1. 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.通过包含头文件"Python.h"，这些Python API可以在C源文件中使用。

扩展模块的编译依赖于你打算如何使用这些模块和你系统的设置；details are given in later chapters.

请注意，如果你的目的是调用C库函数和系统调用，你应该考虑使用ctypes模块，而不是自己写C代码。Not only does ctypes let you write Python code to interface with C code, but it is more portable between implementations of Python than writing and compiling an extension module which typically ties you to CPython.

1.1. A Simple Example

我们来创建一个名为spam的扩展模块（spam，monty python粉丝们最喜欢的食物），我们想给C库函数system()创建一个Python的接口。[1] 这个函数以一个空字符结尾的字符串为参数，返回一个整数。我们希望在Python中这样调用这个函数：

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

首先创建名为spammodule.c的文件。（按历史惯例，如果模块叫做spam，那么包含实现的C文件就叫做spammodule.c；如果模块的名字很长，比如spammify，那么模块的名字可以是spammify.c。）

文件的第一行为：

#include <Python.h>

这会引入Python API的声明（如果乐意，你可以加上关于模块的说明和版权信息的注释）。

注意

在某些平台上，Python可能会定义一些预处理器定义，从而影响到标准头文件，所以在包含任何的标准头文件之前必须先包含Python.h。

除了在标准头文件中定义的符号外，Python.h中定义的用户可见的符号以Py或者PY打头。

接下来我们加入C实现函数，当Python表达式spam.system(string)被演算时，该函数会被调用（不久我们会看到它是如何结束调用的）：

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);
}

可以很清楚地看出Python的参数列表（如"ls -l"）是如何转化成C函数的参数的。

对于模块级函数，self参数指向模块对象；如果是方法，self指向对象实例。

args参数是一个指向包含参数的Python元组对象的指针。元组中的每一个元素对应调用参数列表中的一个参数。这些参数是Python对象——为了能在C函数中处理这些Python对象，我们需要将其转成C可操作的值。Python API中的PyArg_ParseTuple()函数检查参数类型，并将其转成对应的C值。它使用字符串模板来决定参数所需要的类型，同时也决定用来存转换后的值的C变量的类型。后面有更多的说明。

如果所有的参数有正确的类型，且值被存入变量（通过传地址）PyArg_ParseTuple()返回真（非0值）。如果传递了不正确的参数，它返回假（0值）。对于后一种情况，它是个异常情况，所以函数立即返回NULL（如同例子所示）。

1.2. 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 Python variables sys.exc_type, sys.exc_value and sys.exc_traceback (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, PyInt_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 (initspam()) with an exception object (leaving out the error checking for now):

PyMODINIT_FUNC
initspam(void)
{
    PyObject *m;

    m = Py_InitModule("spam", SpamMethods);
    if (m == NULL)
        return;

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

注意异常对象的Python名字是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.

The spam.error exception can be raised in your extension module using a call to PyErr_SetString() as shown below:

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

    if (!PyArg_ParseTuple(args, "s", &command))
        return NULL;
    sts = system(command);
    if (sts < 0) {
        PyErr_SetString(SpamError, "System command failed");
        return NULL;
    }
    return PyLong_FromLong(sts);
}

1.3. 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):

Py_INCREF(Py_None);
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.

1.4. 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 passed to the interpreter in the module’s initialization function. The initialization function must be named initname(), where name is the name of the module, and should be the only non-static item defined in the module file:

PyMODINIT_FUNC
initspam(void)
{
    (void) Py_InitModule("spam", SpamMethods);
}

Note that PyMODINIT_FUNC declares the function as void 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, initspam() is called.当Python程序首次导入 spam 将调用initspam() 。（参见下面关于嵌入Python的注释。）initspam()调用Py_InitModule()，它创建一个“模块对象”（该模块对象被插入字典sys.modules，对应的键为"spam"），根据第二个参数——方法表（PyMethodDef结构的数组），将函数对象插入该模块。Py_InitModule()返回它所创建的模块对象的指针（这里没有使用返回值）。某些错误可以使该函数中止，如果模块不能被初始化，它返回NULL。

当嵌入Python时，initspam()不会被自动调用，除非它在_PyImport_Inittab表中有登记。The easiest way to handle this is to statically initialize your statically-linked modules by directly calling initspam() after the call to Py_Initialize():

int
main(int argc, char *argv[])
{
    /* Pass argv[0] to the Python interpreter */
    Py_SetProgramName(argv[0]);

    /* Initialize the Python interpreter.  Required. */
    Py_Initialize();

    /* Add a static module */
    initspam();

    ...

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

Note

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. Note also that the reload() function can be used with extension modules, and will call the module initialization function (initspam() in the example), but will not load the module again if it was loaded from a dynamically loadable object file (.so on Unix, .dll on Windows).

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.

1.5. 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; 参见构建扩展模块相关的章节（Building C and C++ Extensions with distutils），Windows特定的构建信息参见（Building C and C++ Extensions on Windows）。

如果你无法使用动态装载，或者你希望你的模块永久的成为解释器的一部分，你需要修改配置设置并重新构建解释器。幸运的是，这在Unix上非常简单：在解压的源码包中的Modules/目录下放置你的文件（例子中的spammodule.c），在Modules/Setup.local 中添加一行来描述你的文件：

spam spammodule.o

然后在顶层目录下运行make来重新构建解释器。你也可以在Modules/子目录下运行make，但是你得先运行“make Makefile”以重新构建Makefile。（每次修改了Setup文件都要这样做。）

如果模块需要链接额外的库，它们也可以出现在配置文件中，例如：

spam spammodule.o -lX11

1.6.从 C 调用 Python 函数

到目前为止，我们已经重点讨论了如何从Python调用C的函数。反之，从C调用Python的函数也非常有用。这个功能特别是在创建支持“回调函数”的库时候会很便利。 当一个C接口使用回调，提供同样功能的Python代码经常需要为Python程序员提供一个回调机制； 这个执行将从一个C的回调请求调用Python的回调函数。当然，也有很多其它的用途。

幸运的是，Python解释器可以简单的调用递归，并且有一个标准的调用接口去调用Python函数。 （我不打算详细说明怎样用特定的字符串作为输入去呼叫Python的语法分析器 - 如果你有兴趣，可以从Python的源代码查看-c命令行的Modules/main.c选项）

调用Python函数很容易。 Python程序一定会以某种方式传递给你Python的函数对象。 为此，你应该为其提供一个函数（或其它某种接口）。当这个函数被调用，一个指向Python函数对象（小心对其使用Py_INCREF()！）的指针被保存在一个全局变量 - 或其它任何你认为合适的地方。 比如：以下的函数可能成为模块定义的一部分

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" */
        Py_INCREF(Py_None);
        result = Py_None;
    }
    return result;
}

此函数必须使用METH_VARARGS标志注册到解释器；这在模块的方法表和初始化函数中有描述。PyArg_ParseTuple()函数及其参数在提取扩展函数中的参数一节中有说明。

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.

以后当需要调用这个回调函数时，调用 C语言函数PyObject_CallObject()即可。这个函数有两个参数，都是指向任意Python对象的指针：Python函数和参数列表。参数列表必须始终是一个元组对象，其长度是参数的个数。要调用没有参数的Python函数，传入NULL或空元组；用一个参数调用它，传递一个单例元组。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 = PyObject_CallObject(my_callback, arglist);
Py_DECREF(arglist);

PyObject_CallObject() returns a Python object pointer: this is the return value of the Python function. PyObject_CallObject() is “reference-count-neutral” with respect to its arguments. 在示例中，创建了一个新的元组作为参数列表，它在PyObject_CallObject()调用之后直接执行Py_DECREF()

PyObject_CallObject()的返回值是“new”：它是一个全新的对象，或者它是一个引用计数已增加的现有对象。所以，除非你想保存在一个全局变量，你应该无论如何对result使用Py_DECREF()，甚至（特别是！）如果你不感兴趣的价值。

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 PyObject_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...
Py_DECREF(result);

Depending on the desired interface to the Python callback function, you may also have to provide an argument list to PyObject_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 = PyObject_CallObject(my_callback, arglist);
Py_DECREF(arglist);
if (result == NULL)
    return NULL; /* Pass error back */
/* Here maybe use the result */
Py_DECREF(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 PyObject_Call(), which supports arguments and keyword arguments. As in the above example, we use Py_BuildValue() to construct the dictionary.

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

1.7. Extracting Parameters in Extension Functions

函数 PyArg_ParseTuple() 定义如下：

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!

举例如下:

int ok;
int i, j;
long k, l;
const char *s;
int 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')
       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) */
}

1.8. 扩展函数的关键字参数

PyArg_ParseTupleAndKeywords()函数声明如下：

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.

Note

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);

    Py_INCREF(Py_None);

    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 */
};

void
initkeywdarg(void)
{
  /* Create the module and add the functions */
  Py_InitModule("keywdarg", keywdarg_methods);
}

1.9. 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("ss", "hello", "world")    ('hello', 'world')
Py_BuildValue("s#", "hello", 4)          '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]
Py_BuildValue("{s:i,s:i}",
              "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))

1.10. 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) or by removing the definition of WITH_CYCLE_GC in the pyconfig.h header on other platforms. If the cycle detector is disabled in this way, the gc module will not be available.

1.10.1. 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).

1.10.2. 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 PyInt_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, PyInt_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.

1.10.3. 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:

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

    PyList_SetItem(list, 1, PyInt_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:

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

    Py_INCREF(item);
    PyList_SetItem(list, 1, PyInt_FromLong(0L));
    PyObject_Print(item, stdout, 0);
    Py_DECREF(item);
}

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:

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

1.10.4. NULL Pointers

通常，函数接受一个对象作为一个参数是不希望你传给它们一个空指针（NULL）的，如果你这样做的话，会引发触核或者导致触核。函数返回一个对象的引用，一般只有在指示一个异常发生时才会返回空指针（NULL ）。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.

1.11. 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).

1.12. 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: Capsules. A Capsule is a Python data type which stores a pointer (void *). Capsules 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 Capsule.

There are many ways in which Capsules can be used to export the C API of an extension module. Each function could get its own Capsule, or all C API pointers could be stored in an array whose address is published in a Capsule. 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.

Whichever method you choose, it’s important to name your Capsules properly. The function PyCapsule_New() takes a name parameter (const char *); you’re permitted to pass in a NULL name, but we strongly encourage you to specify a name. Properly named Capsules provide a degree of runtime type-safety; there is no feasible way to tell one unnamed Capsule from another.

In particular, Capsules used to expose C APIs should be given a name following this convention:

modulename.attributename

The convenience function PyCapsule_Import() makes it easy to load a C API provided via a Capsule, but only if the Capsule’s name matches this convention. This behavior gives C API users a high degree of certainty that the Capsule they load contains the correct C API.

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 Capsule. 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:

#define SPAM_MODULE
#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:

PyMODINIT_FUNC
initspam(void)
{
    PyObject *m;
    static void *PySpam_API[PySpam_API_pointers];
    PyObject *c_api_object;

    m = Py_InitModule("spam", SpamMethods);
    if (m == NULL)
        return;

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

    /* Create a Capsule containing the API pointer array's address */
    c_api_object = PyCapsule_New((void *)PySpam_API, "spam._C_API", NULL);

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

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

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

#ifndef Py_SPAMMODULE_H
#define Py_SPAMMODULE_H
#ifdef __cplusplus
extern "C" {
#endif

/* 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


#ifdef SPAM_MODULE
/* This section is used when compiling spammodule.c */

static PySpam_System_RETURN PySpam_System PySpam_System_PROTO;

#else
/* 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 on error, 0 on success.
 * PyCapsule_Import will set an exception if there's an error.
 */
static int
import_spam(void)
{
    PySpam_API = (void **)PyCapsule_Import("spam._C_API", 0);
    return (PySpam_API != NULL) ? 0 : -1;
}

#endif

#ifdef __cplusplus
}
#endif

#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:

PyMODINIT_FUNC
initclient(void)
{
    PyObject *m;

    m = Py_InitModule("client", ClientMethods);
    if (m == NULL)
        return;
    if (import_spam() < 0)
        return;
    /* 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 Capsules offer additional functionality, which is especially useful for memory allocation and deallocation of the pointer stored in a Capsule. The details are described in the Python/C API Reference Manual in the section Capsules and in the implementation of Capsules (files Include/pycapsule.h and Objects/pycapsule.c in the Python source code distribution).

Footnotes

[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.