The structure of this directory maps almost directly to top-level package attributes. Add new top-level packages to Nixpkgs using this mechanism whenever possible.
Packages found in the name-based structure are automatically included, without needing to be added to all-packages.nix
. However if the implicit attribute defaults need to be changed for a package, this must still be declared in all-packages.nix
.
The top-level package pkgs.some-package
may be declared by setting up this file structure:
pkgs
└── by-name
├── so
┊ ├── some-package
┊ └── package.nix
Where some-package
is the attribute name corresponding to the package, and so
is the lowercase 2-letter prefix of the attribute name.
The package.nix
may look like this:
# A function taking an attribute set as an argument
{
# Get access to top-level attributes for use as dependencies
lib,
stdenv,
libbar,
# Make this derivation configurable using `.override { enableBar = true }`
enableBar ? false,
}:
# The return value must be a derivation
stdenv.mkDerivation {
# ...
buildInputs =
lib.optional enableBar libbar;
}
You can also split up the package definition into more files in the same directory if necessary.
Once defined, the package can be built from the Nixpkgs root directory using:
nix-build -A some-package
See the general package conventions for more information on package definitions.
The above expression is called using these arguments by default:
{
lib = pkgs.lib;
stdenv = pkgs.stdenv;
libbar = pkgs.libbar;
}
But the package might need pkgs.libbar_2
instead.
While the function could be changed to take libbar_2
directly as an argument,
this would change the .override
interface, breaking code like .override { libbar = ...; }
.
So instead it is preferable to use the same generic parameter name libbar
and override its value in pkgs/top-level/all-packages.nix
:
{
libfoo = callPackage ../by-name/so/some-package/package.nix {
libbar = libbar_2;
};
}
Most packages are still defined in all-packages.nix
and the category hierarchy.
Since it would take a lot of contributor and reviewer time to migrate all packages manually,
an automated migration is planned,
though it is expected to still take some time to get done.
If you're interested in helping out with this effort,
please see this ticket.
Since only PRs to packages in pkgs/by-name
can be automatically merged,
if package maintainers would like to use this feature, they are welcome to migrate their packages to pkgs/by-name
.
To lessen PR traffic, they're encouraged to also perform some more general maintenance on the package in the same PR,
though this is not required and must not be expected.
Note that definitions in all-packages.nix
with custom arguments should not be removed.
That is a backwards-incompatible change because it changes the .override
interface.
Such packages may still be moved to pkgs/by-name
however, while keeping the definition in all-packages.nix
.
See also changing implicit attribute defaults.
There's some limitations as to which packages can be defined using this structure:
-
Only packages defined using
pkgs.callPackage
. This excludes packages defined usingpkgs.python3Packages.callPackage ...
.Instead:
- Either change the package definition to work with
pkgs.callPackage
. - Or use the category hierarchy.
- Either change the package definition to work with
-
Only top-level packages. This excludes packages for other package sets like
pkgs.pythonPackages.*
.Refer to the definition and documentation of the respective package set to figure out how such packages can be declared.
CI performs certain checks on the pkgs/by-name
structure.
This is done using the nixpkgs-vet
tool.
You can locally emulate the CI check using
$ ./ci/nixpkgs-vet.sh master
See here for more info.
These checks of the pkgs/by-name
structure can cause problems in combination:
- New top-level packages using
callPackage
must be defined viapkgs/by-name
. - Packages in
pkgs/by-name
cannot refer to files outside their own directory.
This means that outside pkgs/by-name
, multiple already-present top-level packages can refer to some common file.
If you open a PR to another instance of such a package, CI will fail check 1,
but if you try to move the package to pkgs/by-name
, it will fail check 2.
This is often the case for packages with multiple versions, such as
{
foo_1 = callPackage ../tools/foo/1.nix { };
foo_2 = callPackage ../tools/foo/2.nix { };
}
The best way to resolve this is to not use callPackage
directly, such that check 1 doesn't trigger.
This can be done by using inherit
on a local package set:
{
inherit
({
foo_1 = callPackage ../tools/foo/1.nix { };
foo_2 = callPackage ../tools/foo/2.nix { };
})
foo_1
foo_2
;
}
While this may seem pointless, this can in fact help with future package set refactorings, because it establishes a clear connection between related attributes.
This is not required, but the above solution also allows refactoring the definitions into a separate file:
{
inherit (import ../tools/foo pkgs)
foo_1 foo_2;
}
# pkgs/tools/foo/default.nix
pkgs: {
foo_1 = callPackage ./1.nix { };
foo_2 = callPackage ./2.nix { };
}
Alternatively using callPackages
if callPackage
isn't used underneath and you want the same .override
arguments for all attributes:
{
inherit (callPackages ../tools/foo { })
foo_1 foo_2;
}
# pkgs/tools/foo/default.nix
{
stdenv
}: {
foo_1 = stdenv.mkDerivation { /* ... */ };
foo_2 = stdenv.mkDerivation { /* ... */ };
}
This is not required, but the above solution also allows exposing the package set as an attribute:
{
foo-versions = import ../tools/foo pkgs;
# Or using callPackages
# foo-versions = callPackages ../tools/foo { };
inherit (foo-versions) foo_1 foo_2;
}