PyConfig is not part of the limited C API

When the first versions of PEP 587 “Python Initialization Configuration” were discussed, there was a private field _config_version (int): the configuration version, used for ABI compatibility. It was decided that if an application embeds Python, it sticks to a Python version anyway, and so there is no need to bother with the ABI compatibility.

The final PyConfig API of PEP 587 is excluded from the limited C API since its main PyConfig structure is not versioned. Python cannot guarantee ABI backward and forward compatibility, it’s incompatible with the stable ABI.

Since PyConfig was added to Python 3.8, the limited C API and the stable ABI are getting more popular. For example, Rust bindings such as the PyO3 project can target the limited C API to embed Python in Rust (but it’s not the default). In practice, PyO3 can use non-limited C API for specific needs, but using them avoids the stable ABI advantages.

Limitations of the legacy API

The legacy API to configure the Python initialization is based on the legacy Py_Initialize() function. It is now mostly deprecated:

• Set the initialization configuration such as Py_SetPath(): deprecated in Python 3.11 and removed in Python 3.13.
• Global configuration variables such as Py_VerboseFlag: deprecated in Python 3.12 and scheduled for removal in Python 3.14.
• Get the current configuration such as Py_GetPath(): deprecated in Python 3.13 and scheduled for removal in Python 3.15.

The legacy API doesn’t support the “Python Configuration” and the “Isolated Configuration” of PEP 587 PyConfig API. It only provides a “legacy configuration” of Py_Initialize() which is in the between, and also uses the legacy global configuration variables (such as Py_VerboseFlag).

Some options can only by set by environment variables, such as home set by the PYTHONHOME environment variable. The problem is that environment variables are inherited by child processes which can be a surprising and unwanted behavior.

Some configuration options, such as configure_locale, simply cannot be set.

Limitations of the limited C API

The limited C API is a subset of the legacy API. For example, global configuration variables, such as Py_VerboseFlag, are not part of the limited C API.

While some functions were removed from the limited C API version 3.13, they are still part of the stable ABI. For example, building a application with the limited C API version 3.12 can still run with Python 3.13 stable ABI.

Get the runtime configuration

PEP 587 has no API to get the current runtime configuration, only to configure the Python initialization.

For example, the global configuration variable Py_UnbufferedStdioFlag was deprecated in Python 3.12 and using PyConfig.buffered_stdio is recommended instead. It only works to configure Python, there is no public API to get PyConfig.buffered_stdio.

Users of the limited C API are asking for a public API to get the current runtime configuration.

Cython needs to get the optimization_level configuration option: issue.

When global configuration variables were deprecated in 2022, Marc-André Lemburg requested a C API to access these configuration variables at runtime (not only during Python initialization).

Security fix

To fix CVE-2020-10735, a denial-of-service when converting very a large string to an integer (in base 10), it was discussed to add a new PyConfig member to stable branches which affects the ABI.

Gregory P. Smith proposed a different API using text based configuration file to not be limited by PyConfig members: FR: Allow private runtime config to enable extending without breaking the PyConfig ABI (August 2022).

In the end, it was decided to not add a new PyConfig member to stable branches, but only add a new PyConfig.int_max_str_digits member to the development branch (which became Python 3.12). A dedicated private global variable (unrelated to PyConfig) is used in stable branches.

Redundancy between PyPreConfig and PyConfig

The Python preinitialization uses the PyPreConfig structure and the Python initialization uses the PyConfig structure. Both structures have four duplicated members: dev_mode, parse_argv, isolated and use_environment.

The redundancy is caused by the fact that the two structures are separated, whereas some PyConfig members are needed by the preinitialization.

Embedding Python

Usage of a stable ABI

Ronald Oussoren:

For tools like py2app/py2exe/pyinstaller, it is pretty inconvenient to have to rebuild the launcher executable that’s used to start the packaged application when there’s a bug fix release of Python.

Gregory P. Smith:

You can’t extend a struct and assume embedding people all rebuild. They don’t. Real world embedding uses exist that use an installed Python minor version as a shared library. Update that to use a different sized struct in a public API and someone is going to have a bad time. That’s why I consider the struct frozen at rc1 time, even when only for use in the embedding / writing their own launcher case.

Colton Murphy:

I am trying to embed the Python interpreter using a non C language. I have to stick with the limited API and private structures for configuration in headers files is a no-no. Basically, I need to be able to allocate and configure everything using only exportable functions and the heap… no private structure details.

(…)

I am strictly limited to what’s in the shared library (DLL). I don’t have headers, I can’t statically “recompile” every time a new version of python comes out. That’s unmaintainable for me.

Quotes of Milian Wolff’s message:

Our application is a large complex C++ code base with lots of dependencies targeting all three major desktop platforms.

Originally, we hoped to be able to use the stable python ABI to allow biologists to “bring your own python”. The idea was that they probably have a custom set of python libraries and code that they would like to continue using. Our integrated API - so we thought - was a tiny addition that should work with any Python out there, so we used the stable ABI.

This turned out to be a dead end, and I believe we can (should?) now use the non-stable ABI of python. Allowing end users to BYO Python caused far too much setup problems and support issues for us that it was not worth it in the end. Instead, we now rather want to ship a custom Python with a custom prefix that they can pip install custom libraries into as needed.

The problems we faced are not directly related to the stable ABI - quite the contrary. Rather, it was due to thirdparty python libraries that we shipped which themselves are not compatible across python version increments. E.g. for the integrated console we use qtconsole/jupyter, which worked in an archaic version with python 3.9 but requires newer versions for python 3.11+.

The ton of dependencies pulled in by UMAP was even worse, with numba and pydnndescent and llvmlite often taking months to support newer Python versions.

David Hewitt of the PyO3 project:

I think making the configuration structure opaque and using an API to set/get configuration by name is a welcome simplification:

• It’s a smaller API for language bindings like PyO3 to wrap and re-expose, and
• It’s easier for people to support multiple Python versions to embed into their application; no need to conditionally compile structure field access, can just use normal error handling if configuration values are not available for a specific version at runtime.

Quotes of Paul P. message:

I cannot agree more, it is the same story everywhere/every time CPython must be embedded. I maintened a runtime+ecosystem for Android 4.4+ for some time (in order more comfortably use Panda3D standalone than with Kivy), patching CPython and making a CI for it was ok.

But I had to give up, because I had often to recompile every known modules: this is not sustainable for one individual.

So I dropped the Android arch to only go WebAssembly (Emscripten). But same (hard and boring) problem as always: have to rebuild numerous packages that are commonly used with 2D/3D framework. (…)

Except for ONE, Harfang3d. I did not rebuild this one since Python 3.11 initial port… Guess why? it is a limited C API - abi3 module!

Limited API abi3 are fresh air, fast and portable. And associated with a stable config runtime, it would be just perfect way!

See also issue gh-116139 building an application embedding Python 3.11 and attempting to run it with Python 3.10: it does crash because the PyConfig structure ABI is not stable between two Python 3.x minor releases.

Set the runtime configuration

Marc-André Lemburg requested a C API to set the value of some configuration options at runtime:

• optimization_level
• verbose
• parser_debug
• inspect
• write_bytecode

Previously, it was possible to set directly global configuration variables:

• Py_OptimizeFlag
• Py_VerboseFlag
• Py_DebugFlag
• Py_InspectFlag
• Py_DontWriteBytecodeFlag

But these configuration flags were deprecated in Python 3.12 and are scheduled for removal in Python 3.14.

Configuration Options

Configuration options are named after PyPreConfig and PyConfig structure members. See the PyPreConfig documentation and the PyConfig documentation.

Deprecating and removing configuration options is out of the scope of the PEP and should be discussed on a case by case basis.

Preinitialization

Calling Py_PreInitializeFromInitConfig() preinitializes Python. For example, it sets the memory allocator, and can configure the LC_CTYPE locale and configure the standard C streams such as stdin and stdout.

The following options can only be set during the Python preinitialization:

• allocator,
• coerce_c_locale,
• coerce_c_locale_warn,
• configure_locale,
• legacy_windows_fs_encoding,
• utf8_mode.

Trying to set these options after Python preinitialization fails with an error.

PyInitConfig_SetStrLocale() and PyInitConfig_SetStrLocaleList() functions call Py_PreInitializeFromInitConfig() if Python is not already preinitialized.

Create Config

PyInitConfig structure:

Opaque structure to configure the Python preinitialization and the Python initialization.

PyInitConfig* PyInitConfig_CreatePython(void):

Create a new initialization configuration using default values of the Python Configuration.

It must be freed with PyInitConfig_Free().

Return NULL on memory allocation failure.

PyInitConfig* PyInitConfig_CreateIsolated(void):

Similar to PyInitConfig_CreatePython(), but use default values of the Isolated Configuration.

void PyInitConfig_Free(PyInitConfig *config):

Free memory of an initialization configuration.

Get Options

The configuration option name parameter must be a non-NULL null-terminated UTF-8 encoded string.

int PyInitConfig_HasOption(PyInitConfig *config, const char *name):

Test if the configuration has an option called name.

Return 1 if the option exists, or return 0 otherwise.

int PyInitConfig_GetInt(PyInitConfig *config, const char *name, int64_t *value):

Get an integer configuration option.

• Set *value, and return 0 on success.
• Set an error in config and return -1 on error.

int PyInitConfig_GetStr(PyInitConfig *config, const char *name, char **value):

Get a string configuration option as a null-terminated UTF-8 encoded string.

• Set *value, and return 0 on success.
• Set an error in config and return -1 on error.

On success, the string must be released with free(value).

int PyInitConfig_GetWStr(PyInitConfig *config, const char *name, wchar_t **value):

Get a string configuration option as a null-terminated wide string.

• Set *value and return 0 on success.
• Set an error in config and return -1 on error.

On success, the string must be released with free(value).

int PyInitConfig_GetStrList(PyInitConfig *config, const char *name, size_t *length, char ***items):

Get a string list configuration option as an array of null-terminated UTF-8 encoded strings.

• Set *length and *value, and return 0 on success.
• Set an error in config and return -1 on error.

On success, the string list must be released with PyInitConfig_FreeStrList(length, items).

void PyInitConfig_FreeStrList(size_t length, char **items):

Free memory of a string list created by PyInitConfig_GetStrList().

int PyInitConfig_GetWStrList(PyInitConfig *config, const char *name, size_t *length, wchar_t ***items):

Get a string list configuration option as an array of null-terminated wide strings.

• Set *length and *value, and return 0 on success.
• Set an error in config and return -1 on error.

On success, the string list must be released with PyInitConfig_FreeWStrList(length, items).

void PyInitConfig_FreeWStrList(size_t length, wchar_t **items):

Free memory of a string list created by PyInitConfig_GetWStrList().

Set Options

The configuration option name parameter must be a non-NULL null-terminated UTF-8 encoded string.

Some configuration options have side effects on other options. This logic is only implemented when Py_InitializeFromInitConfig() is called, not by the “Set” functions below. For example, setting dev_mode to 1 does not set faulthandler to 1.

int PyInitConfig_SetInt(PyInitConfig *config, const char *name, int64_t value):

Set an integer configuration option.

• Return 0 on success.
• Set an error in config and return -1 on error.

int PyInitConfig_SetStr(PyInitConfig *config, const char *name, const char *value):

Set a string configuration option from a null-terminated UTF-8 encoded string. The string is copied.

• Return 0 on success.
• Set an error in config and return -1 on error.

int PyInitConfig_SetStrLocale(PyInitConfig *config, const char *name, const char *value):

Set a string configuration option from a null-terminated bytes string encoded in the locale encoding. The string is copied.

The bytes string is decoded by Py_DecodeLocale(). Py_PreInitializeFromInitConfig() must be called before calling this function.

• Return 0 on success.
• Set an error in config and return -1 on error.

int PyInitConfig_SetWStr(PyInitConfig *config, const char *name, const wchar_t *value):

Set a string configuration option from a null-terminated wide string. The string is copied.

• Return 0 on success.
• Set an error in config and return -1 on error.

int PyInitConfig_SetStrList(PyInitConfig *config, const char *name, size_t length, char * const *items):

Set a string list configuration option from an array of null-terminated UTF-8 encoded strings. The string list is copied.

• Return 0 on success.
• Set an error in config and return -1 on error.

int PyInitConfig_SetStrLocaleList(PyInitConfig *config, const char *name, size_t length, char * const *items):

Set a string list configuration option from an array of null-terminated bytes strings encoded in the locale encoding. The string list is copied.

The bytes string is decoded by Py_DecodeLocale(). Py_PreInitializeFromInitConfig() must be called before calling this function.

• Return 0 on success.
• Set an error in config and return -1 on error.

int PyInitConfig_SetWStrList(PyInitConfig *config, const char *name, size_t length, wchar_t * const *items):

Set a string list configuration option from an error of null-terminated wide strings. The string list is copied.

• Return 0 on success.
• Set an error in config and return -1 on error.

int PyInitConfig_AddModule(PyInitConfig *config, const char *name, PyObject* (*initfunc)(void)):

Add a built-in extension module to the table of built-in modules.

The new module can be imported by the name name, and uses the function initfunc as the initialization function called on the first attempted import.

• Return 0 on success.
• Set an error in config and return -1 on error.

If Python is initialized multiple times, PyInitConfig_AddModule() must be called at each Python initialization.

Similar to the PyImport_AppendInittab() function.

Initialize Python

int Py_PreInitializeFromInitConfig(PyInitConfig *config):

Preinitialize Python from the initialization configuration.

• Return 0 on success.
• Set an error in config and return -1 on error.

int Py_InitializeFromInitConfig(PyInitConfig *config):

Initialize Python from the initialization configuration.

• Return 0 on success.
• Set an error in config and return -1 on error.
• Set an exit code in config and return -1 if Python wants to exit.

See PyInitConfig_GetExitcode() for the exitcode case.

Error Handling

int PyInitConfig_GetError(PyInitConfig* config, const char **err_msg):

Get the config error message.

• Set *err_msg and return 1 if an error is set.
• Set *err_msg to NULL and return 0 otherwise.

An error message is an UTF-8 encoded string.

If config has an exit code, format the exit code as an error message.

The error message remains valid until another PyInitConfig function is called with config. The caller doesn’t have to free the error message.

int PyInitConfig_GetExitcode(PyInitConfig* config, int *exitcode):

Get the config exit code.

• Set *exitcode and return 1 if Python wants to exit.
• Return 0 if config has no exit code set.

Only the Py_InitializeFromInitConfig() function can set an exit code if the parse_argv option is non-zero. For example, an isolated configuration cannot set an exit code by default, since parse_argv is zero by default.

An exit code can be set when parsing the command line failed (exit code 2) or when a command line option asks to display the command line help (exit code 0).

Get and Set the Runtime Configuration

The configuration option name parameter must be a non-NULL null-terminated UTF-8 encoded string.

PyObject* PyConfig_Get(const char *name):

Get the current runtime value of a configuration option as an object.

• Return a new reference on success.
• Set an exception and return NULL on error.

The object type depends on the option: see Configuration Options tables.

Other options are get from internal PyPreConfig and PyConfig structures.

The caller must hold the GIL. The function cannot be called before Python initialization nor after Python finalization.

int PyConfig_GetInt(const char *name, int *value):

Similar to PyConfig_Get(), but get the value as an integer.

• Set *value and return 0 success.
• Set an exception and return -1 on error.

PyObject* PyConfig_Names(void):

Get all configuration option names as a frozenset.

Set an exception and return NULL on error.

The caller must hold the GIL.

PyObject* PyConfig_Set(const char *name, PyObject *value):

Set the current runtime value of a configuration option.

• Raise a ValueError if there is no option name.
• Raise a ValueError if value is an invalid value.
• Raise a ValueError if the option is read-only: cannot be set.
• Raise a TypeError if value has not the proper type.

Read-only configuration options cannot be set.

The caller must hold the GIL. The function cannot be called before Python initialization nor after Python finalization.

Scope of the stable ABI

The limited C API and the stable ABI added by this PEP only provide a stable interface to program the Python initialization.

The behavior of options, the default option values, and the Python behavior can change at each Python version: they are not “stable”.

Moreover, configuration options can be added, deprecated and removed following the usual PEP 387 deprecation process.

Initialize Python

Example initializing Python, set configuration options of different types, return -1 on error:

int init_python(void)
{
    PyInitConfig *config = PyInitConfig_CreatePython();
    if (config == NULL) {
        printf("PYTHON INIT ERROR: memory allocation failed\n");
        return -1;
    }

    // Set an integer (dev mode)
    if (PyInitConfig_SetInt(config, "dev_mode", 1) < 0) {
        goto error;
    }

    // Set a list of wide strings (argv)
    wchar_t *argv[] = {L"my_program", L"-c", L"pass"};
    if (PyInitConfig_SetWStrList(config, "argv",
                                 Py_ARRAY_LENGTH(argv), argv) < 0) {
        goto error;
    }

    // Set a wide string (program name)
    if (PyInitConfig_SetWStr(config, "program_name", L"my_program") < 0) {
        goto error;
    }

    // Set a list of bytes strings (xoptions).
    // Preinitialize implicitly Python to decode the bytes string.
    char* xoptions[] = {"faulthandler"};
    if (PyInitConfig_SetStrList(config, "xoptions",
                                Py_ARRAY_LENGTH(xoptions), xoptions) < 0) {
        goto error;
    }

    // Initialize Python with the configuration
    if (Py_InitializeFromInitConfig(config) < 0) {
        goto error;
    }
    PyInitConfig_Free(config);
    return 0;

error:
    // Display the error message
    const char *err_msg;
    (void)PyInitConfig_GetError(config, &err_msg);
    printf("PYTHON INIT ERROR: %s\n", err_msg);
    PyInitConfig_Free(config);

    return -1;
}

Increase initialization bytes_warning option

Example increasing the bytes_warning option of an initialization configuration:

int config_bytes_warning(PyInitConfig *config)
{
    int bytes_warning;
    if (PyInitConfig_GetInt(config, "bytes_warning", &bytes_warning)) {
        return -1;
    }
    bytes_warning += 1;
    if (PyInitConfig_SetInt(config, "bytes_warning", bytes_warning)) {
        return -1;
    }
    return 0;
}

Get the runtime verbose option

Example getting the current runtime value of the configuration option verbose:

int get_verbose(void)
{
    int verbose;
    if (PyConfig_GetInt("verbose", &verbose) < 0) {
        // Silently ignore the error
        PyErr_Clear();
        return -1;
    }
    return verbose;
}

On error, the function silently ignores the error and returns -1. In practice, getting the verbose option cannot fail, unless a future Python version removes the option.

Configuration as text

It was proposed to provide the configuration as text to make the API compatible with the stable ABI and to allow custom options.

Example:

# integer
bytes_warning = 2

# string
filesystem_encoding = "utf8"   # comment

# list of strings
argv = ['python', '-c', 'code']

The API would take the configuration as a string, not as a file. Example with a hypothetical PyInit_SetConfig() function:

void stable_abi_init_demo(int set_path)
{
    PyInit_SetConfig(
        "isolated = 1\n"
        "argv = ['python', '-c', 'code']\n"
        "filesystem_encoding = 'utf-8'\n"
    );
    if (set_path) {
        PyInit_SetConfig("pythonpath = '/my/path'");
    }
}

The example ignores error handling to make it easier to read.

The problem is that generating such configuration text requires adding quotes to strings and to escape quotes in strings. Formatting an array of strings becomes non-trivial.

Providing an API to format a string or an array of strings is not really worth it, whereas Python can provide directly an API to set a configuration option where the value is passed directly as a string or an array of strings. It avoids giving special meaning to some characters, such as newline characters, which would have to be escaped.

Refer to an option with an integer

Using strings to refer to a configuration option requires comparing strings which can be slower than comparing integers.

Use integers, similar to type “slots” such as Py_tp_doc, to refer to a configuration option. The const char *name parameter is replaced with int option.

Accepting custom options is more likely to cause conflicts when using integers, since it’s harder to maintain “namespaces” (ranges) for integer options. Using strings, a simple prefix with a colon separator can be used.

Integers also requires maintaining a list of integer constants and so make the C API and the Python API larger.

Python 3.13 only has around 62 configuration options, and so performance is not really a blocker issue. If better performance is needed later, a hash table can be used to get an option by its name.

If getting a configuration option is used in hot code, the value can be read once and cached. By the way, most configuration options cannot be changed at runtime.

Fully remove the preinitialization

Multi-phase initialization (similar to PEP 432)

Eric Snow expressed concerns that this proposal might reinforce with embedders the idea that initialization is a single monolithic step. He argued that initialization involves 5 distinct phases and even suggested that the API should reflect this explicitly. Eric proposed that, at the very least, the implementation of initialization should reflect the phases, in part for improved code health. Overall, his explanation has some similarities with PEP 432 and PEP 587.

Another of Eric’s key points relevant to this PEP was that, ideally, the config passed to Py_InitializeFromConfig() should be complete before that function is called, whereas currently initialization actually modifies the config.

While Eric wasn’t necessarily suggesting an alternative to PEP 741, any proposal to add a granular initialization API around phases is effectively the opposite of what this PEP is trying to accomplish. Such API is more complicated, it requires adding new public structures and new public functions. It makes the Python initialization more complicated, rather than this PEP tries to unifiy existing APIs and make them simpler (the opposite). Having multiple structures for similar purpose can lead to duplicate members, similar issue than duplicated members between existing PyPreConfig and PyConfig structures.