We now have the information to properly determine the role the
cc-wrapper dependency has, by taking advantage of `offset`. No longer
use the soon-to-be-deprecated crossConfig environment variable, the
temp hack used before this change.
4 far-reaching changes: Smaller PATH, New vars, different propagation
logic, and different hook logic
Smaller PATH
------------
`buildInputs` no longer go on the PATH at build time, as they cannot be
run when cross compiling and we don't want to special case. Simply make
a `nativeBuildInput` too if one needs them on the PATH. Fixes#21191.
Many new depedendency variables
-------------------------------
See the stdenv chapter of the nixpkgs manual. I pulled out the existing
documentation of dependency specification into a new section, and added
language for these two (and their propagated equivalents) along side
the others'.
More complex propagation logic
------------------------------
Before a propagated*XXX*Input always acted as if it was specified
directly as a *XXX*Input downstream. That's simple enough, but violates
the intended roles of each sort of dep, which has functional and not
just stylistic consequences.
The new algorithm is detailed in the manual, and ensures everything
ends up in the right place. I tried to give both an informal and formal
description, but I suspect in practice it will not make much sense
until one tries cross compiling, after which it will immediately make
sense as the only sane option.
Simplified hook logic
---------------------
Rather than `envHook` and `crossEnvHook`, whose behavior differs
depending on whether we are cross compiling or not, there is now one
hook per sort (or rather non-propagated and propagated pair of sorts)
of dependency. These new hooks have the same meaning regardless of
cross compilation. See the setup hook section of stdenv chapter of the
Nixpkgs manual for more details.
stdenvNoCC should not inject any C++ standard library, just as it
doesn't inject any C standard library. stdenv still does, but only
indirectly through stdenv.cc. Wrapped clangs can be simplified now that
they don't need to worry about clobbering CoreFoundation when replacing
the C++ standard library implementation.
This generally-good cleanup should assist with debugging some C++
failures in #26805.
- tracing seems annoying enough
- we get errors for all packages instead of aborting on the first one
- easier to differentiate from unwanted packages (broken, unfree, etc.)
This continues #23374, which always kept around both attributes, by
always including both propagated files: `propgated-native-build-inputs`
and `propagated-build-inputs`. `nativePkgs` and `crossPkgs` are still
defined as before, however, so this change should only barely
observable.
This is an incremental step to fully keeping the dependencies separate
in all cases.
I find the separation of concerns, accumulating, then processing, easier
to follow. Also, with my yet-to-be-merged cross work, the accumulation
part will become more complex.
One should do this when needed executables at run time. It is more
honest and cross-friendly than refering to binutils directly, if one
neeeds the default binary tools for the target platform, rather than
binutils in particular.
This requires some small changes in the stdenv, then working around the
weird choice LLVM made to hardcode @rpath in its install name, and then
lets us remove a ton of annoying workaround hacks in many of our Go
packages. With any luck this will mean less hackery going forward.
cc-wrapper may wrap a cc-compiler, but it doesn't need one to build
itself. (c.f. expand-response-params is a separate derivation.) This
helps avoid cycles on the cross stuff, in addition to removing a
useless dependency edge.
I could have been super careful with overrides in the stdenv to avoid
the mass rebuild, but I don't think it's worth it.
Why 6? It seems a decently high number, giving us room for more degrees
of debugging before the `set -x` sledgehammer without incurring a
mass-rebuild.
By default, all previous overrides are discarded as before, as they
would only apply to the old host platform. But sometimes it is useful to
add some new ones, and this optional parameter allows that.
This is needed when cross-compiling for iOS (Aarch64 + Darwin). I also
changed the syntax of the Linux stdenv for visual consistency, though
that has no effect on semantics as the os is already guaranteed to be
Linux.
This reverts commit 0a944b345e, reversing
changes made to 61733ed6cc.
I dislike these massive stdenv changes with unclear motivation,
especially when they involve gratuitous mass renames like NIX_CC ->
NIX_BINUTILS. The previous such rename (NIX_GCC -> NIX_CC) caused
months of pain, so let's not do that again.
Environment variable filter in substituteAll was not precise and produced
undefined and invalid variable names. Vladimír Čunát tried to fix that in [1],
but `env -0` did not work during Darwin bootstrap, so [2] reverted this change
and replaced an error due to invalid variables with a warning. Recently in #28057
John Ericson added `set -u` to `setup.sh` and undefined variables made the setup
fail during e.g. `nix-build -A gnat` with `setup: line 519: !varName: unbound
variable`.
[1] 62fc8859c1
[2] 81df035429
This becomes necessary if more wrappers besides cc-wrapper start
supporting hardening flags. Also good to make the warning into an
error.
Also ensure interface is being used right: Not as a string, not just in
bash.
- Don't build with libsigsegv by default. The build apparently attempted
to link against it, but it never retained the reference anyway...
- Side effect: stdenv bootstrapping needs no libsigsegv anymore.
- Run checks, but only in the interactive gawk by default on Linux,
so that stdenv bootstrap isn't slowed down (by glibc locales, etc.).
- xz should be no longer needed in inputs, as we have it in stdenvs now.
The whole change was triggered by some used kernel versions still
breaking libsigsegv tests #28464.
Older bash version, like those in the bootstrap tools and on macOS,
currently confuse variables defined as an empty array with undefined
variables. `${foo+"${foo[@]}"}` will prevent `set -u` problems with
empty arrays and older without making a single '' in the empty case.
Care is taken to `set +u` when running hooks so as to not break existing
packages.
This reverts commit eeabf85780.
This change suddenly makes tons of stdenv internals visible in
nativeBuildInputs of every derivation, which doesn't seem desirable.
E.g:
````
nix-repl> hello.nativeBuildInputs
[ «derivation /nix/store/bcfkyf6bhssxd2vzwgzmsbn7b5b9rpxc-patchelf-0.9.drv»
«derivation /nix/store/4wnshnz9wwanpfzcrdd76rri7pyqn9sk-paxctl-0.9.drv»
<< snip 10+ lines >>
«derivation /nix/store/d35pgh1lcg5nm0x28d899pxj30b8c9b2-gcc-wrapper-6.4.0.drv»
]
````
Additionally, instead of pulling them from `setup.sh`, route them via
Nix. This gets us one step closer to making stdenv be a plain attribute
set instead of a derivation.
Currently the logic of generating nixpkgs Hydra jobs is to walk through
the pkgs evaluated for system = "x86_64-linux", collect any derivations
and their meta.platforms values. However, that doesn't work for
packages whose meta.platforms doesn't include x86_64-linux, as just
evaluating their meta attribute raises an error so they get skipped
completely.
As a less-intrusive fix (i.e. anything than rewriting the current package
enumeration logic), allow passing `config.allowUnsupportedSystem = true`
to permit evaluating packages regardless of their platform and use that
in the package listing phase.
Fixes#25200
@vcunat and others rightly point out that it's easier to quote always,
than learn Bash's idiosyncrasies enough to know when it doesn't make a
difference.
This reverts commit 2743078f66, which
removes quotes that don't do anything, and then goes further adding
even more quotes.
The logic was made pure for the normal libSystem, but this change never
made it to the bootstrap tools. Deduplication the logic as the comment
suggests would have prevented this, but here's a stop-gap until we do
so.
It's better than the eval solution this is adding back, but until we can
rely on a particular version of bash in nix-shell, this just breaks too
much stuff.
See c94f3d5575
and https://github.com/NixOS/nix/pull/1483 for the better long-term
solution.
As @oxij points out in [1], this breakage is especially serious because
it changes the contents of built environments without a corresonding
change in their hashes. Also, the revert is easier than I thought.
This reverts commit 3cb745d5a6.
[1]: https://github.com/NixOS/nixpkgs/pull/27427#issuecomment-317293040
`nix-build pkgs/top-level/release.nix -A tarball` now succeeds.
`configureFlags = null` lead to a type error, and one overrideDrv
needed to be converted to to append a configureFlags list instead of
string due to the normalization.
Thanks @vcunat for alerting me to the issues---sorry I did not catch
them before merging my own PR.
Eventually the adapter will be removed. Moved is
- Name suffix from hostPlatform
- configurePlatforms
To not cause more breakage, the default is currently [], but
eventually it will be [ "build" "host" ], as the cross adapter makes
it today.
I took some liberties with the flags-echoing code to make it more
concise and correct. Also, a few warnings in findInputs and friends I
skipped because I am going to rewrite those anyways.
Thanks @grahamc for telling me about this great linter!
The main changes are in libSystem, which lost the coretls component in 10.13
and some hardening changes that quietly crash any program that uses %n in
a non-constant format string, so we've needed to patch a lot of programs that
use gnulib.
This makes those files a bit easier to read. Also, for what it's worth,
it brings us one baby step closer to handling spaces in store paths.
Also, I optimized handling of many transitive deps with read. Probably,
not very beneficial, but nice to enforce the pkg-per-line structure.
Doing so let me find much dubious code and fix it.
Two misc notes:
- `propagated-user-env-packages` also needed to be adjusted as
sometimes it is copied to/from the propagated input files.
- `local fd` should ensure that file descriptors aren't clobbered
during recursion.
Only cosmetic changes are done otherwise.
Real refactoring is left for later.
There's a small slow-down on my machine:
$ time nix-env -qa -P >/dev/null
gets from ~2.8 to ~3.5 seconds (negligible change in RAM).
That's most likely caused by sharing less computation between different
mkDerivation calls, and I plan to improve that soon.
This is a bit simpler now, but more importantly it scales better when I
double the number of sorts of dependencies as part of my cross
compilation work.
Packages get --host and --target by default, but can explicitly request
any subset to be passed as needed. See docs for more info.
rustc: Avoid hash breakage by using the old (ignored)
dontSetConfigureCross when not cross building
This is especially useful when not cross compiling. It means we can
remove the `stdenv.isGlibc` predicate too.
Additionally, use this to simplify the logic to choose the
appropriate libiconv derivation.
`pkgsNoParams` was removed by me, but then #25035 was merged using it,
leading to an unbound identifier.
It would be nice to get travis to do build release-cross.nix or
something to catch these things.
In the extremely unlikely case that our store hash path ends in several
digits (as is the case right now), the Darwin ld will try to interpret
those digits as a version number and barf. To avoid that, we pass in the
SDK version explicitly to stop it from trying to figure it out from iffy
context.
See previous commit for what was done to `binutils` to make this
possible.
There were some uses of `forcedNativePackages` added. The
combination of overrides with that attribute is highly spooky: it's
often important that if an overridden package comes from it, the
replaced arguments for that package come from it. Long term this
package set and all the spookiness should be gone and irrelevant:
"Move along, nothing to see here!"
No hashes should be changed with this commit
Use `buildPackages.binutils` to get build = host != target binutils,
i.e. the old `binutilsCross`, and use
`buildPackages.buildPackages.binutils` to get build = host = target
binutils, i.e. the old `binutils`.
`buildPackages` chains like this are supposed to remove the need for
all such `*Cross` derivations. We start with binutils because it's
comparatively easy.
No hashes of cross-tests should be changed
stdenv.cross is a silly attribute that needs to go leaving the well-defined hostPlatform and targetPlatform. This PR doesn't remove it, but changes its definition: before it tracked the target platform which is sometimes more useful for compilers, and now it tracks the host platform which is more useful for everything else. Most usages are libraries, falling in the "everything else" category, so changing the definition makes sense to appease the majority. The few compiler (gcc in particular) uses that exist I remove to use targetPlatform --- preserving correctness and becoming more explicit in the process.
I would also update the documentation aside mentioning stdenv.cross as deprecated, but the definition given actually erroneously assumes this PR is already merged!
Before all overrides were also pruned in the previous stage, now
only gcc and binutils are, because they alone care about about the
target platform. The rest of the overrides don't, so it's better to
preserve them in order to avoid spurious rebuilds.
Each bootstrapping stage ought to just depend on the previous stage, but
poorly-written compilers break this elegence. This provides an easy-enough
way to depend on the next stage: targetPackages. PLEASE DO NOT USE IT
UNLESS YOU MUST!
I'm hoping someday in a pleasant future I can revert this commit :)
When not cross compiling, nativeBuildInputs and buildInputs have
identical behaviour. Currently that is implemented by having
mkDerivation do a concatenation of those variables in Nix code and pass
that to the builder via the nativeBuildInputs attribute.
However, that has some annoying side effects, like `foo.buildInputs`
evaluating to `[ ]` even if buildInputs were specified in the nix
expression for foo.
Instead, pass buildInputs and nativeBuildInputs in separate variables as
usual, and move the logic of cross compilation vs. native compilation to
the stdenv builder script. This is probably a tiny bit uglier but
fixes the previous problem.
Issue #4855.
If a package's meta has `knownVulnerabilities`, like so:
stdenv.mkDerivation {
name = "foobar-1.2.3";
...
meta.knownVulnerabilities = [
"CVE-0000-00000: remote code execution"
"CVE-0000-00001: local privilege escalation"
];
}
and a user attempts to install the package, they will be greeted with
a warning indicating that maybe they don't want to install it:
error: Package ‘foobar-1.2.3’ in ‘...default.nix:20’ is marked as insecure, refusing to evaluate.
Known issues:
- CVE-0000-00000: remote code execution
- CVE-0000-00001: local privilege escalation
You can install it anyway by whitelisting this package, using the
following methods:
a) for `nixos-rebuild` you can add ‘foobar-1.2.3’ to
`nixpkgs.config.permittedInsecurePackages` in the configuration.nix,
like so:
{
nixpkgs.config.permittedInsecurePackages = [
"foobar-1.2.3"
];
}
b) For `nix-env`, `nix-build`, `nix-shell` or any other Nix command you can add
‘foobar-1.2.3’ to `permittedInsecurePackages` in
~/.config/nixpkgs/config.nix, like so:
{
permittedInsecurePackages = [
"foobar-1.2.3"
];
}
Adding either of these configurations will permit this specific
version to be installed. A third option also exists:
NIXPKGS_ALLOW_INSECURE=1 nix-build ...
though I specifically avoided having a global file-based toggle to
disable this check. This way, users don't disable it once in order to
get a single package, and then don't realize future packages are
insecure.
If a package's meta has `knownVulnerabilities`, like so:
stdenv.mkDerivation {
name = "foobar-1.2.3";
...
meta.knownVulnerabilities = [
"CVE-0000-00000: remote code execution"
"CVE-0000-00001: local privilege escalation"
];
}
and a user attempts to install the package, they will be greeted with
a warning indicating that maybe they don't want to install it:
error: Package ‘foobar-1.2.3’ in ‘...default.nix:20’ is marked as insecure, refusing to evaluate.
Known issues:
- CVE-0000-00000: remote code execution
- CVE-0000-00001: local privilege escalation
You can install it anyway by whitelisting this package, using the
following methods:
a) for `nixos-rebuild` you can add ‘foobar-1.2.3’ to
`nixpkgs.config.permittedInsecurePackages` in the configuration.nix,
like so:
{
nixpkgs.config.permittedInsecurePackages = [
"foobar-1.2.3"
];
}
b) For `nix-env`, `nix-build`, `nix-shell` or any other Nix command you can add
‘foobar-1.2.3’ to `permittedInsecurePackages` in
~/.config/nixpkgs/config.nix, like so:
{
permittedInsecurePackages = [
"foobar-1.2.3"
];
}
Adding either of these configurations will permit this specific
version to be installed. A third option also exists:
NIXPKGS_ALLOW_INSECURE=1 nix-build ...
though I specifically avoided having a global file-based toggle to
disable this check. This way, users don't disable it once in order to
get a single package, and then don't realize future packages are
insecure.
This fixes the "sliding window" principle:
0. Run packages: build = native; host = foreign; target = foreign;
1. Build packages: build = native; host = native; target = foreign;
2. Vanilla packages: build = native; host = native; target = native;
3. Vanilla packages: build = native; host = native; target = native;
n+3. ...
Each stage's build dependencies are resolved against the previous stage,
and the "foreigns" are shifted accordingly. Vanilla packages alone are
built against themsevles, since there are no more "foreign"s to shift away.
Before, build packages' build dependencies were resolved against
themselves:
0. Run packages: build = native; host = foreign; target = foreign;
1. Build packages: build = native; host = native; target = foreign;
2. Build packages: build = native; host = native; target = foreign;
n+2. ...
This is wrong because that principle is violated by the target
platform staying foreign.
This will change the hashes of many build packages and run packages, but
that is OK. This is an unavoidable cost of fixing cross compiling.
The cross compilation docs have been updated to reflect this fix.
Our bootstrap tools are actually broken right now due to busybox not
working when invoked directly from a store path. (It says e.g.
"0qqqw19y4gmknajw8vg4fvhx9gxdqlhz-busybox: applet not found").
Make this test actually fail in such case, the next commit will fix the
problem with busybox.
This is required for Aarch64 since a lot of source tarballs ship with
outdated configure scripts that don't recognize aarch64. Simply
replacing the config.guess and config.sub with new versions from
upstream makes them build again.
This same approach is used by at least Buildroot and Fedora. In
principle this could be enabled for all architectures but
conditionalizing this on aarch64 avoids a mass rebuild on x86.
