Cannot build numpy in specific situation

[rmartin16]

Describe the bug

When the number of characters for the parent file path for the mobile-forge repo is sufficiently low, the numpy build fails.

File "/opt/homebrew/Cellar/python@3.11/3.11.9_1/Frameworks/Python.framework/Versions/3.11/lib/python3.11/subprocess.py", line 548, in run
with Popen(*popenargs, **kwargs) as process:
^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/opt/homebrew/Cellar/python@3.11/3.11.9_1/Frameworks/Python.framework/Versions/3.11/lib/python3.11/subprocess.py", line 1026, in __init__
self._execute_child(args, executable, preexec_fn, close_fds,
File "/opt/homebrew/Cellar/python@3.11/3.11.9_1/Frameworks/Python.framework/Versions/3.11/lib/python3.11/subprocess.py", line 1955, in _execute_child
raise child_exception_type(errno_num, err_msg, err_filename)
OSError: [Errno 8] Exec format error: '/Users/marko/Documents/Repos/mobileforge2/build/cp311/numpy/1.26.2/venv3.11-ios_13_0_arm64_iphoneos/cross/bin/ninja'

ERROR Backend subprocess exited when trying to invoke get_requires_for_build_wheel
--------------------------------------------------------------------------------
<<< Return code: 1
********************************************************************************
Failed build: numpy 1.26.2 for iphoneos 13.0 on arm64
********************************************************************************
Traceback (most recent call last):
  File "/Users/marko/Documents/Repos/mobileforge2/src/forge/build.py", line 296, in build
    self._build()
  File "/Users/marko/Documents/Repos/mobileforge2/src/forge/build.py", line 563, in _build
    self.cross_venv.run(
  File "/Users/marko/Documents/Repos/mobileforge2/src/forge/cross.py", line 375, in run
    return subprocess.run(logfile, *args, **self.cross_kwargs(kwargs))
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/marko/Documents/Repos/mobileforge2/src/forge/subprocess.py", line 49, in run
    raise stdlib_subprocess.CalledProcessError(return_code, args)
subprocess.CalledProcessError: Command '(['python', '-m', 'build', '--no-isolation', '--wheel', '--outdir', '/Users/marko/Documents/Repos/mobileforge2/dist'],)' returned non-zero exit status 1.

Steps to reproduce

Clone mobile-forge in to /tmp and attempt to build numpy.

Expected behavior

Numpy can build regardless of where mobile-forge exists in the file system.

The underlying problem is the shebang that pip creates for ninja which is used by numpy's build system. Numpy invokes ninja with this file:

> cat mobile-forge/build/cp311/numpy/1.26.2/venv3.11-ios_13_0_arm64_iphoneos/cross/bin/ninja

#!/Users/user/github/beeware/mobile-forge/build/cp311/numpy/1.26.2/venv3.11-ios_13_0_arm64_iphoneos/cross/bin/python3.11
# -*- coding: utf-8 -*-
import re
import sys
from ninja import ninja
if __name__ == '__main__':
    sys.argv[0] = re.sub(r'(-script\.pyw|\.exe)?$', '', sys.argv[0])
    sys.exit(ninja())

The problem here is that venv3.11-ios_13_0_arm64_iphoneos/cross/bin/python3.11 is not a binary; it is another python script that eventually calls os.execv() for venv3.11-ios_13_0_arm64_iphoneos/build/bin/python3.11....and this python3.11 is a binary.

pip creates this shebang when installing ninja. As you can see there, if the shebang is long enough, it creates a shebang that has /bin/sh run venv3.11-ios_13_0_arm64_iphoneos/build/bin/python3.11. This avoids this issue altogether.

pip does a lot of introspection to determine what to use for the shebang....and a lot of the possible values come from sysconfig. Given how much crossenv is messing with sysconfig, this could possibly be considered a bug in crossenv.

Screenshots

No response

Environment

• Operating System: Sonoma
• Python version: 3.11
• Software versions:
  • mobile-forge: eb36d27

Logs

numpy-1.26.2-cp311-ios_13_0_x86_64_iphonesimulator.log
numpy-1.26.2-cp311-ios_13_0_arm64_iphonesimulator.log
numpy-1.26.2-cp311-ios_13_0_arm64_iphoneos.log

Additional context

No response

[freakboy3742]

Confirmed that I can reproduce this - and, interestingly, I also see the "forge failure causes shell exit" bug in the /tmp-based environment.

[freakboy3742]

I've resolved the "forge failure causes shell exit" problem - but now I'm seeing this build numpy failure on every build... still trying to understand what is going on here. It appears to be a discrepancy between what happens when you invoke /path/to/ninja at a venv prompt, and what happens when you run subprocess.run(["/path/to/ninja"]) in a python shell in the same venv.

[rmartin16]

My thought was it's because subprocess.run() is assuming a binary at some point to invoke; it appears to respect the shebang when the first index of the command list for run() isn't a binary itself....but it doesn't then keep recursively reading the shebangs down this chain. For instance, I think if you add shell=true, then it will handle this non-binary shebang.

[freakboy3742]

shell=True doesn't appear to work, because it assumes that the script being referenced is a shell script, so it breaks because the script content is actually Python code.

The problem appears to be a problem with the OS-level execv call, which is rejecting the initial binary because the shebang is script, not an actual binary... but I have no idea why this was working just 2 weeks ago. My only assumption is that some dependency in the chain that leads to ninja being installed has changed... but I can't work out what dependency that would be. pip, wheel and ninja itself have all been stable for ~2 months; the only package that has changes is setuptools... but setuptools shouldn't be involved in either building or installing ninja.

[rmartin16]

Perhaps just for my own edification, how do you know this wasn't happening two weeks ago? I guess more to the point, are you confident that testing was using a short enough parent file path to mobile-forge for the issue to have had a chance to present?

My impression is this has been a latent issue that I just happened to finally encounter.

[freakboy3742]

I know it was working 2 weeks ago because I was able to publish numpy versions using this script; I'm now seeing this problem in every build, not just the /tmp ones.

However.... those builds were done prior to the cross rename... and the older name was considerably longer...

[freakboy3742]

Yup. It's the path rewriting thing you flagged.

Prior to the cross rename, the code worked because I was building in /Users/rkm/beeware/mobile-forge/build/cp312/numpy/1.26.2/venv3.12-ios_13_0_arm64_iphoneos/venv3.12-ios_13_0_arm64_iphoneos/bin/python3.12 (138 chars); it now uses /Users/rkm/beeware/mobile-forge/build/cp312/numpy/1.26.2/venv3.12-ios_13_0_arm64_iphoneos/cross/bin/python3.12 (111 chars), so it no longer uses the exec shim, causing the build to fail.

/me swears extensively.

So - there's a fix needed in crossenv to ensure that the exec shim is always used; with an alternate (more permanent) fix upstream in pip as well.

[freakboy3742]

(also, put another tally in the "RKM, just pay attention to what Russell Martin says, and your life will be a lot easier" column... :-) )

[rmartin16]

haha, careful! I can't vouch for everything I say... FWIW, I am still curious if this is a valid file path for crossenv to put in to sysconfig; while we've only seen an issue when the file path is put in to a shebang, perhaps other use cases also assume it to be directly executable or something.

[freakboy3742]

The path that is in the shebang is the "host Python" - however, the "host Python" can't be an executable (at least, not easily), because it's a shell wrapper around the build Python. The only way that you could make it an actual executable is by writing a binary wrapper that does what a shell script does... which is possible, but not really nice to work with.

Given that the script works fine in every context except os.execv(), I think it's not an unreasonable value for the shebang.

I've pushed an update to my branch of crossenv that patches the problem. I'm working on a bug report and patch for distutils (which will eventually turn in to a vendored patch in pip).

[freakboy3742]

The patch for distlib is now available: pypa/distlib#231

This patch has also been incorporated into the patch for pip: pypa/pip#12962

Once the patch for pip has been landed, the crossenv modification will no longer be required; however, there will be a need to modify mobile-forge ensure that the version of pip is up to date.

As soon as someone can confirm that numpy builds are working for them, I think this bug can be closed.

[freakboy3742]

A patch to enforce pip updates: #64