Welcome to the first post of the "Nix in pills" series. Nix is a purely functional package manager and deployment system for POSIX.

There's a lot of documentation that describes what Nix, NixOS and related projects are. But the purpose of this post is to convince you to give Nix a try. Installing NixOS is not required, but sometimes I may refer to NixOS as a real world example of Nix usage for building a whole operating system.

The Nix, Nixpkgs, and NixOS manuals; and wiki are excellent resources for explaining how Nix/NixOS works, how you can use it, and the number of cool things being done with it. However, at the beginning you may feel that some of the magic which happens behind the scenes is hard to grasp.

This series aims to complement the existing explanations from the more formal documents.

The following is a description of Nix. Just as with pills, I'll try to be as short as possible.

Most, if not all, widely used package managers (dpkg, rpm, ...) mutate the global state of the system. If a package foo-1.0 installs a program to /usr/bin/foo, you cannot install foo-1.1 as well, unless you change the installation paths or the binary name. But changing the binary names means breaking users of that binary.

There are some attempts to mitigate this problem. Debian, for example, partially solves the problem with the alternatives system.

So while in theory it's possible with some current systems to install multiple versions of the same package, in practice it's very painful.

Let's say you need an nginx service and also an nginx-openresty service. You have to create a new package that changes all the paths to have, for example, an -openresty suffix.

Or suppose that you want to run two different instances of mysql: 5.2 and 5.5. The same thing applies, plus you have to also make sure the two mysqlclient libraries do not collide.

This is not impossible but it is very inconvenient. If you want to install two whole stacks of software like GNOME 3.10 and GNOME 3.12, you can imagine the amount of work.

From an administrator's point of view: you can use containers. The typical solution nowadays is to create a container per service, especially when different versions are needed. That somewhat solves the problem, but at a different level and with other drawbacks. For example, needing orchestration tools, setting up a shared cache of packages, and new machines to monitor rather than simple services.

From a developer's point of view: you can use virtualenv for python, or jhbuild for gnome, or whatever else. But then how do you mix the two stacks? How do you avoid recompiling the same thing when it could instead be shared? Also you need to set up your development tools to point to the different directories where libraries are installed. Not only that, there's the risk that some of the software incorrectly uses system libraries.

And so on. Nix solves all this at the packaging level and solves it well. A single tool to rule them all.

Nix makes no assumptions about the global state of the system. This has many advantages, but also some drawbacks of course. The core of a Nix system is the Nix store, usually installed under /nix/store, and some tools to manipulate the store. In Nix there is the notion of a derivation rather than a package. The difference can be subtle at the beginning, so I will often use the words interchangeably.

Derivations/packages are stored in the Nix store as follows: /nix/store/hash-name, where the hash uniquely identifies the derivation (this isn't quite true, it's a little more complex), and the name is the name of the derivation.

Let's take a bash derivation as an example: /nix/store/s4zia7hhqkin1di0f187b79sa2srhv6k-bash-4.2-p45/. This is a directory in the Nix store which contains bin/bash.

What that means is that there's no /bin/bash, there's only that self-contained build output in the store. The same goes for coreutils and everything else. To make them convenient to use from the shell, Nix will arrange for binaries to appear in your PATH as appropriate.

What we have is basically a store of all packages (with different versions occupying different locations), and everything in the Nix store is immutable.

In fact, there's no ldconfig cache either. So where does bash find libc?

$ ldd  `which bash`
libc.so.6 => /nix/store/94n64qy99ja0vgbkf675nyk39g9b978n-glibc-2.19/lib/libc.so.6 (0x00007f0248cce000)

It turns out that when bash was built, it was built against that specific version of glibc in the Nix store, and at runtime it will require exactly that glibc version.

Don't be confused by the version in the derivation name: it's only a name for us humans. You may end up having two derivations with the same name but different hashes: it's the hash that really matters.

What does all this mean? It means that you could run mysql 5.2 with glibc-2.18, and mysql 5.5 with glibc-2.19. You could use your python module with python 2.7 compiled with gcc 4.6 and the same python module with python 3 compiled with gcc 4.8, all in the same system.

In other words: no dependency hell, not even a dependency resolution algorithm. Straight dependencies from derivations to other derivations.

From an administrator's point of view: if you want an old PHP version for one application, but want to upgrade the rest of the system, that's not painful any more.

From a developer's point of view: if you want to develop webkit with llvm 3.4 and 3.3, that's not painful any more.

When upgrading a library, most package managers replace it in-place. All new applications run afterwards with the new library without being recompiled. After all, they all refer dynamically to libc6.so.

Since Nix derivations are immutable, upgrading a library like glibc means recompiling all applications, because the glibc path to the Nix store has been hardcoded.

So how do we deal with security updates? In Nix we have some tricks (still pure) to solve this problem, but that's another story.

Another problem is that unless software has in mind a pure functional model, or can be adapted to it, it can be hard to compose applications at runtime.

Let's take Firefox for example. On most systems, you install flash, and it starts working in Firefox because Firefox looks in a global path for plugins.

In Nix, there's no such global path for plugins. Firefox therefore must know explicitly about the path to flash. The way we handle this problem is to wrap the Firefox binary so that we can setup the necessary environment to make it find flash in the nix store. That will produce a new Firefox derivation: be aware that it takes a few seconds, and it makes composition harder at runtime.

There are no upgrade/downgrade scripts for your data. It doesn't make sense with this approach, because there's no real derivation to be upgraded. With Nix you switch to using other software with its own stack of dependencies, but there's no formal notion of upgrade or downgrade when doing so.

If there is a data format change, then migrating to the new data format remains your own responsibility.

Nix lets you compose software at build time with maximum flexibility, and with builds being as reproducible as possible. Not only that, due to its nature deploying systems in the cloud is so easy, consistent, and reliable that in the Nix world all existing self-containment and orchestration tools are deprecated by NixOps.

It however currently falls short when working with dynamic composition at runtime or replacing low level libraries, due to the need to rebuild dependencies.

That may sound scary, however after running NixOS on both a server and a laptop desktop, I'm very satisfied so far. Some of the architectural problems just need some man-power, other design problems still need to be solved as a community.

Considering Nixpkgs (github link) is a completely new repository of all the existing software, with a completely fresh concept, and with few core developers but overall year-over-year increasing contributions, the current state is more than acceptable and beyond the experimental stage. In other words, it's worth your investment.

...we will install Nix on top of your current system (I assume GNU/Linux, but we also have OSX users) and start inspecting the installed software.

To install Nix, run curl -L https://nixos.org/nix/install | sh as a non-root user and follow the instructions. Alternatively, you may prefer to download the installation script and verify its integrity using GPG signatures. Instructions for doing so can be found here: https://nixos.org/nix/download.html.

These articles are not a tutorial on using Nix. Instead, we're going to walk through the Nix system to understand the fundamentals.

The first thing to note: derivations in the Nix store refer to other derivations which are themselves in the Nix store. They don't use libc from our system or anywhere else. It's a self-contained store of all the software we need to bootstrap up to any particular package.

Start looking at the output of the install command:

copying Nix to /nix/store..........................

That's the /nix/store we were talking about in the first article. We're copying in the necessary software to bootstrap a Nix system. You can see bash, coreutils, the C compiler toolchain, perl libraries, sqlite and Nix itself with its own tools and libnix.

You may have noticed that /nix/store can contain not only directories, but also files, still always in the form hash-name.

Right after copying the store, the installation process initializes a database:

initialising Nix database...

Yes, Nix also has a database. It's stored under /nix/var/nix/db. It is a sqlite database that keeps track of the dependencies between derivations.

The schema is very simple: there's a table of valid paths, mapping from an auto increment integer to a store path.

Then there's a dependency relation from path to paths upon which they depend.

You can inspect the database by installing sqlite (nix-env -iA sqlite -f '<nixpkgs>') and then running sqlite3 /nix/var/nix/db/db.sqlite.

Next in the installation, we encounter the concept of the profile:

creating /home/nix/.nix-profile
installing 'nix-2.1.3'
building path(s) `/nix/store/a7p1w3z2h8pl00ywvw6icr3g5l9vm5r7-user-environment'
created 7 symlinks in user environment

A profile in Nix is a general and convenient concept for realizing rollbacks. Profiles are used to compose components that are spread among multiple paths under a new unified path. Not only that, but profiles are made up of multiple "generations": they are versioned. Whenever you change a profile, a new generation is created.

Generations can be switched and rolled back atomically, which makes them convenient for managing changes to your system.

Let's take a closer look at our profile:

$ ls -l ~/.nix-profile/
bin -> /nix/store/ig31y9gfpp8pf3szdd7d4sf29zr7igbr-nix-2.1.3/bin
[...]
manifest.nix -> /nix/store/q8b5238akq07lj9gfb3qb5ycq4dxxiwm-env-manifest.nix
[...]
share -> /nix/store/ig31y9gfpp8pf3szdd7d4sf29zr7igbr-nix-2.1.3/share

That nix-2.1.3 derivation in the Nix store is Nix itself, with binaries and libraries. The process of "installing" the derivation in the profile basically reproduces the hierarchy of the nix-2.1.3 store derivation in the profile by means of symbolic links.

The contents of this profile are special, because only one program has been installed in our profile, therefore e.g. the bin directory points to the only program which has been installed (Nix itself).

But that's only the contents of the latest generation of our profile. In fact, ~/.nix-profile itself is a symbolic link to /nix/var/nix/profiles/default.

In turn, that's a symlink to default-1-link in the same directory. Yes, that means it's the first generation of the default profile.

Finally, default-1-link is a symlink to the nix store "user-environment" derivation that you saw printed during the installation process.

We'll talk about manifest.nix more in the next article.

More output from the installer:

downloading Nix expressions from `http://releases.nixos.org/nixpkgs/nixpkgs-14.10pre46060.a1a2851/nixexprs.tar.xz'...
unpacking channels...
created 2 symlinks in user environment
modifying /home/nix/.profile...

Nix expressions are used to describe packages and how to build them. Nixpkgs is the repository containing all of the expressions: https://github.com/NixOS/nixpkgs.

The installer downloaded the package descriptions from commit a1a2851.

The second profile we discover is the channels profile. ~/.nix-defexpr/channels points to /nix/var/nix/profiles/per-user/nix/channels which points to channels-1-link which points to a Nix store directory containing the downloaded Nix expressions.

Channels are a set of packages and expressions available for download. Similar to Debian stable and unstable, there's a stable and unstable channel. In this installation, we're tracking nixpkgs-unstable.

Don't worry about Nix expressions yet, we'll get to them later.

Finally, for your convenience, the installer modified ~/.profile to automatically enter the Nix environment. What ~/.nix-profile/etc/profile.d/nix.sh really does is simply to add ~/.nix-profile/bin to PATH and ~/.nix-defexpr/channels/nixpkgs to NIX_PATH. We'll discuss NIX_PATH later.

Read nix.sh, it's short.

You can, but there's a good reason to keep using /nix instead of a different directory. All the derivations depend on other derivations by using absolute paths. We saw in the first article that bash referenced a glibc under a specific absolute path in /nix/store.

You can see for yourself, don't worry if you see multiple bash derivations:

$ ldd /nix/store/*bash*/bin/bash
[...]

Keeping the store in /nix means we can grab the binary cache from nixos.org (just like you grab packages from debian mirrors) otherwise:

After all /nix is a sensible place for the store.

We've installed Nix on our system, fully isolated and owned by the nix user as we're still coming to terms with this new system.

We learned some new concepts like profiles and channels. In particular, with profiles we're able to manage multiple generations of a composition of packages, while with channels we're able to download binaries from nixos.org.

The installation put everything under /nix, and some symlinks in the Nix user home. That's because every user is able to install and use software in her own environment.

I hope I left nothing uncovered so that you think there's some kind of magic going on behind the scenes. It's all about putting components in the store and symlinking these components together.

...we will enter the Nix environment and learn how to interact with the store.

If you're using NixOS, you can skip to the next step.

In the previous article we created a Nix user, so let's start by switching to it with su - nix. If your ~/.profile got evaluated, then you should now be able to run commands like nix-env and nix-store.

If that's not the case:

$ source ~/.nix-profile/etc/profile.d/nix.sh

To remind you, ~/.nix-profile/etc points to the nix-2.1.3 derivation. At this point, we are in our Nix user profile.

Finally something practical! Installation into the Nix environment is an interesting process. Let's install hello, a simple CLI tool which prints Hello world and is mainly used to test compilers and package installations.

Back to the installation:

$ nix-env -i hello
installing 'hello-2.10'
[...]
building '/nix/store/0vqw0ssmh6y5zj48yg34gc6macr883xk-user-environment.drv'...
created 36 symlinks in user environment

Now you can run hello. Things to notice:

We can list generations without walking through the /nix hierarchy:

$ nix-env --list-generations
   1   2014-07-24 09:23:30
   2   2014-07-25 08:45:01   (current)

Listing installed derivations:

$ nix-env -q
nix-2.1.3
hello-2.10

So, where did hello really get installed? which hello is ~/.nix-profile/bin/hello which points to the store. We can also list the derivation paths with nix-env -q --out-path. So that's what those derivation paths are called: the output of a build.

At this point you probably want to run man to get some documentation. Even if you already have man system-wide outside of the Nix environment, you can install and use it within Nix with nix-env -i man-db. As usual, a new generation will be created, and ~/.nix-profile will point to it.

Lets inspect the profile a bit:

$ ls -l ~/.nix-profile/
dr-xr-xr-x 2 nix nix 4096 Jan  1  1970 bin
lrwxrwxrwx 1 nix nix   55 Jan  1  1970 etc -> /nix/store/ig31y9gfpp8pf3szdd7d4sf29zr7igbr-nix-2.1.3/etc
[...]

Now that's interesting. When only nix-2.1.3 was installed, bin was a symlink to nix-2.1.3. Now that we've actually installed some things (man, hello), it's a real directory, not a symlink.

$ ls -l ~/.nix-profile/bin/
[...]
man -> /nix/store/83cn9ing5sc6644h50dqzzfxcs07r2jn-man-1.6g/bin/man
[...]
nix-env -> /nix/store/ig31y9gfpp8pf3szdd7d4sf29zr7igbr-nix-2.1.3/bin/nix-env
[...]
hello -> /nix/store/58r35bqb4f3lxbnbabq718svq9i2pda3-hello-2.10/bin/hello
[...]

Okay, that's clearer now. nix-env merged the paths from the installed derivations. which man points to the Nix profile, rather than the system man, because ~/.nix-profile/bin is at the head of $PATH.

The last command installed man. We should be at generation 3, unless you changed something in the middle. Let's say we want to rollback to the old generation:

$ nix-env --rollback
switching from generation 3 to 2

Now nix-env -q does not list man anymore. ls -l `which man` should now be your system copy.

Enough with the rollback, let's go back to the most recent generation:

$ nix-env -G 3
switching from generation 2 to 3

I invite you to read the manpage of nix-env. nix-env requires an operation to perform, then there are common options for all operations, as well as options specific to each operation.

You can of course also uninstall and upgrade packages.

So far we learned how to query and manipulate the environment. But all of the environment components point to the store.

To query and manipulate the store, there's the nix-store command. We can do some interesting things, but we'll only see some queries for now.

To show the direct runtime dependencies of hello:

$ nix-store -q --references `which hello`
/nix/store/fg4yq8i8wd08xg3fy58l6q73cjy8hjr2-glibc-2.27
/nix/store/58r35bqb4f3lxbnbabq718svq9i2pda3-hello-2.10

The argument to nix-store can be anything as long as it points to the Nix store. It will follow symlinks.

It may not make sense to you right now, but let's print reverse dependencies of hello:

$ nix-store -q --referrers `which hello`
/nix/store/58r35bqb4f3lxbnbabq718svq9i2pda3-hello-2.10
/nix/store/fhvy2550cpmjgcjcx5rzz328i0kfv3z3-env-manifest.nix
/nix/store/mp987abm20c70pl8p31ljw1r5by4xwfw-user-environment

Was it what you expected? It turns out that our environments depend upon hello. Yes, that means that the environments are in the store, and since they contain symlinks to hello, therefore the environment depends upon hello.

Two environments were listed, generation 2 and generation 3, since these are the ones that had hello installed in them.

The manifest.nix file contains metadata about the environment, such as which derivations are installed. So that nix-env can list, upgrade or remove them. And yet again, the current manifest.nix can be found at ~/.nix-profile/manifest.nix.

The closures of a derivation is a list of all its dependencies, recursively, including absolutely everything necessary to use that derivation.

$ nix-store -qR `which man`
[...]

Copying all those derivations to the Nix store of another machine makes you able to run man out of the box on that other machine. That's the base of deployment using Nix, and you can already foresee the potential when deploying software in the cloud (hint: nix-copy-closures and nix-store --export).

A nicer view of the closure:

$ nix-store -q --tree `which man`
[...]

With the above command, you can find out exactly why a runtime dependency, be it direct or indirect, exists for a given derivation.

The same applies to environments. As an exercise, run nix-store -q --tree ~/.nix-profile, and see that the first children are direct dependencies of the user environment: the installed derivations, and the manifest.nix.

There isn't anything like apt which solves a SAT problem in order to satisfy dependencies with lower and upper bounds on versions. There's no need for this because all the dependencies are static: if a derivation X depends on a derivation Y, then it always depends on it. A version of X which depended on Z would be a different derivation.

$ nix-env -e '*'
uninstalling 'hello-2.10'
uninstalling 'nix-2.1.3'
[...]

Oops, that uninstalled all derivations from the environment, including Nix. That means we can't even run nix-env, what now?

Previously we got nix-env from the environment. Environments are a convenience for the user, but Nix is still there in the store!

First, pick one nix-2.1.3 derivation: ls /nix/store/*nix-2.1.3, say /nix/store/ig31y9gfpp8pf3szdd7d4sf29zr7igbr-nix-2.1.3.

The first option is to rollback:

$ /nix/store/ig31y9gfpp8pf3szdd7d4sf29zr7igbr-nix-2.1.3/bin/nix-env --rollback

The second option is to install Nix, thus creating a new generation:

$ /nix/store/ig31y9gfpp8pf3szdd7d4sf29zr7igbr-nix-2.1.3/bin/nix-env -i /nix/store/ig31y9gfpp8pf3szdd7d4sf29zr7igbr-nix-2.1.3/bin/nix-env

So where are we getting packages from? We said something about this already in the second article. There's a list of channels from which we get packages, although usually we use a single channel. The tool to manage channels is nix-channel.

$ nix-channel --list
nixpkgs http://nixos.org/channels/nixpkgs-unstable

If you're using NixOS, you may not see any output from the above command (if you're using the default), or you may see a channel whose name begins with "nixos-" instead of "nixpkgs".

That's essentially the contents of ~/.nix-channels.

To update the channel run nix-channel --update. That will download the new Nix expressions (descriptions of the packages), create a new generation of the channels profile and unpack it under ~/.nix-defexpr/channels.

This is quite similar to apt-get update. (See this table for a rough mapping between Ubuntu and NixOS package management.)

We learned how to query the user environment and to manipulate it by installing and uninstalling software. Upgrading software is also straightforward, as you can read in the manual (nix-env -u will upgrade all packages in the environment).

Everytime we change the environment, a new generation is created. Switching between generations is easy and immediate.

Then we learned how to query the store. We inspected the dependencies and reverse dependencies of store paths.

We saw how symlinks are used to compose paths from the Nix store, a useful trick.

A quick analogy with programming languages: you have the heap with all the objects, that corresponds to the Nix store. You have objects that point to other objects, those correspond to derivations. This is a suggestive metaphor, but will it be the right path?

...we will learn the basics of the Nix language. The Nix language is used to describe how to build derivations, and it's the basis for everything else, including NixOS. Therefore it's very important to understand both the syntax and the semantics of the language.

Nix 2.0 contains a command named nix repl which is a simple command line tool for playing with the Nix language. In fact, Nix is a pure, lazy, functional language, not only a set of tools to manage derivations. The nix repl syntax is slightly different to Nix syntax when it comes to assigning variables, but it shouldn't be confusing so long as you bear it in mind. I prefer to start with nix repl before cluttering your mind with more complex expressions.

Launch nix repl. First of all, Nix supports basic arithmetic operations: +, -, * and /. (To exit nix repl, use the command :q. Help is available through the :? command.)

nix-repl> 1+3
4

nix-repl> 7-4
3

nix-repl> 3*2
6

Attempting to perform division in Nix can lead to some surprises.

nix-repl> 6/3
/home/nix/6/3

What happened? Recall that Nix is not a general purpose language, it's a domain-specific language for writing packages. Integer division isn't actually that useful when writing package expressions. Nix parsed 6/3 as a relative path to the current directory. To get Nix to perform division instead, leave a space after the /. Alternatively, you can use builtins.div.

nix-repl> 6/ 3
2

nix-repl> builtins.div 6 3
2

Other operators are ||, && and ! for booleans, and relational operators such as !=, ==, <, >, <=, >=. In Nix, <, >, <= and >= are not much used. There are also other operators we will see in the course of this series.

Nix has integer, floating point, string, path, boolean and null simple types. Then there are also lists, sets and functions. These types are enough to build an operating system.

Nix is strongly typed, but it's not statically typed. That is, you cannot mix strings and integers, you must first do the conversion.

As demonstrated above, expressions will be parsed as paths as long as there's a slash not followed by a space. Therefore to specify the current directory, use ./. In addition, Nix also parses urls specially.

Not all urls or paths can be parsed this way. If a syntax error occurs, it's still possible to fallback to plain strings. Literal urls and paths are convenient for additional safety.

There's not much to say here, except that dash (-) is allowed in identifiers. That's convenient since many packages use dash in their names. In fact:

nix-repl> a-b
error: undefined variable `a-b' at (string):1:1
nix-repl> a - b
error: undefined variable `a' at (string):1:1

As you can see, a-b is parsed as identifier, not as a subtraction.

It's important to understand the syntax for strings. When learning to read Nix expressions, you may find dollars ($) ambiguous, but they are very important . Strings are enclosed by double quotes ("), or two single quotes ('').

nix-repl> "foo"
"foo"
nix-repl> ''foo''
"foo"

In other languages like Python you can also use single quotes for strings (e.g. 'foo'), but not in Nix.

It's possible to interpolate whole Nix expressions inside strings with the ${...} syntax and only that syntax, not $foo or {$foo} or anything else.

nix-repl> foo = "strval"
nix-repl> "$foo"
"$foo"
nix-repl> "${foo}"
"strval"
nix-repl> "${2+3}"
error: cannot coerce an integer to a string, at (string):1:2

Note: ignore the foo = "strval" assignment, special syntax in nix repl.

As said previously, you cannot mix integers and strings. You need to explicitly include conversions. We'll see this later: function calls are another story.

Using the syntax with two single quotes is useful for writing double quotes inside strings without needing to escape them:

nix-repl> ''test " test''
"test \" test"
nix-repl> ''${foo}''
"strval"

Escaping ${...} within double quoted strings is done with the backslash. Within two single quotes, it's done with '':

nix-repl> "\${foo}"
"${foo}"
nix-repl> ''test ''${foo} test''
"test ${foo} test"

Lists are a sequence of expressions delimited by space (not comma):

nix-repl> [ 2 "foo" true (2+3) ]
[ 2 "foo" true 5 ]

Lists, like everything else in Nix, are immutable. Adding or removing elements from a list is possible, but will return a new list.

Attribute sets are an association between string keys and a Nix values. Keys can only be strings. When writing attribute sets you can also use unquoted identifiers as keys.

nix-repl> s = { foo = "bar"; a-b = "baz"; "123" = "num"; }
nix-repl> s
{ "123" = "num"; a-b = "baz"; foo = "bar"; }

For those reading Nix expressions from nixpkgs: do not confuse attribute sets with argument sets used in functions.

To access elements in the attribute set:

nix-repl> s.a-b
"baz"
nix-repl> s."123"
"num"

Yes, you can use strings to address keys which aren't valid identifiers.

Inside an attribute set you cannot normally refer to elements of the same attribute set:

nix-repl> { a = 3; b = a+4; }
error: undefined variable `a' at (string):1:10

To do so, use recursive attribute sets:

nix-repl> rec { a = 3; b = a+4; }
{ a = 3; b = 7; }

This is very convenient when defining packages, which tend to be recursive attribute sets.

These are expressions, not statements.

nix-repl> a = 3
nix-repl> b = 4
nix-repl> if a > b then "yes" else "no"
"no"

You can't have only the then branch, you must specify also the else branch, because an expression must have a value in all cases.

This kind of expression is used to define local variables for inner expressions.

nix-repl> let a = "foo"; in a
"foo"

The syntax is: first assign variables, then in, then an expression which can use the defined variables. The value of the whole let expression will be the value of the expression after the in.

nix-repl> let a = 3; b = 4; in a + b
7

Let's write two let expressions, one inside the other:

nix-repl> let a = 3; in let b = 4; in a + b
7

With let you cannot assign twice to the same variable. However, you can shadow outer variables:

nix-repl> let a = 3; a = 8; in a
error: attribute `a' at (string):1:12 already defined at (string):1:5
nix-repl> let a = 3; in let a = 8; in a
8

You cannot refer to variables in a let expression outside of it:

nix-repl> let a = (let b = 3; in b); in b
error: undefined variable `b' at (string):1:31

You can refer to variables in the let expression when assigning variables, like with recursive attribute sets:

nix-repl> let a = 4; b = a + 5; in b
9

So beware when you want to refer to a variable from the outer scope, but it's also defined in the current let expression. The same applies to recursive attribute sets.

This kind of expression is something you rarely see in other languages. You can think of it like a more granular version of using from C++, or from module import * from Python. You decide per-expression when to include symbols into the scope.

nix-repl> longName = { a = 3; b = 4; }
nix-repl> longName.a + longName.b
7
nix-repl> with longName; a + b
7

That's it, it takes an attribute set and includes symbols from it in the scope of the inner expression. Of course, only valid identifiers from the keys of the set will be included. If a symbol exists in the outer scope and would also be introduced by the with, it will not be shadowed. You can however still refer to the attribute set:

nix-repl> let a = 10; in with longName; a + b
14
nix-repl> let a = 10; in with longName; longName.a + b
7

Nix evaluates expressions only when needed. This is a great feature when working with packages.

nix-repl> let a = builtins.div 4 0; b = 6; in b
6

Since a is not needed, there's no error about division by zero, because the expression is not in need to be evaluated. That's why we can have all the packages defined on demand, yet have access to specific packages very quickly.

...we will talk about functions and imports. In this pill I've tried to avoid function calls as much as possible, otherwise the post would have been too long.

Functions are anonymous (lambdas), and only have a single parameter. The syntax is extremely simple. Type the parameter name, then ":", then the body of the function.

nix-repl> x: x*2
«lambda»

So here we defined a function that takes a parameter x, and returns x*2. The problem is that we cannot use it in any way, because it's unnamed... joke!

We can store functions in variables.

nix-repl> double = x: x*2
nix-repl> double
«lambda»
nix-repl> double 3
6

As usual, please ignore the special syntax for assignments inside nix repl. So, we defined a function x: x*2 that takes one parameter x, and returns x*2. This function is then assigned to the variable double. Finally we did our first function call: double 3.

Big note: it's not like many other programming languages where you write double(3). It really is double 3.

In summary: to call a function, name the variable, then space, then the argument. Nothing else to say, it's as easy as that.

How do we create a function that accepts more than one parameter? For people not used to functional programming, this may take a while to grasp. Let's do it step by step.

nix-repl> mul = a: (b: a*b)
nix-repl> mul
«lambda»
nix-repl> mul 3
«lambda»
nix-repl> (mul 3) 4
12

We defined a function that takes the parameter a, the body returns another function. This other function takes a parameter b and returns a*b. Therefore, calling mul 3 returns this kind of function: b: 3*b. In turn, we call the returned function with 4, and get the expected result.

You don't have to use parenthesis at all, Nix has sane priorities when parsing the code:

nix-repl> mul = a: b: a*b
nix-repl> mul
«lambda»
nix-repl> mul 3
«lambda»
nix-repl> mul 3 4
12
nix-repl> mul (6+7) (8+9)
221

Much more readable, you don't even notice that functions only receive one argument. Since the argument is separated by a space, to pass more complex expressions you need parenthesis. In other common languages you would write mul(6+7, 8+9).

Given that functions have only one parameter, it is straightforward to use partial application:

nix-repl> foo = mul 3
nix-repl> foo 4
12
nix-repl> foo 5
15

We stored the function returned by mul 3 into a variable foo, then reused it.

Now this is a very cool feature of Nix. It is possible to pattern match over a set in the parameter. We write an alternative version of mul = a: b: a*b first by using a set as argument, then using pattern matching.

nix-repl> mul = s: s.a*s.b
nix-repl> mul { a = 3; b = 4; }
12
nix-repl> mul = { a, b }: a*b
nix-repl> mul { a = 3; b = 4; }
12

In the first case we defined a function that accepts a single parameter. We then access attributes a and b from the given set. Note how the parenthesis-less syntax for function calls is very elegant in this case, instead of doing mul({ a=3; b=4; }) in other languages.

In the second case we defined an arguments set. It's like defining a set, except without values. We require that the passed set contains the keys a and b. Then we can use those a and b in the function body directly.

nix-repl> mul = { a, b }: a*b
nix-repl> mul { a = 3; b = 4; c = 6; }
error: anonymous function at (string):1:2 called with unexpected argument `c', at (string):1:1
nix-repl> mul { a = 3; }
error: anonymous function at (string):1:2 called without required argument `b', at (string):1:1

Only a set with exactly the attributes required by the function is accepted, nothing more, nothing less.

It is possible to specify default values of attributes in the arguments set:

nix-repl> mul = { a, b ? 2 }: a*b
nix-repl> mul { a = 3; }
6
nix-repl> mul { a = 3; b = 4; }
12

Also you can allow passing more attributes (variadic) than the expected ones:

nix-repl> mul = { a, b, ... }: a*b
nix-repl> mul { a = 3; b = 4; c = 2; }

However, in the function body you cannot access the "c" attribute. The solution is to give a name to the given set with the @-pattern:

nix-repl> mul = s@{ a, b, ... }: a*b*s.c
nix-repl> mul { a = 3; b = 4; c = 2; }
24

That's it, you give a name to the whole parameter with name@ before the set pattern.

Advantages of using argument sets:

Disadvantages:

You may find similarities with Python **kwargs.

The import function is built-in and provides a way to parse a .nix file. The natural approach is to define each component in a .nix file, then compose by importing these files.

Let's start with the bare metal.

a.nix:

3

b.nix:

4

mul.nix:

a: b: a*b

nix-repl> a = import ./a.nix
nix-repl> b = import ./b.nix
nix-repl> mul = import ./mul.nix
nix-repl> mul a b
12

Yes it's really that simple. You import a file, and it gets parsed as expression. Note that the scope of the imported file does not inherit the scope of the importer.

test.nix:

x

nix-repl> let x = 5; in import ./test.nix
error: undefined variable `x' at /home/lethal/test.nix:1:1

So how do we pass information to the module? Use functions, like we did with mul.nix. A more complex example:

test.nix:

{ a, b ? 3, trueMsg ? "yes", falseMsg ? "no" }:
if a > b
  then builtins.trace trueMsg true
  else builtins.trace falseMsg false

nix-repl> import ./test.nix { a = 5; trueMsg = "ok"; }
trace: ok
true

Explaining:

So when is the message shown? Only when it's in need to be evaluated.

...we will finally write our first derivation.

The derivation built-in function is used to create derivations. I invite you to read the link in the Nix manual about the derivation built-in. A derivation from a Nix language view point is simply a set, with some attributes. Therefore you can pass the derivation around with variables like anything else.

That's where the real power comes in.

The derivation function receives a set as first argument. This set requires at least the following three attributes:

First of all, what's the name of our system as seen by nix?

nix-repl> builtins.currentSystem
"x86_64-linux"

Let's try to fake the name of the system:

nix-repl> d = derivation { name = "myname"; builder = "mybuilder"; system = "mysystem"; }
nix-repl> d
«derivation /nix/store/z3hhlxbckx4g3n9sw91nnvlkjvyw754p-myname.drv»

Oh oh, what's that? Did it build the derivation? No it didn't, but it did create the .drv file. nix repl does not build derivations unless you tell to do so.

What's that .drv file? It is the specification of how to build the derivation, without all the Nix language fuzz.

Before continuing, some analogies with the C language:

Both drv paths and out paths are stored in the nix store as you can see.

What's in that .drv file? You can read it, but it's better to pretty print it:

$ nix show-derivation /nix/store/z3hhlxbckx4g3n9sw91nnvlkjvyw754p-myname.drv
{
  "/nix/store/z3hhlxbckx4g3n9sw91nnvlkjvyw754p-myname.drv": {
    "outputs": {
      "out": {
        "path": "/nix/store/40s0qmrfb45vlh6610rk29ym318dswdr-myname"
      }
    },
    "inputSrcs": [],
    "inputDrvs": {},
    "platform": "mysystem",
    "builder": "mybuilder",
    "args": [],
    "env": {
      "builder": "mybuilder",
      "name": "myname",
      "out": "/nix/store/40s0qmrfb45vlh6610rk29ym318dswdr-myname",
      "system": "mysystem"
    }
  }
}

Ok we can see there's an out path, but it does not exist yet. We never told Nix to build it, but we know beforehand where the build output will be. Why?

Think, if Nix ever built the derivation just because we accessed it in Nix, we would have to wait a long time if it was, say, Firefox. That's why Nix let us know the path beforehand and keep evaluating the Nix expressions, but it's still empty because no build was ever made.

Important: the hash of the out path is based solely on the input derivations in the current version of Nix, not on the contents of the build product. It's possible however to have content-addressable derivations for e.g. tarballs as we'll see later on.

Many things are empty in that .drv, however I'll write a summary of the .drv format for you:

That's it, the minimum necessary information to build our derivation.

Important note: the environment variables passed to the builder are just those you see in the .drv plus some other Nix related configuration (number of cores, temp dir, ...). The builder will not inherit any variable from your running shell, otherwise builds would suffer from non-determinism.

Back to our fake derivation.

Let's build our really fake derivation:

nix-repl> d = derivation { name = "myname"; builder = "mybuilder"; system = "mysystem"; }
nix-repl> :b d
[...]
these derivations will be built:
  /nix/store/z3hhlxbckx4g3n9sw91nnvlkjvyw754p-myname.drv
building path(s) `/nix/store/40s0qmrfb45vlh6610rk29ym318dswdr-myname'
error: a `mysystem' is required to build `/nix/store/z3hhlxbckx4g3n9sw91nnvlkjvyw754p-myname.drv', but I am a `x86_64-linux'

The :b is a nix repl specific command to build a derivation. You can see more commands with :? . So in the output you can see that it takes the .drv as information on how to build the derivation. Then it says it's trying to produce our out path. Finally the error we were waiting for: that derivation can't be built on our system.

We're doing the build inside nix repl, but what if we don't want to use nix repl? You can realise a .drv with:

$ nix-store -r /nix/store/z3hhlxbckx4g3n9sw91nnvlkjvyw754p-myname.drv

You will get the same output as before.

Let's fix the system attribute:

nix-repl> d = derivation { name = "myname"; builder = "mybuilder"; system = builtins.currentSystem; }
nix-repl> :b d
[...]
build error: invalid file name `mybuilder'

A step forward: of course, that mybuilder executable does not really exist. Stop for a moment.

It is useful to start by inspecting the return value from the derivation function. In this case, the returned value is a plain set:

nix-repl> d = derivation { name = "myname"; builder = "mybuilder"; system = "mysystem"; }
nix-repl> builtins.isAttrs d
true
nix-repl> builtins.attrNames d
[ "all" "builder" "drvAttrs" "drvPath" "name" "out" "outPath" "outputName" "system" "type" ]

You can guess what builtins.isAttrs does, it returns true if the argument is a set. While builtins.attrNames returns a list of keys of the given set. Some kind of reflection, you might say.

Start from drvAttrs:

nix-repl> d.drvAttrs
{ builder = "mybuilder"; name = "myname"; system = "mysystem"; }

That's basically the input we gave to the derivation function. Also d.name, d.system and d.builder attributes are straight the ones we gave as input.

nix-repl> (d == d.out)
true

So out is just the derivation itself, it seems weird but the reason is that we only have one output from the derivation. That's also the reason why d.all is a singleton. We'll see multiple outputs later.

The d.drvPath is the path of the .drv file: /nix/store/z3hhlxbckx4g3n9sw91nnvlkjvyw754p-myname.drv.

Something interesting is the type attribute. It's "derivation". Nix does add a little of magic to sets with type derivation, but not that much. To let you understand, you can create yourself a set with that type, it's a simple set:

nix-repl> { type = "derivation"; }
«derivation ???»

Of course it has no other information, so Nix doesn't know what to say :-) But you get it, the type = "derivation" is just a convention for Nix and for us to understand the set is a derivation.

When writing packages, we are interested in the outputs. The other metadata is needed for Nix to know how to create the drv path and the out path.

The outPath attribute is the build path in the nix store: /nix/store/40s0qmrfb45vlh6610rk29ym318dswdr-myname.

Just like dependencies in other package managers, how do we refer to other packages? How do we refer to other derivations in terms of files on the disk? We use the outPath. The outPath tells where the files are of that derivation. To make it more convenient, Nix is able to do a conversion from a derivation set to a string.

nix-repl> d.outPath
"/nix/store/40s0qmrfb45vlh6610rk29ym318dswdr-myname"
nix-repl> builtins.toString d
"/nix/store/40s0qmrfb45vlh6610rk29ym318dswdr-myname"

Nix does the "set to string conversion" as long as there is the outPath attribute (much like a toString method in other languages):

nix-repl> builtins.toString { outPath = "foo"; }
"foo"
nix-repl> builtins.toString { a = "b"; }
error: cannot coerce a set to a string, at (string):1:1

Say we want to use binaries from coreutils (ignore the nixpkgs etc.):

nix-repl> :l <nixpkgs>
Added 3950 variables.
nix-repl> coreutils
«derivation /nix/store/1zcs1y4n27lqs0gw4v038i303pb89rw6-coreutils-8.21.drv»
nix-repl> builtins.toString coreutils
"/nix/store/8w4cbiy7wqvaqsnsnb3zvabq1cp2zhyz-coreutils-8.21"

Apart the nixpkgs stuff, just think we added to the scope a series of variables. One of them is coreutils. It is the derivation of the coreutils package you all know of from other Linux distributions. It contains basic binaries for GNU/Linux systems (you may have multiple derivations of coreutils in the nix store, no worries):

$ ls /nix/store/*coreutils*/bin
[...]

I remind you, inside strings it's possible to interpolate Nix expressions with ${...}:

nix-repl> "${d}"
"/nix/store/40s0qmrfb45vlh6610rk29ym318dswdr-myname"
nix-repl> "${coreutils}"
"/nix/store/8w4cbiy7wqvaqsnsnb3zvabq1cp2zhyz-coreutils-8.21"

That's very convenient, because then we could refer to e.g. the bin/true binary like this:

nix-repl> "${coreutils}/bin/true"
"/nix/store/8w4cbiy7wqvaqsnsnb3zvabq1cp2zhyz-coreutils-8.21/bin/true"

In the previous attempt we used a fake builder, mybuilder which obviously does not exist. But we can use for example bin/true, which always exits with 0 (success).

nix-repl> :l <nixpkgs>
nix-repl> d = derivation { name = "myname"; builder = "${coreutils}/bin/true"; system = builtins.currentSystem; }
nix-repl> :b d
[...]
builder for `/nix/store/qyfrcd53wmc0v22ymhhd5r6sz5xmdc8a-myname.drv' failed to produce output path `/nix/store/ly2k1vswbfmswr33hw0kf0ccilrpisnk-myname'

Another step forward, it executed the builder (bin/true), but the builder did not create the out path of course, it just exited with 0.

Obvious note: everytime we change the derivation, a new hash is created.

Let's examine the new .drv now that we referred to another derivation:

$ nix show-derivation /nix/store/qyfrcd53wmc0v22ymhhd5r6sz5xmdc8a-myname.drv
{
  "/nix/store/qyfrcd53wmc0v22ymhhd5r6sz5xmdc8a-myname.drv": {
    "outputs": {
      "out": {
        "path": "/nix/store/ly2k1vswbfmswr33hw0kf0ccilrpisnk-myname"
      }
    },
    "inputSrcs": [],
    "inputDrvs": {
      "/nix/store/hixdnzz2wp75x1jy65cysq06yl74vx7q-coreutils-8.29.drv": [
        "out"
      ]
    },
    "platform": "x86_64-linux",
    "builder": "/nix/store/qrxs7sabhqcr3j9ai0j0cp58zfnny0jz-coreutils-8.29/bin/true",
    "args": [],
    "env": {
      "builder": "/nix/store/qrxs7sabhqcr3j9ai0j0cp58zfnny0jz-coreutils-8.29/bin/true",
      "name": "myname",
      "out": "/nix/store/ly2k1vswbfmswr33hw0kf0ccilrpisnk-myname",
      "system": "x86_64-linux"
    }
  }
}

Aha! Nix added a dependency to our myname.drv, it's the coreutils.drv. Before doing our build, Nix should build the coreutils.drv. But since coreutils is already in our nix store, no build is needed, it's already there with out path /nix/store/qrxs7sabhqcr3j9ai0j0cp58zfnny0jz-coreutils-8.29.

Nix does not build derivations during evaluation of Nix expressions. In fact, that's why we have to do ":b drv" in nix repl, or use nix-store -r in the first place.

An important separation is made in Nix:

Think of it as of compile time and link time like with C/C++ projects. You first compile all source files to object files. Then link object files in a single executable.

In Nix, first the Nix expression (usually in a .nix file) is compiled to .drv, then each .drv is built and the product is installed in the relative out paths.

Is it that complicated to create a package for Nix? No it's not.

We're walking through the fundamentals of Nix derivations, to understand how they work, how they are represented. Packaging in Nix is certainly easier than that, but we're not there yet in this post. More Nix pills are needed.

With the derivation function we provide a set of information on how to build a package, and we get back the information about where the package was built. Nix converts a set to a string when there's an outPath, that's very convenient. With that, it's easy to refer to other derivations.

When Nix builds a derivation, it first creates a .drv file from a derivation expression, and uses it to build the output. It does so recursively for all the dependencies (inputs). It "executes" the .drv files like a machine. Not much magic after all.

...we will finally write our first working derivation. Yes, this post is about "our first derivation", but I never said it was a working one ;)

Welcome to the seventh nix pill. In the previous sixth pill we introduced the notion of derivation in the Nix language — how to define a raw derivation and how to (try to) build it.

In this post we continue along the path, by creating a derivation that actually builds something. Then, we try to package a real program: we compile a simple C file and create a derivation out of it, given a blessed toolchain.

I remind you how to enter the Nix environment: source ~/.nix-profile/etc/profile.d/nix.sh

What's the easiest way to run a sequence of commands for building something? A bash script. We write a custom bash script, and we want it to be our builder. Given a builder.sh, we want the derivation to run bash builder.sh.

We don't use hash bangs in builder.sh, because at the time we are writing it we do not know the path to bash in the nix store. Yes, even bash is in the nix store, everything is there.

We don't even use /usr/bin/env, because then we lose the cool stateless property of Nix. Not to mention that PATH gets cleared when building, so it wouldn't find bash anyway.

In summary, we want the builder to be bash, and pass it an argument, builder.sh. Turns out the derivation function accepts an optional args attribute which is used to pass arguments to the builder executable.

First of all, let's write our builder.sh in the current directory:

declare -xp
echo foo > $out

The command declare -xp lists exported variables (declare is a builtin bash function). As we covered in the previous pill, Nix computes the output path of the derivation. The resulting .drv file contains a list of environment variables passed to the builder. One of these is $out.

What we have to do is create something in the path $out, be it a file or a directory. In this case we are creating a file.

In addition, we print out the environment variables during the build process. We cannot use env for this, because env is part of coreutils and we don't have a dependency to it yet. We only have bash for now.

Like for coreutils in the previous pill, we get a blessed bash for free from our magic nixpkgs stuff:

nix-repl> :l <nixpkgs>
Added 3950 variables.
nix-repl> "${bash}"
"/nix/store/ihmkc7z2wqk3bbipfnlh0yjrlfkkgnv6-bash-4.2-p45"

So with the usual trick, we can refer to bin/bash and create our derivation:

nix-repl> d = derivation { name = "foo"; builder = "${bash}/bin/bash"; args = [ ./builder.sh ]; system = builtins.currentSystem; }
nix-repl> :b d
these derivations will be built:
  /nix/store/i76pr1cz0za3i9r6xq518bqqvd2raspw-foo.drv
building '/nix/store/i76pr1cz0za3i9r6xq518bqqvd2raspw-foo.drv'...
declare -x HOME="/homeless-shelter"
declare -x NIX_BUILD_CORES="4"
declare -x NIX_BUILD_TOP="/tmp/nix-build-foo.drv-0"
declare -x NIX_LOG_FD="2"
declare -x NIX_STORE="/nix/store"
declare -x OLDPWD
declare -x PATH="/path-not-set"
declare -x PWD="/tmp/nix-build-foo.drv-0"
declare -x SHLVL="1"
declare -x TEMP="/tmp/nix-build-foo.drv-0"
declare -x TEMPDIR="/tmp/nix-build-foo.drv-0"
declare -x TMP="/tmp/nix-build-foo.drv-0"
declare -x TMPDIR="/tmp/nix-build-foo.drv-0"
declare -x builder="/nix/store/q1g0rl8zfmz7r371fp5p42p4acmv297d-bash-4.4-p19/bin/bash"
declare -x name="foo"
declare -x out="/nix/store/gczb4qrag22harvv693wwnflqy7lx5pb-foo"
declare -x system="x86_64-linux"
warning: you did not specify '--add-root'; the result might be removed by the garbage collector
/nix/store/gczb4qrag22harvv693wwnflqy7lx5pb-foo

this derivation produced the following outputs:
  out -> /nix/store/gczb4qrag22harvv693wwnflqy7lx5pb-foo

We did it! The contents of /nix/store/w024zci0x1hh1wj6gjq0jagkc1sgrf5r-foo is really foo. We've built our first derivation.

Note that we used ./builder.sh and not "./builder.sh". This way, it is parsed as a path, and Nix performs some magic which we will cover later. Try using the string version and you will find that it cannot find builder.sh. This is because it tries to find it relative to the temporary build directory.

Let's inspect those environment variables printed during the build process.

And that's how we were able to use $out in our derivation and put stuff in it. It's like Nix reserved a slot in the nix store for us, and we must fill it.

In terms of autotools, $out will be the --prefix path. Yes, not the make DESTDIR, but the --prefix. That's the essence of stateless packaging. You don't install the package in a global common path under /, you install it in a local isolated path under your nix store slot.

We added something else to the derivation this time: the args attribute. Let's see how this changed the .drv compared to the previous pill:

$ nix show-derivation /nix/store/i76pr1cz0za3i9r6xq518bqqvd2raspw-foo.drv
{
  "/nix/store/i76pr1cz0za3i9r6xq518bqqvd2raspw-foo.drv": {
    "outputs": {
      "out": {
        "path": "/nix/store/gczb4qrag22harvv693wwnflqy7lx5pb-foo"
      }
    },
    "inputSrcs": [
      "/nix/store/lb0n38r2b20r8rl1k45a7s4pj6ny22f7-builder.sh"
    ],
    "inputDrvs": {
      "/nix/store/hcgwbx42mcxr7ksnv0i1fg7kw6jvxshb-bash-4.4-p19.drv": [
        "out"
      ]
    },
    "platform": "x86_64-linux",
    "builder": "/nix/store/q1g0rl8zfmz7r371fp5p42p4acmv297d-bash-4.4-p19/bin/bash",
    "args": [
      "/nix/store/lb0n38r2b20r8rl1k45a7s4pj6ny22f7-builder.sh"
    ],
    "env": {
      "builder": "/nix/store/q1g0rl8zfmz7r371fp5p42p4acmv297d-bash-4.4-p19/bin/bash",
      "name": "foo",
      "out": "/nix/store/gczb4qrag22harvv693wwnflqy7lx5pb-foo",
      "system": "x86_64-linux"
    }
  }
}

Much like the usual .drv, except that there's a list of arguments in there passed to the builder (bash) with builder.sh… In the nix store..? Nix automatically copies files or directories needed for the build into the store to ensure that they are not changed during the build process and that the deployment is stateless and independent of the building machine. builder.sh is not only in the arguments passed to the builder, it's also in the input derivations.

Given that builder.sh is a plain file, it has no .drv associated with it. The store path is computed based on the filename and on the hash of its contents. Store paths are covered in detail in a later pill.

Start off by writing a simple C program called simple.c:

void main() {
  puts("Simple!");
}

And its simple_builder.sh:

export PATH="$coreutils/bin:$gcc/bin"
mkdir $out
gcc -o $out/simple $src

Don't worry too much about where those variables come from yet; let's write the derivation and build it:

nix-repl> :l <nixpkgs>
nix-repl> simple = derivation { name = "simple"; builder = "${bash}/bin/bash"; args = [ ./simple_builder.sh ]; gcc = gcc; coreutils = coreutils; src = ./simple.c; system = builtins.currentSystem; }
nix-repl> :b simple
this derivation produced the following outputs:

  out -> /nix/store/ni66p4jfqksbmsl616llx3fbs1d232d4-simple

Now you can run /nix/store/ni66p4jfqksbmsl616llx3fbs1d232d4-simple/simple in your shell.

We added two new attributes to the derivation call, gcc and coreutils. In gcc = gcc;, the name on the left is the name in the derivation set, and the name on the right refers to the gcc derivation from nixpkgs. The same applies for coreutils.

We also added the src attribute, nothing magical — it's just a name, to which the path ./simple.c is assigned. Like simple-builder.sh, simple.c will be added to the store.

The trick: every attribute in the set passed to derivation will be converted to a string and passed to the builder as an environment variable. This is how the builder gains access to coreutils and gcc: when converted to strings, the derivations evaluate to their output paths, and appending /bin to these leads us to their binaries.

The same goes for the src variable. $src is the path to simple.c in the nix store. As an exercise, pretty print the .drv file. You'll see simple_builder.sh and simple.c listed in the input derivations, along with bash, gcc and coreutils .drv files. The newly added environment variables described above will also appear.

In simple_builder.sh we set the PATH for gcc and coreutils binaries, so that our build script can find the necessary utilities like mkdir and gcc.

We then create $out as a directory and place the binary inside it. Note that gcc is found via the PATH environment variable, but it could equivalently be referenced explicitly using $gcc/bin/gcc.

Drop out of nix repl and write a file simple.nix:

with (import <nixpkgs> {});
derivation {
  name = "simple";
  builder = "${bash}/bin/bash";
  args = [ ./simple_builder.sh ];
  inherit gcc coreutils;
  src = ./simple.c;
  system = builtins.currentSystem;
}

Now you can build it with nix-build simple.nix. This will create a symlink result in the current directory, pointing to the out path of the derivation.

nix-build does two jobs:

Finally, it creates the symlink.

In the first line of simple.nix, we have an import function call nested in a with statement. Recall that import accepts one argument, a nix file to load. In this case, the contents of the file evaluated to a function.

Afterwards, we call the function with the empty set. We saw this already in the fifth pill. To reiterate: import <nixpkgs> {} is calling two functions, not one. Reading it as (import <nixpkgs>) {} makes this clearer.

The value returned by the nixpkgs function is a set. More specifically, it's a set of derivations. Using the with expression we bring them into scope. This is equivalent to the :l <nixpkgs> we used in nix repl; it allows us to easily access derivations such as bash, gcc, and coreutils.

Then we meet the inherit keyword. inherit foo; is equivalent to foo = foo;. Similarly, inherit foo bar; is equivalent to foo = foo; bar = bar;.

This syntax only makes sense inside sets. There's no magic involved, it's simply a convenience to avoid repeating the same name for both the attribute name and the value in scope.

We will generalize the builder. You may have noticed that we wrote two separate builder.sh scripts in this post. We would like to have a generic builder script instead, especially since each build script goes in the nix store: a bit of a waste.

Is it really that hard to package stuff in Nix? No, here we're studying the fundamentals of Nix.

In the previous pill we packaged a simple .c file, which was being compiled with a raw gcc call. That's not a good example of a project. Many use autotools, and since we're going to generalize our builder, it would be better do it with the most used build system.

GNU hello world, despite its name, is a simple yet complete project which uses autotools. Fetch the latest tarball here: http://ftp.gnu.org/gnu/hello/hello-2.10.tar.gz.

Let's create a builder script for GNU hello world, hello_builder.sh:

export PATH="$gnutar/bin:$gcc/bin:$gnumake/bin:$coreutils/bin:$gawk/bin:$gzip/bin:$gnugrep/bin:$gnused/bin:$binutils/bin"
tar -xzf $src
cd hello-2.10
./configure --prefix=$out
make
make install

And the derivation hello.nix:

with (import <nixpkgs> {});
derivation {
  name = "hello";
  builder = "${bash}/bin/bash";
  args = [ ./hello_builder.sh ];
  inherit gnutar gzip gnumake gcc coreutils gawk gnused gnugrep;
  binutils = binutils-unwrapped;
  src = ./hello-2.10.tar.gz;
  system = builtins.currentSystem;
}

with (import <nixpkgs> {});
derivation {
  name = "hello";
  builder = "${bash}/bin/bash";
  args = [ ./hello_builder.sh ];
  inherit gnutar gzip gnumake coreutils gawk gnused gnugrep;
  gcc = clang;
  binutils = clang.bintools.bintools_bin;
  src = ./hello-2.10.tar.gz;
  system = builtins.currentSystem;
}

Now build it with nix-build hello.nix and you can launch result/bin/hello. Nothing easier, but do we have to create a builder.sh for each package? Do we always have to pass the dependencies to the derivation function?

Please note the --prefix=$out we were talking about in the previous pill.

Let's create a generic builder.sh for autotools projects:

set -e
unset PATH
for p in $buildInputs; do
  export PATH=$p/bin${PATH:+:}$PATH
done

tar -xf $src

for d in *; do
  if [ -d "$d" ]; then
    cd "$d"
    break
  fi
done

./configure --prefix=$out
make
make install

What do we do here?

As you can see, there's no reference to "hello" in the builder anymore. It still does several assumptions, but it's certainly more generic.

Now let's rewrite hello.nix:

with (import <nixpkgs> {});
derivation {
  name = "hello";
  builder = "${bash}/bin/bash";
  args = [ ./builder.sh ];
  buildInputs = [ gnutar gzip gnumake gcc binutils-unwrapped coreutils gawk gnused gnugrep ];
  src = ./hello-2.10.tar.gz;
  system = builtins.currentSystem;
}

All clear, except that buildInputs. However it's easier than any black magic you are thinking in this moment.

Nix is able to convert a list to a string. It first converts the elements to strings, and then concatenates them separated by a space:

nix-repl> builtins.toString 123
"123"
nix-repl> builtins.toString [ 123 456 ]
"123 456"

Recall that derivations can be converted to a string, hence:

nix-repl> :l <nixpkgs>
Added 3950 variables.
nix-repl> builtins.toString gnugrep
"/nix/store/g5gdylclfh6d224kqh9sja290pk186xd-gnugrep-2.14"
nix-repl> builtins.toString [ gnugrep gnused ]
"/nix/store/g5gdylclfh6d224kqh9sja290pk186xd-gnugrep-2.14 /nix/store/krgdc4sknzpw8iyk9p20lhqfd52kjmg0-gnused-4.2.2"

Simple! The buildInputs variable is a string with out paths separated by space, perfect for bash usage in a for loop.

We managed to write a builder that can be used for multiple autotools projects. But in the hello.nix expression we are specifying tools that are common to more projects; we don't want to pass them everytime.

A natural approach would be to create a function that accepts an attribute set, similar to the one used by the derivation function, and merge it with another attribute set containing values common to many projects.

Create autotools.nix:

pkgs: attrs:
  with pkgs;
  let defaultAttrs = {
    builder = "${bash}/bin/bash";
    args = [ ./builder.sh ];
    baseInputs = [ gnutar gzip gnumake gcc binutils-unwrapped coreutils gawk gnused gnugrep ];
    buildInputs = [];
    system = builtins.currentSystem;
  };
  in
  derivation (defaultAttrs // attrs)

Ok now we have to remember a little about Nix functions. The whole nix expression of this autotools.nix file will evaluate to a function. This function accepts a parameter pkgs, then returns a function which accepts a parameter attrs.

The body of the function is simple, yet at first sight it might be hard to grasp:

The // operator is an operator between two sets. The result is the union of the two sets. In case of conflicts between attribute names, the value on the right set is preferred.

So we use defaultAttrs as base set, and add (or override) the attributes from attrs.

A couple of examples ought to be enough to clear out the behavior of the operator:

nix-repl> { a = "b"; } // { c = "d"; }
{ a = "b"; c = "d"; }
nix-repl> { a = "b"; } // { a = "c"; }
{ a = "c"; }

Exercise: Complete the new builder.sh by adding $baseInputs in the for loop together with $buildInputs. As you noticed, we passed that new variable in the derivation. Instead of merging buildInputs with the base ones, we prefer to preserve buildInputs as seen by the caller, so we keep them separated. Just a matter of choice.

Then we rewrite hello.nix as follows:

let
  pkgs = import <nixpkgs> {};
  mkDerivation = import ./autotools.nix pkgs;
in mkDerivation {
  name = "hello";
  src = ./hello-2.10.tar.gz;
}

Finally! We got a very simple description of a package! Below are a couple of remarks that you may find useful as you're continuing to understand the nix language:

Note we didn't use any other library. Special C flags may be needed to find include files of other libraries at compile time, and ld flags at link time.

Nix gives us the bare metal tools for creating derivations, setting up a build environment and storing the result in the nix store.

Out of this pill we managed to create a generic builder for autotools projects, and a function mkDerivation that composes by default the common components used in autotools projects instead of repeating them in all the packages we would write.

We are familiarizing ourselves with the way a Nix system grows up: it's about creating and composing derivations with the Nix language.

Analogy: in C you create objects in the heap, and then you compose them inside new objects. Pointers are used to refer to other objects.

In Nix you create derivations stored in the nix store, and then you compose them by creating new derivations. Store paths are used to refer to other derivations.

...we will talk a little about runtime dependencies. Is the GNU hello world package self-contained? What are its runtime dependencies? We only specified build dependencies by means of using other derivations in the "hello" derivation.

Let's start analyzing build dependencies for our GNU hello world package:

$ nix-instantiate hello.nix
/nix/store/z77vn965a59irqnrrjvbspiyl2rph0jp-hello.drv
$ nix-store -q --references /nix/store/z77vn965a59irqnrrjvbspiyl2rph0jp-hello.drv
/nix/store/0q6pfasdma4as22kyaknk4kwx4h58480-hello-2.10.tar.gz
/nix/store/1zcs1y4n27lqs0gw4v038i303pb89rw6-coreutils-8.21.drv
/nix/store/2h4b30hlfw4fhqx10wwi71mpim4wr877-gnused-4.2.2.drv
/nix/store/39bgdjissw9gyi4y5j9wanf4dbjpbl07-gnutar-1.27.1.drv
/nix/store/7qa70nay0if4x291rsjr7h9lfl6pl7b1-builder.sh
/nix/store/g6a0shr58qvx2vi6815acgp9lnfh9yy8-gnugrep-2.14.drv
/nix/store/jdggv3q1sb15140qdx0apvyrps41m4lr-bash-4.2-p45.drv
/nix/store/pglhiyp1zdbmax4cglkpz98nspfgbnwr-gnumake-3.82.drv
/nix/store/q9l257jn9lndbi3r9ksnvf4dr8cwxzk7-gawk-4.1.0.drv
/nix/store/rgyrqxz1ilv90r01zxl0sq5nq0cq7v3v-binutils-2.23.1.drv
/nix/store/qzxhby795niy6wlagfpbja27dgsz43xk-gcc-wrapper-4.8.3.drv
/nix/store/sk590g7fv53m3zp0ycnxsc41snc2kdhp-gzip-1.6.drv

It has exactly the derivations referenced in the derivation function, nothing more, nothing less. Some of them might not be used at all, however given that our generic mkDerivation function always pulls such dependencies (think of it like build-essential of Debian), for every package you build from now on, you will have these packages in the nix store.

Why are we looking at .drv files? Because the hello.drv file is the representation of the build action to perform in order to build the hello out path, and as such it also contains the input derivations needed to be built before building hello.

NAR is the Nix ARchive. First question: why not tar? Why another archiver? Because commonly used archivers are not deterministic. They add padding, they do not sort files, they add timestamps, etc.. Hence NAR, a very simple deterministic archive format being used by Nix for deployment. NARs are also used extensively within Nix itself as we'll see below.

For the rationale and implementation details you can find more in the Dolstra's PhD Thesis.

To create NAR archives, it's possible to use nix-store --dump and nix-store --restore. Those two commands work regardless of /nix/store.

Something is different for runtime dependencies however. Build dependencies are automatically recognized by Nix once they are used in any derivation call, but we never specify what are the runtime dependencies for a derivation.

There's really black magic involved. It's something that at first glance makes you think "no, this can't work in the long term", but at the same it works so well that a whole operating system is built on top of this magic.

In other words, Nix automatically computes all the runtime dependencies of a derivation, and it's possible thanks to the hash of the store paths.

Steps:

You get really all the runtime dependencies, and that's why Nix deployments are so easy.

$ nix-instantiate hello.nix
/nix/store/z77vn965a59irqnrrjvbspiyl2rph0jp-hello.drv
$ nix-store -r /nix/store/z77vn965a59irqnrrjvbspiyl2rph0jp-hello.drv
/nix/store/a42k52zwv6idmf50r9lps1nzwq9khvpf-hello
$ nix-store -q --references /nix/store/a42k52zwv6idmf50r9lps1nzwq9khvpf-hello
/nix/store/94n64qy99ja0vgbkf675nyk39g9b978n-glibc-2.19
/nix/store/8jm0wksask7cpf85miyakihyfch1y21q-gcc-4.8.3
/nix/store/a42k52zwv6idmf50r9lps1nzwq9khvpf-hello

Ok glibc and gcc. Well, gcc really should not be a runtime dependency!

$ strings result/bin/hello|grep gcc
/nix/store/94n64qy99ja0vgbkf675nyk39g9b978n-glibc-2.19/lib:/nix/store/8jm0wksask7cpf85miyakihyfch1y21q-gcc-4.8.3/lib64

Oh Nix added gcc because its out path is mentioned in the "hello" binary. Why is that? That's the ld rpath. It's the list of directories where libraries can be found at runtime. In other distributions, this is usually not abused. But in Nix, we have to refer to particular versions of libraries, thus the rpath has an important role.

The build process adds that gcc lib path thinking it may be useful at runtime, but really it's not. How do we get rid of it? Nix authors have written another magical tool called patchelf, which is able to reduce the rpath to the paths that are really used by the binary.

Even after reducing the rpath, the hello binary would still depend upon gcc because of some debugging information. This unnecesarily increases the size of our runtime dependencies. We'll explore how strip can help us with that in the next section.

We will add a new phase to our autotools builder. The builder has these phases already:

We add a new phase after the installation phase, which we call fixup phase. At the end of the builder.sh follows:

find $out -type f -exec patchelf --shrink-rpath '{}' \; -exec strip '{}' \; 2>/dev/null

That is, for each file we run patchelf --shrink-rpath and strip. Note that we used two new commands here, find and patchelf. Exercise: These two deserve a place in baseInputs of autotools.nix as findutils and patchelf.

Rebuild hello.nix and...:

$ nix-build hello.nix
[...]
$ nix-store -q --references result
/nix/store/94n64qy99ja0vgbkf675nyk39g9b978n-glibc-2.19
/nix/store/md4a3zv0ipqzsybhjb8ndjhhga1dj88x-hello

...only glibc is the runtime dependency. Exactly what we wanted.

The package is self-contained, copy its closure on another machine and you will be able to run it. Remember, only a very few components under the /nix/store are required to run nix. The hello binary will use that exact version of glibc library and interpreter, not the system one:

$ ldd result/bin/hello
 linux-vdso.so.1 (0x00007fff11294000)
 libc.so.6 => /nix/store/94n64qy99ja0vgbkf675nyk39g9b978n-glibc-2.19/lib/libc.so.6 (0x00007f7ab7362000)
 /nix/store/94n64qy99ja0vgbkf675nyk39g9b978n-glibc-2.19/lib/ld-linux-x86-64.so.2 (0x00007f7ab770f000)

Of course, the executable runs fine as long as everything is under the /nix/store path.

Short post compared to previous ones as I'm still on vacation, but I hope you enjoyed it. Nix provides tools with cool features. In particular, Nix is able to compute all runtime dependencies automatically for us. This is not limited to only shared libraries, but also referenced executables, scripts, Python libraries etc..

This makes packages self-contained, ensuring (apart data and configuration) that copying the runtime closure on another machine is sufficient to run the program. That's why Nix has one-click install, or reliable deployment in the cloud. All with one tool.

...we will introduce nix-shell. With nix-build we always build derivations from scratch: the source gets unpacked, configured, built and installed. But this may take a long time, think of WebKit. What if we want to apply some small changes and compile incrementally instead, yet keeping a self-contained environment similar to nix-build?

The nix-shell tool drops us in a shell by setting up the necessary environment variables to hack on a derivation. It does not build the derivation, it only serves as a preparation so that we can run the build steps manually.

I remind you, in a nix environment you don't have access to libraries and programs unless you install them with nix-env. However installing libraries with nix-env is not good practice. We prefer to have isolated environments for development.

$ nix-shell hello.nix
[nix-shell]$ make
bash: make: command not found
[nix-shell]$ echo $baseInputs
/nix/store/jff4a6zqi0yrladx3kwy4v6844s3swpc-gnutar-1.27.1 [...]

First thing to notice, we call nix-shell on a nix expression which returns a derivation. We then enter a new bash shell, but it's really useless. We expected to have the GNU hello world build inputs available in PATH, including GNU make, but it's not the case.

But, we have the environment variables that we set in the derivation, like $baseInputs, $buildInputs, $src and so on.

That means we can source our builder.sh, and it will build the derivation. You may get an error in the installation phase, because the user may not have the permission to write to /nix/store:

[nix-shell]$ source builder.sh
...

It didn't install, but it built. Things to notice:

We're able to cd into hello-2.10 and type make, because now it's available.

In other words, nix-shell drops us in a shell with the same (or almost) environment used to run the builder!

The previous steps are a bit annoying of course, but we can improve our builder to be more nix-shell friendly.

First of all, we were able to source builder.sh because it was in our current directory, but that's not nice. We want the builder.sh that is stored in the nix store, the one that would be used by nix-build. To do so, the right way is to pass the usual environment variable through the derivation.

Note: $builder is already defined, but it's the bash executable, not our builder.sh. Our builder.sh is an argument to bash.

Second, we don't want to run the whole builder, we only want it to setup the necessary environment for manually building the project. So we'll write two files, one for setting up the environment, and the real builder.sh that runs with nix-build.

Additionally, we'll wrap the phases in functions, it may be useful, and move the set -e to the builder instead of the setup. The set -e is annoying in nix-shell.

Here is our modified autotools.nix. Noteworthy is the setup = ./setup.sh; attribute in the derivation, which adds setup.sh to the nix store and as usual, adds a $setup environment variable in the builder.

pkgs: attrs:
  with pkgs;
  let defaultAttrs = {
    builder = "${bash}/bin/bash";
    args = [ ./builder.sh ];
    setup = ./setup.sh;
    baseInputs = [ gnutar gzip gnumake gcc binutils-unwrapped coreutils gawk gnused gnugrep patchelf findutils ];
    buildInputs = [];
    system = builtins.currentSystem;
  };
  in
derivation (defaultAttrs // attrs)

Thanks to that, we can split builder.sh into setup.sh and builder.sh. What builder.sh does is sourcing $setup and calling the genericBuild function. Everything else is just some bash changes.

Here is the modified builder.sh.

set -e
source $setup
genericBuild

Here is the newly added setup.sh.

unset PATH
for p in $baseInputs $buildInputs; do
  export PATH=$p/bin${PATH:+:}$PATH
done

function unpackPhase() {
  tar -xzf $src

  for d in *; do
    if [ -d "$d" ]; then
      cd "$d"
      break
    fi
  done
}

function configurePhase() {
  ./configure --prefix=$out
}

function buildPhase() {
  make
}

function installPhase() {
  make install
}

function fixupPhase() {
  find $out -type f -exec patchelf --shrink-rpath '{}' \; -exec strip '{}' \; 2>/dev/null
}

function genericBuild() {
  unpackPhase
  configurePhase
  buildPhase
  installPhase
  fixupPhase
}

Finally, here is hello.nix.

let
  pkgs = import <nixpkgs> {};
  mkDerivation = import ./autotools.nix pkgs;
in mkDerivation {
  name = "hello";
  src = ./hello-2.10.tar.gz;
}

Now back to nix-shell:

$ nix-shell hello.nix
[nix-shell]$ source $setup
[nix-shell]$

Now you can run, for example, unpackPhase which unpacks $src and enters the directory. And you can run commands like ./configure, make etc. manually, or run phases with their respective functions.

It's that straightforward, nix-shell builds the .drv file and its input dependencies, then drops into a shell by setting up the environment variables necessary to build the .drv, in particular those passed to the derivation function.

With nix-shell we're able to drop into an isolated environment for developing a project, with the necessary dependencies just like nix-build does. Additionally, we can build and debug the project manually, step by step like you would do in any other operating system. Note that we never installed gcc, make, etc. system-wide. These tools and libraries are available per-build.

...we will clean up the nix store. We wrote and built derivations, added stuff to nix store, but until now we never worried about cleaning up the used space in the store. It's time to collect some garbage.

Other package managers, like dpkg, have ways of removing unused software. Nix is much more precise in its garbage collection compared to these other systems.

I bet with dpkg, rpm or similar traditional packaging systems, you end up having some unnecessary packages installed or dangling files. With nix this does not happen.

How do we determine whether a store path is still needed? The same way programming languages with a garbage collector decide whether an object is still alive.

Programming languages with a garbage collector have an important concept in order to keep track of live objects: GC roots. A GC root is an object that is always alive (unless explicitly removed as GC root). All objects recursively referred to by a GC root are live.

Therefore, the garbage collection process starts from GC roots, and recursively mark referenced objects as live. All other objects can be collected and deleted.

In Nix there's this same concept. Instead of being objects, of course, GC roots are store paths. The implementation is very simple and transparent to the user. GC roots are stored under /nix/var/nix/gcroots. If there's a symlink to a store path, then that store path is a GC root.

Nix allows this directory to have subdirectories: it will simply recurse directories in search of symlinks to store paths.

So we have a list of GC roots. At this point, deleting dead store paths is as easy as you can imagine. We have the list of all live store paths, hence the rest of the store paths are dead.

In particular, Nix first moves dead store paths to /nix/store/trash which is an atomic operation. Afterwards, the trash is emptied.

Before playing with the GC, first run the nix garbage collector once, so that we have a clean playground for our experiments:

$ nix-collect-garbage
finding garbage collector roots...
[...]
deleting unused links...
note: currently hard linking saves -0.00 MiB
1169 store paths deleted, 228.43 MiB freed

Perfect, if you run it again it won't find anything new to delete, as expected.

What's left in the nix store is everything being referenced from the GC roots.

Let's install for a moment bsd-games:

$ nix-env -iA nixpkgs.bsdgames
$ readlink -f `which fortune`
/nix/store/b3lxx3d3ggxcggvjw5n0m1ya1gcrmbyn-bsd-games-2.17/bin/fortune
$ nix-store -q --roots `which fortune`
/nix/var/nix/profiles/default-9-link
$ nix-env --list-generations
[...]
   9   2014-08-20 12:44:14   (current)

The nix-store command can be used to query the GC roots that refer to a given derivation. In this case, our current user environment does refer to bsd-games.

Now remove it, collect garbage and note that bsd-games is still in the nix store:

$ nix-env -e bsd-games
uninstalling `bsd-games-2.17'
$ nix-collect-garbage
[...]
$ ls /nix/store/b3lxx3d3ggxcggvjw5n0m1ya1gcrmbyn-bsd-games-2.17
bin  share

This is because the old generation is still in the nix store because it's a GC root. As we'll see below, all profiles and their generations are GC roots.

Removing a GC root is simple. Let's try deleting the generation that refers to bsd-games, collect garbage, and note that now bsd-games is no longer in the nix store:

$ rm /nix/var/nix/profiles/default-9-link
$ nix-env --list-generations
[...]
   8   2014-07-28 10:23:24
  10   2014-08-20 12:47:16   (current)
$ nix-collect-garbage
[...]
$ ls /nix/store/b3lxx3d3ggxcggvjw5n0m1ya1gcrmbyn-bsd-games-2.17
ls: cannot access /nix/store/b3lxx3d3ggxcggvjw5n0m1ya1gcrmbyn-bsd-games-2.17: No such file or directory

Note: nix-env --list-generations does not rely on any particular metadata. It is able to list generations based solely on the file names under the profiles directory.

However we removed the link from /nix/var/nix/profiles, not from /nix/var/nix/gcroots. Turns out, that /nix/var/nix/gcroots/profiles is a symlink to /nix/var/nix/profiles. That is very handy. It means any profile and its generations are GC roots.

It's as simple as that, anything under /nix/var/nix/gcroots is a GC root. And anything not being garbage collected is because it's referred from one of the GC roots.

Remember that building the GNU hello world package with nix-build produces a result symlink in the current directory. Despite the collected garbage done above, the hello program is still working: therefore it has not been garbage collected. Clearly, since there's no other derivation that depends upon the GNU hello world package, it must be a GC root.

In fact, nix-build automatically adds the result symlink as a GC root. Yes, not the built derivation, but the symlink. These GC roots are added under /nix/var/nix/gcroots/auto.

$ ls -l /nix/var/nix/gcroots/auto/
total 8
drwxr-xr-x 2 nix nix 4096 Aug 20 10:24 ./
drwxr-xr-x 3 nix nix 4096 Jul 24 10:38 ../
lrwxrwxrwx 1 nix nix   16 Jul 31 10:51 xlgz5x2ppa0m72z5qfc78b8wlciwvgiz -> /home/nix/result/

Don't care about the name of the symlink. What's important is that a symlink exists that point to /home/nix/result. This is called an indirect GC root. That is, the GC root is effectively specified outside of /nix/var/nix/gcroots. Whatever result points to, it will not be garbage collected.

How do we remove the derivation then? There are two possibilities:

In the first case, the derivation will be deleted from the nix store, and result becomes a dangling symlink. In the second case, the derivation is removed as well as the indirect root in /nix/var/nix/gcroots/auto.

Running nix-collect-garbage after deleting the GC root or the indirect GC root, will remove the derivation from the store.

What's the main source of software duplication in the nix store? Clearly, GC roots due to nix-build and profile generations. Doing a nix-build results in a GC root for a build that somehow will refer to a specific version of glibc, and other libraries. After an upgrade, if that build is not deleted by the user, it will not be garbage collected. Thus the old dependencies referred to by the build will not be deleted either.

Same goes for profiles. Manipulating the nix-env profile will create further generations. Old generations refer to old software, thus increasing duplication in the nix store after an upgrade.

What are the basic steps for upgrading and removing everything old, including old generations? In other words, do an upgrade similar to other systems, where they forget everything about the older state:

$ nix-channel --update
$ nix-env -u --always
$ rm /nix/var/nix/gcroots/auto/*
$ nix-collect-garbage -d

First, we download a new version of the nixpkgs channel, which holds the description of all the software. Then we upgrade our installed packages with nix-env -u. That will bring us into a fresh new generation with all updated software.

Then we remove all the indirect roots generated by nix-build: beware, this will result in dangling symlinks. You may be smarter and also remove the target of those symlinks.

Finally, the -d option of nix-collect-garbage is used to delete old generations of all profiles, then collect garbage. After this, you lose the ability to rollback to any previous generation. So make sure the new generation is working well before running the command.

Garbage collection in Nix is a powerful mechanism to cleanup your system. The nix-store commands allow us to know why a certain derivation is in the nix store.

Cleaning up everything down to the oldest bit of software after an upgrade seems a bit contrived, but that's the price of having multiple generations, multiple profiles, multiple versions of software, thus rollbacks etc.. The price of having many possibilities.

...we will package another project and introduce what I call the "inputs" design pattern. We've only played with a single derivation until now, however we'd like to start organizing a small repository of software. The "inputs" pattern is widely used in nixpkgs; it allows us to decouple derivations from the repository itself and increase customization opportunities.

Nix is a tool for build and deployment, it does not enforce any particular repository format. A repository of packages is the main usage for Nix, but not the only possibility. It's more like a consequence due to the need of organizing packages.

Nix is a language, and it is powerful enough to let you choose the format of your own repository. In this sense, it is not declarative, but functional.

There is no preset directory structure or preset packaging policy. It's all about you and Nix.

The nixpkgs repository has a certain structure, which evolved and evolves with the time. Like other languages, Nix has its own history and therefore I'd like to say that it also has its own design patterns. Especially when packaging, you often do the same task again and again except for different software. It's inevitable to identify patterns during this process. Some of these patterns get reused if the community thinks it's a good way to package the software.

Some of the patterns I'm going to show do not apply only to Nix, but to other systems of course.

Before introducing the "inputs" pattern, we can start talking about another pattern first which I'd like to call "single repository" pattern.

Systems like Debian scatter packages in several small repositories. This can make it hard to track interdependent changes and to contribute to new packages.

Alternatively, systems like Gentoo put package descriptions all in a single repository.

The nix reference for packages is nixpkgs, a single repository of all descriptions of all packages. I find this approach very natural and attractive for new contributions.

For the rest of this chapter, we will adopt the single repository technique. The natural implementation in Nix is to create a top-level Nix expression, and one expression for each package. The top-level expression imports and combines all expressions in a giant attribute set with name -> package pairs.

But isn't that heavy? It isn't, because Nix is a lazy language, it evaluates only what's needed! And that's why nixpkgs is able to maintain such a big software repository in a giant attribute set.

We have packaged GNU hello world, imagine you would like to package something else for creating at least a repository of two projects :-) . I chose graphviz, which uses the standard autotools build system, requires no patching and dependencies are optional.

Download graphviz from here. The graphviz.nix expression is straightforward:

let
  pkgs = import <nixpkgs> {};
  mkDerivation = import ./autotools.nix pkgs;
in mkDerivation {
  name = "graphviz";
  src = ./graphviz-2.38.0.tar.gz;
}

Build with nix-build graphviz.nix and you will get runnable binaries under result/bin. Notice how we reused the same autotools.nix of hello.nix. Let's create a simple png:

$ echo 'graph test { a -- b }'|result/bin/dot -Tpng -o test.png
Format: "png" not recognized. Use one of: canon cmap [...]

Oh of course... graphviz doesn't know about png. It built only the output formats it supports natively, without using any extra library.

Remember, in autotools.nix there's a buildInputs variable which gets concatenated to baseInputs. That would be the perfect place to add a build dependency. We created that variable exactly for this reason to be overridable from package expressions.

This 2.38 version of graphviz has several plugins to output png. For simplicity, we will use libgd.

The gd, jpeg, fontconfig and bzip2 libraries (dependencies of gd) don't use pkg-config to specify which flags to pass to the compiler. Since there's no global location for libraries, we need to tell gcc and ld where to find includes and libs.

The nixpkgs provides gcc and binutils, which we are currently using for our packaging. It also provides wrappers for them which allow passing extra arguments to gcc and ld, bypassing the project build systems:

What can we do about it? We can employ the same trick we did for PATH: automatically filling the variables from buildInputs. This is the relevant snippet of setup.sh:

for p in $baseInputs $buildInputs; do
  if [ -d $p/bin ]; then
    export PATH="$p/bin${PATH:+:}$PATH"
  fi
  if [ -d $p/include ]; then
    export NIX_CFLAGS_COMPILE="-I $p/include${NIX_CFLAGS_COMPILE:+ }$NIX_CFLAGS_COMPILE"
  fi
  if [ -d $p/lib ]; then
    export NIX_LDFLAGS="-rpath $p/lib -L $p/lib${NIX_LDFLAGS:+ }$NIX_LDFLAGS"
  fi
done

Now adding derivations to buildInputs will add their lib, include and bin paths automatically in setup.sh.

The [-rpath] flag in ld is needed because at runtime, the executable must use exactly that version of the library.

If unneeded paths are specified, the fixup phase will shrink the rpath for us!

Finish the expression for graphviz with gd support (note the use of the with expression in buildInputs to avoid repeating pkgs):

let
  pkgs = import <nixpkgs> {};
  mkDerivation = import ./autotools.nix pkgs;
in mkDerivation {
  name = "graphviz";
  src = ./graphviz-2.38.0.tar.gz;
  buildInputs = with pkgs; [ gd fontconfig libjpeg bzip2 ];
}

Now you can create the png! Ignore any error from fontconfig, especially if you are in a chroot.

Now that we have two packages, what's a good way to put them together in a single repository? We'll do something like nixpkgs does. With nixpkgs, we import it and then we pick derivations by accessing the giant attribute set.

For us nixers, this is a good technique, because it abstracts from the file names. We don't refer to a package by REPO/some/sub/dir/package.nix but by importedRepo.package (or pkgs.package in our examples).

Create a default.nix in the current directory:

{
  hello = import ./hello.nix;
  graphviz = import ./graphviz.nix;
}

Ready to use! Try it with nix repl:

$ nix repl
nix-repl> :l default.nix
Added 2 variables.
nix-repl> hello
«derivation /nix/store/dkib02g54fpdqgpskswgp6m7bd7mgx89-hello.drv»
nix-repl> graphviz
«derivation /nix/store/zqv520v9mk13is0w980c91z7q1vkhhil-graphviz.drv»

With nix-build:

$ nix-build default.nix -A hello
[...]
$ result/bin/hello
Hello, world!

The [-A] argument is used to access an attribute of the set from the given .nix expression.

Important: why did we choose the default.nix? Because when a directory (by default the current directory) has a default.nix, that default.nix will be used (see import here). In fact you can run nix-build -A hello without specifying default.nix.

For pythoners, it is similar to __init__.py.

With nix-env, install the package into your user environment:

$ nix-env -f . -iA graphviz
[...]
$ dot -V

The [-f] option is used to specify the expression to use, in this case the current directory, therefore ./default.nix.

The [-i] stands for installation.

The [-A] is the same as above for nix-build.

We reproduced the very basic behavior of nixpkgs.

After a long preparation, we finally arrived. I know you're having a big doubt in this moment. It's about the hello.nix and graphviz.nix. They are very, very dependent on nixpkgs:

The current answers to the above questions are: change the expression to match your needs (or change the callee to match your needs).

With the inputs pattern, we decided to provide another answer: let the user change the inputs of the expression (or change the caller to pass different inputs).

By inputs of an expression, we refer to the set of derivations needed to build that expression. In this case:

The src is also an input but it's pointless to change the source from the caller. For version bumps, in nixpkgs we prefer to write another expression (e.g. because patches are needed or different inputs are needed).

Goal: make package expressions independent of the repository.

How do we achieve that? The answer is simple: use functions to declare inputs for a derivation. Doing it for graphviz.nix, will make the derivation independent of the repository and customizable:

{ mkDerivation, gdSupport ? true, gd, fontconfig, libjpeg, bzip2 }:

mkDerivation {
  name = "graphviz";
  src = ./graphviz-2.38.0.tar.gz;
  buildInputs = if gdSupport then [ gd fontconfig libjpeg bzip2 ] else [];
}

I recall that "{...}: ..." is the syntax for defining functions accepting an attribute set as argument.

We made gd and its dependencies optional. If gdSupport is true (by default), we will fill buildInputs and thus graphviz will be built with gd support, otherwise it won't.

Now back to default.nix:

let
  pkgs = import <nixpkgs> {};
  mkDerivation = import ./autotools.nix pkgs;
in with pkgs; {
  hello = import ./hello.nix { inherit mkDerivation; };
  graphviz = import ./graphviz.nix { inherit mkDerivation gd fontconfig libjpeg bzip2; };
  graphvizCore = import ./graphviz.nix {
    inherit mkDerivation gd fontconfig libjpeg bzip2;
    gdSupport = false;
  };
}

So we factorized the import of nixpkgs and mkDerivation, and also added a variant of graphviz with gd support disabled. The result is that both hello.nix (exercise for the reader) and graphviz.nix are independent of the repository and customizable by passing specific inputs.

If you wanted to build graphviz with a specific version of gd, it would suffice to pass gd = ...;.

If you wanted to change the toolchain, you may pass a different mkDerivation function.

Clearing up the syntax:

You can find the whole repository at the pill 12 gist.

The "inputs" pattern allows our expressions to be easily customizable through a set of arguments. These arguments could be flags, derivations, or whatever else. Our package expressions are functions, don't think there's any magic in there.

It also makes the expressions independent of the repository. Given that all the needed information is passed through arguments, it is possible to use that expression in any other context.

...we will talk about the "callPackage" design pattern. It is tedious to specify the names of the inputs twice, once in the top-level default.nix, and once in the package expression. With callPackage, we will implicitly pass the necessary inputs from the top-level expression.

In the previous pill, we underlined the fact that the inputs pattern is great to decouple packages from the repository, in that we can pass manually the inputs to the derivation. The derivation declares its inputs, and the caller passes the arguments.

However as with usual programming languages, we declare parameter names, and then we have to pass arguments. We do the job twice. With package management, we often see common patterns. In the case of nixpkgs it's the following.

Some package derivation:

{ input1, input2, ... }:
...

Repository derivation:

rec {
  lib1 = import package1.nix { inherit input1 input2 ...; };
  program2 = import package1.nix { inherit inputX inputY lib1 ...; };
}

Where inputs may even be packages in the repository itself (note the rec keyword). The pattern here is clear, often inputs have the same name of the attributes in the repository itself. Our desire is to pass those inputs from the repository automatically, and in case be able to specify a particular argument (that is, override the automatically passed default argument).

To achieve this, we will define a callPackage function with the following synopsis:

{
  lib1 = callPackage package1.nix { };
  program2 = callPackage package2.nix { someoverride = overriddenDerivation; };
}

What should it do?

First of all, we need a way to introspect (reflection or whatever) at runtime the argument names of a function. That's because we want to automatically pass such arguments.

Then callPackage requires access to the whole packages set, because it needs to find the packages to pass automatically.

We start off simple with :

nix-repl> add = { a ? 3, b }: a+b
nix-repl> builtins.functionArgs add
{ a = true; b = false; }

Nix provides a builtin function to introspect the names of the arguments of a function. In addition, for each argument, it tells whether the argument has a default value or not. We don't really care about default values in our case. We are only interested in the argument names.

Now we need a set with all the values, let's call it values. And a way to intersect the attributes of values with the function arguments:

nix-repl> values = { a = 3; b = 5; c = 10; }
nix-repl> builtins.intersectAttrs values (builtins.functionArgs add)
{ a = true; b = false; }
nix-repl> builtins.intersectAttrs (builtins.functionArgs add) values
{ a = 3; b = 5; }

Perfect, note from the example above that the intersectAttrs returns a set whose names are the intersection, and the attribute values are taken from the second set.

We're done, we have a way to get argument names from a function, and match with an existing set of attributes. This is our simple implementation of callPackage:

nix-repl> callPackage = set: f: f (builtins.intersectAttrs (builtins.functionArgs f) set)
nix-repl> callPackage values add
8
nix-repl> with values; add { inherit a b; }
8

Clearing up the syntax:

In the code above, I've also shown that the callPackage call is equivalent to directly calling add a b.

We achieved what we wanted. Automatically call functions given a set of possible arguments. If an argument is not found in the set, that's nothing special. It's a function call with a missing parameter, and that's an error (unless the function has varargs ... as explained in the 5th pill).

Or not. We missed something. Being able to override some of the parameters. We may not want to always call functions with values taken from the big set. Then we add a further parameter, which takes a set of overrides:

nix-repl> callPackage = set: f: overrides: f ((builtins.intersectAttrs (builtins.functionArgs f) set) // overrides)
nix-repl> callPackage values add { }
8
nix-repl> callPackage values add { b = 12; }
15

Apart from the increasing number of parenthesis, it should be clear that we simply do a set union between the default arguments, and the overriding set.

Given our brand new tool, we can simplify the repository expression (default.nix).

Let me write it down first:

let
  nixpkgs = import <nixpkgs> {};
  allPkgs = nixpkgs // pkgs;
  callPackage = path: overrides:
    let f = import path;
    in f ((builtins.intersectAttrs (builtins.functionArgs f) allPkgs) // overrides);
  pkgs = with nixpkgs; {
    mkDerivation = import ./autotools.nix nixpkgs;
    hello = callPackage ./hello.nix { };
    graphviz = callPackage ./graphviz.nix { };
    graphvizCore = callPackage ./graphviz.nix { gdSupport = false; };
  };
in pkgs

Wow, there's a lot to say here:

Note how easy is to override arguments in the case of graphviz without gd. But most importantly, how easy it was to merge two repositories: nixpkgs and our pkgs!

The reader should notice a magic thing happening. We're defining pkgs in terms of callPackage, and callPackage in terms of pkgs. That magic is possible thanks to lazy evaluation.

The "callPackage" pattern has simplified our repository a lot. We're able to import packages that require some named arguments and call them automatically, given the set of all packages.

We've also introduced some useful builtin functions that allows us to introspect Nix functions and manipulate attributes. These builtin functions are not usually used when packaging software, rather to provide tools for packaging. That's why they are not documented in the nix manual.

Writing a repository in nix is an evolution of writing convenient functions for combining the packages. This demonstrates even more how nix is a generic tool to build and deploy something, and how suitable it is to create software repositories with your own conventions.

...we will talk about the "override" design pattern. The graphvizCore seems straightforward. It starts from graphviz.nix and builds it without gd. Now I want to give you another point of view: what if we instead wanted to start from pkgs.graphviz and disable gd?

Functional languages are known for being able to compose functions. In particular, you gain a lot from functions that are able to manipulate the original value into a new value having the same structure. So that in the end we're able to call multiple functions to have the desired modifications.

In Nix we mostly talk about functions that accept inputs in order to return derivations. In our world we want nice utility functions that are able to manipulate those structures. These utilities add some useful properties to the original value, and we must be able to apply more utilities on top of it.

For example let's say we have an initial derivation drv and we want it to be a drv with debugging information and also to apply some custom patches:

debugVersion (applyPatches [ ./patch1.patch ./patch2.patch ] drv)

The final result will be still the original derivation plus some changes. That's both interesting and very different from other packaging approaches, which is a consequence of using a functional language to describe packages.

Designing such utilities is not trivial in a functional language that is not statically typed, because understanding what can or cannot be composed is difficult. But we try to do the best.

In the pill 12 we introduced the inputs design pattern. We do not return a derivation picking dependencies directly from the repository, rather we declare the inputs and let the callers pass the necessary arguments.

In our repository we have a set of attributes that import the expressions of the packages and pass these arguments, getting back a derivation. Let's take for example the graphviz attribute:

graphviz = import ./graphviz.nix { inherit mkDerivation gd fontconfig libjpeg bzip2; };

If we wanted to produce a derivation of graphviz with a customized gd version, we would have to repeat most of the above plus specifying an alternative gd:

mygraphviz = import ./graphviz.nix {
  inherit mkDerivation fontconfig libjpeg bzip2;
  gd = customgd;
};

That's hard to maintain. Using callPackage it would be easier:

mygraphviz = callPackage ./graphviz.nix { gd = customgd; };

But we may still be diverging from the original graphviz in the repository.

We would like to avoid specifying the nix expression again, instead reuse the original graphviz attribute in the repository and add our overrides like this:

mygraphviz = graphviz.override { gd = customgd; };

The difference is obvious, as well as the advantages of this approach.

Note: that .override is not a "method" in the OO sense as you may think. Nix is a functional language. That .override is simply an attribute of a set.

I remind you, the graphviz attribute in the repository is the derivation returned by the function imported from graphviz.nix. We would like to add a further attribute named "override" to the returned set.

Let's start simple by first creating a function "makeOverridable" that takes a function and a set of original arguments to be passed to the function.

Contract: the wrapped function must return a set.

Let's write a lib.nix:

{
  makeOverridable = f: origArgs:
    let
      origRes = f origArgs;
    in
      origRes // { override = newArgs: f (origArgs // newArgs); };
}

So makeOverridable takes a function and a set of original arguments. It returns the original returned set, plus a new override attribute.

This override attribute is a function taking a set of new arguments, and returns the result of the original function called with the original arguments unified with the new arguments. What a mess.

Let's try it with nix repl:

$ nix repl
nix-repl> :l lib.nix
Added 1 variables.
nix-repl> f = { a, b }: { result = a+b; }
nix-repl> f { a = 3; b = 5; }
{ result = 8; }
nix-repl> res = makeOverridable f { a = 3; b = 5; }
nix-repl> res
{ override = «lambda»; result = 8; }
nix-repl> res.override { a = 10; }
{ result = 15; }

Note that the function f does not return the plain sum but a set, because of the contract. You didn't forget already, did you? :-)

The variable res is the result of the function call without any override. It's easy to see in the definition of makeOverridable. In addition you can see the new override attribute being a function.

Calling that .override with a set will invoke the original function with the overrides, as expected.

But: we can't override again! Because the returned set with result 15 does not have an override attribute!

That's bad, it breaks further compositions.

The solution is simple, the .override function should make the result overridable again:

rec {
  makeOverridable = f: origArgs:
    let
      origRes = f origArgs;
    in
      origRes // { override = newArgs: makeOverridable f (origArgs // newArgs); };
}

Please note the rec keyword. It's necessary so that we can refer to makeOverridable from makeOverridable itself.

Now let's try overriding twice:

nix-repl> :l lib.nix
Added 1 variables.
nix-repl> f = { a, b }: { result = a+b; }
nix-repl> res = makeOverridable f { a = 3; b = 5; }
nix-repl> res2 = res.override { a = 10; }
nix-repl> res2
{ override = «lambda»; result = 15; }
nix-repl> res2.override { b = 20; }
{ override = «lambda»; result = 30; }

Success! The result is 30, as expected because a is overridden to 10 in the first override, and b to 20.

Now it would be nice if callPackage made our derivations overridable. That was the goal of this pill after all. This is an exercise for the reader.

The "override" pattern simplifies the way we customize packages starting from an existing set of packages. This opens a world of possibilities about using a central repository like nixpkgs, and defining overrides on our local machine without even modifying the original package.

Dream of a custom isolated nix-shell environment for testing graphviz with a custom gd:

debugVersion (graphviz.override { gd = customgd; })

Once a new version of the overridden package comes out in the repository, the customized package will make use of it automatically.

The key in Nix is to find powerful yet simple abstractions in order to let the user customize his environment with highest consistency and lowest maintenance time, by using predefined composable components.

...we will talk about Nix search paths. By search path I mean a place in the file system where Nix looks for expressions. You may have wondered, where does that holy <nixpkgs> come from?

The NIX_PATH environment variable is very important. It's very similar to the PATH environment variable. The syntax is similar, several paths are separated by a colon :. Nix will then search for something in those paths from left to right.

Who uses NIX_PATH? The nix expressions! Yes, NIX_PATH is not of much use by the nix tools themselves, rather it's used when writing nix expressions.

In the shell for example, when you execute the command ping, it's being searched in the PATH directories. The first one found is the one being used.

In nix it's exactly the same, however the syntax is different. Instead of just typing ping you have to type <ping>. Yes, I know... you are already thinking of <nixpkgs>. However don't stop reading here, let's keep going.

What's NIX_PATH good for? Nix expressions may refer to an "abstract" path such as <nixpkgs>, and it's possible to override it from the command line.

For ease we will use nix-instantiate --eval to do our tests. I remind you, nix-instantiate is used to evaluate nix expressions and generate the .drv files. Here we are not interested in building derivations, so evaluation is enough. It can be used for one-shot expressions.

It's useless from a nix view point, but I think it's useful for your own understanding. Let's use PATH itself as NIX_PATH, and try to locate ping (or another binary if you don't have it).

$ nix-instantiate --eval -E '<ping>'
error: file `ping' was not found in the Nix search path (add it using $NIX_PATH or -I)
$ NIX_PATH=$PATH nix-instantiate --eval -E '<ping>'
/bin/ping
$ nix-instantiate -I /bin --eval -E '<ping>'
/bin/ping

Great. At first attempt nix obviously said could not be found anywhere in the search path. Note that the [-I] option accepts a single directory. Paths added with [-I] take precedence over NIX_PATH.

The NIX_PATH also accepts a different yet very handy syntax: "somename=somepath". That is, instead of searching inside a directory for a name, we specify exactly the value of that name.

$ NIX_PATH="ping=/bin/ping" nix-instantiate --eval -E '<ping>'
/bin/ping
$ NIX_PATH="ping=/bin/foo" nix-instantiate --eval -E '<ping>'
error: file `ping' was not found in the Nix search path (add it using $N

Note in the second case how Nix checks whether the path exists or not.

You are out of curiosity, right?

$ nix-instantiate --eval -E '<nixpkgs>'
/home/nix/.nix-defexpr/channels/nixpkgs
$ echo $NIX_PATH
nixpkgs=/home/nix/.nix-defexpr/channels/nixpkgs

You may have a different path, depending on how you added channels etc.. Anyway that's the whole point. The <nixpkgs> stranger that we used in our nix expressions, is referring to a path in the filesystem specified by NIX_PATH.

You can list that directory and realize it's simply a checkout of the nixpkgs repository at a specific commit (hint: .version-suffix).

The NIX_PATH variable is exported by nix.sh, and that's the reason why I always asked you to source nix.sh at the beginning of my posts.

You may wonder: then I can also specify a different nixpkgs path to, e.g., a git checkout of nixpkgs? Yes, you can and I encourage doing that. We'll talk about this in the next pill.

Let's define a path for our repository, then! Let's say all the default.nix, graphviz.nix etc. are under /home/nix/mypkgs:

$ export NIX_PATH=mypkgs=/home/nix/mypkgs:$NIX_PATH
$ nix-instantiate --eval '<mypkgs>'
{ graphviz = <code>; graphvizCore = <code>; hello = <code>; mkDerivation = <code>; }

Yes, nix-build also accepts paths with angular brackets. We first evaluate the whole repository (default.nix) and then peek the graphviz attribute.

The nix-env command is a little different than nix-instantiate and nix-build. Whereas nix-instantiate and nix-build require a starting nix expression, nix-env does not.

You may be crippled by this concept at the beginning, you may think nix-env uses NIX_PATH to find the nixpkgs repository. But that's not it.

The nix-env command uses ~/.nix-defexpr, which is also part of NIX_PATH by default, but that's only a coincidence. If you empty NIX_PATH, nix-env will still be able to find derivations because of ~/.nix-defexpr.

So if you run nix-env -i graphviz inside your repository, it will install the nixpkgs one. Same if you set NIX_PATH to point to your repository.

In order to specify an alternative to ~/.nix-defexpr it's possible to use the [-f] option:

$ nix-env -f '<mypkgs>' -i graphviz
warning: there are multiple derivations named `graphviz'; using the first one
replacing old `graphviz'
installing `graphviz'

Oh why did it say there's another derivation named graphviz? Because both graphviz and graphvizCore attributes in our repository have the name "graphviz" for the derivation:

$ nix-env -f '<mypkgs>' -qaP
graphviz      graphviz
graphvizCore  graphviz
hello         hello

By default nix-env parses all derivations and use the derivation names to interpret the command line. So in this case "graphviz" matched two derivations. Alternatively, like for nix-build, one can use [-A] to specify an attribute name instead of a derivation name:

$ nix-env -f '<mypkgs>' -i -A graphviz
replacing old `graphviz'
installing `graphviz'

This form, other than being more precise, it's also faster because nix-env does not have to parse all the derivations.

For completeness: you must install graphvizCore with [-A,] since without the [-A] switch it's ambiguous.

In summary, it may happen when playing with nix that nix-env peeks a different derivation than nix-build. In such case you probably specified NIX_PATH, but nix-env is instead looking into ~/.nix-defexpr.

Why is nix-env having this different behavior? I don't know specifically by myself either, but the answers could be:

It may also happen to you that you cannot match a derivation name when installing, because of the derivation name vs [-A] switch described above. Maybe nix-env wanted to be more friendly in this case for default user setups.

It may or may not make sense for you, or it's like that for historical reasons, but that's how it works currently, unless somebody comes up with a better idea.

The NIX_PATH variable is the search path used by nix when using the angular brackets syntax. It's possible to refer to "abstract" paths inside nix expressions and define the "concrete" path by means of NIX_PATH, or the usual [-I] flag in nix tools.

We've also explained some of the uncommon nix-env behaviors for newcomers. The nix-env tool does not use NIX_PATH to search for packages, but rather for ~/.nix-defexpr. Beware of that!

In general do not abuse NIX_PATH, when possible use relative paths when writing your own nix expressions. Of course, in the case of <nixpkgs> in our repository, that's a perfectly fine usage of NIX_PATH. Instead, inside our repository itself, refer to expressions with relative paths like ./hello.nix.

...we will finally dive into nixpkgs. Most of the techniques we have developed in this series are already in nixpkgs, like mkDerivation, callPackage, override, etc., but of course better. With time, those base utilities get enhanced by the community with more features in order to handle more and more use cases and in a more general way.

We will not start inspecting packages at the beginning, rather the general structure of nixpkgs.

In our custom repository we created a default.nix which composed the expressions of the various packages.

Also nixpkgs has its own default.nix, which is the one being loaded when referring to <nixpkgs>. It does a simple thing: check whether the nix version is at least 1.7 (at the time of writing this blog post). Then import pkgs/top-level/all-packages.nix. From now on, we will refer to this set of packages as pkgs.

The all-packages.nix is then the file that composes all the packages. Note the pkgs/ subdirectory, while nixos is in the nixos/ subdirectory.

The all-packages.nix is a bit contrived. First of all, it's a function. It accepts a couple of interesting parameters:

The system parameter, as per comment in the expression, it's the system for which the packages will be built. It allows for example to install i686 packages on amd64 machines.

The config parameter is a simple attribute set. Packages can read some of its values and change the behavior of some derivations.

You will find this parameter in many other .nix expressions (e.g. release expressions). The reason is that, given pkgs accepts a system parameter, then whenever you want to import pkgs you also want to pass through the value of system. E.g.:

myrelease.nix:

{ system ? builtins.currentSystem }:

let pkgs = import <nixpkgs> { inherit system; };
...

Why is it useful? With this parameter it's very easy to select a set of packages for a particular system. For example:

nix-build -A psmisc --argstr system i686-linux

This will build the psmisc derivation for i686-linux instead of x86_64-linux. This concept is very similar to multi-arch of Debian.

The setup for cross compiling is also in nixpkgs, however it's a little contrived to talk about it and I don't know much of it either.

I'm sure on the wiki or other manuals you've read about ~/.nixpkgs/config.nix and I'm sure you've wondered whether that's hardcoded in nix. It's not, it's in nixpkgs.

The all-packages.nix expression accepts the config parameter. If it's null, then it reads the NIXPKGS_CONFIG environment variable. If not specified, nixpkgs will peek $HOME/.nixpkgs/config.nix.

After determining config.nix, it will be imported as nix expression, and that will be the value of config (in case it hasn't been passed as parameter to import <nixpkgs>).

The config is available in the resulting repository:

$ nix repl
nix-repl> pkgs = import <nixpkgs> {}
nix-repl> pkgs.config
{ }
nix-repl> pkgs = import <nixpkgs> { config = { foo = "bar"; }; }
nix-repl> pkgs.config
{ foo = "bar"; }

What attributes go in config is a matter of convenience and conventions.

For example, config.allowUnfree is an attribute that forbids building packages that have an unfree license by default. The config.pulseaudio setting tells whether to build packages with pulseaudio support or not where applicable and when the derivation obeys to the setting.

A .nix file contains a nix expression. Thus it can also be a function. I remind you that nix-build expects the expression to return a derivation. Therefore it's natural to return straight a derivation from a .nix file. However, it's also very natural for the .nix file to accept some parameters, in order to tweak the derivation being returned.

In this case, nix does a trick:

For example you can nix-build the .nix file below:

{ pkgs ? import <nixpkgs> {} }:

pkgs.psmisc

Nix is able to call the function because the pkgs parameter has a default value. This allows you to pass a different value for pkgs using the --arg option.

Does it work if you have a function returning a function that returns a derivation? No, Nix only calls the function it encounters once.

We've unleashed the <nixpkgs> repository. It's a function that accepts some parameters, and returns the set of all packages. Due to laziness, only the accessed derivations will be built.

You can use this repository to build your own packages as we've seen in the previous pill when creating our own repository.

Lately I'm a little busy with the NixOS 14.11 release and other stuff, and I'm also looking toward migrating from blogger to a more coder-oriented blogging platform. So sorry for the delayed and shorter pills :)

...we will talk about overriding packages in the nixpkgs repository. What if you want to change some options of a library and let all other packages pick the new library? One possibility is to use, like described above, the config parameter when applicable. The other possibility is to override derivations.

Recall the override design pattern from the nix pill 14. Instead of calling a function with parameters directly, we make the call (function + parameters) overridable.

We put the override function in the returned attribute set of the original function call.

Take for example graphviz. It has an input parameter xorg. If it's null, then graphviz will build without X support.

$ nix repl
nix-repl> :l <nixpkgs>
Added 4360 variables.
nix-repl> :b graphviz.override { xorg = null; }

This will build graphviz without X support, it's as simple as that.

However let's say a package P depends on graphviz, how do we make P depend on the new graphviz without X support?

...you could do something like this:

pkgs = import <nixpkgs> {};
pkgs.graphviz = pkgs.graphviz.override { xorg = null; };
build(pkgs.P)

Given pkgs.P depends on pkgs.graphviz, it's easy to build P with the replaced graphviz. On a pure functional language it's not that easy because you can assign to variables only once.

The fixed point with lazy evaluation is crippling but about necessary in a language like Nix. It lets us achieve something similar to what we'd do imperatively.

Follows the definition of fixed point in nixpkgs:

# Take a function and evaluate it with its own returned value.
fix = f: let result = f result; in result;

It's a function that accepts a function f, calls f result on the result just returned by f result and returns it. In other words it's f(f(f(....

At first sight, it's an infinite loop. With lazy evaluation it isn't, because the call is done only when needed.

nix-repl> fix = f: let result = f result; in result
nix-repl> pkgs = self: { a = 3; b = 4; c = self.a+self.b; }
nix-repl> fix pkgs
{ a = 3; b = 4; c = 7; }

Without the rec keyword, we were able to refer to a and b of the same set.

The trick is that c is not needed to be evaluated in the inner call, thus it doesn't go in an infinite loop.

Won't go further with the explanation here. A good post about fixed point and Nix can be found here.

nix-repl> overrides = { a = 1; b = 2; }
nix-repl> let newpkgs = pkgs (newpkgs // overrides); in newpkgs
{ a = 3; b = 4; c = 3; }
nix-repl> let newpkgs = pkgs (newpkgs // overrides); in newpkgs // overrides
{ a = 1; b = 2; c = 3; }

We've seen how to override attributes in a set such that they get recursively picked by dependant attributes. This approach can be used for derivations too, after all nixpkgs is a giant set of attributes that depend on each other.

To do this, nixpkgs offers config.packageOverrides. So nixpkgs returns a fixed point of the package set, and packageOverrides is used to inject the overrides.

Create a config.nix file like this somewhere:

{
  packageOverrides = pkgs: {
    graphviz = pkgs.graphviz.override { xorg = null; };
  };
}

Now we can build e.g. asciidocFull and it will automatically use the overridden graphviz:

nix-repl> pkgs = import <nixpkgs> { config = import ./config.nix; }
nix-repl> :b pkgs.asciidocFull

Note how we pass the config with packageOverrides when importing nixpkgs. Then pkgs.asciidocFull is a derivation that has graphviz input (pkgs.asciidoc is the lighter version and doesn't use graphviz at all).

Since there's no version of asciidoc with graphviz without X support in the binary cache, Nix will recompile the needed stuff for you.

In the previous pill we already talked about this file. The above config.nix that we just wrote could be the content of ~/.nixpkgs/config.nix.

Instead of passing it explicitly whenever we import nixpkgs, it will be automatically imported by nixpkgs.

We've learned about a new design pattern: using fixed point for overriding packages in a package set.

Whereas in an imperative setting, like with other package managers, a library is installed replacing the old version and applications will use it, in Nix it's not that straight and simple. But it's more precise.

Nix applications will depend on specific versions of libraries, hence the reason why we have to recompile asciidoc to use the new graphviz library.

The newly built asciidoc will depend on the new graphviz, and old asciidoc will keep using the old graphviz undisturbed.

...we will stop diving nixpkgs for a moment and talk about store paths. How does Nix compute the path in the store where to place the result of builds? How to add files to the store for which we have an integrity hash?

Let's start simple. You know nix allows relative paths to be used, such that the file or directory is stored in the nix store, that is ./myfile gets stored into /nix/store/....... We want to understand how is the store path generated for such a file:

$ echo mycontent > myfile

I remind you, the simplest derivation you can write has a name, a builder and the system:

$ nix repl
nix-repl> derivation { system = "x86_64-linux"; builder = ./myfile; name = "foo"; }
«derivation /nix/store/y4h73bmrc9ii5bxg6i7ck6hsf5gqv8ck-foo.drv»

Now inspect the .drv to see where is ./myfile being stored:

$ nix show-derivation /nix/store/y4h73bmrc9ii5bxg6i7ck6hsf5gqv8ck-foo.drv
{
  "/nix/store/y4h73bmrc9ii5bxg6i7ck6hsf5gqv8ck-foo.drv": {
    "outputs": {
      "out": {
        "path": "/nix/store/hs0yi5n5nw6micqhy8l1igkbhqdkzqa1-foo"
      }
    },
    "inputSrcs": [
      "/nix/store/xv2iccirbrvklck36f1g7vldn5v58vck-myfile"
    ],
    "inputDrvs": {},
    "platform": "x86_64-linux",
    "builder": "/nix/store/xv2iccirbrvklck36f1g7vldn5v58vck-myfile",
    "args": [],
    "env": {
      "builder": "/nix/store/xv2iccirbrvklck36f1g7vldn5v58vck-myfile",
      "name": "foo",
      "out": "/nix/store/hs0yi5n5nw6micqhy8l1igkbhqdkzqa1-foo",
      "system": "x86_64-linux"
    }
  }
}

Great, how did nix decide to use xv2iccirbrvklck36f1g7vldn5v58vck ? Keep looking at the nix comments.

Note: doing nix-store --add myfile will store the file in the same store path.

$ nix-hash --type sha256 myfile
2bfef67de873c54551d884fdab3055d84d573e654efa79db3c0d7b98883f9ee3

$ nix-store --dump myfile|sha256sum
2bfef67de873c54551d884fdab3055d84d573e654efa79db3c0d7b98883f9ee3

$ echo -n "source:sha256:2bfef67de873c54551d884fdab3055d84d573e654efa79db3c0d7b98883f9ee3:/nix/store:myfile" > myfile.str

$ nix-hash --type sha256 --truncate --base32 --flat myfile.str
xv2iccirbrvklck36f1g7vldn5v58vck

Output paths are usually generated for derivations. We use the above example because it's simple. Even if we didn't build the derivation, nix knows the out path hs0yi5n5nw6micqhy8l1igkbhqdkzqa1. This is because the out path only depends on inputs.

It's computed in a similar way to source paths, except that the .drv is hashed and the type of derivation is output:out. In case of multiple outputs, we may have different output:<id>.

At the time nix computes the out path, the .drv contains an empty string for each out path. So what we do is getting our .drv and replacing the out path with an empty string:

$ cp -f /nix/store/y4h73bmrc9ii5bxg6i7ck6hsf5gqv8ck-foo.drv myout.drv
$ sed -i 's,/nix/store/hs0yi5n5nw6micqhy8l1igkbhqdkzqa1-foo,,g' myout.drv

The myout.drv is the .drv state in which nix is when computing the out path for our derivation:

$ sha256sum myout.drv
1bdc41b9649a0d59f270a92d69ce6b5af0bc82b46cb9d9441ebc6620665f40b5  myout.drv
$ echo -n "output:out:sha256:1bdc41b9649a0d59f270a92d69ce6b5af0bc82b46cb9d9441ebc6620665f40b5:/nix/store:foo" > myout.str
$ nix-hash --type sha256 --truncate --base32 --flat myout.str
hs0yi5n5nw6micqhy8l1igkbhqdkzqa1

Then nix puts that out path in the .drv, and that's it.

In case the .drv has input derivations, that is it references other .drv, then such .drv paths are replaced by this same algorithm which returns a hash.

In other words, you get a final .drv where every other .drv path is replaced by its hash.

Finally, the other most used kind of path is when we know beforehand an integrity hash of a file. This is usual for tarballs.

A derivation can take three special attributes: outputHashMode, outputHash and outputHashAlgo which are well documented in the nix manual.

The builder must create the out path and make sure its hash is the same as the one declared with outputHash.

Let's say our builder should create a file whose contents is mycontent:

$ echo mycontent > myfile
$ sha256sum myfile
f3f3c4763037e059b4d834eaf68595bbc02ba19f6d2a500dce06d124e2cd99bb  myfile
nix-repl> derivation { name = "bar"; system = "x86_64-linux"; builder = "none"; outputHashMode = "flat"; outputHashAlgo = "sha256"; outputHash = "f3f3c4763037e059b4d834eaf68595bbc02ba19f6d2a500dce06d124e2cd99bb"; }
«derivation /nix/store/ymsf5zcqr9wlkkqdjwhqllgwa97rff5i-bar.drv»

Inspect the .drv and see that it also stored the fact that it's a fixed-output derivation with sha256 algorithm, compared to the previous examples:

$ nix show-derivation /nix/store/ymsf5zcqr9wlkkqdjwhqllgwa97rff5i-bar.drv
{
  "/nix/store/ymsf5zcqr9wlkkqdjwhqllgwa97rff5i-bar.drv": {
    "outputs": {
      "out": {
        "path": "/nix/store/a00d5f71k0vp5a6klkls0mvr1f7sx6ch-bar",
        "hashAlgo": "sha256",
        "hash": "f3f3c4763037e059b4d834eaf68595bbc02ba19f6d2a500dce06d124e2cd99bb"
      }
    },
[...]
}

It doesn't matter which input derivations are being used, the final out path must only depend on the declared hash.

What nix does is to create an intermediate string representation of the fixed-output content:

$ echo -n "fixed:out:sha256:f3f3c4763037e059b4d834eaf68595bbc02ba19f6d2a500dce06d124e2cd99bb:" > mycontent.str
$ sha256sum mycontent.str
423e6fdef56d53251c5939359c375bf21ea07aaa8d89ca5798fb374dbcfd7639  myfile.str

Then proceed as it was a normal derivation output path:

$ echo -n "output:out:sha256:423e6fdef56d53251c5939359c375bf21ea07aaa8d89ca5798fb374dbcfd7639:/nix/store:bar" > myfile.str
$ nix-hash --type sha256 --truncate --base32 --flat myfile.str
a00d5f71k0vp5a6klkls0mvr1f7sx6ch

Hence, the store path only depends on the declared fixed-output hash.

There are other types of store paths, but you get the idea. Nix first hashes the contents, then creates a string description, and the final store path is the hash of this string.

Also we've introduced some fundamentals, in particular the fact that Nix knows beforehand the out path of a derivation since it only depends on the inputs. We've also introduced fixed-output derivations which are especially used by the nixpkgs repository for downloading and verifying source tarballs.

...we will introduce stdenv. In the previous pills we rolled our own mkDerivation convenience function for wrapping the builtin derivation, but the nixpkgs repository also has its own convenience functions for dealing with autotools projects and other build systems.

First of all stdenv is a derivation. And it's a very simple one:

$ nix-build '<nixpkgs>' -A stdenv
/nix/store/k4jklkcag4zq4xkqhkpy156mgfm34ipn-stdenv
$ ls -R result/
result/:
nix-support/  setup

result/nix-support:
propagated-user-env-packages

It has just two files: /setup and /nix-support/propagated-user-env-packages. Don't care about the latter; it's empty, in fact. The important file is /setup.

How can this simple derivation pull in all the toolchain and basic tools needed to compile packages? Let's look at the runtime dependencies:

$ nix-store -q --references result
/nix/store/3a45nb37s0ndljp68228snsqr3qsyp96-bzip2-1.0.6
/nix/store/a457ywa1haa0sgr9g7a1pgldrg3s798d-coreutils-8.24
/nix/store/zmd4jk4db5lgxb8l93mhkvr3x92g2sx2-bash-4.3-p39
/nix/store/47sfpm2qclpqvrzijizimk4md1739b1b-gcc-wrapper-4.9.3
...

How can it be? The package must be referring to those package somehow. In fact, they are hardcoded in the /setup file:

$ head result/setup
export SHELL=/nix/store/zmd4jk4db5lgxb8l93mhkvr3x92g2sx2-bash-4.3-p39/bin/bash
initialPath="/nix/store/a457ywa1haa0sgr9g7a1pgldrg3s798d-coreutils-8.24 ..."
defaultNativeBuildInputs="/nix/store/sgwq15xg00xnm435gjicspm048rqg9y6-patchelf-0.8 ..."

Remember our generic builder.sh in Pill 8? It sets up a basic PATH, unpacks the source and runs the usual autotools commands for us.

The stdenv setup file is exactly that. It sets up several environment variables like PATH and creates some helper bash functions to build a package. I invite you to read it, it's only 860 lines at the time of this writing.

The hardcoded toolchain and utilities are used to initially fill up the environment variables so that it's more pleasant to run common commands, similar to what we did with our builder with baseInputs and buildInputs.

The build with stdenv works in phases. Phases are like unpackPhase, configurePhase, buildPhase, checkPhase, installPhase, fixupPhase. You can see the default list in the genericBuild function.

What genericBuild does is just run these phases. Default phases are just bash functions, you can easily read them.

Every phase has hooks to run commands before and after the phase has been executed. Phases can be overwritten, reordered, whatever, it's just bash code.

How to use this file? Like our old builder. To test it, we enter a fake empty derivation, source the stdenv setup, unpack the hello sources and build it:

$ nix-shell -E 'derivation { name = "fake"; builder = "fake"; system = "x86_64-linux"; }'
nix-shell$ unset PATH
nix-shell$ source /nix/store/k4jklkcag4zq4xkqhkpy156mgfm34ipn-stdenv/setup
nix-shell$ tar -xf hello-2.10.tar.gz
nix-shell$ cd hello-2.10
nix-shell$ configurePhase
...
nix-shell$ buildPhase
...

I unset PATH to further show that the stdenv is enough self-contained to build autotools packages that have no other dependencies.

So we ran the configurePhase function and buildPhase function and they worked. These bash functions should be self-explanatory, you can read the code in the setup file.

Until now we worked with plain bash scripts. What about the Nix side? The nixpkgs repository offers a useful function, like we did with our old builder. It is a wrapper around the raw derivation function which pulls in the stdenv for us, and runs genericBuild. It's stdenv.mkDerivation.

Note how stdenv is a derivation but it's also an attribute set which contains some other attributes, like mkDerivation. Nothing fancy here, just convenience.

Let's write a hello.nix expression using this new discovered stdenv:

with import <nixpkgs> {};
stdenv.mkDerivation {
  name = "hello";
  src = ./hello-2.10.tar.gz;
}

Don't be scared by the with expression. It pulls the nixpkgs repository into scope, so we can directly use stdenv. It looks very similar to the hello expression in Pill 8.

It builds, and runs fine:

$ nix-build hello.nix
...
/nix/store/6y0mzdarm5qxfafvn2zm9nr01d1j0a72-hello
$ result/bin/hello
Hello, world!

Let's take a look at the builder used by mkDerivation. You can read the code here in nixpkgs:

{
  ...
  builder = attrs.realBuilder or shell;
  args = attrs.args or ["-e" (attrs.builder or ./default-builder.sh)];
  stdenv = result;
  ...
}

Also take a look at our old derivation wrapper in previous pills! The builder is bash (that shell variable), the argument to the builder (bash) is default-builder.sh, and then we add the environment variable $stdenv in the derivation which is the stdenv derivation.

You can open default-builder.sh and see what it does:

source $stdenv/setup
genericBuild

It's what we did in Pill 10 to make the derivations nix-shell friendly. When entering the shell, the setup file only sets up the environment without building anything. When doing nix-build, it actually runs the build process.

To get a clear understanding of the environment variables, look at the .drv of the hello derivation:

$ nix show-derivation $(nix-instantiate hello.nix)
warning: you did not specify '--add-root'; the result might be removed by the garbage collector
{
  "/nix/store/abwj50lycl0m515yblnrvwyydlhhqvj2-hello.drv": {
    "outputs": {
      "out": {
        "path": "/nix/store/6y0mzdarm5qxfafvn2zm9nr01d1j0a72-hello"
      }
    },
    "inputSrcs": [
      "/nix/store/9krlzvny65gdc8s7kpb6lkx8cd02c25b-default-builder.sh",
      "/nix/store/svc70mmzrlgq42m9acs0prsmci7ksh6h-hello-2.10.tar.gz"
    ],
    "inputDrvs": {
      "/nix/store/hcgwbx42mcxr7ksnv0i1fg7kw6jvxshb-bash-4.4-p19.drv": [
        "out"
      ],
      "/nix/store/sfxh3ybqh97cgl4s59nrpi78kgcc8f3d-stdenv-linux.drv": [
        "out"
      ]
    },
    "platform": "x86_64-linux",
    "builder": "/nix/store/q1g0rl8zfmz7r371fp5p42p4acmv297d-bash-4.4-p19/bin/bash",
    "args": [
      "-e",
      "/nix/store/9krlzvny65gdc8s7kpb6lkx8cd02c25b-default-builder.sh"
    ],
    "env": {
      "buildInputs": "",
      "builder": "/nix/store/q1g0rl8zfmz7r371fp5p42p4acmv297d-bash-4.4-p19/bin/bash",
      "configureFlags": "",
      "depsBuildBuild": "",
      "depsBuildBuildPropagated": "",
      "depsBuildTarget": "",
      "depsBuildTargetPropagated": "",
      "depsHostBuild": "",
      "depsHostBuildPropagated": "",
      "depsTargetTarget": "",
      "depsTargetTargetPropagated": "",
      "name": "hello",
      "nativeBuildInputs": "",
      "out": "/nix/store/6y0mzdarm5qxfafvn2zm9nr01d1j0a72-hello",
      "propagatedBuildInputs": "",
      "propagatedNativeBuildInputs": "",
      "src": "/nix/store/svc70mmzrlgq42m9acs0prsmci7ksh6h-hello-2.10.tar.gz",
      "stdenv": "/nix/store/6kz2vbh98s2r1pfshidkzhiy2s2qdw0a-stdenv-linux",
      "system": "x86_64-linux"
    }
  }
}

So short I decided to paste it entirely above. The builder is bash, with -e default-builder.sh arguments. Then you can see the src and stdenv environment variables.

Last bit, the unpackPhase in the setup is used to unpack the sources and enter the directory, again like we did in our old builder.

The stdenv is the core of the nixpkgs repository. All packages use the stdenv.mkDerivation wrapper instead of the raw derivation. It does a bunch of operations for us and also sets up a pleasant build environment.

The overall process is simple:

That's it, everything you need to know about the stdenv phases is in the setup file.

Really, take your time to read that file. Don't forget that juicy docs are also available in the nixpkgs manual.

...we will talk about how to add dependencies to our packages with buildInputs and propagatedBuildInputs, and influence downstream builds with setup hooks and env hooks. These concepts are crucial to how nixpkgs packages are composed.

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.

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.

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.

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.

...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!