Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

Supporting Free-Threaded CPython

CPython 3.13 introduces an experimental "free-threaded" build of CPython that does not rely on the global interpreter lock (often referred to as the GIL) for thread safety. As of version 0.23, PyO3 also has preliminary support for building Rust extensions for the free-threaded Python build and support for calling into free-threaded Python from Rust.

If you want more background on free-threaded Python in general, see the what's new entry in the 3.13 release notes, the free-threading HOWTO guide in the CPython docs, the extension porting guide in the community-maintained Python free-threading guide, and PEP 703, which provides the technical background for the free-threading implementation in CPython.

In the GIL-enabled build, the global interpreter lock serializes access to the Python runtime. The GIL is therefore a fundamental limitation to parallel scaling of multithreaded Python workflows, due to Amdahl's law, because any time spent executing a parallel processing task on only one execution context fundamentally cannot be sped up using parallelism.

The free-threaded build removes this limit on multithreaded Python scaling. This means it's much more straightforward to achieve parallelism using the Python threading module. If you have ever needed to use multiprocessing to achieve a parallel speedup for some Python code, free-threading will likely allow the use of Python threads instead for the same workflow.

PyO3's support for free-threaded Python will enable authoring native Python extensions that are thread-safe by construction, with much stronger safety guarantees than C extensions. Our goal is to enable "fearless concurrency" in the native Python runtime by building on the Rust Send and Sync traits.

This document provides advice for porting Rust code using PyO3 to run under free-threaded Python.

Supporting free-threaded Python with PyO3

Many simple uses of PyO3, like exposing bindings for a "pure" Rust function with no side-effects or defining an immutable Python class, will likely work "out of the box" on the free-threaded build. All that will be necessary is to annotate Python modules declared by rust code in your project to declare that they support free-threaded Python, for example by declaring the module with #[pymodule(gil_used = false)].

More complicated #[pyclass] types may need to deal with thread-safety directly; there is a dedicated section of the guide to discuss this.

At a low-level, annotating a module sets the Py_MOD_GIL slot on modules defined by an extension to Py_MOD_GIL_NOT_USED, which allows the interpreter to see at runtime that the author of the extension thinks the extension is thread-safe. You should only do this if you know that your extension is thread-safe. Because of Rust's guarantees, this is already true for many extensions, however see below for more discussion about how to evaluate the thread safety of existing Rust extensions and how to think about the PyO3 API using a Python runtime with no GIL.

If you do not explicitly mark that modules are thread-safe, the Python interpreter will re-enable the GIL at runtime while importing your module and print a RuntimeWarning with a message containing the name of the module causing it to re-enable the GIL. You can force the GIL to remain disabled by setting the PYTHON_GIL=0 as an environment variable or passing -Xgil=0 when starting Python (0 means the GIL is turned off).

If you are sure that all data structures exposed in a PyModule are thread-safe, then pass gil_used = false as a parameter to the pymodule procedural macro declaring the module or call PyModule::gil_used on a PyModule instance. For example:

use pyo3::prelude::*;

/// This module supports free-threaded Python
#[pymodule(gil_used = false)]
fn my_extension(m: &Bound<'_, PyModule>) -> PyResult<()> {
    // add members to the module that you know are thread-safe
    Ok(())
}

Or for a module that is set up without using the pymodule macro:

use pyo3::prelude::*;

#[allow(dead_code)]
fn register_child_module(parent_module: &Bound<'_, PyModule>) -> PyResult<()> {
    let child_module = PyModule::new(parent_module.py(), "child_module")?;
    child_module.gil_used(false)?;
    parent_module.add_submodule(&child_module)
}

For now you must explicitly opt in to free-threading support by annotating modules defined in your extension. In a future version of PyO3, we plan to make gil_used = false the default.

See the string-sum example for how to declare free-threaded support using raw FFI calls for modules using single-phase initialization and the sequential example for modules using multi-phase initialization.

If you would like to use conditional compilation to trigger different code paths under the free-threaded build, you can use the Py_GIL_DISABLED attribute once you have configured your crate to generate the necessary build configuration data. See the guide section for more details about supporting multiple different Python versions, including the free-threaded build.

Special considerations for the free-threaded build

The free-threaded interpreter does not have a GIL. Many existing extensions providing mutable data structures relied on the GIL to lock Python objects and make interior mutability thread-safe. Historically, PyO3's API was designed around the same strong assumptions, but is transitioning towards more general APIs applicable for both builds.

Calling into the CPython C API is only legal when an OS thread is explicitly attached to the interpreter runtime. In the GIL-enabled build, this happens when the GIL is acquired. In the free-threaded build there is no GIL, but the same C macros that release or acquire the GIL in the GIL-enabled build instead ask the interpreter to attach the thread to the Python runtime, and there can be many threads simultaneously attached. See PEP 703 for more background about how threads can be attached and detached from the interpreter runtime, in a manner analogous to releasing and acquiring the GIL in the GIL-enabled build.

In the GIL-enabled build, PyO3 uses the Python<'py> type and the 'py lifetime to signify that the global interpreter lock is held. In the freethreaded build, holding a 'py lifetime means only that the thread is currently attached to the Python interpreter -- other threads can be simultaneously interacting with the interpreter.

Attaching to the runtime

You still need to obtain a 'py lifetime to interact with Python objects or call into the CPython C API. If you are not yet attached to the Python runtime, you can register a thread using the Python::attach function. Threads created via the Python threading module do not need to do this, and pyo3 will handle setting up the Python<'py> token when CPython calls into your extension.

Detaching to avoid hangs and deadlocks

The free-threaded build triggers global synchronization events in the following situations:

  • During garbage collection in order to get a globally consistent view of reference counts and references between objects
  • In Python 3.13, when the first background thread is started in order to mark certain objects as immortal
  • When either sys.settrace or sys.setprofile are called in order to instrument running code objects and threads
  • During a call to os.fork(), to ensure a process-wide consistent state.

This is a non-exhaustive list and there may be other situations in future Python versions that can trigger global synchronization events.

This means that you should detach from the interpreter runtime using Python::detach in exactly the same situations as you should detach from the runtime in the GIL-enabled build: when doing long-running tasks that do not require the CPython runtime or when doing any task that needs to re-attach to the runtime (see the guide section that covers this). In the former case, you would observe a hang on threads that are waiting on the long-running task to complete, and in the latter case you would see a deadlock while a thread tries to attach after the runtime triggers a global synchronization event, but the spawning thread prevents the synchronization event from completing.

Exceptions and panics for multithreaded access of mutable pyclass instances

Data attached to pyclass instances is protected from concurrent access by a RefCell-like pattern of runtime borrow checking. Like a RefCell, PyO3 will raise exceptions (or in some cases panic) to enforce exclusive access for mutable borrows. It was always possible to generate panics like this in PyO3 in code that releases the GIL with Python::detach or calling a python method accepting &self from a &mut self (see the docs on interior mutability,) but now in free-threaded Python there are more opportunities to trigger these panics from Python because there is no GIL to lock concurrent access to mutably borrowed data from Python.

The most straightforward way to trigger this problem is to use the Python threading module to simultaneously call a rust function that mutably borrows a pyclass in multiple threads. For example, consider the following implementation:

use pyo3::prelude::*;
#[pyclass]
#[derive(Default)]
struct ThreadIter {
    count: usize,
}

#[pymethods]
impl ThreadIter {
    #[new]
    pub fn new() -> Self {
        Default::default()
    }

    fn __next__(&mut self, py: Python<'_>) -> usize {
        self.count += 1;
        self.count
    }
}

And then if we do something like this in Python:

import concurrent.futures
from my_module import ThreadIter

i = ThreadIter()

def increment():
    next(i)

with concurrent.futures.ThreadPoolExecutor(max_workers=16) as tpe:
    futures = [tpe.submit(increment) for _ in range(100)]
    [f.result() for f in futures]

We will see an exception:

Traceback (most recent call last)
  File "example.py", line 5, in <module>
    next(i)
RuntimeError: Already borrowed

We may allow user-selectable semantics for mutable pyclass definitions in a future version of PyO3, allowing some form of opt-in locking to emulate the GIL if that is needed. For now you should explicitly add locking, possibly using conditional compilation or using the critical section API, to avoid creating deadlocks with the GIL.

Cannot build extensions using the limited API

The free-threaded build uses a completely new ABI and there is not yet an equivalent to the limited API for the free-threaded ABI. That means if your crate depends on PyO3 using the abi3 feature or an an abi3-pyxx feature, PyO3 will print a warning and ignore that setting when building extensions using the free-threaded interpreter.

This means that if your package makes use of the ABI forward compatibility provided by the limited API to upload only one wheel for each release of your package, you will need to update your release procedure to also upload a version-specific free-threaded wheel.

See the guide section for more details about supporting multiple different Python versions, including the free-threaded build.

Thread-safe single initialization

To initialize data exactly once, use the PyOnceLock type, which is a close equivalent to std::sync::OnceLock that also helps avoid deadlocks by detaching from the Python interpreter when threads are blocking waiting for another thread to complete intialization. If already using OnceLock and it is impractical to replace with a PyOnceLock, there is the OnceLockExt extension trait which adds OnceLockExt::get_or_init_py_attached to detach from the interpreter when blocking in the same fashion as PyOnceLock. Here is an example using PyOnceLock to single-initialize a runtime cache holding a Py<PyDict>:

use pyo3::prelude::*;
use pyo3::sync::PyOnceLock;
use pyo3::types::PyDict;

let cache: PyOnceLock<Py<PyDict>> = PyOnceLock::new();

Python::attach(|py| {
    // guaranteed to be called once and only once
    cache.get_or_init(py, || PyDict::new(py).unbind())
});

In cases where a function must run exactly once, you can bring the OnceExt trait into scope. The OnceExt trait adds OnceExt::call_once_py_attached and OnceExt::call_once_force_py_attached functions to the api of std::sync::Once, enabling use of Once in contexts where the thread is attached to the Python interpreter. These functions are analogous to Once::call_once, Once::call_once_force except they accept a Python<'py> token in addition to an FnOnce. All of these functions detach from the interpreter before blocking and re-attach before executing the function, avoiding deadlocks that are possible without using the PyO3 extension traits. Here the same example as above built using a Once instead of a PyOnceLock:

use pyo3::prelude::*;
use std::sync::Once;
use pyo3::sync::OnceExt;
use pyo3::types::PyDict;

struct RuntimeCache {
    once: Once,
    cache: Option<Py<PyDict>>
}

let mut cache = RuntimeCache {
    once: Once::new(),
    cache: None
};

Python::attach(|py| {
    // guaranteed to be called once and only once
    cache.once.call_once_py_attached(py, || {
        cache.cache = Some(PyDict::new(py).unbind());
    });
});

GILProtected is not exposed

GILProtected is a (deprecated) PyO3 type that allows mutable access to static data by leveraging the GIL to lock concurrent access from other threads. In free-threaded Python there is no GIL, so you will need to replace this type with some other form of locking. In many cases, a type from std::sync::atomic or a std::sync::Mutex will be sufficient.

Before:

#![allow(deprecated)]
fn main() {
#[cfg(not(Py_GIL_DISABLED))] {
use pyo3::prelude::*;
use pyo3::sync::GILProtected;
use pyo3::types::{PyDict, PyNone};
use std::cell::RefCell;

static OBJECTS: GILProtected<RefCell<Vec<Py<PyDict>>>> =
    GILProtected::new(RefCell::new(Vec::new()));

Python::attach(|py| {
    // stand-in for something that executes arbitrary Python code
    let d = PyDict::new(py);
    d.set_item(PyNone::get(py), PyNone::get(py)).unwrap();
    OBJECTS.get(py).borrow_mut().push(d.unbind());
});
}}

After:

use pyo3::prelude::*;
fn main() {
use pyo3::types::{PyDict, PyNone};
use std::sync::Mutex;

static OBJECTS: Mutex<Vec<Py<PyDict>>> = Mutex::new(Vec::new());

Python::attach(|py| {
    // stand-in for something that executes arbitrary Python code
    let d = PyDict::new(py);
    d.set_item(PyNone::get(py), PyNone::get(py)).unwrap();
    // as with any `Mutex` usage, lock the mutex for as little time as possible
    // in this case, we do it just while pushing into the `Vec`
    OBJECTS.lock().unwrap().push(d.unbind());
});
}

If you are executing arbitrary Python code while holding the lock, then you should import the MutexExt trait and use the lock_py_attached method instead of lock. This ensures that global synchronization events started by the Python runtime can proceed, avoiding possible deadlocks with the interpreter.