The old config generation system used impure shell scripts and could break in specific circumstances (see #1234).
### Writing good commit messages
In addition to writing properly formatted commit messages, it's important to include relevant information so other developers can later understand *why* a change was made. While this information usually can be found by digging code, mailing list/Discourse archives, pull request discussions or upstream changes, it may require a lot of work.
Package version upgrades usually allow for simpler commit messages, including attribute name, old and new version, as well as a reference to the relevant release notes/changelog. Every once in a while a package upgrade requires more extensive changes, and that subsequently warrants a more verbose message.
Pull requests should not be squash merged in order to keep complete commit messages and GPG signatures intact and must not be when the change doesn't make sense as a single commit.
This means that, when addressing review comments in order to keep the pull request in an always mergeable status, you will sometimes need to rewrite your branch's history and then force-push it with `git push --force-with-lease`.
Useful git commands that can help a lot with this are `git commit --patch --amend` and `git rebase --interactive`. For more details consult the git man pages or online resources like [git-rebase.io](https://git-rebase.io/) or [The Pro Git Book](https://git-scm.com/book/en/v2/Git-Tools-Rewriting-History).
#### Rebasing between branches (i.e. from master to staging)
From time to time, changes between branches must be rebased, for example, if the
number of new rebuilds they would cause is too large for the target branch. When
rebasing, care must be taken to include only the intended changes, otherwise
many CODEOWNERS will be inadvertently requested for review. To achieve this,
rebasing should not be performed directly on the target branch, but on the merge
base between the current and target branch. As an additional precautionary measure,
you should temporarily mark the PR as draft for the duration of the operation.
This reduces the probability of mass-pinging people. (OfBorg might still
request a couple of persons for reviews though.)
In the following example, we assume that the current branch, called `feature`,
is based on `master`, and we rebase it onto the merge base between
`master` and `staging` so that the PR can eventually be retargeted to
`staging` without causing a mess. The example uses `upstream` as the remote for `NixOS/nixpkgs.git`
while `origin` is the remote you are pushing to.
```console
# Rebase your commits onto the common merge base
git rebase --onto upstream/staging... upstream/master
# Force push your changes
git push origin feature --force-with-lease
```
The syntax `upstream/staging...` is equivalent to `upstream/staging...HEAD` and
stands for the merge base between `upstream/staging` and `HEAD` (hence between
`upstream/staging` and `upstream/master`).
Then change the base branch in the GitHub PR using the *Edit* button in the upper
right corner, and switch from `master` to `staging`. *After* the PR has been
retargeted it might be necessary to do a final rebase onto the target branch, to
resolve any outstanding merge conflicts.
```console
# Rebase onto target branch
git rebase upstream/staging
# Review and fixup possible conflicts
git status
# Force push your changes
git push origin feature --force-with-lease
```
##### Something went wrong and a lot of people were pinged
It happens. Remember to be kind, especially to new contributors.
There is no way back, so the pull request should be closed and locked
(if possible). The changes should be re-submitted in a new PR, in which the people
originally involved in the conversation need to manually be pinged again.
No further discussion should happen on the original PR, as a lot of people
are now subscribed to it.
The following message (or a version thereof) might be left when closing to
describe the situation, since closing and locking without any explanation
is kind of rude:
```markdown
It looks like you accidentally mass-pinged a bunch of people, which are now subscribed
and getting notifications for everything in this pull request. Unfortunately, they
cannot be automatically unsubscribed from the issue (removing review request does not
unsubscribe), therefore development cannot continue in this pull request anymore.
Please open a new pull request with your changes, link back to this one and ping the
people actually involved in here over there.
In order to avoid this in the future, there are instructions for how to properly
rebase between branches in our [contribution guidelines](https://github.com/NixOS/nixpkgs/blob/master/CONTRIBUTING.md#rebasing-between-branches-ie-from-master-to-staging).
Setting your pull request to draft prior to rebasing is strongly recommended.
In draft status, you can preview the list of people that are about to be requested
for review, which allows you to sidestep this issue.
This is not a bulletproof method though, as OfBorg still does review requests even on draft PRs.
```
### Backporting changes
Follow these steps to backport a change into a release branch in compliance with the [commit policy](https://nixos.org/nixpkgs/manual/#submitting-changes-stable-release-branches).
You can add a label such as `backport release-23.05` to a PR, so that merging it will
automatically create a backport (via [a GitHub Action](.github/workflows/backport.yml)).
This also works for pull requests that have already been merged, and might take a couple of minutes to trigger.
You can also create the backport manually:
1. Take note of the commits in which the change was introduced into `master` branch.
2. Check out the target _release branch_, e.g. `release-23.05`. Do not use a _channel branch_ like `nixos-23.05` or `nixpkgs-23.05-darwin`.
3. Create a branch for your change, e.g. `git checkout -b backport`.
4. When the reason to backport is not obvious from the original commit message, use `git cherry-pick -xe <original commit>` and add a reason. Otherwise use `git cherry-pick -x <original commit>`. That's fine for minor version updates that only include security and bug fixes, commits that fixes an otherwise broken package or similar. Please also ensure the commits exists on the master branch; in the case of squashed or rebased merges, the commit hash will change and the new commits can be found in the merge message at the bottom of the master pull request.
5. Push to GitHub and open a backport pull request. Make sure to select the release branch (e.g. `release-23.05`) as the target branch of the pull request, and link to the pull request in which the original change was committed to `master`. The pull request title should be the commit title with the release version as prefix, e.g. `[23.05]`.
6. When the backport pull request is merged and you have the necessary privileges you can also replace the label `9.needs: port to stable` with `8.has: port to stable` on the original pull request. This way maintainers can keep track of missing backports easier.
#### Criteria for Backporting changes
Anything that does not cause user or downstream dependency regressions can be backported. This includes:
- New Packages / Modules
- Security / Patch updates
- Version updates which include new functionality (but no breaking changes)
- Services which require a client to be up-to-date regardless. (E.g. `spotify`, `steam`, or `discord`)
### File naming and organisation {#sec-organisation}
Names of files and directories should be in lowercase, with dashes between words — not in camel case. For instance, it should be `all-packages.nix`, not `allPackages.nix` or `AllPackages.nix`.
### Syntax {#sec-syntax}
- Use 2 spaces of indentation per indentation level in Nix expressions, 4 spaces in shell scripts.
- Do not use tab characters, i.e. configure your editor to use soft tabs. For instance, use `(setq-default indent-tabs-mode nil)` in Emacs. Everybody has different tab settings so it’s asking for trouble.
- Use `lowerCamelCase` for variable names, not `UpperCamelCase`. Note, this rule does not apply to package attribute names, which instead follow the rules in [](#sec-package-naming).
- Function calls with attribute set arguments are written as
```nix
foo {
arg = ...;
}
```
not
```nix
foo
{
arg = ...;
}
```
Also fine is
```nix
foo { arg = ...; }
```
if it's a short call.
- In attribute sets or lists that span multiple lines, the attribute names or list elements should be aligned:
```nix
# A long list.
list = [
elem1
elem2
elem3
];
# A long attribute set.
attrs = {
attr1 = short_expr;
attr2 =
if true then big_expr else big_expr;
};
# Combined
listOfAttrs = [
{
attr1 = 3;
attr2 = "fff";
}
{
attr1 = 5;
attr2 = "ggg";
}
];
```
- Short lists or attribute sets can be written on one line:
```nix
# A short list.
list = [ elem1 elem2 elem3 ];
# A short set.
attrs = { x = 1280; y = 1024; };
```
- Breaking in the middle of a function argument can give hard-to-read code, like
```nix
someFunction { x = 1280;
y = 1024; } otherArg
yetAnotherArg
```
(especially if the argument is very large, spanning multiple lines).
Better:
```nix
someFunction
{ x = 1280; y = 1024; }
otherArg
yetAnotherArg
```
or
```nix
let res = { x = 1280; y = 1024; };
in someFunction res otherArg yetAnotherArg
```
- The bodies of functions, asserts, and withs are not indented to prevent a lot of superfluous indentation levels, i.e.
```nix
{ arg1, arg2 }:
assert system == "i686-linux";
stdenv.mkDerivation { ...
```
not
```nix
{ arg1, arg2 }:
assert system == "i686-linux";
stdenv.mkDerivation { ...
```
- Function formal arguments are written as:
```nix
{ arg1, arg2, arg3 }:
```
but if they don't fit on one line they're written as:
```nix
{ arg1, arg2, arg3
, arg4, ...
, # Some comment...
argN
}:
```
- Functions should list their expected arguments as precisely as possible. That is, write
```nix
{ stdenv, fetchurl, perl }: ...
```
instead of
```nix
args: with args; ...
```
or
```nix
{ stdenv, fetchurl, perl, ... }: ...
```
For functions that are truly generic in the number of arguments (such as wrappers around `mkDerivation`) that have some required arguments, you should write them using an `@`-pattern:
... if args ? doCoverageAnalysis && args.doCoverageAnalysis then "bla" else "" ...
})
```
- Unnecessary string conversions should be avoided. Do
```nix
rev = version;
```
instead of
```nix
rev = "${version}";
```
- Building lists conditionally _should_ be done with `lib.optional(s)` instead of using `if cond then [ ... ] else null` or `if cond then [ ... ] else [ ]`.
```nix
buildInputs = lib.optional stdenv.isDarwin iconv;
```
instead of
```nix
buildInputs = if stdenv.isDarwin then [ iconv ] else null;
```
As an exception, an explicit conditional expression with null can be used when fixing a important bug without triggering a mass rebuild.
If this is done a follow up pull request _should_ be created to change the code to `lib.optional(s)`.
- Arguments should be listed in the order they are used, with the exception of `lib`, which always goes first.
