classification
Title: Python crashes upon exit if importing g++ compiled mod after importing gcc compiled mod
Type: Stage:
Components: Documentation, Extension Modules, Windows Versions: Python 2.7
process
Status: open Resolution:
Dependencies: Superseder:
Assigned To: docs@python Nosy List: docs@python, matham, ncoghlan, r.david.murray, steve.dower, tim.golden, zach.ware
Priority: normal Keywords:

Created on 2015-03-23 05:10 by matham, last changed 2015-03-26 02:19 by ncoghlan.

Messages (6)
msg238981 - (view) Author: (matham) Date: 2015-03-23 05:10
I have encountered a situation where python crashes when python exits if one imports a c compiled extension followed by a cpp compiled extension (but not if imported in the reverse order). This is on windows with mingw (current using mingw-get install gcc g++ msys-make) and py2.7.9.

This is basically the issue reported at https://mail.python.org/pipermail/python-win32/2013-April/012798.html a while back by someone else, but I'm filing it in bug form so it can't be ignored :D

Here's how to replicate it:

cplayground.c:

#include <Python.h>

static PyObject*
say_hello(PyObject* self, PyObject* args)
{
    Py_RETURN_NONE;
}

static PyMethodDef HelloMethods[] =
{
     {"say_hello", say_hello, METH_VARARGS, "Greet somebody."},
     {NULL, NULL, 0, NULL}
};

PyMODINIT_FUNC
initcplayground(void)
{
     (void) Py_InitModule("cplayground", HelloMethods);
}


cplayground.cpp:

#include <Python.h>

static PyObject*
say_hello(PyObject* self, PyObject* args)
{
    Py_RETURN_NONE;
}

static PyMethodDef HelloMethods[] =
{
     {"say_hello", say_hello, METH_VARARGS, "Greet somebody."},
     {NULL, NULL, 0, NULL}
};

PyMODINIT_FUNC
initcppplayground(void)
{
     (void) Py_InitModule("cppplayground", HelloMethods);
}


setup.py:

from distutils.core import setup
from distutils.extension import Extension
import Cython.Compiler.Options
from Cython.Distutils import build_ext
from os.path import join, sep, dirname, abspath

def get_extensions_from_sources():
    ext_modules = []
    ext_modules.append(Extension('src.cplayground', sources=['src/cplayground.c']))
    ext_modules.append(Extension('src.cppplayground', sources=['src/cplayground.cpp']))
    return ext_modules

setup(
    name='Playground',
    version='.1',
    author='Matthew Einhorn',
    ext_modules=get_extensions_from_sources(),
    cmdclass={'build_ext': build_ext},
    packages=['src']
      )


Here's a demonstration of the issue:

G:\Python\libs\Playground\src>python
Python 2.7.9 (default, Dec 10 2014, 12:24:55) [MSC v.1500 32 bit (Intel)] on win
32
Type "help", "copyright", "credits" or "license" for more information.
>>> import cplayground
>>> import cppplayground
>>> exit()

This application has requested the Runtime to terminate it in an unusual way.
Please contact the application's support team for more information.

G:\Python\libs\Playground\src>python
Python 2.7.9 (default, Dec 10 2014, 12:24:55) [MSC v.1500 32 bit (Intel)] on win
32
Type "help", "copyright", "credits" or "license" for more information.
>>> import cppplayground
>>> import cplayground
>>> exit()

Here's my config:

G:\Python\libs\Playground\src>gcc -v
Using built-in specs.
COLLECT_GCC=gcc
COLLECT_LTO_WRAPPER=g:/python/env/test/py279_x86/mingw/bin/../libexec/gcc/mingw3
2/4.8.1/lto-wrapper.exe
Target: mingw32
Configured with: ../gcc-4.8.1/configure --prefix=/mingw --host=mingw32 --build=m
ingw32 --without-pic --enable-shared --enable-static --with-gnu-ld --enable-lto
--enable-libssp --disable-multilib --enable-languages=c,c++,fortran,objc,obj-c++
,ada --disable-sjlj-exceptions --with-dwarf2 --disable-win32-registry --enable-l
ibstdcxx-debug --enable-version-specific-runtime-libs --with-gmp=/usr/src/pkg/gm
p-5.1.2-1-mingw32-src/bld --with-mpc=/usr/src/pkg/mpc-1.0.1-1-mingw32-src/bld --
with-mpfr= --with-system-zlib --with-gnu-as --enable-decimal-float=yes --enable-
libgomp --enable-threads --with-libiconv-prefix=/mingw32 --with-libintl-prefix=/
mingw --disable-bootstrap LDFLAGS=-s CFLAGS=-D_USE_32BIT_TIME_T
Thread model: win32
gcc version 4.8.1 (GCC)

G:\Python\libs\Playground\src>g++ -v
Using built-in specs.
COLLECT_GCC=g++
COLLECT_LTO_WRAPPER=g:/python/env/test/py279_x86/mingw/bin/../libexec/gcc/mingw3
2/4.8.1/lto-wrapper.exe
Target: mingw32
Configured with: ../gcc-4.8.1/configure --prefix=/mingw --host=mingw32 --build=m
ingw32 --without-pic --enable-shared --enable-static --with-gnu-ld --enable-lto
--enable-libssp --disable-multilib --enable-languages=c,c++,fortran,objc,obj-c++
,ada --disable-sjlj-exceptions --with-dwarf2 --disable-win32-registry --enable-l
ibstdcxx-debug --enable-version-specific-runtime-libs --with-gmp=/usr/src/pkg/gm
p-5.1.2-1-mingw32-src/bld --with-mpc=/usr/src/pkg/mpc-1.0.1-1-mingw32-src/bld --
with-mpfr= --with-system-zlib --with-gnu-as --enable-decimal-float=yes --enable-
libgomp --enable-threads --with-libiconv-prefix=/mingw32 --with-libintl-prefix=/
mingw --disable-bootstrap LDFLAGS=-s CFLAGS=-D_USE_32BIT_TIME_T
Thread model: win32
gcc version 4.8.1 (GCC)
msg239122 - (view) Author: R. David Murray (r.david.murray) * (Python committer) Date: 2015-03-24 14:04
Oh, I'm afraid it can still be ignored :).  What filing it here means is that it won't be *forgotten*, and hopefully there will eventually be someone with the time and interest to address it.
msg239208 - (view) Author: (matham) Date: 2015-03-25 01:30
Well, we won't have to wait until interested solves it. After hours of debugging including compiling Python with VS2013 and looking at windows dump files, it seems that hard work pays off - sorry, I mean asking someone who knows the answer pays off :P.

I asked on #mingw and it was suggested that I use mingw compiled with the SJLJ exception model. And indeed, using sjlj fixed this so that it does not crash anymore. It seems that mingw from mingw.org compiles with `"--disable-sjlj-exceptions`, so one has to use mingw-w64's distribution which offers a sjlj build (for x86 and SEH for x64).

I am wondering whether this information should be somewhere on the python website so that other people do not run into this issue. I'm not sure what the proper resolution is here so I'll live it open. Or maybe it should be open until the docs is updated to warn about this issue.

Anyway, here's the pertinent parts of my #mingw chat:

<jon_y> matham: I suggest going for SJLJ on 32bit
<jon_y> SEH on 64bit
<jon_y> avoid dw2 exception like the plague
<matham> ok, the one I got from mingw-get says "--disable-sjlj-exceptions --with-dwarf2"
<matham> I'll try with sjlj
<matham> from mingw-w64
<matham> jon_y, \o/\o/\o/\o/\o/\o/\o/ omg I'm so happy :D it looks like using sjlj fixed it
<jon_y> matham: I mean --disable-sjlj-exceptions
<jon_y> that is the C++ exception part
<jon_y> dwarf2 as debug is fine
<jon_y> --with-dwarf2 specifies the debug data format
<jon_y> yeah, dw2 (when --disable-sjlj-exceptions is used) has lots of issues with python
<matham> so the problem is when sjlj is disabled?
<matham> I thought it was either/or with sjlj vs dw2
<jon_y> dw2 doesn't quite work right for windows
<jon_y> it requires ALL dll code use the same shared libgcc dll
<jon_y> failure to do so leads to random crashes, especially during program termination
<matham> so why does mingw-get automatically do --disable-sjlj-exceptions?
<jon_y> mingw.org only provides toolchain that uses dw2 exceptions
<jon_y> this has been so ever since the 4.2.1 days
<jon_y> (that was the first widely used Windows gcc 4.x version)
<jon_y> matham: Cygwin had the same problem with python and their dw2 gcc recently
<jon_y> but had some local patches done in to fix it
msg239255 - (view) Author: R. David Murray (r.david.murray) * (Python committer) Date: 2015-03-25 13:49
Well, clearly you are the interested party that solved it :)  Thanks.

Unfortunately I don't know that we have a place where this could be documented, since currently mingw isn't a fully supported platform.  Perhaps in the distutils docs somewhere?
msg239265 - (view) Author: Steve Dower (steve.dower) * (Python committer) Date: 2015-03-25 15:53
I think these days as soon as a repro includes "import distutils" or "import setuptools" the docs go straight to http://pypa.io :)
msg239301 - (view) Author: Nick Coghlan (ncoghlan) * (Python committer) Date: 2015-03-26 02:19
pypa.io hosts the internal dev docs for PyPA, I believe you meant packaging.python.org, and that would indeed be the right home.

However, this kind of C/C++ linker related problem is way too advanced for the current state of the binary extension docs which basically say "C/C++ is hard relative to Python, let someone else deal with creating your extension modules for you": https://packaging.python.org/en/latest/extensions.html

For the scientific stack in particular we go even further and say to get someone else to build and link the extension modules for you: https://packaging.python.org/en/latest/science.html

Folks encountering the kind of problem described in this issue is thus mostly useful as a data point in *why* we give those instructions: because DIY cross-platform C/C++ development is a teeming morass of "fun" interoperability bugs like this one just waiting to be discovered :)
History
Date User Action Args
2015-03-26 02:19:18ncoghlansetmessages: + msg239301
2015-03-25 15:53:06steve.dowersetnosy: + ncoghlan
messages: + msg239265
2015-03-25 13:49:51r.david.murraysetnosy: + docs@python
messages: + msg239255

assignee: docs@python
components: + Documentation
type: crash ->
2015-03-25 01:30:37mathamsetmessages: + msg239208
2015-03-24 14:04:26r.david.murraysetnosy: + r.david.murray
messages: + msg239122
2015-03-23 05:10:34mathamcreate