Python with ctypes for an ancient machine

[CsatiZoltan]

I want to install a recent Python 3 (preferably Python 3.13) to be used with Nuitka on a cluster. That cluster has glibc 2.12 from 2010.
The Python installation must have the ctypes module available due to Nuitka. ctypes requires the libffi library, which I didn't manage to install because of the ancient toolchain.
So I wanted to give uv a try. I installed uv, then told it to install a Python 3.13:

$ uv python install 3.13
error: Failed to install cpython-3.13.5-linux-x86_64-gnu
  Caused by: Failed to download https://github.com/astral-sh/python-build-standalone/releases/download/20250612/cpython-3.13.5%2B20250612-x86_64-unknown-linux-gnu-install_only_stripped.tar.gz
  ...

which makes sense because the cluster does not connect to the internet (I installed musl uv manually).
At least I know which Python version uv wanted to install. Then I downloaded that version and launched the interpreter:

$ ./python
./python: /lib64/libc.so.6: version `GLIBC_2.17' not found (required by ./python)
./python: /lib64/libc.so.6: version `GLIBC_2.13' not found (required by ./python)
./python: /lib64/libc.so.6: version `GLIBC_2.14' not found (required by ./python)
./python: /lib64/libc.so.6: version `GLIBC_2.15' not found (required by ./python)

It also makes sense, uv wanted to download a version linked with glibc.
My next attempt was to download a static musl version, namely cpython-3.13.5+20250612-x86_64_v4-unknown-linux-musl-noopt+static-full.tar.zst. Well, this Python has no runtime dependency, so it worked. However, ctypes couldn't be imported:

ModuleNotFoundError: No module named '_ctypes'

so probably this musl version was compiled without libffi.

---

I am running out of ideas. Is there a way to make Python 3 (with ctypes and Nuitka) work on this old machine?

[zanieb]

Hm, I downloaded that version and it does include libffi

❯ find python -name "libffi.a"
python/build/lib/libffi.a

and I don't see an obvious exclusion of ctypes for that platform

python-build-standalone/cpython-unix/extension-modules.yml

Lines 110 to 145 in d2030cb

	_ctypes:
	sources:
	- _ctypes/_ctypes.c
	- _ctypes/callbacks.c
	- _ctypes/callproc.c
	- _ctypes/stgdict.c
	- _ctypes/cfield.c
	sources-conditional:
	- source: _ctypes/darwin/dlfcn_simple.c
	targets:
	- .*-apple-.*
	# Functionality removed in 3.12.
	maximum-python-version: "3.11"
	- source: _ctypes/malloc_closure.c
	targets:
	- .*-apple-.*
	defines:
	- HAVE_FFI_PREP_CIF_VAR=1
	- HAVE_FFI_PREP_CLOSURE_LOC=1
	- HAVE_FFI_CLOSURE_ALLOC=1
	defines-conditional:
	- define: MACOSX
	targets:
	- .*-apple-.*
	- define: USING_MALLOC_CLOSURE_DOT_C=1
	targets:
	- .*-apple-.*
	includes-deps:
	- include
	includes-conditional:
	- path: _ctypes/darwin
	targets:
	- .*-apple-.*
	links:
	- ffi
	- dl

we even have a patch for some ctypes problem with the static builds

python-build-standalone/cpython-unix/build-cpython.sh

Lines 213 to 216 in eb29135

	# Code that runs at ctypes module import time does not work with
	# non-dynamic binaries. Patch Python to work around this.
	# See https://bugs.python.org/issue37060.
	patch -p1 -i ${ROOT}/patch-ctypes-static-binary.patch

[zanieb]

I actually can import ctypes

$ wget https://github.com/astral-sh/python-build-standalone/releases/download/20250612/cpython-3.13.5+20250612-x86_64-unknown-linux-musl-noopt+static-full.tar.zst
$ tar -xvf cpython-3.13.5+20250612-x86_64-unknown-linux-musl-noopt+static-full.tar.zst
$ docker run -it --platform linux/amd64 -v ./python:/python alpine:latest /python/install/bin/python -c "import ctypes"

(I can't use the v4 variant because my CPU does not support that)

[CsatiZoltan]

What is the difference between the variants?

[zanieb]

They're x86-64 microarchitecture variants, https://github.com/astral-sh/uv/blob/26db29caac6965359b375fb1733e591b89068ec3/crates/uv-python/src/platform.rs#L20-L32

[CsatiZoltan]

It turned out that our Intel Xeon E5-2620 does not have the required instruction sets, so that's why the v4 variant didn't work.
The version you showed works fine and has ctypes.
However, it is not compatible with Nuitka:

b'Traceback (most recent call last):'
b'# .../python/install/lib/python3.13/__pycache__/ast.cpython-313.pyc matches .../python/install/lib/python3.13/ast.py'
b"# code object from .../python/install/lib/python3.13/__pycache__/ast.cpython-313.pyc'"
b"import '_ast' # <class '_frozen_importlib.BuiltinImporter'>"
b"import 'ast' # <_frozen_importlib_external.SourceFileLoader object at 0x7f6828d1bb90>"
b'  File "<string>", line 1, in <module>'
b'    import sys;exec(sys.stdin.read())'
b'               ~~~~^^^^^^^^^^^^^^^^^^'
b'  File "<string>", line 20, in <module>'
b'  File "<frozen importlib._bootstrap>", line 1360, in _find_and_load'
b'  File "<frozen importlib._bootstrap>", line 1331, in _find_and_load_unlocked'
b'  File "<frozen importlib._bootstrap>", line 935, in _load_unlocked'
b'  File "<frozen importlib._bootstrap_external>", line 1026, in exec_module'
b'  File "<frozen importlib._bootstrap>", line 488, in _call_with_frames_removed'
b'  File ".../python/install/lib/python3.13/_pyrepl/_minimal_curses.py", line 31, in <module>'
b'    clib = ctypes.cdll.LoadLibrary(_clibpath)'
b'  File ".../python/install/lib/python3.13/ctypes/__init__.py", line 471, in LoadLibrary'
b'    return self._dlltype(name)'
b'           ~~~~~~~~~~~~~^^^^^^'
b'  File ".../python/install/lib/python3.13/ctypes/__init__.py", line 390, in __init__'
b'    self._handle = _dlopen(self._name, mode)'
b'                   ~~~~~~~^^^^^^^^^^^^^^^^^^'

when the --standalone Nuitka flag is used. Without this flag, I get the linker error

/usr/bin/ld: BFD version 2.20.51.0.2-5.36.el6 20100205 internal error, aborting at reloc.c line 443 in bfd_get_reloc_size

Our compiler and linker:

$ ldd --version
ldd (GNU libc) 2.12
$ CC --version
gcc (GCC) 9.2.0

[zanieb]

You can't use dlopen from a fully statically linked binary like that, I presume that's what's happening? I'm not sure if it's expected that Nuitka will work in that scenario.

[CsatiZoltan]

That makes sense. Anyway, we moved away from this old machine to a newer one, where the non-static Pythons downloaded by uv work fine with Nuitka. Thanks for the explanations.