call Python callback from a non-python thread in a C extension module (in response to zed's demo, https://gist.github.com/3471699, and this question http://stackoverflow.com/questions/12116390/python-code-calls-c-library)

c_extension_module.c

#include "Python.h" // should be before any standard headers
#include <errno.h>
#include <pthread.h>

static void*
non_python_thread(void *python_callback) {
  //XXX without PyEval_InitThreads() it produces:
  // Fatal Python error: PyEval_SaveThread: NULL tstate
  //XXX with PyEval_InitThreads() it deadlocks!!
  //XXX without PyGILState_*() (with/without PyEval_InitThreads()) it produces:
  // Exception KeyError: KeyError(139794956588800,)
  // in <module 'threading' from '/usr/lib/python2.7/threading.pyc'> ignored
  PyGILState_STATE state = PyGILState_Ensure();
  
  // print something
  PyObject* stdout = PySys_GetObject("stdout");
  if (stdout == NULL) {
    PyErr_SetString(PyExc_RuntimeError,  "no stdout");
    goto done;
  }
  Py_INCREF(stdout);
  if (PyFile_WriteString("non-python thread\n", stdout) != 0) {
    // a write error
    Py_DECREF(stdout);
    goto done;
  }
  else {
#if PY_MAJOR_VERSION < 3
      PyFile_SoftSpace(stdout, 0); // softspace = False after a newline
#endif
    Py_DECREF(stdout);
  }

  // call python callback
  PyObject* ignored = PyObject_CallFunctionObjArgs(python_callback, NULL);
  Py_XDECREF(ignored); // `ignored` may be NULL
  // fall through
 done:
  PyGILState_Release(state);
  return NULL; //NOTE: it doesn't propagate exceptions
}

static PyObject *
spawn_non_python_thread(PyObject *self, PyObject *args) {
 /* it should be safe to call PyEval_InitThreads(): current function holds
    GIL or it is started from the main thread */
  //XXX see the above comments near PyGILState_Ensure
  
  // I would suggest calling this in the module initializer, rather than on a
  // per-function basis. It is a matter of efficiency, not safety, though.
  // -- by LIU Yu <liuyu@openjudge.net>
  //// PyEval_InitThreads();
  
  PyObject* ret = NULL;
  PyObject* python_callback = NULL;
  if (!PyArg_ParseTuple(args, "O:spawn_non_python_thread", &python_callback))
    return NULL; // propagate exception
  Py_INCREF(python_callback); // hold on to it until the thread is finished
  
  // Give up GIL when doing C thread creation, this prevents the deadlock
  // caused by double-locking the GIL.
  // -- by LIU Yu <liuyu@openjudge.net>
  Py_BEGIN_ALLOW_THREADS
  
  pthread_t tid;
  int s = 0;
  // start a single non-python thread
  if ((s = pthread_create(&tid, NULL, non_python_thread, python_callback))!= 0) {
    errno = s; // errno!!
    // Get back GIL so that we can safely call Python API's
    // -- by LIU Yu <liuyu@openjudge.net>
    Py_BLOCK_THREADS
    PyErr_SetFromErrno(PyExc_OSError); goto done;
  }
  // join it
  else if ((s = pthread_join(tid, NULL)) != 0) {
    errno = s; // errno!!
    Py_BLOCK_THREADS
    PyErr_SetFromErrno(PyExc_OSError); goto done;
  }
  // check whether non_python_thread raised any Python exceptions
  else {
    Py_BLOCK_THREADS
    if (PyErr_Occurred() == NULL) {  // no error
      ret = Py_None; Py_INCREF(ret);
    }
  }
  // fall through
 done:
  Py_UNBLOCK_THREADS
  Py_END_ALLOW_THREADS
  Py_DECREF(python_callback);
  return ret;
}

static PyMethodDef
module_functions[] = {
    { "spawn_non_python_thread", spawn_non_python_thread,
      METH_VARARGS, "func docstring" },
    { NULL }
};

// http://python3porting.com/cextensions.html
#if PY_MAJOR_VERSION >= 3
  #define MOD_ERROR_VAL NULL
  #define MOD_SUCCESS_VAL(val) val
  #define MOD_INIT(name) PyMODINIT_FUNC PyInit_##name(void)
  #define MOD_DEF(ob, name, doc, methods) \
          static struct PyModuleDef moduledef = { \
            PyModuleDef_HEAD_INIT, name, doc, -1, methods, }; \
          ob = PyModule_Create(&moduledef);
#else
  #define MOD_ERROR_VAL
  #define MOD_SUCCESS_VAL(val)
  #define MOD_INIT(name) void init##name(void)
  #define MOD_DEF(ob, name, doc, methods) \
          ob = Py_InitModule3(name, methods, doc);
#endif

MOD_INIT(c_extension_module)
{
    PyObject *m = NULL;

    MOD_DEF(m, "c_extension_module", "module docstring",  module_functions)

    if (m == NULL)
        return MOD_ERROR_VAL;
    
    // The decref will not destroy the imported module because another reference
    // reference is held by sys.module until the interpreter exits. But it seems
    // PY3K is planning to add module unload support. I have no idea what will
    // happen by that time.
    // -- by LIU Yu <liuyu@openjudge.net>
    PyObject * mod = PyImport_ImportModule("threading");
    Py_XDECREF(mod);
    
    PyEval_InitThreads();
    
    return MOD_SUCCESS_VAL(m);
}

main.py

import c_extension_module

def python_callback():
    print("python callback called")
    def python_thread():
        print("python thread started")
        print("python thread ended")
    # `threading` might implicitly call PyEval_InitThreads()
    import threading
    threading.Thread(target=python_thread).start()

# there should be no GIL yet (we have only one main thread)
c_extension_module.spawn_non_python_thread(python_callback)

# if there is no KeyError in threading module; add time.sleep() to reproduce it:
####import time; time.sleep(1e-3)
#NOTE: if there is no new thread in python_callback or if there are
#multiple Python threads before calling the extension module then
#there is no error
print("exit main thread")

We think the same. :) I've already tried import threading and Py_BEGIN_ALLOW_THREADS before seeing your comment.

yes import threading fixes the KeyError originated here.

But imagine you don't see the definition of python_callback(). You don't know whether it uses threads or not (same for c_extension_module). Why would you add import threading if none of your code uses threading? Either python_callback(), c_extension_module are incorrect or it is a bug in Python.

---

The docs say:

Ensure that the current thread is ready to call the Python C API regardless of the current state of Python, or of the global interpreter lock.

It seems it is a documentation bug in Python because you can't call PyGILState_Ensure() "regardless of the current state of Python, or of the global interpreter lock". Though in this case it is natural to release GIL using Py_BEGIN_ALLOW_THREADS.

---

ceval.h seems to have an incorrect comment:

The function PyEval_InitThreads() should be called only from
init_thread() in "_threadmodule.c".

Looking at the definition of PyEval_InitThreads() it should be safe to call it from a thread that hold GIL or if there is no GIL yet (then current thread will be the main thread).

---

btw, I can't reproduce KeyError using openmp threads though there is a rare .AttributeError'>: 'NoneType' object has no attribute 'write' exception on the interpreter shutdown on Python 2.x.

I tracked the KeyError issue into the standard threading module. And I am pretty sure there exists an implementation fault.

Here is roughly the cause.

The threading module internally maintains an _active[] dictionary of all threads (both Python threads and OS threads are tracked), i.e. so as to correctly cleanup module states when os.fork() is called.

Problem is, 1) when the threading module is NOT imported by the MainThread, and 2) when the MainThread exits so soon such that it is never scheduled (i.e. put into sleep mode and wakeup again), then the threading module will NOT be aware of the MainThread until the interpreter exits.

However, upon the exit of the interpreter, each active thread will delete its own identity from this _active[] dictionary -- including the MainThread. And the KeyError actually complained about the missing of the MainThread's identity, that is, when the MainThread removes itself from the _active[] dictionary.

I came up with another workaround to get rid of the KeyError problem. We can import the threading module, i.e. PyImport_ImportModule("threading"), in the initializer of c_extension_module. This can enforce the MainThread to be correctly logged by the threading module.