Embedding CPython: how do you constuct Python call

2020-07-26 14:48发布

问题:

Suppose I am embedding the CPython interpreter into a larger program, written in C. The C component of the program occasionally needs to call functions written in Python, supplying callback functions to them as arguments.

Using the CPython extending and embedding APIs, how do I construct a Python "callable" object that wraps a C pointer-to-function, so that I can pass that object to Python code and have the Python code successfully call back into the C code?

Note: this is a revised version of a question originally posted by user dhanasubbu, which I answered, but which was then deleted. I think it was actually a good question, so I have converted what I wrote into a self-answer to my own statement of the question. Alternative answers are welcome.

回答1:

To define an extension type that is “callable” in the sense Python uses that term, you fill the tp_call slot of the type object, which is the C equivalent of the __call__ special method. The function that goes in that slot will be a glue routine that calls the actual C callback. Here’s code for the simplest case, when the C callback takes no arguments and returns nothing.

typedef struct {
    PyObject_HEAD
    /* Type-specific fields go here. */
    void (*cfun)(void);  /* or whatever parameters it actually takes */
} CallbackObj;

static PyObject *Callback_call(PyObject *self, PyObject *args, PyObject *kw)
{
    /* check that no arguments were passed */
    const char no_kwargs[] = { 0 };
    if (!PyArg_ParseTupleAndKeywords(args, kw, "", no_kwargs))
        return 0;

    CallbackObj *cself = (CallbackObj *)self;
    cself->cfun();
    Py_RETURN_NONE;
}

static PyTypeObject CallbackType = {
    PyVarObject_HEAD_INIT(NULL, 0)
    .tp_name = "mymodule.Callback",
    .tp_doc = "Callback function passed to foo, bar, and baz.",
    .tp_basicsize = sizeof(CallbackObj),
    .tp_itemsize = 0,
    .tp_flags = Py_TPFLAGS_DEFAULT,
    .tp_new = PyType_GenericNew,
    .tp_call = Callback_call,
};

Instantiate the type object with PyType_Ready as usual. Don’t put it in any module visible to Python, though, because Python code can’t correctly create instances of this type. (Because of this, I haven’t bothered with a tp_init function; just make sure you always initialize ->cfun after creating instances from C, or Callback_call will crash.)

Now, suppose the actual function you need to call is named real_callback, and the Python function that you want to pass it to is named function_to_call. First you create one of the callback objects, by calling the type object, as usual, and initialize its ->cfun field:

 PyObject *args = PyTuple_New(0);
 CallbackObj *cb = (CallbackObj *)PyObject_CallObject(
     (PyObject *)CallbackType, args);
 Py_DECREF(args);
 cb->cfun = real_callback;

Then you put cb into an argument tuple, and call the Python function object with that, as usual.

 args = Py_BuildValue("(O)", cb);
 PyObject *ret = PyObject_CallObject(function_to_call, args);
 Py_DECREF(args);
 Py_DECREF(cb);
 // do stuff with ret, here, perhaps
 Py_DECREF(ret);

Extending this to more complex cases, where the C callback needs to take arguments and/or return values and/or raise Python exceptions on error and/or receive “closure” information from the outer context, is left as an exercise.



回答2:

I'd be tempted to use the standard library ctypes module since it already contains appropriate wrappers for C function pointers, and can automatically deal with conversions from Python types to C types for wide variety of arguments.

I've written a working example in Cython since it's an easy way of mixing Python and C, but it should show how to use these objects:

cdef extern from *:
    """
    int f(int x) {
       return x*2;
    }
    """
    int f(int f)

I define an example function f (in the docstring, which Cython incorporates directly into the compiled file).

import ctypes
from libc.stdint cimport intptr_t

def make_f_wrapper():
    func_type = ctypes.CFUNCTYPE(ctypes.c_int, ctypes.c_int)
    cdef intptr_t f_ptr = <intptr_t>&f
    return func_type(f_ptr)

This is a Cython (pretty close to Python syntax) version of the creation of a ctypes pointer to f. I define the arguments of the function, convert the f pointer to an appropriately sized integer, then initialize the wrapper object with that integer.

cdef extern from *:
    """
    PyObject* make_f_wrapper_c_impl() {
        PyObject *ctypes_module = NULL, *CFUNCTYPE = NULL, *c_int = NULL, 
                 *func_type = NULL, *ptr_value = NULL, *result = NULL;

        ctypes_module = PyImport_ImportModule("ctypes");
        if (ctypes_module == NULL) goto cleanup;
        CFUNCTYPE = PyObject_GetAttrString(ctypes_module,"CFUNCTYPE");
        if (CFUNCTYPE == NULL) goto cleanup;
        c_int = PyObject_GetAttrString(ctypes_module,"c_int");
        if (c_int == NULL) goto cleanup;
        func_type = PyObject_CallFunctionObjArgs(CFUNCTYPE,c_int,c_int,NULL);
        if (func_type == NULL) goto cleanup;
        ptr_value = PyLong_FromVoidPtr(&f);
        if (ptr_value == NULL) goto cleanup;
        result = PyObject_CallFunctionObjArgs(func_type,ptr_value,NULL);

        cleanup:
        Py_XDECREF(ptr_value);
        Py_XDECREF(func_type);
        Py_XDECREF(c_int);
        Py_XDECREF(CFUNCTYPE);
        Py_XDECREF(ctypes_module);
        return result;
    }
    """
    object make_f_wrapper_c_impl()

def make_f_wrapper_c():
    return make_f_wrapper_c_impl()

The code above is a C translation of the Pythony code above - it does exactly the same thing but is a bit more convoluted since it uses the C API. It just uses the ctypes module through its Python interface. Once again the C code is embedded in a Cython file through a docstring; however similar code could be used in a directly written C API module.

(All these Cython snippets combine to form one long working example)