Introduction

The Application Programmer’s Interface to Python gives C and C++ programmers access to the Python interpreter at a variety of levels. The API is equally usable from C++, but for brevity it is generally referred to as the Python/C API. There are two fundamentally different reasons for using the Python/C API. The first reason is to write extension modules for specific purposes; these are C modules that extend the Python interpreter. This is probably the most common use. The second reason is to use Python as a component in a larger application; this technique is generally referred to as embedding Python in an application.

Writing an extension module is a relatively well-understood process, where a “cookbook” approach works well. There are several tools that automate the process to some extent. While people have embedded Python in other applications since its early existence, the process of embedding Python is less straightforward than writing an extension.

Many API functions are useful independent of whether you’re embedding or extending Python; moreover, most applications that embed Python will need to provide a custom extension as well, so it’s probably a good idea to become familiar with writing an extension before attempting to embed Python in a real application.

Include Files

All function, type and macro definitions needed to use the Python/C API are included in your code by the following line:

#include "Python.h"

This implies inclusion of the following standard headers: <stdio.h>, <string.h>, <errno.h>, <limits.h>, <assert.h> and <stdlib.h> (if available).

Note

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 names defined by Python.h (except those defined by the included standard headers) have one of the prefixes Py or _Py. Names beginning with _Py are for internal use by the Python implementation and should not be used by extension writers. Structure member names do not have a reserved prefix.

Important: user code should never define names that begin with Py or _Py. This confuses the reader, and jeopardizes the portability of the user code to future Python versions, which may define additional names beginning with one of these prefixes.

The header files are typically installed with Python. On Unix, these are located in the directories prefix/include/pythonversion/ and exec_prefix/include/pythonversion/, where prefix and exec_prefix are defined by the corresponding parameters to Python’s configure script and version is sys.version[:3]. On Windows, the headers are installed in prefix/include, where prefix is the installation directory specified to the installer.

To include the headers, place both directories (if different) on your compiler’s search path for includes. Do not place the parent directories on the search path and then use #include <pythonX.Y/Python.h>; this will break on multi-platform builds since the platform independent headers under prefix include the platform specific headers from exec_prefix.

C++ users should note that though the API is defined entirely using C, the header files do properly declare the entry points to be extern "C", so there is no need to do anything special to use the API from C++.

Objects, Types and Reference Counts

Most Python/C API functions have one or more arguments as well as a return value of type PyObject*. This type is a pointer to an opaque data type representing an arbitrary Python object. Since all Python object types are treated the same way by the Python language in most situations (e.g., assignments, scope rules, and argument passing), it is only fitting that they should be represented by a single C type. Almost all Python objects live on the heap: you never declare an automatic or static variable of type PyObject, only pointer variables of type PyObject* can be declared. The sole exception are the type objects; since these must never be deallocated, they are typically static PyTypeObject objects.

All Python objects (even Python integers) have a type and a reference count. An object’s type determines what kind of object it is (e.g., an integer, a list, or a user-defined function; there are many more as explained in The standard type hierarchy). For each of the well-known types there is a macro to check whether an object is of that type; for instance, PyList_Check(a) is true if (and only if) the object pointed to by a is a Python list.

Reference Counts

The reference count is important because today’s computers have a finite (and often severely limited) memory size; it counts how many different places there are that have a reference to an object. Such a place could be another object, or a global (or static) C variable, or a local variable in some C function. When an object’s reference count becomes zero, the object is deallocated. If it contains references to other objects, their reference count is decremented. Those other objects may be deallocated in turn, if this decrement makes their reference count become zero, and so on. (There’s an obvious problem with objects that reference each other here; for now, the solution is “don’t do that.”)

Reference counts are always manipulated explicitly. The normal way is to use the macro Py_INCREF() to increment an object’s reference count by one, and Py_DECREF() to decrement it by one. The Py_DECREF() macro is considerably more complex than the incref one, since it must check whether the reference count becomes zero and then cause the object’s deallocator to be called. The deallocator is a function pointer contained in the object’s type structure. The type-specific deallocator takes care of decrementing the reference counts for other objects contained in the object if this is a compound object type, such as a list, as well as performing any additional finalization that’s needed. There’s no chance that the reference count can overflow; at least as many bits are used to hold the reference count as there are distinct memory locations in virtual memory (assuming sizeof(Py_ssize_t) >= sizeof(void*)). Thus, the reference count increment is a simple operation.

It is not necessary to increment an object’s reference count for every local variable that contains a pointer to an object. In theory, the object’s reference count goes up by one when the variable is made to point to it and it goes down by one when the variable goes out of scope. However, these two cancel each other out, so at the end the reference count hasn’t changed. The only real reason to use the reference count is to prevent the object from being deallocated as long as our variable is pointing to it. If we know that there is at least one other reference to the object that lives at least as long as our variable, there is no need to increment the reference count temporarily. An important situation where this arises is in objects that are passed as arguments to C functions in an extension module that are called from Python; the call mechanism guarantees to hold a reference to every argument for the duration of the call.

However, a common pitfall is to extract an object from a list and hold on to it for a while without incrementing its reference count. Some other operation might conceivably remove the object from the list, decrementing its reference count and possible deallocating it. The real danger is that innocent-looking operations may invoke arbitrary Python code which could do this; there is a code path which allows control to flow back to the user from a Py_DECREF(), so almost any operation is potentially dangerous.

A safe approach is to always use the generic operations (functions whose name begins with PyObject_, PyNumber_, PySequence_ or PyMapping_). These operations always increment the reference count of the object they return. This leaves the caller with the responsibility to call Py_DECREF() when they are done with the result; this soon becomes second nature.

Reference Count Details

The reference count behavior of functions in the Python/C API is best explained in terms of ownership of references. Ownership pertains to references, never to objects (objects are not owned: they are always shared). “Owning a reference” means being responsible for calling Py_DECREF on it when the reference is no longer needed. Ownership can also be transferred, meaning that the code that receives ownership of the reference then becomes responsible for eventually decref’ing it by calling Py_DECREF() or Py_XDECREF() when it’s no longer needed—or passing on this responsibility (usually to its caller). When a function passes ownership of a reference on to its caller, the caller is said to receive a new reference. When no ownership is transferred, the caller is said to borrow the reference. Nothing needs to be done for a borrowed reference.

Conversely, when a calling function passes in a reference to an object, there are two possibilities: the function steals a reference to the object, or it does not. Stealing a reference means that when you pass a reference to a function, that function assumes that it now owns that reference, and you are not responsible for it any longer.

Few functions steal references; the two notable exceptions are PyList_SetItem() and PyTuple_SetItem(), which steal a reference to the item (but not to the tuple or list into which the item is put!). These functions were designed to steal a reference because of a common idiom for populating a tuple or list with newly created objects; for example, the code to create the tuple (1, 2, "three") could look like this (forgetting about error handling for the moment; a better way to code this is shown below):

PyObject *t;

t = PyTuple_New(3);
PyTuple_SetItem(t, 0, PyLong_FromLong(1L));
PyTuple_SetItem(t, 1, PyLong_FromLong(2L));
PyTuple_SetItem(t, 2, PyUnicode_FromString("three"));

Here, PyLong_FromLong() returns a new reference which is immediately stolen by PyTuple_SetItem(). When you want to keep using an object although the reference to it will be stolen, use Py_INCREF() to grab another reference before calling the reference-stealing function.

Incidentally, PyTuple_SetItem() is the only way to set tuple items; PySequence_SetItem() and PyObject_SetItem() refuse to do this since tuples are an immutable data type. You should only use PyTuple_SetItem() for tuples that you are creating yourself.

Equivalent code for populating a list can be written using PyList_New() and PyList_SetItem().

However, in practice, you will rarely use these ways of creating and populating a tuple or list. There’s a generic function, Py_BuildValue(), that can create most common objects from C values, directed by a format string. For example, the above two blocks of code could be replaced by the following (which also takes care of the error checking):

PyObject *tuple, *list;

tuple = Py_BuildValue("(iis)", 1, 2, "three");
list = Py_BuildValue("[iis]", 1, 2, "three");

It is much more common to use PyObject_SetItem() and friends with items whose references you are only borrowing, like arguments that were passed in to the function you are writing. In that case, their behaviour regarding reference counts is much saner, since you don’t have to increment a reference count so you can give a reference away (“have it be stolen”). For example, this function sets all items of a list (actually, any mutable sequence) to a given item:

int
set_all(PyObject *target, PyObject *item)
{
    Py_ssize_t i, n;

    n = PyObject_Length(target);
    if (n < 0)
        return -1;
    for (i = 0; i < n; i++) {
        PyObject *index = PyLong_FromSsize_t(i);
        if (!index)
            return -1;
        if (PyObject_SetItem(target, index, item) < 0) {
            Py_DECREF(index);
            return -1;
        }
        Py_DECREF(index);
    }
    return 0;
}

The situation is slightly different for function return values. While passing a reference to most functions does not change your ownership responsibilities for that reference, many functions that return a reference to an object give you ownership of the reference. The reason is simple: in many cases, the returned object is created on the fly, and the reference you get is the only reference to the object. Therefore, the generic functions that return object references, like PyObject_GetItem() and PySequence_GetItem(), always return a new reference (the caller becomes the owner of the reference).

It is important to realize that whether you own a reference returned by a function depends on which function you call only — the plumage (the type of the object passed as an argument to the function) doesn’t enter into it! Thus, if you extract an item from a list using PyList_GetItem(), you don’t own the reference — but if you obtain the same item from the same list using PySequence_GetItem() (which happens to take exactly the same arguments), you do own a reference to the returned object.

Here is an example of how you could write a function that computes the sum of the items in a list of integers; once using PyList_GetItem(), and once using PySequence_GetItem().

long
sum_list(PyObject *list)
{
    Py_ssize_t i, n;
    long total = 0, value;
    PyObject *item;

    n = PyList_Size(list);
    if (n < 0)
        return -1; /* Not a list */
    for (i = 0; i < n; i++) {
        item = PyList_GetItem(list, i); /* Can't fail */
        if (!PyLong_Check(item)) continue; /* Skip non-integers */
        value = PyLong_AsLong(item);
        if (value == -1 && PyErr_Occurred())
            /* Integer too big to fit in a C long, bail out */
            return -1;
        total += value;
    }
    return total;
}

long
sum_sequence(PyObject *sequence)
{
    Py_ssize_t i, n;
    long total = 0, value;
    PyObject *item;
    n = PySequence_Length(sequence);
    if (n < 0)
        return -1; /* Has no length */
    for (i = 0; i < n; i++) {
        item = PySequence_GetItem(sequence, i);
        if (item == NULL)
            return -1; /* Not a sequence, or other failure */
        if (PyLong_Check(item)) {
            value = PyLong_AsLong(item);
            Py_DECREF(item);
            if (value == -1 && PyErr_Occurred())
                /* Integer too big to fit in a C long, bail out */
                return -1;
            total += value;
        }
        else {
            Py_DECREF(item); /* Discard reference ownership */
        }
    }
    return total;
}

Types

There are few other data types that play a significant role in the Python/C API; most are simple C types such as int, long, double and char*. A few structure types are used to describe static tables used to list the functions exported by a module or the data attributes of a new object type, and another is used to describe the value of a complex number. These will be discussed together with the functions that use them.

Exceptions

The Python programmer only needs to deal with exceptions if specific error handling is required; unhandled exceptions are automatically propagated to the caller, then to the caller’s caller, and so on, until they reach the top-level interpreter, where they are reported to the user accompanied by a stack traceback.

For C programmers, however, error checking always has to be explicit. All functions in the Python/C API can raise exceptions, unless an explicit claim is made otherwise in a function’s documentation. In general, when a function encounters an error, it sets an exception, discards any object references that it owns, and returns an error indicator. If not documented otherwise, this indicator is either NULL or -1, depending on the function’s return type. A few functions return a Boolean true/false result, with false indicating an error. Very few functions return no explicit error indicator or have an ambiguous return value, and require explicit testing for errors with PyErr_Occurred(). These exceptions are always explicitly documented.

Exception state is maintained in per-thread storage (this is equivalent to using global storage in an unthreaded application). A thread can be in one of two states: an exception has occurred, or not. The function PyErr_Occurred() can be used to check for this: it returns a borrowed reference to the exception type object when an exception has occurred, and NULL otherwise. There are a number of functions to set the exception state: PyErr_SetString() is the most common (though not the most general) function to set the exception state, and PyErr_Clear() clears the exception state.

The full exception state consists of three objects (all of which can be NULL): the exception type, the corresponding exception value, and the traceback. These have the same meanings as the Python result of sys.exc_info(); however, they are not the same: the Python objects represent the last exception being handled by a Python try ... except statement, while the C level exception state only exists while an exception is being passed on between C functions until it reaches the Python bytecode interpreter’s main loop, which takes care of transferring it to sys.exc_info() and friends.

Note that starting with Python 1.5, the preferred, thread-safe way to access the exception state from Python code is to call the function sys.exc_info(), which returns the per-thread exception state for Python code. Also, the semantics of both ways to access the exception state have changed so that a function which catches an exception will save and restore its thread’s exception state so as to preserve the exception state of its caller. This prevents common bugs in exception handling code caused by an innocent-looking function overwriting the exception being handled; it also reduces the often unwanted lifetime extension for objects that are referenced by the stack frames in the traceback.

As a general principle, a function that calls another function to perform some task should check whether the called function raised an exception, and if so, pass the exception state on to its caller. It should discard any object references that it owns, and return an error indicator, but it should not set another exception — that would overwrite the exception that was just raised, and lose important information about the exact cause of the error.

A simple example of detecting exceptions and passing them on is shown in the sum_sequence() example above. It so happens that this example doesn’t need to clean up any owned references when it detects an error. The following example function shows some error cleanup. First, to remind you why you like Python, we show the equivalent Python code:

def incr_item(dict, key):
    try:
        item = dict[key]
    except KeyError:
        item = 0
    dict[key] = item + 1

Here is the corresponding C code, in all its glory:

int
incr_item(PyObject *dict, PyObject *key)
{
    /* Objects all initialized to NULL for Py_XDECREF */
    PyObject *item = NULL, *const_one = NULL, *incremented_item = NULL;
    int rv = -1; /* Return value initialized to -1 (failure) */

    item = PyObject_GetItem(dict, key);
    if (item == NULL) {
        /* Handle KeyError only: */
        if (!PyErr_ExceptionMatches(PyExc_KeyError))
            goto error;

        /* Clear the error and use zero: */
        PyErr_Clear();
        item = PyLong_FromLong(0L);
        if (item == NULL)
            goto error;
    }
    const_one = PyLong_FromLong(1L);
    if (const_one == NULL)
        goto error;

    incremented_item = PyNumber_Add(item, const_one);
    if (incremented_item == NULL)
        goto error;

    if (PyObject_SetItem(dict, key, incremented_item) < 0)
        goto error;
    rv = 0; /* Success */
    /* Continue with cleanup code */

 error:
    /* Cleanup code, shared by success and failure path */

    /* Use Py_XDECREF() to ignore NULL references */
    Py_XDECREF(item);
    Py_XDECREF(const_one);
    Py_XDECREF(incremented_item);

    return rv; /* -1 for error, 0 for success */
}

This example represents an endorsed use of the goto statement in C! It illustrates the use of PyErr_ExceptionMatches() and PyErr_Clear() to handle specific exceptions, and the use of Py_XDECREF() to dispose of owned references that may be NULL (note the 'X' in the name; Py_DECREF() would crash when confronted with a NULL reference). It is important that the variables used to hold owned references are initialized to NULL for this to work; likewise, the proposed return value is initialized to -1 (failure) and only set to success after the final call made is successful.

Embedding Python

The one important task that only embedders (as opposed to extension writers) of the Python interpreter have to worry about is the initialization, and possibly the finalization, of the Python interpreter. Most functionality of the interpreter can only be used after the interpreter has been initialized.

The basic initialization function is Py_Initialize(). This initializes the table of loaded modules, and creates the fundamental modules builtins, __main__, and sys. It also initializes the module search path (sys.path).

Py_Initialize() does not set the “script argument list” (sys.argv). If this variable is needed by Python code that will be executed later, it must be set explicitly with a call to PySys_SetArgvEx(argc, argv, updatepath) after the call to Py_Initialize().

On most systems (in particular, on Unix and Windows, although the details are slightly different), Py_Initialize() calculates the module search path based upon its best guess for the location of the standard Python interpreter executable, assuming that the Python library is found in a fixed location relative to the Python interpreter executable. In particular, it looks for a directory named lib/pythonX.Y relative to the parent directory where the executable named python is found on the shell command search path (the environment variable PATH).

For instance, if the Python executable is found in /usr/local/bin/python, it will assume that the libraries are in /usr/local/lib/pythonX.Y. (In fact, this particular path is also the “fallback” location, used when no executable file named python is found along PATH.) The user can override this behavior by setting the environment variable PYTHONHOME, or insert additional directories in front of the standard path by setting PYTHONPATH.

The embedding application can steer the search by calling Py_SetProgramName(file) before calling Py_Initialize(). Note that PYTHONHOME still overrides this and PYTHONPATH is still inserted in front of the standard path. An application that requires total control has to provide its own implementation of Py_GetPath(), Py_GetPrefix(), Py_GetExecPrefix(), and Py_GetProgramFullPath() (all defined in Modules/getpath.c).

Sometimes, it is desirable to “uninitialize” Python. For instance, the application may want to start over (make another call to Py_Initialize()) or the application is simply done with its use of Python and wants to free memory allocated by Python. This can be accomplished by calling Py_Finalize(). The function Py_IsInitialized() returns true if Python is currently in the initialized state. More information about these functions is given in a later chapter. Notice that Py_Finalize() does not free all memory allocated by the Python interpreter, e.g. memory allocated by extension modules currently cannot be released.

Debugging Builds

Python can be built with several macros to enable extra checks of the interpreter and extension modules. These checks tend to add a large amount of overhead to the runtime so they are not enabled by default.

A full list of the various types of debugging builds is in the file Misc/SpecialBuilds.txt in the Python source distribution. Builds are available that support tracing of reference counts, debugging the memory allocator, or low-level profiling of the main interpreter loop. Only the most frequently-used builds will be described in the remainder of this section.

Compiling the interpreter with the Py_DEBUG macro defined produces what is generally meant by “a debug build” of Python. Py_DEBUG is enabled in the Unix build by adding --with-pydebug to the ./configure command. It is also implied by the presence of the not-Python-specific _DEBUG macro. When Py_DEBUG is enabled in the Unix build, compiler optimization is disabled.

In addition to the reference count debugging described below, the following extra checks are performed:

• Extra checks are added to the object allocator.
• Extra checks are added to the parser and compiler.
• Downcasts from wide types to narrow types are checked for loss of information.
• A number of assertions are added to the dictionary and set implementations. In addition, the set object acquires a test_c_api() method.
• Sanity checks of the input arguments are added to frame creation.
• The storage for ints is initialized with a known invalid pattern to catch reference to uninitialized digits.
• Low-level tracing and extra exception checking are added to the runtime virtual machine.
• Extra checks are added to the memory arena implementation.
• Extra debugging is added to the thread module.

There may be additional checks not mentioned here.

Defining Py_TRACE_REFS enables reference tracing. When defined, a circular doubly linked list of active objects is maintained by adding two extra fields to every PyObject. Total allocations are tracked as well. Upon exit, all existing references are printed. (In interactive mode this happens after every statement run by the interpreter.) Implied by Py_DEBUG.

Please refer to Misc/SpecialBuilds.txt in the Python source distribution for more detailed information.