A C extension for CPython is a shared library (for example, a .so file on Linux, .pyd DLL on Windows), which is loadable into the Python process (for example, it is compiled with compatible compiler settings), and which exports an export hook function (or an old-style
).
To be importable by default (that is, by
importlib.machinery.ExtensionFileLoader
), the shared library must be available on
, and must be named after the module name plus an extension listed in
importlib.machinery.EXTENSION_SUFFIXES
.
Extension export hook
Added in version 3.15: Support for the PyModExport_<name> export hook was added in Python 3.15. The older way of defining modules is still available: consult either the
section or earlier versions of this documentation if you plan to support earlier Python versions.
The export hook must be an exported function with the following signature:
*PyModExport_modulename(void)
For modules with ASCII-only names, the
must be named PyModExport_<name>, with <name> replaced by the module’s name.
For non-ASCII module names, the export hook must instead be named PyModExportU_<name> (note the U), with <name> encoded using Python’s punycode encoding with hyphens replaced by underscores. In Python:
defhook_name(name):try:suffix=b'_'+name.encode('ascii')exceptUnicodeEncodeError:suffix=b'U_'+name.encode('punycode').replace(b'-',b'_')returnb'PyModExport'+suffixThe export hook returns an array of
entries, terminated by an entry with a slot ID of 0. These slots describe how the module should be created and initialized.
This array must remain valid and constant until interpreter shutdown. Typically, it should use static storage. Prefer using the
and
slots for any dynamic behavior.
The export hook may return NULL with an exception set to signal failure.
It is recommended to define the export hook function using a helper macro:
PyMODEXPORT_FUNC
Part of the
since version 3.15.Declare an extension module export hook. This macro:
specifies the
* return type,
adds any special linkage declarations required by the platform, and
for C++, declares the function as extern"C".
For example, a module called spam would be defined like this:
PyABIInfo_VAR(abi_info);staticPySlotspam_slots[]={PySlot_STATIC_DATA(Py_mod_abi,&abi_info),PySlot_STATIC_DATA(Py_mod_name,"spam"),PySlot_FUNC(Py_mod_init,spam_init_function),...PySlot_END};PyMODEXPORT_FUNCPyModExport_spam(void){returnspam_slots;}The export hook is typically the only non-static item defined in the module’s C source.
The hook should be kept short. If it does more than return a static array, several caveats apply:
If you need to use any Python C API, it is recommended to call
first to raise an exception, rather than crash, in common cases of ABI mismatch.
Code in the export hook must never rely on the
:
of Python can only check the
slot (or the lack of it) after the hook returns,
Similarly, the hook may be called in any subinterpreter, since the
slot (or lack of it) is only checked after the hook returns.
For example:
PyMODEXPORT_FUNCPyModExport_modulename(void){if(PyABIInfo_Check(&abi_info,"modulename")<0){/* ABI mismatch. It's not safe to examine the raised exception. */returnNULL;}/* use Python API (as little as possible); don't rely on GIL */returnmodulename_slots;}Note
It is possible to export multiple modules from a single shared library by defining multiple export hooks. However, importing them requires a custom importer or suitably named copies/links of the extension file, because Python’s import machinery only finds the function corresponding to the filename. See the
Multiple modules in one library
section in
for details.
Multi-phase initialization
The process of creating an extension module follows several phases:
Python finds and calls the export hook to get information on how to create the module.
Before any substantial code is executed, Python can determine which capabilities the module supports, and it can adjust the environment or refuse loading an incompatible extension. Slots like
,
and
influence this step.
By default, Python itself then creates the module object – that is, it does the equivalent of calling
when creating an object. This step can be overridden using the
slot.
Python sets initial module attributes like
and
, and inserts the module object into
.
Afterwards, the module object is initialized in an extension-specific way – the equivalent of
when creating an object, or of executing top-level code in a Python-language module. The behavior is specified using the
slot.
This is called multi-phase initialization to distinguish it from the legacy (but still supported)
, where an initialization function returns a fully constructed module.
Changed in version 3.5: Added support for multi-phase initialization (
).
Multiple module instances
By default, extension modules are not singletons. For example, if the
entry is removed and the module is re-imported, a new module object is created and, typically, populated with fresh method and type objects. The old module is subject to normal garbage collection. This mirrors the behavior of pure-Python modules.
Additional module instances may be created in
or after Python runtime reinitialization (
and
). In these cases, sharing Python objects between module instances would likely cause crashes or undefined behavior.
To avoid such issues, each instance of an extension module should be isolated: changes to one instance should not implicitly affect the others, and all state owned by the module, including references to Python objects, should be specific to a particular module instance. See
for more details and a practical guide.
A simpler way to avoid these issues is
raising an error on repeated initialization
.
All modules are expected to support
, or otherwise explicitly signal a lack of support. This is usually achieved by isolation or blocking repeated initialization, as above. A module may also be limited to the main interpreter using the
slot.
PyInit function
since version 3.15: This functionality will not get new features, but there are no plans to remove it.
Instead of
, an extension module can define an older-style initialization function with the signature:
*PyInit_modulename(void)
Its name should be PyInit_<name>, with <name> replaced by the name of the module. For non-ASCII module names, use PyInitU_<name> instead, with <name> encoded in the same way as for the
(that is, using Punycode with underscores).
If a module exports both PyInit_<name> and PyModExport_<name>, the PyInit_<name> function is ignored.
Like with
, it is recommended to define the initialization function using a helper macro:
PyMODINIT_FUNC
Declare an extension module initialization function. This macro:
specifies the
* return type,
adds any special linkage declarations required by the platform, and
for C++, declares the function as extern"C".
Normally, the initialization function (PyInit_modulename) returns a
instance with non-NULL
. This allows Python to use
.
Before it is returned, the PyModuleDef instance must be initialized using the following function:
*PyModuleDef_Init(
*def)
Part of the
since version 3.5.Ensure a module definition is a properly initialized Python object that correctly reports its type and a reference count.
Return def cast to PyObject*, or NULL if an error occurred.
Calling this function is required before returning a
from a module initialization function. It should not be used in other contexts.
Note that Python assumes that PyModuleDef structures are statically allocated. This function may return either a new reference or a borrowed one; this reference must not be released.
Added in version 3.5.
For example, a module called spam would be defined like this:
staticstructPyModuleDefspam_module={.m_base=PyModuleDef_HEAD_INIT,.m_name="spam",...};PyMODINIT_FUNCPyInit_spam(void){returnPyModuleDef_Init(&spam_module);}Legacy single-phase initialization
since version 3.15: Single-phase initialization is a legacy mechanism to initialize extension modules, with known drawbacks and design flaws. Extension module authors are encouraged to use multi-phase initialization instead.
However, there are no plans to remove support for it.
In single-phase initialization, the old-style
(PyInit_modulename) should create, populate and return a module object. This is typically done using
and functions like
.
Single-phase initialization differs from the
in the following ways:
Single-phase modules are, or rather contain, “singletons”.
When the module is first initialized, Python saves the contents of the module’s __dict__ (that is, typically, the module’s functions and types).
For subsequent imports, Python does not call the initialization function again. Instead, it creates a new module object with a new __dict__, and copies the saved contents to it. For example, given a single-phase module _testsinglephase
that defines a function sum and an exception class error:
>>> importsys>>> import_testsinglephaseasone>>> delsys.modules['_testsinglephase']>>> import_testsinglephaseastwo>>> oneistwoFalse>>> one.__dict__istwo.__dict__False>>> one.sumistwo.sumTrue>>> one.erroristwo.errorTrueThe exact behavior should be considered a CPython implementation detail.
To work around the fact that PyInit_modulename does not take a spec argument, some state of the import machinery is saved and applied to the first suitable module created during the PyInit_modulename call. Specifically, when a sub-module is imported, this mechanism prepends the parent package name to the name of the module.
A single-phase PyInit_modulename function should create “its” module object as soon as possible, before any other module objects can be created.
Non-ASCII module names (PyInitU_modulename) are not supported.
Single-phase modules support module lookup functions like
.
The module’s
must be NULL.