Welcome to the 20th Nix pill. In the previous 19th pill we introduced Nixpkgs' stdenv, including setup.sh script, default-builder.sh helper script, and stdenv.mkDerivation builder. We focused on how stdenv is put together, and how it's used, and a bit about the phases of genericBuild.

This time, we'll focus on the interaction of packages built with stdenv.mkDerivation. Packages need to depend on each other, of course. For this we have buildInputs and propagatedBuildInputs attributes. We've also found that dependencies sometimes need to influence their dependents in ways the dependents can't or shouldn't predict. For this we have setup hooks and env hooks. Together, these 4 concepts support almost all build-time package interactions.

Note: The complexity of the dependencies and hooks infrastructure has increased, over time, to support cross compilation. Once you learn the core concepts, you will be able to understand the extra complexity. As a starting point, you might want to refer to nixpkgs commit 6675f0a5, the last version of stdenv without cross-compilation complexity.

20.1. The buildInputs Attribute

For the simplest dependencies where the current package directly needs another, we use the buildInputs attribute. This is exactly the pattern in taught with our builder in Pill 8. To demo this, lets build GNU Hello, and then another package which provides a shell script that execs it.

let

  nixpkgs = import <nixpkgs> {};

  inherit (nixpkgs) stdenv fetchurl which;

  actualHello = stdenv.mkDerivation {
    name = "hello-2.3";

    src = fetchurl {
      url = mirror://gnu/hello/hello-2.3.tar.bz2;
      sha256 = "0c7vijq8y68bpr7g6dh1gny0bff8qq81vnp4ch8pjzvg56wb3js1";
    };
  };

  wrappedHello = stdenv.mkDerivation {
    name = "hello-wrapper";

    buildInputs = [ actualHello which ];

    unpackPhase = "true";

    installPhase = ''
      mkdir -p "$out/bin"
      echo "#! ${stdenv.shell}" >> "$out/bin/hello"
      echo "exec $(which hello)" >> "$out/bin/hello"
    '';
  };

in wrappedHello

Notice that the wrappedHello derivation finds the hello binary from the PATH. This works because stdenv contains something like:

pkgs=""
for i in $buildInputs; do
    findInputs $i
done

where findInputs is defined like:

findInputs() {
    local pkg=$1

    ## Don't need to repeat already processed package
    case $pkgs in
        *\ $pkg\ *)
            return 0
            ;;
    esac

    pkgs="$pkgs $pkg "

    ## More goes here in reality that we can ignore for now.
}

then after this is run:

for i in $pkgs; do
    addToEnv $i
done

where addToEnv is defined like:

addToEnv() {
    local pkg=$1

    if test -d $1/bin; then
        addToSearchPath _PATH $1/bin
    fi

    ## More goes here in reality that we can ignore for now.
}

The addToSearchPath call adds $1/bin to _PATH if the former exists (code here). Once all the packages in buildInputs have been processed, then content of _PATH is added to PATH, as follows:

PATH="${_PATH-}${_PATH:+${PATH:+:}}$PATH"

With the real hello on the PATH, the installPhase should hopefully make sense.

20.2. The propagatedBuildInputs Attribute

The buildInputs covers direct dependencies, but what about indirect dependencies where one package needs a second package which needs a third? Nix itself handles this just fine, understanding various dependency closures as covered in previous builds. But what about the conveniences that buildInputs provides, namely accumulating in pkgs environment variable and inclusion of pkg/bin directories on the PATH? For this, stdenv provides the propagatedBuildInputs:

let

  nixpkgs = import <nixpkgs> {};

  inherit (nixpkgs) stdenv fetchurl which;

  actualHello = stdenv.mkDerivation {
    name = "hello-2.3";

    src = fetchurl {
      url = mirror://gnu/hello/hello-2.3.tar.bz2;
      sha256 = "0c7vijq8y68bpr7g6dh1gny0bff8qq81vnp4ch8pjzvg56wb3js1";
    };
  };

  intermediary = stdenv.mkDerivation {
    name = "middle-man";

    propagatedBuildInputs = [ actualHello ];

    unpackPhase = "true";

    installPhase = ''
      mkdir -p "$out"
    '';
  };

  wrappedHello = stdenv.mkDerivation {
    name = "hello-wrapper";

    buildInputs = [ intermediary which ];

    unpackPhase = "true";

    installPhase = ''
      mkdir -p "$out/bin"
      echo "#! ${stdenv.shell}" >> "$out/bin/hello"
      echo "exec $(which hello)" >> "$out/bin/hello"
    '';
  };

in wrappedHello

See how the intermediate package has a propagatedBuildInputs dependency, but the wrapper only needs a buildInputs dependency on the intermediary.

How does this work? You might think we do something in Nix, but actually its done not at eval time but at build time in bash. lets look at part of the fixupPhase of stdenv:

fixupPhase() {

    ## Elided

    if test -n "$propagatedBuildInputs"; then
        mkdir -p "$out/nix-support"
        echo "$propagatedBuildInputs" > "$out/nix-support/propagated-build-inputs"
    fi

    ## Elided

}

This dumps the propagated build inputs in a so-named file in $out/nix-support/. Then, back in findInputs look at the lines at the bottom we elided before:

findInputs() {
    local pkg=$1

    ## More goes here in reality that we can ignore for now.

    if test -f $pkg/nix-support/propagated-build-inputs; then
        for i in $(cat $pkg/nix-support/propagated-build-inputs); do
            findInputs $i
        done
    fi
}

See how findInputs is actually recursive, looking at the propagated build inputs of each dependency, and those dependencies' propagated build inputs, etc.

We actually simplified the findInputs call site from before; propagatedBuildInputs is also looped over in reality:

pkgs=""
for i in $buildInputs $propagatedBuildInputs; do
    findInputs $i
done

This demonstrates an important point. For the current package alone, it doesn't matter whether a dependency is propagated or not. It will be processed the same way: called with findInputs and addToEnv. (The packages discovered by findInputs, which are also accumulated in pkgs and passed to addToEnv, are also the same in both cases.) Downstream however, it certainly does matter because only the propagated immediate dependencies are put in the $out/nix-support/propagated-build-inputs.

20.3. Setup Hooks

As we mentioned above, sometimes dependencies need to influence the packages that use them in ways other than just being a dependency. [1] propagatedBuildInputs can actually be seen as an example of this: packages using that are effectively "injecting" those dependencies as extra buildInputs in their downstream dependents. But in general, a dependency might affect the packages it depends on in arbitrary ways. Arbitrary is the key word here. We could teach setup.sh things about upstream packages like pkg/nix-support/propagated-build-inputs, but not arbitrary interactions.

Setup hooks are the basic building block we have for this. In nixpkgs, a "hook" is basically a bash callback, and a setup hook is no exception. Let's look at the last part of findInputs we haven't covered:

findInputs() {
    local pkg=$1

    ## More goes here in reality that we can ignore for now.

    if test -f $pkg/nix-support/setup-hook; then
        source $pkg/nix-support/setup-hook
    fi

    ## More goes here in reality that we can ignore for now.

}

If a package includes the path pkg/nix-support/setup-hook, it will be sourced by any stdenv-based build including that as a dependency.

This is strictly more general than any of the other mechanisms introduced in this chapter. For example, try writing a setup hook that has the same effect as a propagatedBuildInputs entry. One can almost think of this as an escape hatch around Nix's normal isolation guarantees, and the principle that dependencies are immutable and inert. We're not actually doing something unsafe or modifying dependencies, but we are allowing arbitrary ad-hoc behavior. For this reason, setup-hooks should only be used as a last resort.

20.4. Environment Hooks

As a final convenience, we have environment hooks. Recall in Pill 12 how we created NIX_CFLAGS_COMPILE for -I flags and NIX_LDFLAGS for -L flags, in a similar manner to how we prepared the PATH. One point of ugliness was how anti-modular this was. It makes sense to build the PATH in generic builder, because the PATH is used by the shell, and the generic builder is intrinsically tied to the shell. But -I and -L flags are only relevant to the C compiler. The stdenv isn't wedded to including a C compiler (though it does by default), and there are other compilers too which may take completely different flags.

As a first step, we can move that logic to a setup hook on the C compiler; indeed that's just what we do in CC Wrapper. [2] But this pattern comes up fairly often, so somebody decided to add some helper support to reduce boilerplate.

The other half of addToEnv is:

addToEnv() {
    local pkg=$1

    ## More goes here in reality that we can ignore for now.

    # Run the package-specific hooks set by the setup-hook scripts.
    for i in "${envHooks[@]}"; do
        $i $pkg
    done
}

Functions listed in envHooks are applied to every package passed to addToEnv. One can write a setup hook like:

anEnvHook() {
    local pkg=$1

    echo "I'm depending on \"$pkg\""
}

envHooks+=(anEnvHook)

and if one dependency has that setup hook then all of them will be so echoed. Allowing dependencies to learn about their sibling dependencies is exactly what compilers need.

20.5. Next pill...

...I'm not sure! We could talk about the additional dependency types and hooks which cross compilation necessitates, building on our knowledge here to cover stdenv as it works today. We could talk about how nixpkgs is bootstrapped. Or we could talk about how localSystem and crossSystem are elaborated into the buildPlatform, hostPlatform, and targetPlatform each bootstrapping stage receives. Let us know which most interests you!




[1] We can now be precise and consider what addToEnv does alone the minimal treatment of a dependency: i.e. a package that is just a dependency would only have addToEnv applied to it.

[2] It was called GCC Wrapper in the version of nixpkgs suggested for following along in this pill; Darwin and Clang support hadn't yet motivated the rename.