Segmentation fault when starting Python interpreter by a A.dll function and then calling Python functions in B.dll

I made several Windows shared libraries (dlls) to hide an embed python under the hood.

• A.dll exports two C functions: start_python, stop_python which calls Py_Initialize and Py_Finalize respectively;
• B.dll exports a B_func which calls some cython function declared by cdef public.
• An application loads A.dll and B.dll by LoadLibrary and gets these exported functions, then the following codes cause segmentation fault:

A_start_python();  // in A.dll
B_func();          // in B.dll
A_stop_python();   // in A.dll

B_func will crash due to some NULL pointer access. It looks like the underlying cython function is called without a living python interpreter. Why?

The following works as expected:

B_start_python();  // in B.dll
B_func();          // in B.dll
B_stop_python();   // in B.dll

{
    A_start_python();
    A_stop_python();

    B_start_python();
    B_func();        // segmentation fault
    B_stop_python();
}

{
    B_start_python();
    B_stop_python();

    B_start_python();
    B_func();         // AttributeError: 'NoneType' object has no attribute 'current_thread'
    B_stop_python();
}

{
    B_start_python();
    B_func();
    B_stop_python();
}

{
    B_start_python();
    B_func();         // AttributeError: 'NoneType' object has no attribute 'current_thread'
    B_stop_python();
}

This is A Simple yet Full Demo which reproduces these problems. An embed python included, no other dependencies required. Its a CLion project which is managed by a CMakeLists.txt.

This is just the Source Codes without any binaries

Some implementation details:

    bool start_python() {
        int status = PyImport_AppendInittab("py_funcs", PyInit_py_funcs);
        if (status == -1)
            return false;
        Py_Initialize();
        auto m = PyImport_ImportModule("py_funcs");
        return bool(m);
    }

    void stop_python() {
        Py_FinalizeEx();
    }

    void call_py_func() {
        std::cout << "To call_py_func. is Python initialized: " << Py_IsInitialized() << std::endl;
        py_func();   // cython cdef public function
    }