segfault for python importing vtk but not for vtkpython

Hi all,
I am building VTK with the latest conda-build v3 on mac. vtkpython seems to work fine, however importing vtk from python gives me a segfault. Any ideas? (See the attached information)

Python 3.7.1 | packaged by conda-forge | (default, Nov 13 2018, 10:30:07)
[Clang 4.0.1 (tags/RELEASE_401/final)] :: Anaconda, Inc. on darwin
Type “help”, “copyright”, “credits” or “license” for more information.
>>> import vtk
Segmentation fault: 11

The last 5 stack frames are:

  • frame #0: 0x0000000111486f07 libpython3.7m.dylibPyModule_Create2 + 7 frame #1: 0x00000001113f77a8 libvtkRenderingMatplotlibPython37D-8.2.1.dylibreal_initvtkRenderingMatplotlibPython + 24
    frame #2: 0x000000010019e19d python_PyImport_LoadDynamicModuleWithSpec + 557 frame #3: 0x000000010019d5ac python_imp_create_dynamic + 220
    frame #4: 0x000000010002e1a3 python_PyMethodDef_RawFastCallDict + 259 frame #5: 0x000000010002f871 pythonPyCFunction_Call + 65

It seems that the latest python from conda is build statically, which means there are going to be two version for the C++ library. Could this be a reason?

otool -L /Users/danlipsa/anaconda3/envs/test-vtk/bin/python
/Users/danlipsa/anaconda3/envs/test-vtk/bin/python:
/usr/lib/libSystem.B.dylib (compatibility version 1.0.0, current version 1197.1.1)

An older python from conda links with the following libraries (and that works fine for importing vtk)

(nightly) [~]$ otool -L ~/anaconda3/envs/nightly/bin/python
/Users/danlipsa/anaconda3/envs/nightly/bin/python:
/usr/lib/libc++.1.dylib (compatibility version 1.0.0, current version 120.0.0)
/System/Library/Frameworks/CoreFoundation.framework/Versions/A/CoreFoundation (compatibility version 150.0.0, current version 1153.18.0)
@rpath/libpython3.6m.dylib (compatibility version 3.6.0, current version 3.6.0)
/usr/lib/libSystem.B.dylib (compatibility version 1.0.0, current version 1213.0.0)
(nightly) [~]$ otool -L ~/anaconda3/envs/nightly/bin/python
/Users/danlipsa/anaconda3/envs/nightly/bin/python:
/usr/lib/libc++.1.dylib (compatibility version 1.0.0, current version 120.0.0)
/System/Library/Frameworks/CoreFoundation.framework/Versions/A/CoreFoundation (compatibility version 150.0.0, current version 1153.18.0)
@rpath/libpython3.6m.dylib (compatibility version 3.6.0, current version 3.6.0)
/usr/lib/libSystem.B.dylib (compatibility version 1.0.0, current version 1213.0.0)

Try importing smaller chunks. The fact that it failed on RenderingMatplotlib suggest that it managed to load some of the other modules:

import vtk.vtkmodules.vtkCommonCore
etc.

This might also have something to do with my new code for handling run-time dependencies between the python modules. For example, if you try importing vtkComonDataModel, the new code will try to import vtkCommonCore for you, and perhaps this works differently for python vs. vtkpython. The old (VTK 8.2) code relied on the system’s dynamic linker to handle these dependencies via the libvtkCommonCore36D.dylib etc. libraries which are no longer present in VTK.

Hi David,
Thanks for your answer. Strangely it fails in the same place even if I load what you suggested.

(lldb) r

Process 94986 launched: ‘/Users/danlipsa/anaconda3/envs/test/bin/python’ (x86_64)

Python 3.7.2 (default, Dec 29 2018, 00:00:04)

[Clang 4.0.1 (tags/RELEASE_401/final)] :: Anaconda, Inc. on darwin

Type “help”, “copyright”, “credits” or “license” for more information.

import vtk.vtkmodule.vtkCommonCore

Process 94986 stopped

  • thread #1, queue = ‘com.apple.main-thread’, stop reason = EXC_BAD_ACCESS (code=1, address=0x10)

    frame #0: 0x000000010f1af0f7 libpython3.7m.dylib`PyModule_Create2 + 7

libpython3.7m.dylib`PyModule_Create2:

-> 0x10f1af0f7 <+7>: movq 0x10(%rax), %rax

0x10f1af0fb <+11>: cmpq   $0x0, 0x28(%rax)

0x10f1af100 <+16>: je     0x10f1af107               ; <+23>

0x10f1af102 <+18>: jmp    0x10f1af120               ; _PyModule_CreateInitialized

Target 0: (python) stopped.

(lldb) up

frame #1: 0x000000010f11f7a8 libvtkRenderingMatplotlibPython37D-8.2.1.dylib`real_initvtkRenderingMatplotlibPython + 24

libvtkRenderingMatplotlibPython37D-8.2.1.dylib`real_initvtkRenderingMatplotlibPython:

0x10f11f7a8 <+24>: movq   %rax, %r14

0x10f11f7ab <+27>: movq   %r14, %rdi

0x10f11f7ae <+30>: callq  0x10f11f8be               ; symbol stub for: PyModule_GetDict

0x10f11f7b3 <+35>: movq   %rax, %rbx

(lldb) bt

  • thread #1, queue = ‘com.apple.main-thread’, stop reason = EXC_BAD_ACCESS (code=1, address=0x10)

    frame #0: 0x000000010f1af0f7 libpython3.7m.dylib`PyModule_Create2 + 7

    • frame #1: 0x000000010f11f7a8 libvtkRenderingMatplotlibPython37D-8.2.1.dylib`real_initvtkRenderingMatplotlibPython + 24

      frame #2: 0x000000010019a73d python`_PyImport_LoadDynamicModuleWithSpec + 557

      frame #3: 0x0000000100199b53 python`_imp_create_dynamic + 243

      frame #4: 0x000000010002e9d3 python`_PyMethodDef_RawFastCallDict + 259

      frame #5: 0x00000001000300a1 python`PyCFunction_Call + 65

I should have mentioned, I am building an older VTK, before the new module system

SHA

63f048ad7a9bd7915639a4d19d0e7e70487a2333

from Dec 10. This VTK worked fine with older conda packages, but they introduced a new way of compiling packages with an embedded compiler and this is causing these issues.

I chased a few mismatched libc++ (linked either from conda or from the system libraries). Are there any python bidings related libraries/headers I should worry about.

Thanks,

Dan

Dan, I made a horrid typo. Try this instead:

import vtkmodules.vtkCommonCore

I typed import vtk.vtkmodules.xxx which of course is going to import vtk before doing anything else.

David,

Yes, that works! Sorry for not catching this typo! I am not used to including only individual modules.

This is what I get:

(test) [~/anaconda3/conda-bld/vtk-cdat_1548382855937/work (context2d-contour-label-improvements %=)]$ python

Python 3.7.2 (default, Dec 29 2018, 00:00:04)

[Clang 4.0.1 (tags/RELEASE_401/final)] :: Anaconda, Inc. on darwin

Type “help”, “copyright”, “credits” or “license” for more information.

import vtkmodules.vtkCommonCore

import vtkmodules.vtkCommonDataModel

import vtkmodules.vtkCommonTransforms

import vtkmodules.vtkImagingCore

import vtkmodules.vtkPython

Traceback (most recent call last):

File “”, line 1, in

ModuleNotFoundError: No module named ‘vtkmodules.vtkPython’

import vtkmodules.vtkRenderingCore

import vtkmodules.vtkWrappingPythonCore

Traceback (most recent call last):

File “”, line 1, in

ModuleNotFoundError: No module named ‘vtkmodules.vtkWrappingPythonCore’

import vtkmodules.vtkRenderingFreeType

import vtkmodules.vtkPythonInterpreter

Traceback (most recent call last):

File “”, line 1, in

ModuleNotFoundError: No module named ‘vtkmodules.vtkPythonInterpreter’

import vtkmodules.vtkRenderingMatplotlib

Segmentation fault: 11

(test) [~/anaconda3/conda-bld/

All the “ModuleNotFound” errors are expected, since those aren’t wrapped modules.

So that leaves vtkRenderingMatplotlib, which is special because it links to libvtkPythonInterpreter, and is the only python-wrapped module to do so.

I suspect that you are correct that static linking is the culprit. Really vtkPythonInterpreter is the guts of the vtkpython executable, so when you load vtkRenderingMatplotlib into regular statically-linked ‘python’ you end up with two copies of libpython: the one that python links to, and the one that vtkRenderingMatplotlib.so links to. When you use vtkpython there’s no problem, because both vtkpython and vtkRenderingMatplotlib.so get libpython via libvtkPythonInterpreter.dylib, i.e. only one copy of libpython is used.

The easiest fix is to disable vtkRenderingMatplotlib. I think the only other option is to remove its dependence on vtkPythonInterpreter.

Since yesterday, I dug some more and found a few more details. I initially thought that just using vtkpython instead of python would solve our problem.

All tests pass fine, but we just realized that ipython would not work using vtkpython instead of python?

Do you know why that would be the case? I thought python and vtkpython are very similar: just some paths added to find VTK.

I am getting:

Python 3.6.7 | packaged by conda-forge | (default, Nov 20 2018, 18:39:03)

Type ‘copyright’, ‘credits’ or ‘license’ for more information

IPython 7.2.0 – An enhanced Interactive Python. Type ‘?’ for help.

In [1]: import sys

File “”, line 1

import

      ^

SyntaxError: invalid syntax

In [2]: ---------------------------------------------------------------------------

NameError Traceback (most recent call last)

in

----> 1 sys

NameError: name ‘sys’ is not defined

In [3]:

All the “ModuleNotFound” errors are expected, since those aren’t wrapped modules.

So that leaves vtkRenderingMatplotlib, which is special because it links to libvtkPythonInterpreter, and is the only python-wrapped module to do so.

I suspect that you are correct that static linking is the culprit. Really vtkPythonInterpreter is the guts of the vtkpython executable, so when you load vtkRenderingMatplotlib into regular statically-linked ‘python’ you end up with two copies of libpython: the one that python links to, and the one that vtkRenderingMatplotlib.so links to. When you use vtkpython there’s no problem, because both vtkpython and vtkRenderingMatplotlib.so get libpython via libvtkPythonInterpreter.dylib, i.e. only one copy of libpython is used.

Yes, that seems to be the case. Interestingly, python2 links dynamically with libpython.

The same happens on Linux (dynamic for python2 and static for python3), however VTK mathplotlib works fine on linux python3.

For python2 it works on both linux and mac.

The easiest fix is to disable vtkRenderingMatplotlib.

Unfortunately we would loose math text which we would like to keep.

I think the only other option is to remove its dependence on vtkPythonInterpreter.

Not sure how would this work. Can you describe this in more detail?

Thank you very much for your insights!

Dan

I’m actually not sure how vtkPythonInterpreter could be removed from vtkRenderingMatplotlib. Allison Vacanti might know of a way, but in any case it wouldn’t be easy and might not be possible.

A third option is to link weakly to libpython, which basically means that all Python API symbols would end up linking to the running python interpreter. This isn’t easy either, though (if it was easy we already would have done it, see https://gitlab.kitware.com/vtk/vtk/issues/17214).

Actually, VTK might already have some of the magic needed for weak linking. For example, the module.cmake for Utilities/PythonInterpreter has the OPTIONAL_PYTHON_LINK option, but the module.cmake for Rendering/Matplotlib does not. Perhaps when OPTIONAL_PYTHON_LINK was added to VTK, the Matplotlib module was simply overlooked? The relevant MR is here:

https://gitlab.kitware.com/vtk/vtk/merge_requests/1713

Thanks for the great pointers David. We’ll look into it.

Dan

Hi David,

We were able to temporarily fix our problem by using OPTIONAL_PYTHON_LINK and removing libpython from the link command line for

vtkRenderingMatplotlib, vtkRenderingMatplotlibPython, vtkPythonInterpreter

We are going to look into adding this to VTK master. For now, I built those libraries by hand. Also, likely, there is another step needed ( = loading libpython if none is provided) to make things work from both C++ and python. With the current solution they only work for python - which is what we care about in our project.

We’ll post an update once we have a fix for VTK.

Thanks,

Dan

Hi all,
We have an update on VTK_PYTHON_OPTIONAL_LINK.

To summarize the issue. Recently conda linked python3 statically, so all python symbols are included in the executable instead of being brought in by libpython. This created a problem with VTK used from python, because VTK links with libpython (it uses matplotlib for math text). So, you had python code brought in by the pyhton executable and by libpyton which resulted in a segfault for tests that used python and math text.

In the new module system, with Ben’s work, there is an option called VTK_PYTHON_OPTIONAL_LINK that is disabled by default. Patching VTK to enable this option and using cmake >= 3.13 would enable this option and compile VTK so that it works correctly with a python statically linked with libpython.

The reason why this option is still disable is that it does not work correctly with ParaView libraries because (we believe) more cleanup is needed there.

However, I used the patch VTK in our project using conda and statically linked python3 it works correctly.

Dan