wroclaw-git-backups/nix
1
0
Fork 0
mirror of https://github.com/NixOS/nix synced 2025-07-10 04:43:53 +02:00
Code Activity

Expand manual on derivation outputs

Browse source 
Note, this includes some text adapted from from Eelco's dissertation

(cherry picked from commit 2aa6e0f084)
This commit is contained in:
John Ericson 2025-02-10 01:08:00 -05:00 • committed by Mergify
parent 7e61e9650c
commit f278a631b0
12 changed files with 508 additions and 174 deletions
Download patch file Download diff file Expand all files Collapse all files

192
doc/manual/source/store/derivation/outputs/content-address.md Normal file
View file

@ -0,0 +1,192 @@
# Content-addressing derivation outputs
The content-addressing of an output only depends on that store object itself, not any other information external (such has how it was made, when it was made, etc.).
As a consequence, a store object will be content-addressed the same way regardless of whether it was manually inserted into the store, outputted by some derivation, or outputted by a some other derivation.
The output spec for a content-addressed output must contains the following field:
- *method*: how the data of the store object is digested into a content address
The possible choices of *method* are described in the [section on content-addressing store objects](@docroot@/store/store-object/content-address.md).
Given the method, the output's name (computed from the derivation name and output spec mapping as described above), and the data of the store object, the output's store path will be computed as described in that section.
## Fixed-output content-addressing {#fixed}
In this case the content-address of the *fixed* in advanced by the derivation itself.
In other words, when the derivation has finished [building](@docroot@/store/building.md), and the provisional output' content-address is computed as part of the process to turn it into a *bona fide* store object, the calculated content address must much that given in the derivation, or the build of that derivation will be deemed a failure.
The output spec for an output with a fixed content addresses additionally contains:
- *hash*, the hash expected from digesting the store object's file system objects.
This hash may be of a freely-chosen hash algorithm (that Nix supports)
> **Design note**
>
> In principle, the output spec could also specify the references the store object should have, since the references and file system objects are equally parts of a content-addressed store object proper that contribute to its content-addressed.
> However, at this time, the references are not not done because all fixed content-addressed outputs are required to have no references (including no self-reference).
>
> Also in principle, rather than specifying the references and file system object data with separate hashes, a single hash that constraints both could be used.
> This could be done with the final store path's digest, or better yet, the hash that will become the store path's digest before it is truncated.
>
> These possible future extensions are included to elucidate the core property of fixed-output content addressing --- that all parts of the output must be cryptographically fixed with one or more hashes --- separate from the particulars of the currently-supported store object content-addressing schemes.
### Design rationale
What is the purpose of fixing an output's content address in advanced?
In abstract terms, the answer is carefully controlled impurity.
Unlike a regular derivation, the [builder] executable of a derivation that produced fixed outputs has access to the network.
The outputs' guaranteed content-addresses are supposed to mitigate the risk of the builder being given these capabilities;
regardless of what the builder does *during* the build, it cannot influence downstream builds in unanticipated ways because all information it passed downstream flows through the outputs whose content-addresses are fixed.
[builder]: @docroot@/store/derivation/index.md#builder
In concrete terms, the purpose of this feature is fetching fixed input data like source code from the network.
For example, consider a family of "fetch URL" derivations.
These derivations download files from given URL.
To ensure that the downloaded file has not been modified, each derivation must also specify a cryptographic hash of the file.
For example,
```jsonc
{
"outputs: {
"out": {
"method": "nar",
"hashAlgo": "sha256",
"hash: "1md7jsfd8pa45z73bz1kszpp01yw6x5ljkjk2hx7wl800any6465",
},
},
"env": {
"url": "http://ftp.gnu.org/pub/gnu/hello/hello-2.1.1.tar.gz"
// ...
},
// ...
}
```
It sometimes happens that the URL of the file changes,
e.g., because servers are reorganised or no longer available.
In these cases, we then must update the call to `fetchurl`, e.g.,
```diff
"env": {
- "url": "http://ftp.gnu.org/pub/gnu/hello/hello-2.1.1.tar.gz"
+ "url": "ftp://ftp.nluug.nl/pub/gnu/hello/hello-2.1.1.tar.gz"
// ...
},
```
If a `fetchurl` derivation's outputs were [input-addressed][input addressing], the output paths of the derivation and of *all derivations depending on it* would change.
For instance, if we were to change the URL of the Glibc source distribution in Nixpkgs (a package on which almost all other packages depend on Linux) massive rebuilds would be needed.
This is unfortunate for a change which we know cannot have a real effect as it propagates upwards through the dependency graph.
For content-addressed outputs (fixed or floating), on the other hand, the outputs' store path only depends on the derivation's name, data, and the `method` of the outputs' specs.
The rest of the derivation is ignored for the purpose of computing the output path.
> **History Note**
>
> Fixed content-addressing is especially important both today and historically as the *only* form of content-addressing that is stabilized.
> This is why the rationale above contrasts it with [input addressing].
## (Floating) Content-Addressing {#floating}
> **Warning**
> This is part of an [experimental feature](@docroot@/development/experimental-features.md).
>
> To use this type of output addressing, you must enable the
> [`ca-derivations`][xp-feature-ca-derivations] experimental feature.
> For example, in [nix.conf](@docroot@/command-ref/conf-file.md) you could add:
>
> ```
> extra-experimental-features = ca-derivations
> ```
With this experimemental feature enabled, derivation outputs can also be content-addressed *without* fixing in the output spec what the outputs' content address must be.
### Purity
Because the derivation output is not fixed (just like with [input addressing]), the [builder] is not given any impure capabilities [^purity].
> **Configuration note**
>
> Strictly speaking, the extent to which sandboxing and deprivilaging is possible varies with the environment Nix is running in.
> Nix's configuration settings indicate what level of sandboxing is required or enabled.
> Builds of derivations will fail if they request an absense of sandboxing which is not allowed.
> Builds of derivations will also fail if the level of sandboxing specified in the configure exceeds what is possible in teh given environment.
>
> (The "environment", in this case, consists of attributes such as the Operating System Nix runs atop, along with the operating-system-specific privilages that Nix has been granted.
> Because of how conventional operating systems like macos, Linux, etc. work, granting builders *fewer* privilages may ironically require that Nix be run with *more* privilages.)
That said, derivations producing floating content-addressed outputs may declare their builders as impure (like the builders of derivations producing producing fixed outputs).
This is provisionally supported as part of the [`impure-derivations`][xp-feature-impure-derivations] experimental feature.
### Compatibility negotiation
Any derivation producing a floating content-addresssed output implicitly requires the `ca-derivations` [system feature](@docroot@/command-ref/conf-file.md#conf-system-features).
This prevents scheduling the building of the derivation on a machine without the experimental feature enabled.
Even once the experimental feature is stabilized, this is still useful in order to be allow using remote builder running odler versions of Nix, or alternative implementations that do not support floating content addressing.
### Determinism
In the earlier [discussion of how self-references are handled when content-addressing store objects](@docroot@/store/store-object/content-address.html#self-references), it was pointed out that methods of producing store objects ought to be deterministic regardless of the choice of provisional store path.
For store objects produced by manually inserting into the store to create a store object, the "method of production" is an informally concept --- formally, Nix has no idea where the store object came from, and content-addressing is crucial in order to ensure that the derivation is *intrinsically* tamper-proof.
But for store objects produced by derivation, the "method is quite formal" --- the whole point of derivations is to be a formal notion of building, after all.
In this case, we can elevate this informal property to a formal one.
A *determinstic* content-addressing derivation should produce outputs with the same content addresses:
1. Every time the builder is run
This is because either the builder is completely sandboxed, or because all any remaining impurities that leak inside the build sandbox are ignored by the builder and do not influence its behavior.
2. Regardless of the choice of any provisional outputs paths
Provisional store paths must be chosen for any output that has a self-reference.
The choice of provisional store path can be thought of as an impurity, since it is an arbitrary choice.
If provisional outputs paths are deterministically chosen, we are in the first branch of part (1).
The builder the data it produces based on it in arbitrary ways, but this gets us closer to to [input addressing].
Deterministically choosing the provisional path may be considered "complete sandboxing" by removing an impurity, but this is unsatisfactory
<!--
TODO
(Both these points will be expanded-upon below.)
-->
If provisional outputs paths are randomly chosen, we are in the second branch of part (1).
The builder *must* not let the random input affect the final outputs it produces, and multiple builds may be performed and the compared in order to ensure that this is in fact the case.
### Floating versus Fixed
While the destinction between content- and input-addressing is one of *mechanism*, the distinction between fixed and floating content addression is more one of *policy*.
A fixed output that passes its content address check is just like a floating output.
It is only in the potential for that check to fail that they are different.
> **Design Note**
>
> In a future world where floating content-addressing is also stable, we in principle no longer need separate [fixed](#fixed) content-addressing.
> Instead, we could always use floating content-addressing, and separately assert the precise value content address of a given store object to be used as an input (of another derivation).
> A stand-alone assertion object of this sort is not yet implemented, but its possible creation is tracked in [Issue #11955](https://github.com/NixOS/nix/issues/11955).
>
> In the current version of Nix, fixed outputs which fail their hash check are still registered as valid store objects, just not registered as outputs of the derivation which produced them.
> This is an optimization that means if the wrong output hash is specified in a derivation, and then the derivation is recreated with the right output hash, derivation does not need to be rebuilt --- avoiding downloading potentially large amounts of data twice.
> This optimisation prefigures the design above:
> If the output hash assertion was removed outside the derivation itself, Nix could additionally not only register that outputted store object like today, but could also make note that derivation did in fact successfully download some data.
For example, for the "fetch URL" example above, making such a note is tantamount to recording what data is available at the time of download at the given URL.
> It would only be when Nix subsequently tries to build something with that (refining our example) downloaded source code that Nix would be forced to check the output hash assertion, preventing it from e.g. building compromised malware.
>
> Recapping, Nix would
>
> 1. successfully download data
> 2. insert that data into the store
> 3. associate (presumably with some sort of expiration policy) the downloaded data with the derivation that downloaded it
>
> But only use the downloaded store object in subsequent derivations that depended upon the assertion if the assertion passed.
>
> This possible future extension is included to illustrate this distinction:
[input addressing]: ./input-address.md
[xp-feature-ca-derivations]: @docroot@/development/experimental-features.md#xp-feature-ca-derivations
[xp-feature-git-hashing]: @docroot@/development/experimental-features.md#xp-feature-git-hashing
[xp-feature-impure-derivations]: @docroot@/development/experimental-features.md#xp-feature-impure-derivations

97
doc/manual/source/store/derivation/outputs/index.md Normal file
View file

@ -0,0 +1,97 @@
# Derivation Outputs and Types of Derivations
As stated on the [main pages on derivations](../index.md#store-derivation),
a derivation produces [store objects], which are known as the *outputs* of the derivation.
Indeed, the entire point of derivations is to produce these outputs, and to reliably and reproducably produce these derivations each time the derivation is run.
One of the parts of a derivation is its *outputs specification*, which specifies certain information about the outputs the derivation produces when run.
The outputs specification is a map, from names to specifications for individual outputs.
## Output Names {#outputs}
Output names can be any string which is also a valid [store path] name.
The name mapped to each output specification is not actually the name of the output.
In the general case, the output store object has name `derivationName + "-" + outputSpecName`, not any other metadata about it.
However, an output spec named "out" describes and output store object whose name is just the derivation name.
> **Example**
>
> A derivation is named `hello`, and has two outputs, `out`, and `dev`
>
> - The derivation's path will be: `/nix/store/<hash>-hello.drv`.
>
> - The store path of `out` will be: `/nix/store/<hash>-hello`.
>
> - The store path of `dev` will be: `/nix/store/<hash>-hello-dev`.
The outputs are the derivations are the [store objects][store object] it is obligated to produce.
> **Note**
>
> The formal terminology here is somewhat at adds with everyday communication in the Nix community today.
> "output" in casual usage tends to refer to either to the actual output store object, or the notional output spec, depending on context.
>
> For example "hello's `dev` output" means the store object referred to by the store path `/nix/store/<hash>-hello-dev`.
> It is unusual to call this the "`hello-dev` output", even though `hello-dev` is the actual name of that store object.
## Types of output addressing
The main information contained in an output specification is how the derivation output is addressed.
In particular, the specification decides:
- whether the output is [content-addressed](./content-address.md) or [input-addressed](./input-address.md)
- if the content is content-addressed, how is it content addressed
- if the content is content-addressed, [what is its content address](./content-address.md#fixed-content-addressing) (and thus what is its [store path])
## Types of derivations
The sections on each type of derivation output addressing ended up discussing other attributes of the derivation besides its outputs, such as purity, scheduling, determinism, etc.
This is no concidence; for the type of a derivation is in fact one-for-one with the type of its outputs:
- A derivation that produces *xyz-addressed* outputs is an *xyz-addressing* derivations.
The rules for this are fairly concise:
- All the outputs must be of the same type / use the same addressing
- The derivation must have at least one output
- Additionally, if the outputs are fixed content-addressed, there must be exactly one output, whose specification is mapped from the name `out`.
(The name `out` is special, according to the rules described above.
Having only one output and calling its specification `out` means the single output is effectively anonymous; the store path just has the derivation name.)
(This is an arbitrary restriction that could be lifted.)
- The output is either *fixed* or *floating*, indicating whether the its store path is known prior to building it.
- With fixed content-addressing it is fixed.
> A *fixed content-addressing* derivation is also called a *fixed-output derivation*, since that is the only currently-implemented form of fixed-output addressing
- With floating content-addressing or input-addressing it is floating.
> Thus, historically with Nix, with no experimental features enabled, *all* outputs are fixed.
- The derivation may be *pure* or *impure*, indicating what read access to the outside world the [builder](../index.md#builder) has.
- An input-addressing derivation *must* be pure.
> If it is impure, we would have a large problem, because an input-addressed derivation always produces outputs with the same paths.
- A content-addressing derivation may be pure or impure
- If it is impure, it may be be fixed (typical), or it may be floating if the additional [`impure-derivations`][xp-feature-impure-derivations] experimental feature is enabled.
- If it is pure, it must be floating.
- Pure, fixed content-addressing derivations are not suppported
> There is no use for this forth combination.
> The sole purpose of an output's store path being fixed is to support the derivation being impure.
[xp-feature-ca-derivations]: @docroot@/development/experimental-features.md#xp-feature-ca-derivations
[xp-feature-git-hashing]: @docroot@/development/experimental-features.md#xp-feature-git-hashing
[xp-feature-impure-derivations]: @docroot@/development/experimental-features.md#xp-feature-impure-derivations

31
doc/manual/source/store/derivation/outputs/input-address.md Normal file
View file

@ -0,0 +1,31 @@
# Input-addressing derivation outputs
[input addressing]: #input-addressing
"Input addressing" means the address the store object by the *way it was made* rather than *what it is*.
That is to say, an input-addressed output's store path is a function not of the output itself, but the derivation that produced it.
Even if two store paths have the same contents, if they are produced in different ways, and one is input-addressed, then they will have different store paths, and thus guaranteed to not be the same store object.
<!---
### Modulo fixed-output derivations
**TODO hash derivation modulo.**
So how do we compute the hash part of the output path of a derivation?
This is done by the function `hashDrv`, shown in Figure 5.10.
It distinguishes between two cases.
If the derivation is a fixed-output derivation, then it computes a hash over just the `outputHash` attributes.
If the derivation is not a fixed-output derivation, we replace each element in the derivation’s inputDrvs with the result of a call to `hashDrv` for that element.
(The derivation at each store path in `inputDrvs` is converted from its on-disk ATerm representation back to a `StoreDrv` by the function `parseDrv`.) In essence, `hashDrv` partitions store derivations into equivalence classes, and for hashing purpose it replaces each store path in a derivation graph with its equivalence class.
The recursion in Figure 5.10 is inefficient:
it will call itself once for each path by which a subderivation can be reached, i.e., `O(V k)` times for a derivation graph with `V` derivations and with out-degree of at most `k`.
In the actual implementation, memoisation is used to reduce this to `O(V + E)` complexity for a graph with E edges.
-->
[xp-feature-ca-derivations]: @docroot@/development/experimental-features.md#xp-feature-ca-derivations
[xp-feature-git-hashing]: @docroot@/development/experimental-features.md#xp-feature-git-hashing
[xp-feature-impure-derivations]: @docroot@/development/experimental-features.md#xp-feature-impure-derivations