See [the Nixpkgs manual]. This allows modules to declare themselves
as being only for NixOS or nix-darwin, reducing compatibility
risks. Unmarked modules will continue to behave as before. Technically
a breaking change, but any configuration importing a module explicitly
marked as NixOS-specific was probably broken already.
[the Nixpkgs manual]: https://nixos.org/manual/nixpkgs/unstable/#module-system-lib-evalModules-param-class
This is a big change that disentangles a lot of mistaken assumptions
about mixing multiple versions of Nixpkgs, treating external flake
inputs as gospel for the source of Nixpkgs and nix-darwin, etc.;
the end result should be much simpler conceptually, but it will be a
breaking change for anyone using `eval-config.nix` directly. Hopefully
that shouldn't be a big issue, as it is more of an internal API and
it's quite likely that existing uses may have been broken in the same
way the internal ones were.
It was previously easy to get into a state where your `lib` comes
from nix-darwin's `nixpkgs` input or a global channel and your
`pkgs` comes from another major version of Nixpkgs. This is pretty
fundamentally broken due to the coupling of `pkgs` to its corresponding
`lib`, but the brokenness was hidden much of the time until something
surfaced it. Now there is exactly one mandatory `lib` input to system
evaluation, and the handling of various additional options like `pkgs`
and `system` can be done modularly; maintaining backwards compatibility
with the previous calling convention is punted to the `default.nix`
and `lib.darwinSystem` entry points. `inputs` is no longer read by
nix-darwin or special in any way, merely a convention for user code,
and the argument is retained in the entry points only for backwards
compatibility.
All correct invocations of the entry points should keep working
after this change, and some previously-broken ones should be fixed
too. The documentation and template have been adjusted to show the
newly-recommended modular way of specifying various things, but no
deprecation warnings have been introduced yet by this change.
There is one potential, mostly cosmetic regression:
`system.nixpkgsRevision` and related options are less likely to be
set than before, in cases where it is not possible to determine the
origin of the package set. Setting `nixpkgs.source` explicitly will
make this work again, and I hope to look into sending changes upstream
to Nixpkgs to make `lib.trivial.revisionWithDefault` behave properly
under flakes, which would fix this regression and potentially allow
reducing some of the complexity.
Fixes: #669
This is based on the current NixOS `nixpkgs` module, adjusted for the
nix-darwin context and without adding the options due for deprecation
in NixOS.
This gives us the ability to set the package set modularly through
`nixpkgs.pkgs` and builds up infrastructure for handling user-specified
Nixpkgs instantiations more robustly.
The cross-compilation options are currently not very useful due to
even Darwin->Darwin cross-compilation not being wholly functional
yet, but it looks feasible to build an `aarch64-darwin` system from
`x86_64-darwin` with some patching and it should be possible to make
cross-compilation more widely supported after the Darwin SDK situation
in Nixpkgs improves.
One casualty is the error for setting `nixpkgs.*` options when
overriding the package set. That could be ported over to this new
scheme, but it'd increase divergence with the NixOS module and reduce
cross-compatibility of configurations, so I lean towards adding it
upstream to NixOS if anything. (But if people want to keep it I can add
it back.)
This is useful for flake users as they will usually already have an
instantiated Nixpkgs e.g.
let
pkgs = import nixpkgs {
config.allowUnfree = true;
overlays = [ ... ];
}
in darwin.lib.darwinSystem {
inherit pkgs;
}
This change makes `nix-darwin` match the behaviour of NixOS and
`home-manager`.
