Segmentation fault on import; even tests fail

[jjaraalm]

I'm trying to get pybind11 working on a new system. I have tried multiple versions of pybind11, along with both python 3.8.8 and a 3.9.0. No matter what I do, even simple examples cause a segmentation fault on import. This happens on the tests as well and they fail.

Example cmake outptut:

-- The CXX compiler identification is GNU 8.4.1
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Check for working CXX compiler: /usr/bin/c++ - skipped
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- pybind11 v2.7.1 
-- CMake 3.19.7
-- Found PythonInterp: /usr/licensed/anaconda3/2021.5/bin/python3.8 (found version "3.8.8") 
-- Found PythonLibs: /usr/licensed/anaconda3/2021.5/lib/libpython3.8.so
-- PYTHON 3.8.8
-- Performing Test HAS_FLTO
-- Performing Test HAS_FLTO - Success
-- pybind11::lto enabled
-- pybind11::thin_lto enabled
-- Setting tests build type to MinSizeRel as none was specified
-- Building tests with Eigen v3.3.4
-- Found Boost: /usr/include (found suitable version "1.66.0", minimum required is "1.56")  
-- Found pytest 6.2.3
-- Downloading catch v2.13.2...
-- Building interpreter tests using Catch v2.13.2
-- Looking for C++ include pthread.h
-- Looking for C++ include pthread.h - found
-- Performing Test CMAKE_HAVE_LIBC_PTHREAD
-- Performing Test CMAKE_HAVE_LIBC_PTHREAD - Failed
-- Looking for pthread_create in pthreads
-- Looking for pthread_create in pthreads - not found
-- Looking for pthread_create in pthread
-- Looking for pthread_create in pthread - found
-- Found Threads: TRUE  
-- pybind11 v2.7.1 
-- pybind11 v2.7.1 
-- Configuring done
-- Generating done
-- Build files have been written to: /foo/pybind11-2.7.1/build

Compile line

cd /foo/pybind11-2.7.1/build/tests && /usr/bin/c++ -DPYBIND11_TEST_BOOST -DPYBIND11_TEST_EIGEN -Dpybind11_tests_EXPORTS -I/foo/pybind11-2.7.1/include -isystem /usr/licensed/anaconda3/2021.5/include/python3.8 -isystem /usr/include/eigen3 -Os -DNDEBUG -fPIC -fvisibility=hidden -Wall -Wextra -Wconversion -Wcast-qual -Wdeprecated -Wundef -Wnon-virtual-dtor -flto -fno-fat-lto-objects -o CMakeFiles/pybind11_tests.dir/pybind11_tests.cpp.o -c /foo/pybind11-2.7.1/tests/pybind11_tests.cpp

Abbreviated link line

cd /foo/pybind11-2.7.1/build/tests && /usr/local/cmake/3.19.7/bin/cmake -E cmake_link_script CMakeFiles/pybind11_tests.dir/link.txt --verbose=1
/usr/bin/c++ -fPIC -Os -DNDEBUG -flto -shared  -o pybind11_tests.cpython-38-x86_64-linux-gnu.so CMakeFiles/pybind11_tests.dir/pybind11_tests.cpp.o ...  -lstdc++fs 

Test output

$ make check
[  4%] Built target cross_module_gil_utils
[  8%] Built target pybind11_cross_module_tests
[ 89%] Built target pybind11_tests
make[3]: *** [tests/CMakeFiles/pytest.dir/build.make:78: tests/CMakeFiles/pytest] Segmentation fault (core dumped)
make[2]: *** [CMakeFiles/Makefile2:364: tests/CMakeFiles/pytest.dir/all] Error 2
make[1]: *** [CMakeFiles/Makefile2:429: tests/CMakeFiles/check.dir/rule] Error 2
make: *** [Makefile:262: check] Error 2

Any help or advice on what the issue could be would be greatly appreciated.

[henryiii]

Are you using the compiler used for Python? If there's a pretty bad compiler mismatch, that could cause issues building extensions.

(We test on lots of systems, so don't see an obvious problem - the fact you seem to be able to pick Pythons makes that the most obvious thing that might have an issue)

[henryiii]

Also, you didn't say what sort of system you are on. What OS, what version of the OS, what compiler, where you got Python, etc.

[henryiii]

No, nevermind, I just saw the anaconda in the path. You can't use your system compiler with an anaconda Python to build an extension (pybind11 or anything else). You need to get the compiler (compilers?) package from conda-forge and use that.

[jjaraalm]

Thanks for the reply! I was actually able to figure this out recently, but did not update it here. This failed even with system python and not conda distributed.

The issue seems to be a bad compiler configuration on the system I was using. It is a RHEL8 cluster and there are two separate gcc modules setup by the admins. I was using the general gcc module, however switching to the RHEL gcc-toolset module fixed the problem. Since it is now working, I didn't investigate further.

[CharlesB2]

I had the same problem, posted as an issue in #3386 and #3380 (sorry for opening too many of them!)

I'd be willing to submit a PR in the documentation to warn users about this, maybe a new section about segmentation faults in https://pybind11.readthedocs.io/en/stable/advanced/misc.html ?