running on LicheePi 4A";
inputs = {
nixpkgs.url = "github:nixos/nixpkgs/nixos-23.05-small";
};
outputs = { self, nixpkgs, ... }: let
# create a new pkgs instance with overlays
pkgs-gcc12 = import nixpkgs {
localSystem = "x86_64-linux";
crossSystem = {
config = "riscv64-unknown-linux-gnu";
};
overlays = [
(self: super: { gcc = self.gcc12; })
];
};
in {
nixosConfigurations.lp4a = nixpkgs.lib.nixosSystem {
system = "x86_64-linux";
specialArgs = {
# pass the new pkgs instance to the module
inherit pkgs-gcc12;
};
modules = [
{
nigpkgs.crossSystem = {
config = "riscv64-unknown-linux-gnu";
};
}
({pkgs-gcc12, ...}: {
# use the custom pkgs instance to build the package hello
environment.systemPackages = [ pkgs-gcc12.hello ];
})
# other moduels ......
];
};
};
}
Through the above method, we can easily customize the build toolchain of some packages without affecting the build of other packages.