Extending Python with C or C++

It is quite easy to add new built-in modules to Python, if you know how to program in C. Such extension modules can do two things that can’t be done directly in Python: they can implement new built-in object types, and they can call C library functions and system calls.

To support extensions, the Python API (Application Programmers Interface) defines a set of functions, macros and variables that provide access to most aspects of the Python run-time system. The Python API is incorporated in a C source file by including the header "Python.h".

The compilation of an extension module depends on its intended use as well as on your system setup; details are given in later chapters.

A Simple Example

Let’s create an extension module called spam (the favorite food of Monty Python fans...) and let’s say we want to create a Python interface to the C library function system. [1] This function takes a null-terminated character string as argument and returns an integer. We want this function to be callable from Python as follows:

>>> import spam
>>> status = spam.system("ls -l")

Begin by creating a file spammodule.c. (Historically, if a module is called spam, the C file containing its implementation is called spammodule.c; if the module name is very long, like spammify, the module name can be just spammify.c.)

The first line of our file can be:

#include <Python.h>

which pulls in the Python API (you can add a comment describing the purpose of the module and a copyright notice if you like).


Since Python may define some pre-processor definitions which affect the standard headers on some systems, you must include Python.h before any standard headers are included.

All user-visible symbols defined by Python.h have a prefix of Py or PY, except those defined in standard header files. For convenience, and since they are used extensively by the Python interpreter, "Python.h" includes a few standard header files: <stdio.h>, <string.h>, <errno.h>, and <stdlib.h>. If the latter header file does not exist on your system, it declares the functions malloc, free and realloc directly.

The next thing we add to our module file is the C function that will be called when the Python expression spam.system(string) is evaluated (we’ll see shortly how it ends up being called):

static PyObject *
spam_system(PyObject *self, PyObject *args)
    const char *command;
    int sts;

    if (!PyArg_ParseTuple(args, "s", &command))
        return NULL;
    sts = system(command);
    return Py_BuildValue("i", sts);

There is a straightforward translation from the argument list in Python (for example, the single expression "ls -l") to the arguments passed to the C function. The C function always has two arguments, conventionally named self and args.

The self argument is only used when the C function implements a built-in method, not a function. In the example, self will always be a NULL pointer, since we are defining a function, not a method. (This is done so that the interpreter doesn’t have to understand two different types of C functions.)

The args argument will be a pointer to a Python tuple object containing the arguments. Each item of the tuple corresponds to an argument in the call’s argument list. The arguments are Python objects — in order to do anything with them in our C function we have to convert them to C values. The function PyArg_ParseTuple in the Python API checks the argument types and converts them to C values. It uses a template string to determine the required types of the arguments as well as the types of the C variables into which to store the converted values. More about this later.

PyArg_ParseTuple returns true (nonzero) if all arguments have the right type and its components have been stored in the variables whose addresses are passed. It returns false (zero) if an invalid argument list was passed. In the latter case it also raises an appropriate exception so the calling function can return NULL immediately (as we saw in the example).

Intermezzo: Errors and Exceptions

An important convention throughout the Python interpreter is the following: when a function fails, it should set an exception condition and return an error value (usually a NULL pointer). Exceptions are stored in a static global variable inside the interpreter; if this variable is NULL no exception has occurred. A second global variable stores the “associated value” of the exception (the second argument to raise). A third variable contains the stack traceback in case the error originated in Python code. These three variables are the C equivalents of the result in Python of sys.exc_info() (see the section on module sys in the Python Library Reference). It is important to know about them to understand how errors are passed around.

The Python API defines a number of functions to set various types of exceptions.

The most common one is PyErr_SetString. Its arguments are an exception object and a C string. The exception object is usually a predefined object like PyExc_ZeroDivisionError. The C string indicates the cause of the error and is converted to a Python string object and stored as the “associated value” of the exception.

Another useful function is PyErr_SetFromErrno, which only takes an exception argument and constructs the associated value by inspection of the global variable errno. The most general function is PyErr_SetObject, which takes two object arguments, the exception and its associated value. You don’t need to Py_INCREF the objects passed to any of these functions.

You can test non-destructively whether an exception has been set with PyErr_Occurred. This returns the current exception object, or NULL if no exception has occurred. You normally don’t need to call PyErr_Occurred to see whether an error occurred in a function call, since you should be able to tell from the return value.

When a function f that calls another function g detects that the latter fails, f should itself return an error value (usually NULL or -1). It should not call one of the PyErr_* functions — one has already been called by g. f‘s caller is then supposed to also return an error indication to its caller, again without calling PyErr_*, and so on — the most detailed cause of the error was already reported by the function that first detected it. Once the error reaches the Python interpreter’s main loop, this aborts the currently executing Python code and tries to find an exception handler specified by the Python programmer.

(There are situations where a module can actually give a more detailed error message by calling another PyErr_* function, and in such cases it is fine to do so. As a general rule, however, this is not necessary, and can cause information about the cause of the error to be lost: most operations can fail for a variety of reasons.)

To ignore an exception set by a function call that failed, the exception condition must be cleared explicitly by calling PyErr_Clear. The only time C code should call PyErr_Clear is if it doesn’t want to pass the error on to the interpreter but wants to handle it completely by itself (possibly by trying something else, or pretending nothing went wrong).

Every failing malloc call must be turned into an exception — the direct caller of malloc (or realloc) must call PyErr_NoMemory and return a failure indicator itself. All the object-creating functions (for example, PyLong_FromLong) already do this, so this note is only relevant to those who call malloc directly.

Also note that, with the important exception of PyArg_ParseTuple and friends, functions that return an integer status usually return a positive value or zero for success and -1 for failure, like Unix system calls.

Finally, be careful to clean up garbage (by making Py_XDECREF or Py_DECREF calls for objects you have already created) when you return an error indicator!

The choice of which exception to raise is entirely yours. There are predeclared C objects corresponding to all built-in Python exceptions, such as PyExc_ZeroDivisionError, which you can use directly. Of course, you should choose exceptions wisely — don’t use PyExc_TypeError to mean that a file couldn’t be opened (that should probably be PyExc_IOError). If something’s wrong with the argument list, the PyArg_ParseTuple function usually raises PyExc_TypeError. If you have an argument whose value must be in a particular range or must satisfy other conditions, PyExc_ValueError is appropriate.

You can also define a new exception that is unique to your module. For this, you usually declare a static object variable at the beginning of your file:

static PyObject *SpamError;

and initialize it in your module’s initialization function (PyInit_spam) with an exception object (leaving out the error checking for now):

    PyObject *m;

    m = PyModule_Create(&spammodule);
    if (m == NULL)
        return NULL;

    SpamError = PyErr_NewException("spam.error", NULL, NULL);
    PyModule_AddObject(m, "error", SpamError);
    return m;

Note that the Python name for the exception object is spam.error. The PyErr_NewException function may create a class with the base class being Exception (unless another class is passed in instead of NULL), described in Built-in Exceptions.

Note also that the SpamError variable retains a reference to the newly created exception class; this is intentional! Since the exception could be removed from the module by external code, an owned reference to the class is needed to ensure that it will not be discarded, causing SpamError to become a dangling pointer. Should it become a dangling pointer, C code which raises the exception could cause a core dump or other unintended side effects.

We discuss the use of PyMODINIT_FUNC as a function return type later in this sample.

Back to the Example

Going back to our example function, you should now be able to understand this statement:

if (!PyArg_ParseTuple(args, "s", &command))
    return NULL;

It returns NULL (the error indicator for functions returning object pointers) if an error is detected in the argument list, relying on the exception set by PyArg_ParseTuple. Otherwise the string value of the argument has been copied to the local variable command. This is a pointer assignment and you are not supposed to modify the string to which it points (so in Standard C, the variable command should properly be declared as const char *command).

The next statement is a call to the Unix function system, passing it the string we just got from PyArg_ParseTuple:

sts = system(command);

Our spam.system() function must return the value of sts as a Python object. This is done using the function Py_BuildValue, which is something like the inverse of PyArg_ParseTuple: it takes a format string and an arbitrary number of C values, and returns a new Python object. More info on Py_BuildValue is given later.

return Py_BuildValue("i", sts);

In this case, it will return an integer object. (Yes, even integers are objects on the heap in Python!)

If you have a C function that returns no useful argument (a function returning void), the corresponding Python function must return None. You need this idiom to do so (which is implemented by the Py_RETURN_NONE macro):

return Py_None;

Py_None is the C name for the special Python object None. It is a genuine Python object rather than a NULL pointer, which means “error” in most contexts, as we have seen.

The Module’s Method Table and Initialization Function

I promised to show how spam_system is called from Python programs. First, we need to list its name and address in a “method table”:

static PyMethodDef SpamMethods[] = {
    {"system",  spam_system, METH_VARARGS,
     "Execute a shell command."},
    {NULL, NULL, 0, NULL}        /* Sentinel */

Note the third entry (METH_VARARGS). This is a flag telling the interpreter the calling convention to be used for the C function. It should normally always be METH_VARARGS or METH_VARARGS | METH_KEYWORDS; a value of 0 means that an obsolete variant of PyArg_ParseTuple is used.

When using only METH_VARARGS, the function should expect the Python-level parameters to be passed in as a tuple acceptable for parsing via PyArg_ParseTuple; more information on this function is provided below.

The METH_KEYWORDS bit may be set in the third field if keyword arguments should be passed to the function. In this case, the C function should accept a third PyObject \* parameter which will be a dictionary of keywords. Use PyArg_ParseTupleAndKeywords to parse the arguments to such a function.

The method table must be referenced in the module definition structure:

static struct PyModuleDef spammodule = {
   "spam",   /* name of module */
   spam_doc, /* module documentation, may be NULL */
   -1,       /* size of per-interpreter state of the module,
                or -1 if the module keeps state in global variables. */

This structure, in turn, must be passed to the interpreter in the module’s initialization function. The initialization function must be named PyInit_name, where name is the name of the module, and should be the only non-static item defined in the module file:

    return PyModule_Create(&spammodule);

Note that PyMODINIT_FUNC declares the function as PyObject * return type, declares any special linkage declarations required by the platform, and for C++ declares the function as extern "C".

When the Python program imports module spam for the first time, PyInit_spam is called. (See below for comments about embedding Python.) It calls PyModule_Create, which returns a module object, and inserts built-in function objects into the newly created module based upon the table (an array of PyMethodDef structures) found in the module definition. PyModule_Create returns a pointer to the module object that it creates. It may abort with a fatal error for certain errors, or return NULL if the module could not be initialized satisfactorily. The init function must return the module object to its caller, so that it then gets inserted into sys.modules.

When embedding Python, the PyInit_spam function is not called automatically unless there’s an entry in the PyImport_Inittab table. To add the module to the initialization table, use PyImport_AppendInittab, optionally followed by an import of the module:

main(int argc, char *argv[])
    /* Add a builtin module, before Py_Initialize */
    PyImport_AppendInittab("spam", PyInit_spam);

    /* Pass argv[0] to the Python interpreter */

    /* Initialize the Python interpreter.  Required. */

    /* Optionally import the module; alternatively,
       import can be deferred until the embedded script
       imports it. */

An example may be found in the file Demo/embed/demo.c in the Python source distribution.


Removing entries from sys.modules or importing compiled modules into multiple interpreters within a process (or following a fork without an intervening exec) can create problems for some extension modules. Extension module authors should exercise caution when initializing internal data structures.

A more substantial example module is included in the Python source distribution as Modules/xxmodule.c. This file may be used as a template or simply read as an example. The modulator.py script included in the source distribution or Windows install provides a simple graphical user interface for declaring the functions and objects which a module should implement, and can generate a template which can be filled in. The script lives in the Tools/modulator/ directory; see the README file there for more information.

Compilation and Linkage

There are two more things to do before you can use your new extension: compiling and linking it with the Python system. If you use dynamic loading, the details may depend on the style of dynamic loading your system uses; see the chapters about building extension modules (chapter Building C and C++ Extensions with distutils) and additional information that pertains only to building on Windows (chapter Building C and C++ Extensions on Windows) for more information about this.

If you can’t use dynamic loading, or if you want to make your module a permanent part of the Python interpreter, you will have to change the configuration setup and rebuild the interpreter. Luckily, this is very simple on Unix: just place your file (spammodule.c for example) in the Modules/ directory of an unpacked source distribution, add a line to the file Modules/Setup.local describing your file:

spam spammodule.o

and rebuild the interpreter by running make in the toplevel directory. You can also run make in the Modules/ subdirectory, but then you must first rebuild Makefile there by running ‘make Makefile’. (This is necessary each time you change the Setup file.)

If your module requires additional libraries to link with, these can be listed on the line in the configuration file as well, for instance:

spam spammodule.o -lX11

Calling Python Functions from C

So far we have concentrated on making C functions callable from Python. The reverse is also useful: calling Python functions from C. This is especially the case for libraries that support so-called “callback” functions. If a C interface makes use of callbacks, the equivalent Python often needs to provide a callback mechanism to the Python programmer; the implementation will require calling the Python callback functions from a C callback. Other uses are also imaginable.

Fortunately, the Python interpreter is easily called recursively, and there is a standard interface to call a Python function. (I won’t dwell on how to call the Python parser with a particular string as input — if you’re interested, have a look at the implementation of the -c command line option in Modules/main.c from the Python source code.)

Calling a Python function is easy. First, the Python program must somehow pass you the Python function object. You should provide a function (or some other interface) to do this. When this function is called, save a pointer to the Python function object (be careful to Py_INCREF it!) in a global variable — or wherever you see fit. For example, the following function might be part of a module definition:

static PyObject *my_callback = NULL;

static PyObject *
my_set_callback(PyObject *dummy, PyObject *args)
    PyObject *result = NULL;
    PyObject *temp;

    if (PyArg_ParseTuple(args, "O:set_callback", &temp)) {
        if (!PyCallable_Check(temp)) {
            PyErr_SetString(PyExc_TypeError, "parameter must be callable");
            return NULL;
        Py_XINCREF(temp);         /* Add a reference to new callback */
        Py_XDECREF(my_callback);  /* Dispose of previous callback */
        my_callback = temp;       /* Remember new callback */
        /* Boilerplate to return "None" */
        result = Py_None;
    return result;

This function must be registered with the interpreter using the METH_VARARGS flag; this is described in section The Module’s Method Table and Initialization Function. The PyArg_ParseTuple function and its arguments are documented in section Extracting Parameters in Extension Functions.

The macros Py_XINCREF and Py_XDECREF increment/decrement the reference count of an object and are safe in the presence of NULL pointers (but note that temp will not be NULL in this context). More info on them in section Reference Counts.

Later, when it is time to call the function, you call the C function PyEval_CallObject. This function has two arguments, both pointers to arbitrary Python objects: the Python function, and the argument list. The argument list must always be a tuple object, whose length is the number of arguments. To call the Python function with no arguments, pass in NULL, or an empty tuple; to call it with one argument, pass a singleton tuple. Py_BuildValue returns a tuple when its format string consists of zero or more format codes between parentheses. For example:

int arg;
PyObject *arglist;
PyObject *result;
arg = 123;
/* Time to call the callback */
arglist = Py_BuildValue("(i)", arg);
result = PyEval_CallObject(my_callback, arglist);

PyEval_CallObject returns a Python object pointer: this is the return value of the Python function. PyEval_CallObject is “reference-count-neutral” with respect to its arguments. In the example a new tuple was created to serve as the argument list, which is Py_DECREF-ed immediately after the call.

The return value of PyEval_CallObject is “new”: either it is a brand new object, or it is an existing object whose reference count has been incremented. So, unless you want to save it in a global variable, you should somehow Py_DECREF the result, even (especially!) if you are not interested in its value.

Before you do this, however, it is important to check that the return value isn’t NULL. If it is, the Python function terminated by raising an exception. If the C code that called PyEval_CallObject is called from Python, it should now return an error indication to its Python caller, so the interpreter can print a stack trace, or the calling Python code can handle the exception. If this is not possible or desirable, the exception should be cleared by calling PyErr_Clear. For example:

if (result == NULL)
    return NULL; /* Pass error back */
...use result...

Depending on the desired interface to the Python callback function, you may also have to provide an argument list to PyEval_CallObject. In some cases the argument list is also provided by the Python program, through the same interface that specified the callback function. It can then be saved and used in the same manner as the function object. In other cases, you may have to construct a new tuple to pass as the argument list. The simplest way to do this is to call Py_BuildValue. For example, if you want to pass an integral event code, you might use the following code:

PyObject *arglist;
arglist = Py_BuildValue("(l)", eventcode);
result = PyEval_CallObject(my_callback, arglist);
if (result == NULL)
    return NULL; /* Pass error back */
/* Here maybe use the result */

Note the placement of Py_DECREF(arglist) immediately after the call, before the error check! Also note that strictly speaking this code is not complete: Py_BuildValue may run out of memory, and this should be checked.

You may also call a function with keyword arguments by using PyEval_CallObjectWithKeywords. As in the above example, we use Py_BuildValue to construct the dictionary.

PyObject *dict;
dict = Py_BuildValue("{s:i}", "name", val);
result = PyEval_CallObjectWithKeywords(my_callback, NULL, dict);
if (result == NULL)
    return NULL; /* Pass error back */
/* Here maybe use the result */

Extracting Parameters in Extension Functions

The PyArg_ParseTuple function is declared as follows:

int PyArg_ParseTuple(PyObject *arg, char *format, ...);

The arg argument must be a tuple object containing an argument list passed from Python to a C function. The format argument must be a format string, whose syntax is explained in Parsing arguments and building values in the Python/C API Reference Manual. The remaining arguments must be addresses of variables whose type is determined by the format string.

Note that while PyArg_ParseTuple checks that the Python arguments have the required types, it cannot check the validity of the addresses of C variables passed to the call: if you make mistakes there, your code will probably crash or at least overwrite random bits in memory. So be careful!

Note that any Python object references which are provided to the caller are borrowed references; do not decrement their reference count!

Some example calls:

#define PY_SSIZE_T_CLEAN  /* Make "s#" use Py_ssize_t rather than int. */
#include <Python.h>
int ok;
int i, j;
long k, l;
const char *s;
Py_ssize_t size;

ok = PyArg_ParseTuple(args, ""); /* No arguments */
    /* Python call: f() */
ok = PyArg_ParseTuple(args, "s", &s); /* A string */
    /* Possible Python call: f('whoops!') */
ok = PyArg_ParseTuple(args, "lls", &k, &l, &s); /* Two longs and a string */
    /* Possible Python call: f(1, 2, 'three') */
ok = PyArg_ParseTuple(args, "(ii)s#", &i, &j, &s, &size);
    /* A pair of ints and a string, whose size is also returned */
    /* Possible Python call: f((1, 2), 'three') */
    const char *file;
    const char *mode = "r";
    int bufsize = 0;
    ok = PyArg_ParseTuple(args, "s|si", &file, &mode, &bufsize);
    /* A string, and optionally another string and an integer */
    /* Possible Python calls:
       f('spam', 'w')
       f('spam', 'wb', 100000) */
    int left, top, right, bottom, h, v;
    ok = PyArg_ParseTuple(args, "((ii)(ii))(ii)",
             &left, &top, &right, &bottom, &h, &v);
    /* A rectangle and a point */
    /* Possible Python call:
       f(((0, 0), (400, 300)), (10, 10)) */
    Py_complex c;
    ok = PyArg_ParseTuple(args, "D:myfunction", &c);
    /* a complex, also providing a function name for errors */
    /* Possible Python call: myfunction(1+2j) */

Keyword Parameters for Extension Functions

The PyArg_ParseTupleAndKeywords function is declared as follows:

int PyArg_ParseTupleAndKeywords(PyObject *arg, PyObject *kwdict,
                                char *format, char *kwlist[], ...);

The arg and format parameters are identical to those of the PyArg_ParseTuple function. The kwdict parameter is the dictionary of keywords received as the third parameter from the Python runtime. The kwlist parameter is a NULL-terminated list of strings which identify the parameters; the names are matched with the type information from format from left to right. On success, PyArg_ParseTupleAndKeywords returns true, otherwise it returns false and raises an appropriate exception.


Nested tuples cannot be parsed when using keyword arguments! Keyword parameters passed in which are not present in the kwlist will cause TypeError to be raised.

Here is an example module which uses keywords, based on an example by Geoff Philbrick (philbrick@hks.com):

#include "Python.h"

static PyObject *
keywdarg_parrot(PyObject *self, PyObject *args, PyObject *keywds)
    int voltage;
    char *state = "a stiff";
    char *action = "voom";
    char *type = "Norwegian Blue";

    static char *kwlist[] = {"voltage", "state", "action", "type", NULL};

    if (!PyArg_ParseTupleAndKeywords(args, keywds, "i|sss", kwlist,
                                     &voltage, &state, &action, &type))
        return NULL;

    printf("-- This parrot wouldn't %s if you put %i Volts through it.\n",
           action, voltage);
    printf("-- Lovely plumage, the %s -- It's %s!\n", type, state);


    return Py_None;

static PyMethodDef keywdarg_methods[] = {
    /* The cast of the function is necessary since PyCFunction values
     * only take two PyObject* parameters, and keywdarg_parrot() takes
     * three.
    {"parrot", (PyCFunction)keywdarg_parrot, METH_VARARGS | METH_KEYWORDS,
     "Print a lovely skit to standard output."},
    {NULL, NULL, 0, NULL}   /* sentinel */
  /* Create the module and add the functions */
  Py_InitModule("keywdarg", keywdarg_methods);

Building Arbitrary Values

This function is the counterpart to PyArg_ParseTuple. It is declared as follows:

PyObject *Py_BuildValue(char *format, ...);

It recognizes a set of format units similar to the ones recognized by PyArg_ParseTuple, but the arguments (which are input to the function, not output) must not be pointers, just values. It returns a new Python object, suitable for returning from a C function called from Python.

One difference with PyArg_ParseTuple: while the latter requires its first argument to be a tuple (since Python argument lists are always represented as tuples internally), Py_BuildValue does not always build a tuple. It builds a tuple only if its format string contains two or more format units. If the format string is empty, it returns None; if it contains exactly one format unit, it returns whatever object is described by that format unit. To force it to return a tuple of size 0 or one, parenthesize the format string.

Examples (to the left the call, to the right the resulting Python value):

Py_BuildValue("")                        None
Py_BuildValue("i", 123)                  123
Py_BuildValue("iii", 123, 456, 789)      (123, 456, 789)
Py_BuildValue("s", "hello")              'hello'
Py_BuildValue("y", "hello")              b'hello'
Py_BuildValue("ss", "hello", "world")    ('hello', 'world')
Py_BuildValue("s#", "hello", 4)          'hell'
Py_BuildValue("y#", "hello", 4)          b'hell'
Py_BuildValue("()")                      ()
Py_BuildValue("(i)", 123)                (123,)
Py_BuildValue("(ii)", 123, 456)          (123, 456)
Py_BuildValue("(i,i)", 123, 456)         (123, 456)
Py_BuildValue("[i,i]", 123, 456)         [123, 456]
              "abc", 123, "def", 456)    {'abc': 123, 'def': 456}
Py_BuildValue("((ii)(ii)) (ii)",
              1, 2, 3, 4, 5, 6)          (((1, 2), (3, 4)), (5, 6))

Reference Counts

In languages like C or C++, the programmer is responsible for dynamic allocation and deallocation of memory on the heap. In C, this is done using the functions malloc and free. In C++, the operators new and delete are used with essentially the same meaning and we’ll restrict the following discussion to the C case.

Every block of memory allocated with malloc should eventually be returned to the pool of available memory by exactly one call to free. It is important to call free at the right time. If a block’s address is forgotten but free is not called for it, the memory it occupies cannot be reused until the program terminates. This is called a memory leak. On the other hand, if a program calls free for a block and then continues to use the block, it creates a conflict with re-use of the block through another malloc call. This is called using freed memory. It has the same bad consequences as referencing uninitialized data — core dumps, wrong results, mysterious crashes.

Common causes of memory leaks are unusual paths through the code. For instance, a function may allocate a block of memory, do some calculation, and then free the block again. Now a change in the requirements for the function may add a test to the calculation that detects an error condition and can return prematurely from the function. It’s easy to forget to free the allocated memory block when taking this premature exit, especially when it is added later to the code. Such leaks, once introduced, often go undetected for a long time: the error exit is taken only in a small fraction of all calls, and most modern machines have plenty of virtual memory, so the leak only becomes apparent in a long-running process that uses the leaking function frequently. Therefore, it’s important to prevent leaks from happening by having a coding convention or strategy that minimizes this kind of errors.

Since Python makes heavy use of malloc and free, it needs a strategy to avoid memory leaks as well as the use of freed memory. The chosen method is called reference counting. The principle is simple: every object contains a counter, which is incremented when a reference to the object is stored somewhere, and which is decremented when a reference to it is deleted. When the counter reaches zero, the last reference to the object has been deleted and the object is freed.

An alternative strategy is called automatic garbage collection. (Sometimes, reference counting is also referred to as a garbage collection strategy, hence my use of “automatic” to distinguish the two.) The big advantage of automatic garbage collection is that the user doesn’t need to call free explicitly. (Another claimed advantage is an improvement in speed or memory usage — this is no hard fact however.) The disadvantage is that for C, there is no truly portable automatic garbage collector, while reference counting can be implemented portably (as long as the functions malloc and free are available — which the C Standard guarantees). Maybe some day a sufficiently portable automatic garbage collector will be available for C. Until then, we’ll have to live with reference counts.

While Python uses the traditional reference counting implementation, it also offers a cycle detector that works to detect reference cycles. This allows applications to not worry about creating direct or indirect circular references; these are the weakness of garbage collection implemented using only reference counting. Reference cycles consist of objects which contain (possibly indirect) references to themselves, so that each object in the cycle has a reference count which is non-zero. Typical reference counting implementations are not able to reclaim the memory belonging to any objects in a reference cycle, or referenced from the objects in the cycle, even though there are no further references to the cycle itself.

The cycle detector is able to detect garbage cycles and can reclaim them so long as there are no finalizers implemented in Python (__del__() methods). When there are such finalizers, the detector exposes the cycles through the gc module (specifically, the garbage variable in that module). The gc module also exposes a way to run the detector (the collect() function), as well as configuration interfaces and the ability to disable the detector at runtime. The cycle detector is considered an optional component; though it is included by default, it can be disabled at build time using the --without-cycle-gc option to the configure script on Unix platforms (including Mac OS X). If the cycle detector is disabled in this way, the gc module will not be available.

Reference Counting in Python

There are two macros, Py_INCREF(x) and Py_DECREF(x), which handle the incrementing and decrementing of the reference count. Py_DECREF also frees the object when the count reaches zero. For flexibility, it doesn’t call free directly — rather, it makes a call through a function pointer in the object’s type object. For this purpose (and others), every object also contains a pointer to its type object.

The big question now remains: when to use Py_INCREF(x) and Py_DECREF(x)? Let’s first introduce some terms. Nobody “owns” an object; however, you can own a reference to an object. An object’s reference count is now defined as the number of owned references to it. The owner of a reference is responsible for calling Py_DECREF when the reference is no longer needed. Ownership of a reference can be transferred. There are three ways to dispose of an owned reference: pass it on, store it, or call Py_DECREF. Forgetting to dispose of an owned reference creates a memory leak.

It is also possible to borrow [2] a reference to an object. The borrower of a reference should not call Py_DECREF. The borrower must not hold on to the object longer than the owner from which it was borrowed. Using a borrowed reference after the owner has disposed of it risks using freed memory and should be avoided completely. [3]

The advantage of borrowing over owning a reference is that you don’t need to take care of disposing of the reference on all possible paths through the code — in other words, with a borrowed reference you don’t run the risk of leaking when a premature exit is taken. The disadvantage of borrowing over owning is that there are some subtle situations where in seemingly correct code a borrowed reference can be used after the owner from which it was borrowed has in fact disposed of it.

A borrowed reference can be changed into an owned reference by calling Py_INCREF. This does not affect the status of the owner from which the reference was borrowed — it creates a new owned reference, and gives full owner responsibilities (the new owner must dispose of the reference properly, as well as the previous owner).

Ownership Rules

Whenever an object reference is passed into or out of a function, it is part of the function’s interface specification whether ownership is transferred with the reference or not.

Most functions that return a reference to an object pass on ownership with the reference. In particular, all functions whose function it is to create a new object, such as PyLong_FromLong and Py_BuildValue, pass ownership to the receiver. Even if the object is not actually new, you still receive ownership of a new reference to that object. For instance, PyLong_FromLong maintains a cache of popular values and can return a reference to a cached item.

Many functions that extract objects from other objects also transfer ownership with the reference, for instance PyObject_GetAttrString. The picture is less clear, here, however, since a few common routines are exceptions: PyTuple_GetItem, PyList_GetItem, PyDict_GetItem, and PyDict_GetItemString all return references that you borrow from the tuple, list or dictionary.

The function PyImport_AddModule also returns a borrowed reference, even though it may actually create the object it returns: this is possible because an owned reference to the object is stored in sys.modules.

When you pass an object reference into another function, in general, the function borrows the reference from you — if it needs to store it, it will use Py_INCREF to become an independent owner. There are exactly two important exceptions to this rule: PyTuple_SetItem and PyList_SetItem. These functions take over ownership of the item passed to them — even if they fail! (Note that PyDict_SetItem and friends don’t take over ownership — they are “normal.”)

When a C function is called from Python, it borrows references to its arguments from the caller. The caller owns a reference to the object, so the borrowed reference’s lifetime is guaranteed until the function returns. Only when such a borrowed reference must be stored or passed on, it must be turned into an owned reference by calling Py_INCREF.

The object reference returned from a C function that is called from Python must be an owned reference — ownership is transferred from the function to its caller.

Thin Ice

There are a few situations where seemingly harmless use of a borrowed reference can lead to problems. These all have to do with implicit invocations of the interpreter, which can cause the owner of a reference to dispose of it.

The first and most important case to know about is using Py_DECREF on an unrelated object while borrowing a reference to a list item. For instance:

bug(PyObject *list)
    PyObject *item = PyList_GetItem(list, 0);

    PyList_SetItem(list, 1, PyLong_FromLong(0L));
    PyObject_Print(item, stdout, 0); /* BUG! */

This function first borrows a reference to list[0], then replaces list[1] with the value 0, and finally prints the borrowed reference. Looks harmless, right? But it’s not!

Let’s follow the control flow into PyList_SetItem. The list owns references to all its items, so when item 1 is replaced, it has to dispose of the original item 1. Now let’s suppose the original item 1 was an instance of a user-defined class, and let’s further suppose that the class defined a __del__() method. If this class instance has a reference count of 1, disposing of it will call its __del__() method.

Since it is written in Python, the __del__() method can execute arbitrary Python code. Could it perhaps do something to invalidate the reference to item in bug? You bet! Assuming that the list passed into bug is accessible to the __del__() method, it could execute a statement to the effect of del list[0], and assuming this was the last reference to that object, it would free the memory associated with it, thereby invalidating item.

The solution, once you know the source of the problem, is easy: temporarily increment the reference count. The correct version of the function reads:

no_bug(PyObject *list)
    PyObject *item = PyList_GetItem(list, 0);

    PyList_SetItem(list, 1, PyLong_FromLong(0L));
    PyObject_Print(item, stdout, 0);

This is a true story. An older version of Python contained variants of this bug and someone spent a considerable amount of time in a C debugger to figure out why his __del__() methods would fail...

The second case of problems with a borrowed reference is a variant involving threads. Normally, multiple threads in the Python interpreter can’t get in each other’s way, because there is a global lock protecting Python’s entire object space. However, it is possible to temporarily release this lock using the macro Py_BEGIN_ALLOW_THREADS, and to re-acquire it using Py_END_ALLOW_THREADS. This is common around blocking I/O calls, to let other threads use the processor while waiting for the I/O to complete. Obviously, the following function has the same problem as the previous one:

bug(PyObject *list)
    PyObject *item = PyList_GetItem(list, 0);
    ...some blocking I/O call...
    PyObject_Print(item, stdout, 0); /* BUG! */

NULL Pointers

In general, functions that take object references as arguments do not expect you to pass them NULL pointers, and will dump core (or cause later core dumps) if you do so. Functions that return object references generally return NULL only to indicate that an exception occurred. The reason for not testing for NULL arguments is that functions often pass the objects they receive on to other function — if each function were to test for NULL, there would be a lot of redundant tests and the code would run more slowly.

It is better to test for NULL only at the “source:” when a pointer that may be NULL is received, for example, from malloc or from a function that may raise an exception.

The macros Py_INCREF and Py_DECREF do not check for NULL pointers — however, their variants Py_XINCREF and Py_XDECREF do.

The macros for checking for a particular object type (Pytype_Check()) don’t check for NULL pointers — again, there is much code that calls several of these in a row to test an object against various different expected types, and this would generate redundant tests. There are no variants with NULL checking.

The C function calling mechanism guarantees that the argument list passed to C functions (args in the examples) is never NULL — in fact it guarantees that it is always a tuple. [4]

It is a severe error to ever let a NULL pointer “escape” to the Python user.

Writing Extensions in C++

It is possible to write extension modules in C++. Some restrictions apply. If the main program (the Python interpreter) is compiled and linked by the C compiler, global or static objects with constructors cannot be used. This is not a problem if the main program is linked by the C++ compiler. Functions that will be called by the Python interpreter (in particular, module initialization functions) have to be declared using extern "C". It is unnecessary to enclose the Python header files in extern "C" {...} — they use this form already if the symbol __cplusplus is defined (all recent C++ compilers define this symbol).

Providing a C API for an Extension Module

Many extension modules just provide new functions and types to be used from Python, but sometimes the code in an extension module can be useful for other extension modules. For example, an extension module could implement a type “collection” which works like lists without order. Just like the standard Python list type has a C API which permits extension modules to create and manipulate lists, this new collection type should have a set of C functions for direct manipulation from other extension modules.

At first sight this seems easy: just write the functions (without declaring them static, of course), provide an appropriate header file, and document the C API. And in fact this would work if all extension modules were always linked statically with the Python interpreter. When modules are used as shared libraries, however, the symbols defined in one module may not be visible to another module. The details of visibility depend on the operating system; some systems use one global namespace for the Python interpreter and all extension modules (Windows, for example), whereas others require an explicit list of imported symbols at module link time (AIX is one example), or offer a choice of different strategies (most Unices). And even if symbols are globally visible, the module whose functions one wishes to call might not have been loaded yet!

Portability therefore requires not to make any assumptions about symbol visibility. This means that all symbols in extension modules should be declared static, except for the module’s initialization function, in order to avoid name clashes with other extension modules (as discussed in section The Module’s Method Table and Initialization Function). And it means that symbols that should be accessible from other extension modules must be exported in a different way.

Python provides a special mechanism to pass C-level information (pointers) from one extension module to another one: CObjects. A CObject is a Python data type which stores a pointer (void *). CObjects can only be created and accessed via their C API, but they can be passed around like any other Python object. In particular, they can be assigned to a name in an extension module’s namespace. Other extension modules can then import this module, retrieve the value of this name, and then retrieve the pointer from the CObject.

There are many ways in which CObjects can be used to export the C API of an extension module. Each name could get its own CObject, or all C API pointers could be stored in an array whose address is published in a CObject. And the various tasks of storing and retrieving the pointers can be distributed in different ways between the module providing the code and the client modules.

The following example demonstrates an approach that puts most of the burden on the writer of the exporting module, which is appropriate for commonly used library modules. It stores all C API pointers (just one in the example!) in an array of void pointers which becomes the value of a CObject. The header file corresponding to the module provides a macro that takes care of importing the module and retrieving its C API pointers; client modules only have to call this macro before accessing the C API.

The exporting module is a modification of the spam module from section A Simple Example. The function spam.system() does not call the C library function system directly, but a function PySpam_System, which would of course do something more complicated in reality (such as adding “spam” to every command). This function PySpam_System is also exported to other extension modules.

The function PySpam_System is a plain C function, declared static like everything else:

static int
PySpam_System(const char *command)
    return system(command);

The function spam_system is modified in a trivial way:

static PyObject *
spam_system(PyObject *self, PyObject *args)
    const char *command;
    int sts;

    if (!PyArg_ParseTuple(args, "s", &command))
        return NULL;
    sts = PySpam_System(command);
    return Py_BuildValue("i", sts);

In the beginning of the module, right after the line

#include "Python.h"

two more lines must be added:

#include "spammodule.h"

The #define is used to tell the header file that it is being included in the exporting module, not a client module. Finally, the module’s initialization function must take care of initializing the C API pointer array:

    PyObject *m;
    static void *PySpam_API[PySpam_API_pointers];
    PyObject *c_api_object;

    m = PyModule_Create(&spammodule);
    if (m == NULL)
        return NULL;

    /* Initialize the C API pointer array */
    PySpam_API[PySpam_System_NUM] = (void *)PySpam_System;

    /* Create a CObject containing the API pointer array's address */
    c_api_object = PyCObject_FromVoidPtr((void *)PySpam_API, NULL);

    if (c_api_object != NULL)
        PyModule_AddObject(m, "_C_API", c_api_object);
    return m;

Note that PySpam_API is declared static; otherwise the pointer array would disappear when PyInit_spam() terminates!

The bulk of the work is in the header file spammodule.h, which looks like this:

#ifdef __cplusplus
extern "C" {

/* Header file for spammodule */

/* C API functions */
#define PySpam_System_NUM 0
#define PySpam_System_RETURN int
#define PySpam_System_PROTO (const char *command)

/* Total number of C API pointers */
#define PySpam_API_pointers 1

/* This section is used when compiling spammodule.c */

static PySpam_System_RETURN PySpam_System PySpam_System_PROTO;

/* This section is used in modules that use spammodule's API */

static void **PySpam_API;

#define PySpam_System \
 (*(PySpam_System_RETURN (*)PySpam_System_PROTO) PySpam_API[PySpam_System_NUM])

/* Return -1 and set exception on error, 0 on success. */
static int
    PyObject *module = PyImport_ImportModule("spam");

    if (module != NULL) {
        PyObject *c_api_object = PyObject_GetAttrString(module, "_C_API");
        if (c_api_object == NULL)
            return -1;
        if (PyCObject_Check(c_api_object))
            PySpam_API = (void **)PyCObject_AsVoidPtr(c_api_object);
    return 0;


#ifdef __cplusplus

#endif /* !defined(Py_SPAMMODULE_H) */

All that a client module must do in order to have access to the function PySpam_System is to call the function (or rather macro) import_spam in its initialization function:

    PyObject *m;

    m = Py_InitModule("client", ClientMethods);
    if (m == NULL)
    if (import_spam() < 0)
    /* additional initialization can happen here */

The main disadvantage of this approach is that the file spammodule.h is rather complicated. However, the basic structure is the same for each function that is exported, so it has to be learned only once.

Finally it should be mentioned that CObjects offer additional functionality, which is especially useful for memory allocation and deallocation of the pointer stored in a CObject. The details are described in the Python/C API Reference Manual in the section CObjects and in the implementation of CObjects (files Include/cobject.h and Objects/cobject.c in the Python source code distribution).


[1]An interface for this function already exists in the standard module os — it was chosen as a simple and straightforward example.
[2]The metaphor of “borrowing” a reference is not completely correct: the owner still has a copy of the reference.
[3]Checking that the reference count is at least 1 does not work — the reference count itself could be in freed memory and may thus be reused for another object!
[4]These guarantees don’t hold when you use the “old” style calling convention — this is still found in much existing code.