In Autoconf, some old NetBSD targets like "i686-unknown-netbsd" are
interpreted as a.out, not elf, and virtually nothing supports it. We
need to specify e.g. "i686-unknown-netbsdelf" to get the right
behaviour.
Newer bootloaders for RISC-V (i.e., OpenSBI + U-Boot) support
flat and compressed kernel images but not vmlinux. Therefore,
let's build "Image" like what we do with aarch64.
Also copy DTBs while we are at it.
This will begin the process of breaking up the `useLLVM` monolith. That
is good in general, but I hope will be good for NetBSD and Darwin in
particular.
Co-authored-by: sterni <sternenseemann@systemli.org>
Stating that CPUs and the isCompatible relation forms a category (or
preorder) is correct but overtly technical. We can state it more
clearly for readers unfamiliar with mathematics while retaining some
keywords to be useful to technical readers.
This PR adds a new aarch64 android toolchain, which leverages the
existing crossSystem infrastructure and LLVM builders to generate a
working toolchain with minimal prebuilt components.
The only thing that is prebuilt is the bionic libc. This is because it
is practically impossible to compile bionic outside of an AOSP tree. I
tried and failed, braver souls may prevail. For now I just grab the
relevant binaries from https://android.googlesource.com/.
I also grab the msm kernel sources from there to generate headers. I've
included a minor patch to the existing kernel-headers derivation in
order to expose an internal function.
Everything else, from binutils up, is using stock code. Many thanks to
@Ericson2314 for his help on this, and for building such a powerful
system in the first place!
One motivation for this is to be able to build a toolchain which will
work on an aarch64 linux machine. To my knowledge, there is no existing
toolchain for an aarch64-linux builder and an aarch64-android target.
When a list is passed to isStorePath this is most likely a mistake and
it is therefore better to just return false. There is one case where
this theoretically makes sense (if a list contains a single element for
which isStorePath elem), but since that case is also probably seldomly
intentional, it may save someone from debbuging unclear evaluation
errors.
I am taking the non-invasive parts of #110914 to hopefully help out with #111988.
In particular:
- Use `lib.makeScopeWithSplicing` to make the `darwin` package set have
a proper `callPackage`.
- Adjust Darwin `stdenv`'s overlays keeping things from the previous
stage to not stick around too much.
- Expose `binutilsNoLibc` / `darwin.binutilsNoLibc` to hopefully get us
closer to a unified LLVM and GCC bootstrap.
Previously, an option of type
attrsOf string
wouldn't throw a deprecation warning, even though the string type is
deprecated. This was because the deprecation warning trigger only looked
at the type of the option itself, not any of its subtypes.
This commit fixes this, causing each of the types deprecationMessages to
trigger for the option. This relies on the subtypes mkOptionType
attribute introduced in 26607a5a2e06653fec453c83d063cdfc4b59185f
In 2d45a62899, the submodule type
description was amended with the freeformType description. This causes
all the modules passed to the submodule to be evaluated once on their
own, without any extra definitions from the config section. This means
that the specified modules need to be valid on their own, without any
undeclared options.
This commit adds a test that evaluates a submodules option description,
which would trigger the above problem for one of the tests, if it were
not fixed by this commit as well.
This is done because the next commit makes option evaluation a bit more
strict, which would also trigger this test failure, even though it's not
related to the change at all.
This will be used to issue deprecation warnings recursively in the next
commit
In addition, this allows easily getting nested types of other options, which
is useful when you want to create an option that aliases a part of
another one.
It's a common pattern in Nixpkgs to want to emit a warning in certain
cases, but not actually change behaviours.
This is often expressed as either
if cond then lib.warn "Don't do that thing" x else x
Or
(if cond then lib.warn "Don't do that thing" else lib.id) x
Neither of which really expresses the intent here, because it looks
like 'x' is being chosen conditionally.
To make this clearer, I introduce a "warnIf" function, which makes it
clear that the only thing being affected by the condition is whether
the warning is generated, not the value being returned.
A subflake that can be indidividually accessed without also providing
an interface to the whole of nixpkgs.
Usage:
inputs.nixpkgs-lib.url = "github:NixOS/nixpkgs?dir=lib"
Since shellPackage actually requires the value to be an attribute set
(i. e. an derivation in this case), we cannot re-use the package.check
type checker since it also allows strings or things that are coercible
to strings as long as they look like store paths.
These are all the architectures supported by Nixpkgs on other
platforms, that are also supported by NetBSD. (So I haven't added
any architectures that are new to Nixpkgs here, even though NetBSD
supports some that we don't have.)
The previous mess was partially grouped by OS, and partially grouped
by architecture, which made it very difficult to know where to add new
entries.
I've chosen to group by OS entirely, because OSes are likely to
maintain exhaustive lists of supported architectures, but it's far
less likely we'd be able to find exhaustive lists of supported OSes
for every architecture.
When in the presence of worktrees, it happens that /commondir has a
trailing slash.
In these circumstances, it can lead to `lib.pathType` being passed paths
like `/foo/bar/.git/`, which in turn lead to
`error: attribute '.git' missing`.
With this change, we now make sure send properly-formatted paths to all
other functions.
This, in particular, fixes running NixOS tests on worktrees created by
libgit2 on my machine. (Worktrees created by git itself appear to not
hit the issue.)
I recently wrote some Nix code where I wrongly set a value to an option
which wasn't an actual option, but an attr-set of options. The mistake I
made can be demonstrated with an expression like this:
{
foo = { lib, pkgs, config, ... }: with lib; {
options.foo.bar.baz = mkOption {
type = types.str;
};
config.foo.bar = 23;
};
}
While it wasn't too hard to find the cause of the mistake for me, it was
necessary to have some practice in reading stack traces from the module
system since the eval-error I got was not very helpful:
error: --- TypeError --------------------------------------------------------- nix-build
at: (323:25) in file: /nix/store/3nm31brdz95pj8gch5gms6xwqh0xx55c-source/lib/modules.nix
322| foldl' (acc: module:
323| acc // (mapAttrs (n: v:
| ^
324| (acc.${n} or []) ++ f module v
value is an integer while a set was expected
(use '--show-trace' to show detailed location information)
I figured that such an error can be fairly confusing for someone who's
new to NixOS, so I decided to catch this case in th `byName` function in
`lib/modules.nix` by checking if the value to map through is an actual
attr-set. If not, a different error will be thrown.
Usually we ensure using the mapAttrs call wrapping the license set that
every license has an associated shortName. A change related to legacy
aliases most likely introduced the removal of the shortName attribute
for all the legacy license names by splitting the set into two sets
connected by a record update operator -- leading to mapAttrs only
affecting the first set.
Since it used to be a valid assumption to have that every license had a
shortName attribute, we reintroduce this attribute for the legacy
aliases as well.
Forcing the module to be builtin breaks 5.10, which wants to compile it as a
module (probably due to dependencies). There doesn't seem to be a need to have
it builtin anymore, so we can just remove the override.
In 9c213398b3 kernelPreferBuiltin was
moved/renamed to linux-kernel.preferBuiltin. However, for
armv7l-hf-multiplatform the new option was written with an uppercase P,
which made the kernel build process ignore it.
This reverts commit d9a7d03da8.
Reason for this is that it actually doesn't migitate the issue on nix
stable for another reason: builtins.tryEval doesn't prevent the error
generated by builtins.functionArgs from halting evaluation:
> builtins.tryEval (builtins.functionArgs builtins.functionArgs)
error: 'functionArgs' requires a function, at (string):1:19
Thus it seems that there is no workaround to make
lib.generators.toPretty work with nix stable and primops since there is
no way to distinguish between primops and lambdas in nix.
An high level example case of this problem occuring can be found below:
nix-repl> lib.generators.toPretty {} (lib.concatStringsSep "\n")
error: 'functionArgs' requires a function, at /home/lukas/src/nix/nixpkgs/lib/trivial.nix:334:42
However this does not happen on other partially applied functions:
nix-repl> lib.generators.toPretty {} (lib.concatMapStringsSep "\n")
"<function>"
The issue, as it turns out is that while builtins are functions,
builtins.functionArgs throws if is passed a builtin or a partially
applied builtin:
nix-repl> lib.generators.toPretty {} builtins.toString
error: 'functionArgs' requires a function, at /home/lukas/src/nix/nixpkgs/lib/trivial.nix:334:42
nix-repl> lib.generators.toPretty {} (builtins.foldl' (a: b: a + b))
error: 'functionArgs' requires a function, at /home/lukas/src/nix/nixpkgs/lib/trivial.nix:334:42
I'm pretty sure this qualifies as a nix bug and should be filed
accordingly, but we can work around it in lib.generators.toPretty by
using tryEval and falling back to {} which functionArgs _should_ return
for builtins.
The nix behavior is inconsistent to say the least:
nix-repl> builtins.functionArgs builtins.functionArgs
error: 'functionArgs' requires a function, at (string):1:1
nix-repl> builtins.typeOf builtins.functionArgs
"lambda"
builtins.functionArgs (a: 1 + a)
{ }
nix-repl> builtins.typeOf (a: 1 + a)
"lambda"
PPC64 supports two ABIs: ELF v1 and v2.
ELFv1 is historically what GCC and most packages expect, but this is
changing because musl outright does not work with ELFv1. So any distro
which uses musl must use ELFv2. Many other platforms are moving to ELFv2
too, such as FreeBSD (as of v13) and Gentoo (as of late 2020).
Since we use musl extensively, let's default to ELFv2.
Nix gives us the power to specify this declaratively for the entire
system, so ELFv1 is not dropped entirely. It can be specified explicitly
in the target config, e.g. "powerpc64-unknown-linux-elfv1". Otherwise the
default is "powerpc64-unknown-linux-elfv2". For musl,
"powerpc64-unknown-linux-musl" must use elfv2 internally to function.
Now type checks the resulting function values and allows mkMerge and co.
Also indicates that the type check is done in the function body
Co-Authored-By: Robert Hensing <robert@roberthensing.nl>
Looks like these got left behind in the
kernelArch -> linuxArch migration.
Fixes:
* pkgsCross.powernv.linuxHeaders
* pkgsCross.riscv64.linuxHeaders
* pkgsCross.riscv32.linuxHeaders
and dependees
Immensely helpful when you want to see the changes a function makes to
its value as it passes through.
Example:
```
$ nix-instantiate --strict --eval -E '(with import ./lib; traceFnSeqN 2 "id" (x: x) { a.b.c = 3; })'
trace: {
fn = "id";
from = {
a = {
b = {…};
};
};
to = {
a = {
b = {…};
};
};
}
{ a = { b = { c = 3; }; }; }
```
This reverts commit 4ff1ab5a56.
We need this to type options like:
services.xserver.windowManager.xmonad.extraPackages that specify functions that
take an attribute set containing packages / plugins and return a list containing
a selection of the values in this set.
The reason we need a dedicated type for this is to have the correct merge
behaviour. Without the functionTo type merging multiple function option
definitions results in an evaluation error. The functionTo type merges
definitions by returning a new function that applies the functions of all the
definitions to the given input and merges the result.
(cherry picked from commit 7ed41ff5e7)
The `platform` field is pointless nesting: it's just stuff that happens
to be defined together, and that should be an implementation detail.
This instead makes `linux-kernel` and `gcc` top level fields in platform
configs. They join `rustc` there [all are optional], which was put there
and not in `platform` in anticipation of a change like this.
`linux-kernel.arch` in particular also becomes `linuxArch`, to match the
other `*Arch`es.
The next step after is this to combine the *specific* machines from
`lib.systems.platforms` with `lib.systems.examples`, keeping just the
"multiplatform" ones for defaulting.
For renames like
mkAliasOptionModule [ "services" "compton" ] [ "services" "picom" ]
where the target is an option set (like services.picom) instead of a single
option (like services.picom.enable), previously the renamed option type
was unset, leading to it being `types.unspecified`.
This changes it to be `types.submodule {}` instead, which makes more
sense.
Since 40e7be1 all ARM platforms that didn't have a parsed cpu version
(e.g. arm-none-eabi) would be handled as armv7l-hf-multiplatform which
did break building arm-trusted-platform packages for some targets (e.g.
rk3399).
Using pcBase as fallback, instead of armv7l-hf-multiplatform,
corresponds with the behaviour we had before 40e7be1.
We recently switched to more explicit GPL license names in line
with the SPDX change and GNU Foundation recommendations:
https://www.gnu.org/licenses/identify-licenses-clearly.html
This followed up older change to use the recommended SPDX ID
18a5e8c36b
but using the `-only` variant for these deprecated licenses too
makes it harder to check for them automatically.
Let’s switch to the appropriate SPDX ID again.
Previously the .enable option was used to encode the condition as well,
which lead to some oddness:
- In order to encode an assertion, one had to invert it
- To disable a check, one had to mkForce it
By introducing a separate .check option this is solved because:
- It can be used to encode assertions
- Disabling is done separately with .enable option, whose default can be
overridden without a mkForce
Previously this option was thought to be necessary to avoid infinite
recursion, but it actually isn't, since the check evaluation isn't fed
back into the module fixed-point.
This implements assertions/warnings supported by the module system directly,
instead of just being a NixOS option (see
nixos/modules/misc/assertions.nix).
This has the following benefits:
- It allows cleanly redoing the user interface. The new
implementation specifically allows disabling assertions or
converting them to warnings instead.
- Assertions/warnings can now be thrown easily from within
submodules, which previously wasn't possible and needed workarounds.
Initially https://github.com/NixOS/nixpkgs/pull/82897 prevented
non-visible options from being rendered in the manual, but
visible-but-internal options were still being recursed into. This fixes
this, aligning the recurse condition here with the one in
make-options-doc/default.nix
The last use of `kernelMajor` in Nixpkgs was removed in 2018.
Even then, I'm not positive it was actually in an exercised code path.
AFAIUI this is now totally redundant and useless as it really was meant
for the 2.4 -> 2.6 transition.
split comes from builtins, not lib.
error: attribute 'split' missing, at /nix/path/nixpkgs/lib/sources.nix:4:4
(use '--show-trace' to show detailed location information)
I think there was a silent (i.e. semantic) merge conflict between PR #101139 and
PR #100456. This commit should fix the error, which manifests as follows:
error: undefined variable 'boolToString' at /home/kkini/src/nixpkgs/lib/types.nix:552:42
Nix can perform static scope checking, but whenever code is inside
a `with` expression, the analysis breaks down, because it can't
know statically what's in the attribute set whose attributes were
brought into scope. In those cases, Nix has to assume that
everything works out.
Except it doesnt. Removing `with` from lib/ revealed an undefined
variable in an error message.
If that doesn't convince you that we're better off without `with`,
I can tell you that this PR results in a 3% evaluation performance
improvement because Nix can look up local variables by index.
This adds up with applications like the module system.
Furthermore, removing `with` makes the binding site of each
variable obvious, which helps with comprehension.
Add a friendly function to easily return a flattened list of files
within a directory.
This is useful if you want to easily iterate or concatSep the list of
files all found within a directory.
(i.e. when constructing Java's CLASSPATH)
Style improvements
Co-authored-by: Silvan Mosberger <github@infinisil.com>
Previously if `_file` was specified by a module:
trace: warning: The type `types.string' of option `foo' defined in `/nix/store/yxhm2il5yrb92fldgriw0wyqh2kk9qyc-bug.nix' is deprecated. See https://github.com/NixOS/nixpkgs/pull/66346 for better alternative types.
With this change:
trace: warning: The type `types.string' of option `foo' defined in `/home/infinisil/src/nixpkgs/bug.nix' is deprecated. See https://github.com/NixOS/nixpkgs/pull/66346 for better alternative types.
If multiple definitions are passed, this evaluates them all as if they
were the only one, for a better error message. In particular this won't
show module-internal properties like `_type = "override"` and co.
- These symbols can be confusing for those not familiar with them
- There's no harm in making these more obvious
- Terminals may not print them correctly either
Also changes the function argument printing slightly to be more obvious
This new type has unsurprising merge behavior: Only attribute sets are
merged together (recursively), and only if they don't conflict.
This is in contrast to the existing types:
- types.attrs is problematic because later definitions completely
override attributes of earlier definitions, and it doesn't support
mkIf and co.
- types.unspecified is very similar to types.attrs, but it has smart
merging behavior that often doesn't make sense, and it doesn't support
all types
The vision here is that configuration tools can generate .json or .toml
files, which can be plugged into an existing configuration.
Eg:
{ lib, ... }:
{
imports = [
(lib.modules.importJSON ./hardware-configuration.json)
];
}
Jasper has been marked insecure for a while, and upstream has not
been responsive to CVEs for over a year.
Fixes#55388.
Signed-off-by: David Anderson <dave@natulte.net>
Previously the only way to deprecate a type was using
theType = lib.warn "deprecated" (mkOptionType ...)
This caused the warning to be emitted when the type was evaluated, but
the error didn't include which option actually used that type.
With this commit, types can specify a deprecationMessage, which when
non-null, is printed along with the option that uses the type
> NOTE: This function is not performant and should be avoided.
It's not used at all in-tree now, so we can remove it completely after
any remaining users are given notice.
An easy-to-make mistake when declaring e.g. a submodule is the accidental
confusion of `options` and `config`:
types.submodule {
config = {
foo = mkOption { /* ... */ };
};
}
However the error-message
The option `[definition 1-entry 1].foo' defined in `<expr.nix>' does not exist.
is fairly unhelpful because it seems as the options are declared at the
first sight. In fact, it took a colleague and me a while to track down such
a mistake a few days ago and we both agreed that this should be somehow caught
to save the time we spent debugging the module in question.
At first I decided to catch this error in the `submodules`-type directly
by checking whether `options` is undeclared, however this becomes fairly
complicated as soon as a submodule-declaration e.g. depends on existing
`config`-values which would've lead to some ugly `builtins.tryExec`-heuristic.
This patch now simply checks if the option's prefix has any options
defined if a point in evaluation is reached where it's clear that the
option in question doesn't exist. This means that this patch doesn't
change the logic of the module system, it only provides a more detailed
error in certain cases:
The option `[definition 1-entry 1].foo' defined in `<expr.nix>' does not exist.
However it seems as there are no options defined in [definition 1-entry 1]. Are you sure you've
declared your options properly? This happens if you e.g. declared your options in `types.submodule'
under `config' rather than `options'.
The refactoring in fd75dc8765
introduced a mistake in the error message that doesn't show the full
context anymore. E.g. with this module:
options.foo.bar = lib.mkOption {
type = lib.types.submodule {
baz = 10;
};
default = {};
};
You'd get the error
The option `baz' defined in `/home/infinisil/src/nixpkgs/config.nix' does not exist.
instead of the previous
The option `foo.bar.baz' defined in `/home/infinisil/src/nixpkgs/config.nix' does not exist.
This commit undoes this regression
Submodules that have a freeform type set behave as if that was the type
of the option itself (for values that don't have an option). Since the
submodules options are shown as separate entries in the manual, it makes
sense to show the freeform type as the submodule type.
For programs that have a lot of (Nix-representable) configuration options,
a simple way to represent this in a NixOS module is to declare an
option of a type like `attrsOf str`, representing a key-value mapping
which then gets generated into a config file. However with such a type,
there's no way to add type checking for only some key values.
On the other end of the spectrum, one can declare a single separate
option for every key value that the program supports, ending up with a module
with potentially 100s of options. This has the benefit that every value
gets type checked, catching mistakes at evaluation time already. However
the disadvantage is that the module becomes big, becomes coupled to the
program version and takes a lot of effort to write and maintain.
Previously there was a middle ground between these two
extremes: Declare an option of a type like `attrsOf str`, but declare
additional separate options for the values you wish to have type
checked, and assign their values to the `attrsOf str` option. While this
works decently, it has the problem of duplicated options, since now both
the additional options and the `attrsOf str` option can be used to set a
key value. This leads to confusion about what should happen if both are
set, which defaults should apply, and more.
Now with this change, a middle ground becomes available that solves above
problems: The module system now supports setting a freeform type, which
gets used for all definitions that don't have an associated option. This
means that you can now declare all options you wish to have type
checked, while for the rest a freeform type like `attrsOf str` can be
used.
This fundamentally changes how the module evaluation internally
handles definitions without an associated option.
Previously the values of these definitions were discarded and only
the names were propagated. This was fine because that's all that's
needed for optionally checking whether all definitions have an
associated option with _module.check.
However with the upcoming change of supporting freeform modules,
we *do* need the values of these.
With this change, the module evaluation cleanly separates definitions
that match an option, and ones that do not.
`toHex` converts the given positive integer to a string of the hexadecimal
representation of that integer. For example:
```
toHex 0 => "0"
toHex 16 => "10"
toHex 250 => "FA"
```
`toBase base i` converts the positive integer `i` to a list of it
digits in the given `base`. For example:
```
toBase 10 123 => [ 1 2 3 ]
toBase 2 6 => [ 1 1 0 ]
toBase 16 250 => [ 15 10 ]
```
The previous hash was too short and caused evaluation-time errors like:
invalid SRI hash 'sha256-AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA='
Additionally, since the fact that this is broken implies that nobody
could have been using it, "SRI" is a bit of a vague and obscure term,
`fakeSriHash` would be somewhat of a mouthful, and the relevant fetcher
parameters are just called `hash`, rename it to `fakeHash`.
This is used in in the manual generation for option identifiers that can
be linked. This, unike what the example describes, doesn't preserve
quotes which is needed for these identifiers to be valid.
This reverts commit 124cccbe3b.
124cccbe3b
broke the build of NixOS manual.
It does not make sense to be as strict as with attributes since we
are not limited by the CLI's inability to handle numbers.
Placeholders should not be quoted either as they are not part of Nix
syntax but a meta-level construct.
fix: Adding libtool to allow darwin compiles
Libtool seems to be required for mongodb to compile on darwin.
fix: Marking MongoDB as broken on aarch64
fix: Adding libtools to the pkg imports
Update mongodb to 4.0.4
Currently, not providing `name` to `cleanSourceWith` will use the name
of the imported directory. However, a common case is for this to be the
top level of some repository. In that case, the name will be the name of
the checkout on the current machine, which is not necessarily
reproducible across different settings, and can lead to e.g. cache
misses in CI.
This is documented in the comment on `cleanSourceWith`, but this does
not stop it being a subtle trap for users.
There are different tradeoffs in each case:
1. If `cleanSourceWith` defaults to `"source"`, then we may end up with a
user not knowing what directory a source store path corresponds to.
However, it being called "unnamed" may give them a clue that there is a
way for them to name it, and lead them to the definition of the
function, which has a clear `name` parameter.
2. If `cleanSoureWith` defaults to the directory name, then a user may face
occasional loss of caching, which is hard to notice, and hard to track
down. Tracking it down likely requires use of more advanced tools like
`nix-diff`, and reading the source of a lot of nix code.
I think the downside of the status quo is worse.
This is really another iteration of
https://github.com/NixOS/nix/issues/1305: that led to adding the `name`
argument in the first place, this just makes us use a better default
`name`.
The _module option is added as an internal option set, and it messes up
the results of module evaluations, requiring people to manually filter
_modules out.
If people depend on this, they can still use config._module from inside
the modules, exposing _module as an explicitly declared user option. Or
alternatively with the _module attribute now returned by evalModules.
newlib is the default for most tools when no kernel is provided. Other
exist, but this seems like a safe default.
(cherry picked from commit 8009c20711)
This helps with troubleshooting exceptions in config values, which were hard
to track down for options with many definitions.
The trace will look like:
error: while evaluating the attribute 'config.foo' at undefined position:
[...]
while evaluating the option `foo':
[...]
while evaluating definitions from `/home/user/mymod.nix':
while evaluating 'dischargeProperties' at /home/user/nixpkgs/lib/modules.nix:464:25, called from /home/user/nixpkgs/lib/modules.nix:392:137:
while evaluating the attribute 'value' at /home/user/nixpkgs/lib/modules.nix:277:44:
Value error!
where the `/home/user/mymod.nix` module is
{ lib, ... }: {
options.foo = lib.mkOption {
type = lib.types.lines;
};
config.foo = builtins.throw "Value error!";
}
In 87a19e9048 I merged staging-next into master using the GitHub gui as intended.
In ac241fb7a5 I merged master into staging-next for the next staging cycle, however, I accidentally pushed it to master.
Thinking this may cause trouble, I reverted it in 0be87c7979. This was however wrong, as it "removed" master.
This reverts commit 0be87c7979.
