Set kp_debug, kp_info, kp_warning and kp_error to py::none() when the program terminates.

[thinking-tower]

cc @unexploredtest, @axsaucedo, Superchunk. This should fix #230.

Description

The first instance of segfault occurs with a0f9af9, where

// main.cpp
- py::object kp_logger;
+ py::object kp_debug, kp_info, kp_warning, kp_error;

I suspect that the "attributes" of the Logger class (in this case,

kp_debug    = kp_logger.attr("debug");
kp_info     = kp_logger.attr("info");
kp_warning  = kp_logger.attr("warning");
kp_error    = kp_logger.attr("error");

) are cleaned up once when the program terminates and then the static global py::objects are cleaned up again, leading to segfault.

In PyTorch, there is a similar idea where they use static variables to "cache" py::object attributes. Here is a short summary:

// python_rpc_handler.cpp
void PythonRpcHandler::init() {
    // ...
    py::object rpcInternal = py::module::import(kInternalModule);   
    // ...
    pyRunFunction_ = getFunction(rpcInternal, "_run_function");    
    pySerialize_ = getFunction(rpcInternal, "serialize");    
    pyDeserialize_ = getFunction(rpcInternal, "deserialize");
}

py::object getFunction(const py::object& module, const char* name) {  
    py::object fn = module.attr(name);  
    TORCH_CHECK(      
              py::isinstance<py::function>(fn),
              "attribute ",   
              name,      
              " is not a function"
    );
    return fn;
}

In this PR from PyTorch, the fix was to set them to py:none during cleanup.

The pybind docs suggests using the atexit for cleanup.

It's unclear why this wasn't an issue in python3.8 and below.

Minimal Reproducible Example

• git clone https://github.com/pybind/python_example
• Change src/main.cpp to

#include <pybind11/pybind11.h>

py::object kp_debug;

PYBIND11_MODULE(kp, m) {
    py::module_ logging  = py::module_::import("logging");
    py::object kp_logger = logging.attr("getLogger")("kp");
    kp_debug             = kp_logger.attr("debug");
    /* // Add me to fix!
    auto atexit = py::module_::import("atexit");
    atexit.attr("register")(py::cpp_function([](){
        kp_debug = py::none();
        kp_info = py::none();
        kp_warning = py::none();
        kp_error = py::none();
    }));
    */
}

• python3.9 -m pip install .
• python3.9 -c "import kp"

[unexploredtest]

Neat! I used to have the same issue on my laptop and I can confirm that it fixed the problem.