Take two of #40708 (4fe2898608).
That PR attempted to bidirectionally default `config.nixpkgs.system` and
`config.nixpkgs.localSystem.system` to each be updated by the other. But
this is not possible with the way the module system works. Divergence in
certain cases in inevitable.
This PR is more conservative and just has `system` default `localSystem`
and `localSystem` make the final call as-is. This solves a number of
issues.
- `localSystem` completely overrides `system`, just like with nixpkgs
proper. There is no need to specify `localSystem.system` to clobber the
old system.
- `config.nixpkgs.localSystem` is exactly what is passed to nixpkgs. No
spooky steps.
- `config.nixpkgs.localSystem` is elaborated just as nixpkgs would so
that all attributes are available, not just the ones the user
specified.
The remaining issue is just that `config.nixpkgs.system` doesn't update
based on `config.nixpkgs.localSystem.system`. It should never be
referred to lest it is a bogus stale value because
`config.nixpkgs.localSystem` overwrites it.
Fixes #46320
- `localSystem` is added, it strictly supercedes system
- `crossSystem`'s description mentions `localSystem` (and vice versa).
- No more weird special casing I don't even understand
TEMP
`configuration` seems to be a reference to an argument that was
removed seven years ago in commit 2892aed7.
`configuration.nix` makes it a big more clear what we're referring to.
Ideally the module system could be configured pretty much completely by
the contents of the modules themselves, so add comments about avoiding
complicating it further and possibly removing now-redundant
configurability from the existing interface.
This is useful for adding extra functionality or defaults to _every_
nixos evaluation.
My use case is overriding behaviour for all nixos tests, for example
setting packageOverrides to newer versions and changing some default
dependencies/settings.
By making this accessible through an environment variable, this can now
be fully accomplished externally. No more need to fork
nixos/nixpkgs (which becomes a maintenance burden), just use the channel
instead and plug in via this envvar.
You can now say:
systemd.containers.foo.config =
{ services.openssh.enable = true;
services.openssh.ports = [ 2022 ];
users.extraUsers.root.openssh.authorizedKeys.keys = [ "ssh-dss ..." ];
};
which defines a NixOS instance with the given configuration running
inside a lightweight container.
You can also manage the configuration of the container independently
from the host:
systemd.containers.foo.path = "/nix/var/nix/profiles/containers/foo";
where "path" is a NixOS system profile. It can be created/updated by
doing:
$ nix-env --set -p /nix/var/nix/profiles/containers/foo \
-f '<nixos>' -A system -I nixos-config=foo.nix
The container configuration (foo.nix) should define
boot.isContainer = true;
to optimise away the building of a kernel and initrd. This is done
automatically when using the "config" route.
On the host, a lightweight container appears as the service
"container-<name>.service". The container is like a regular NixOS
(virtual) machine, except that it doesn't have its own kernel. It has
its own root file system (by default /var/lib/containers/<name>), but
shares the Nix store of the host (as a read-only bind mount). It also
has access to the network devices of the host.
Currently, if the configuration of the container changes, running
"nixos-rebuild switch" on the host will cause the container to be
rebooted. In the future we may want to send some message to the
container so that it can activate the new container configuration
without rebooting.
Containers are not perfectly isolated yet. In particular, the host's
/sys/fs/cgroup is mounted (writable!) in the guest.
The major changes are:
* The evaluation is now driven by the declared options. In
particular, this fixes the long-standing problem with lack of
laziness of disabled option definitions. Thus, a configuration like
config = mkIf false {
environment.systemPackages = throw "bla";
};
will now evaluate without throwing an error. This also improves
performance since we're not evaluating unused option definitions.
* The implementation of properties is greatly simplified.
* There is a new type constructor "submodule" that replaces
"optionSet". Unlike "optionSet", "submodule" gets its option
declarations as an argument, making it more like "listOf" and other
type constructors. A typical use is:
foo = mkOption {
type = type.attrsOf (type.submodule (
{ config, ... }:
{ bar = mkOption { ... };
xyzzy = mkOption { ... };
}));
};
Existing uses of "optionSet" are automatically mapped to
"submodule".
* Modules are now checked for unsupported attributes: you get an error
if a module contains an attribute other than "config", "options" or
"imports".
* The new implementation is faster and uses much less memory.
