Segmentation fault when attempting to load `msys-magic-1.dll` from Git SCM

[jdknight]

Running Python from within a Windows Git SCM shell (MINGW64; which includes file support), it appears the import of the magic module would always fail to load. The typical error would be a Segmentation fault; however, sometimes the error state would vary (either hanging forever or another error). This script (just the import):

import magic

Tweaked the magic library to see where it was failing:

    print('lib', lib, os.path.isfile(lib), flush=True)
    try:
      return ctypes.CDLL(lib)
    except OSError:
      pass
    print('failed', flush=True)

Segmentation fault

``` $ python ./my-script lib ./libmagic.dll False failed lib ./magic1.dll False failed lib ./cygmagic-1.dll False failed lib ./libmagic-1.dll False failed lib ./msys-magic-1.dll False failed lib C:\Program Files\Git\usr\bin\msys-magic-1.dll True Segmentation fault ```

Internal error

``` $ python ./my-script lib ./libmagic.dll False failed lib ./magic1.dll False failed lib ./cygmagic-1.dll False failed lib ./libmagic-1.dll False failed lib ./msys-magic-1.dll False failed lib C:\Program Files\Git\usr\bin\msys-magic-1.dll True 0 [main] python 1822 C:\Program Files\Python311\python.exe: *** fatal error - Internal error: TP_NUM_C_BUFS too small: 50 479 [main] python 1822 cygwin_exception::open_stackdumpfile: Dumping stack trace to python.exe.stackdump ```

---

After installing the python-magic-bin package, this module worked as expected.

Loaded up a real MinGW-w64 environment, installed file (and manually deleted their broken magic.py), and installed python-magic — the magic module worked as expected.

---

Originally, I assumed msys-magic-1.dll library provided by Git SCM is built in a way which is invalid for the Python interpreter to load. Checking to make sure it wasn't a specific interpreter version, I made a simple DLL load script to see if the DLL could be loaded and tested on a couple of interpreters (3.7, 3.8, 3.10 and 3.11):

import ctypes
lib = r'C:\Program Files\Git\usr\bin\msys-magic-1.dll'
print('lib', lib, flush=True)
print(ctypes.CDLL(lib))

From these checks, it looked like all interpreters were able to load the DLL:

lib C:\Program Files\Git\usr\bin\msys-magic-1.dll
<CDLL 'C:\Program Files\Git\usr\bin\msys-magic-1.dll', handle 4b3930000 at 0x271234aa8d0>

Hacking around, it looks that I was able to use the magic module in this environment (i.e. without python-magic-bin) by doing two steps:

1. Commenting out _add_compat(globals()) from python-magic's __init__.py script; and,
2. Manually loading the DLL before attempting to load the magic module:

import ctypes
ctypes.CDLL(r'C:\Program Files\Git\usr\bin\msys-magic-1.dll')
import magic

Granted, these changes are not really good for long term. Had to stop investigating after this. Don't know if anyone has any suggestions/comments on this. I'll try to find time later to continue this investigation (unless someone else beats me to it).

---

(python-magic 0.4.27, Git SCM 2.40.1, Windows 22H2)

[ahupp]

Can you explicitly setting PATH to control where the DLL is being searched for? I'm not sure what the right solution is here other than avoiding loading the library entirely.

[ahupp]

Merging into https://github.com/ahupp/python-magic/issues/293