The program stucks when redirecting std::cout to python in openmp parallel for loop

[znshje]

Hi,

I'm writing an algorithm in C++ with OpenMP parallel acceleration, and compile it using pybind11 as a python module. I make some output in the omp parallel for, and redirect the std output to python, but the program just stucked, and I need to send a SIGKILL to stop it.

The C++ code is below:

#include <pybind11/pybind11.h>
#include <pybind11/iostream.h>
#include <mutex>
namespace py = pybind11;

#define _CRT_SECURE_NO_DEPRECATE

#include <omp.h>
#include <iostream>
#include <string>

static std::mutex cout_mutex;

int run() {
#pragma omp parallel for
    for (int i = 0; i < 16; i++) {
//        std::cout << "index: " << i << std::endl;
        {
            // #HelpAppreciated: Work on iostream.h thread safety.
            // Without this lock, the clang ThreadSanitizer (tsan) reliably reports a
            // data race, and this test is predictably flakey on Windows.
            // For more background see the discussion under
            // https://github.com/pybind/pybind11/pull/2982 and
            // https://github.com/pybind/pybind11/pull/2995.
            const std::lock_guard<std::mutex> lock(cout_mutex);
            std::cout << "index: " << i << std::endl;
        }
    }
    return 0;
}

PYBIND11_MODULE(tr, m) {
    m.def("run", &run,
      py::call_guard<py::scoped_ostream_redirect, py::scoped_estream_redirect>());
}

And the python code to call the module is below:

import tr

if __name__ == '__main__':
    tr.run()

I haved tried to lock the std::cout with mutex and #pramga omp critical block, as mentioned in PR2982 PR2995, but failed. Nothing is printed in the console, and the program is stucked.

Is there any mistakes in my code? Or is there other ways to solve the problem?
Thanks.