python-cryptography: fix cross #226120
python-cryptography: fix cross #226120
Conversation
- setuptools-rust and rustPlatform.cargoSetupHook has been moved
from nativeBuildInputs to buildInputs.
- rustPlatform seems to interfere with splicing, so rustc
and cargo are now top-level parameters in order to facilitate that.
- Two environment variables need to be passed explicitly:
- CARGO_BUILD_TARGET because otherwise the final link step
attempts to use the hostPlatform linker
- PYO3_CROSS_LIB_DIR because without it pyo3 throws a fit.
ofborg
bot
added
the
6.topic: cross-compilation
|
@ofborg build pkgsCross.aarch64-multiplatform.python3Packages.cryptography |
| setuptools-rust | ||
| rustPlatform.cargoSetupHook |
There was a problem hiding this comment.
setuptools is consumed by the python that is building as far as I know. setup hooks also usually go to nativeBuildInputs. I assume the hooks themselves require some fixes to fix this properly.
| @@ -21,6 +21,11 @@ | |||
| , py | |||
| , pytz | |||
| , hypothesis | |||
| , buildPackages | |||
There was a problem hiding this comment.
| , buildPackages |
| cargo | ||
| rustc |
There was a problem hiding this comment.
I think we used rustPlaform to make sure all the rust things fit together. Can you describe the problem why you have split this?
There was a problem hiding this comment.
rustPlatform breaks splicing, which is why #212795 needs that actually #212795 does the same thing.CARGO_BUILD_TARGET hack.
We should stop using rustPlatform.cargo and rustPlatform.rustc. The splicing magic only seems to work for arguments that are passed directly rather than by reference through an attrset.
@Artturin can probably explain this better, he actually understands it while I just pretend to.
There was a problem hiding this comment.
It would be similar to #211340 but it's not the issue here because those attrs are spliced correctly pkgsCross.aarch64-multiplatform.__splicedPackages.rustPlatform.rust.rustc.__spliced (because they're not inside a derivation, read through the issue.)
nix-repl> pkgsCross.aarch64-multiplatform.python3Packages.cryptography.nativeBuildInputs
[ ... «derivation /nix/store/7zsx9k68akrnhbznhr2q63fcxv853ksv-cargo-setup-hook.sh.drv» «derivation /nix/store/31rhagc1nqf5d0wdjwvc0hbg3dg71xcl-python3.10-setuptools-rust-1.5.2.drv» «derivation /nix/store/1977acij72524vy0c8ij19ng7nzrf6yv-cargo-1.68.2.drv» «derivation /nix/store/b0lpygng67j4iww09wr9c45gl66vs84d-rustc-1.68.2.drv» ]
no aarch64-unknown-linux-gnu suffix so they should be for build
There was a problem hiding this comment.
diff --git a/pkgs/development/python-modules/cryptography/default.nix b/pkgs/development/python-modules/cryptography/default.nix
index ba2407923f3..105a5dd4dfc 100644
--- a/pkgs/development/python-modules/cryptography/default.nix
+++ b/pkgs/development/python-modules/cryptography/default.nix
@@ -77,6 +77,8 @@ buildPythonPackage rec {
"--disable-pytest-warnings"
];
+ passthru = { inherit rustPlatform; };
+
disabledTestPaths = lib.optionals (stdenv.isDarwin && stdenv.isAarch64) [
# aarch64-darwin forbids W+X memory, but this tests depends on it:
# * https://cffi.readthedocs.io/en/latest/using.html#callbacksnix-repl> pkgsCross.aarch64-multiplatform.__splicedPackages.python3Packages.cryptography.rustPlatform.rust.rustc.__spliced
{ buildBuild = «derivation /nix/store/b0lpygng67j4iww09wr9c45gl66vs84d-rustc-1.68.2.drv»; buildHost = «derivation /nix/store/b0lpygng67j4iww09wr9c45gl66vs84d-rustc-1.68.2.drv»; buildTarget = «derivation /nix/store/b0lpygng67j4iww09wr9c45gl66vs84d-rustc-1.68.2.drv»; hostHost = «derivation /nix/store/g01vq22qlqgpcihqjvg2g79byg04gplv-rustc-1.68.2.drv»; hostTarget = «derivation /nix/store/g01vq22qlqgpcihqjvg2g79byg04gplv-rustc-1.68.2.drv»; }
There was a problem hiding this comment.
It would be similar to #211340 but it's not the issue here because those attrs are spliced correctly ... (because they're not inside a derivation, read through the issue.)
In other words, it is exactly the issue here, but "spliced correctly" has been defined in a way that is totally contrary to peoples' expectations about how this should work, just so this can be declared NOTABUG.
I think it is time to admit that the splicing magic was not a good idea. It misleads people into thinking they know what's going on when they don't.
Your Edit: sorry, I misinterpreted that diff as being a proposed solution rather than an investigatory tool.passthru hack is the same kind of thing: sure, it works, but for reasons that 99% of the people reading the code will not understand. It is a footgun.
There was a problem hiding this comment.
Your passthru hack is the same kind of thing: sure, it works, but for reasons that 99% of the people reading the code will not understand. It is a footgun.
It's not a hack, it's just for inspecting the attrset and showing that the package attrs have __spliced.
|
I think this will be fixed on a higher level and for all packages in #212795 and also for cryptography. I think the linked PR is the more correct solution for your problem, hence I am going to close this one. We would appreciate it, if you could test #212795 and give us feedback, if it fixes your problem. |
ghost
deleted the
This PR was produced by banging my head against the wall. Repeatedly.
Description of changes
Make
pkgsCross.*.python-cryptographybuild.setuptools-rust and rustPlatform.cargoSetupHook has been moved from nativeBuildInputs to buildInputs.
rustPlatform seems to interfere with splicing, so rustc and cargo are now top-level parameters in order to facilitate that.
Two environment variables need to be passed explicitly:
CARGO_BUILD_TARGET because otherwise the final link step attempts to use the hostPlatform linker
PYO3_CROSS_LIB_DIR because without it pyo3 throws a fit.
Things done
x86_64-linux)