pub struct Python<'py>(_);
Expand description
A marker token that represents holding the GIL.
It serves three main purposes:
- It provides a global API for the Python interpreter, such as
Python::eval
. - It can be passed to functions that require a proof of holding the GIL, such as
Py::clone_ref
. - Its lifetime represents the scope of holding the GIL which can be used to create Rust
references that are bound to it, such as
&
PyAny
.
Note that there are some caveats to using it that you might need to be aware of. See the Deadlocks and Releasing and freeing memory paragraphs for more information about that.
Obtaining a Python token
The following are the recommended ways to obtain a Python
token, in order of preference:
- In a function or method annotated with
#[pyfunction]
or#[pymethods]
you can declare it as a parameter, and PyO3 will pass in the token when Python code calls it. - If you already have something with a lifetime bound to the GIL, such as
&
PyAny
, you can use its.py()
method to get a token. - When you need to acquire the GIL yourself, such as when calling Python code from Rust, you
should call
Python::with_gil
to do that and pass your code as a closure to it.
Deadlocks
Note that the GIL can be temporarily released by the Python interpreter during a function call (e.g. importing a module). In general, you don’t need to worry about this because the GIL is reacquired before returning to the Rust code:
`Python` exists |=====================================|
GIL actually held |==========| |================|
Rust code running |=======| |==| |======|
This behaviour can cause deadlocks when trying to lock a Rust mutex while holding the GIL:
- Thread 1 acquires the GIL
- Thread 1 locks a mutex
- Thread 1 makes a call into the Python interpreter which releases the GIL
- Thread 2 acquires the GIL
- Thread 2 tries to locks the mutex, blocks
- Thread 1’s Python interpreter call blocks trying to reacquire the GIL held by thread 2
To avoid deadlocking, you should release the GIL before trying to lock a mutex or await
ing in
asynchronous code, e.g. with Python::allow_threads
.
Releasing and freeing memory
The Python
type can be used to create references to variables owned by the Python
interpreter, using functions such as Python::eval
and PyModule::import
. These
references are tied to a GILPool
whose references are not cleared until it is dropped.
This can cause apparent “memory leaks” if it is kept around for a long time.
use pyo3::prelude::*;
use pyo3::types::PyString;
Python::with_gil(|py| -> PyResult<()> {
for _ in 0..10 {
let hello: &PyString = py.eval("\"Hello World!\"", None, None)?.extract()?;
println!("Python says: {}", hello.to_str()?);
// Normally variables in a loop scope are dropped here, but `hello` is a reference to
// something owned by the Python interpreter. Dropping this reference does nothing.
}
Ok(())
})
// This is where the `hello`'s reference counts start getting decremented.
The variable hello
is dropped at the end of each loop iteration, but the lifetime of the
pointed-to memory is bound to Python::with_gil
’s GILPool
which will not be dropped until
the end of Python::with_gil
’s scope. Only then is each hello
’s Python reference count
decreased. This means that at the last line of the example there are 10 copies of hello
in
Python’s memory, not just one at a time as we might expect from Rust’s scoping rules.
See the Memory Management chapter of the guide for more information about how PyO3 uses
GILPool
to manage memory.
Implementations§
source§impl Python<'_>
impl Python<'_>
sourcepub fn with_gil<F, R>(f: F) -> Rwhere
F: for<'py> FnOnce(Python<'py>) -> R,
pub fn with_gil<F, R>(f: F) -> Rwhere F: for<'py> FnOnce(Python<'py>) -> R,
Acquires the global interpreter lock, allowing access to the Python interpreter. The
provided closure F
will be executed with the acquired Python
marker token.
If implementing #[pymethods]
or #[pyfunction]
,
declare py: Python
as an argument. PyO3 will pass in the token to grant access to the GIL
context in which the function is running, avoiding the need to call with_gil
.
If the auto-initialize
feature is enabled and the Python runtime is not already
initialized, this function will initialize it. See
prepare_freethreaded_python
for details.
If the current thread does not yet have a Python “thread state” associated with it,
a new one will be automatically created before F
is executed and destroyed after F
completes.
Panics
- If the
auto-initialize
feature is not enabled and the Python interpreter is not initialized.
Examples
use pyo3::prelude::*;
Python::with_gil(|py| -> PyResult<()> {
let x: i32 = py.eval("5", None, None)?.extract()?;
assert_eq!(x, 5);
Ok(())
})
sourcepub unsafe fn with_gil_unchecked<F, R>(f: F) -> Rwhere
F: for<'py> FnOnce(Python<'py>) -> R,
pub unsafe fn with_gil_unchecked<F, R>(f: F) -> Rwhere F: for<'py> FnOnce(Python<'py>) -> R,
Like Python::with_gil
except Python interpreter state checking is skipped.
Normally when the GIL is acquired, we check that the Python interpreter is an appropriate state (e.g. it is fully initialized). This function skips those checks.
Safety
If Python::with_gil
would succeed, it is safe to call this function.
In most cases, you should use Python::with_gil
.
A justified scenario for calling this function is during multi-phase interpreter
initialization when Python::with_gil
would fail before
_Py_InitializeMain
is called because the interpreter is only partially initialized.
Behavior in other scenarios is not documented.
source§impl<'py> Python<'py>
impl<'py> Python<'py>
sourcepub fn allow_threads<T, F>(self, f: F) -> Twhere
F: Ungil + FnOnce() -> T,
T: Ungil,
pub fn allow_threads<T, F>(self, f: F) -> Twhere F: Ungil + FnOnce() -> T, T: Ungil,
Temporarily releases the GIL, thus allowing other Python threads to run. The GIL will be
reacquired when F
’s scope ends.
If you don’t need to touch the Python interpreter for some time and have other Python threads around, this will let you run Rust-only code while letting those other Python threads make progress.
Only types that implement Ungil
can cross the closure. See the
module level documentation for more information.
If you need to pass Python objects into the closure you can use Py
<T>
to create a
reference independent of the GIL lifetime. However, you cannot do much with those without a
Python
token, for which you’d need to reacquire the GIL.
Example: Releasing the GIL while running a computation in Rust-only code
use pyo3::prelude::*;
#[pyfunction]
fn sum_numbers(py: Python<'_>, numbers: Vec<u32>) -> PyResult<u32> {
// We release the GIL here so any other Python threads get a chance to run.
py.allow_threads(move || {
// An example of an "expensive" Rust calculation
let sum = numbers.iter().sum();
Ok(sum)
})
}
Please see the Parallelism chapter of the guide for a thorough discussion of using
Python::allow_threads
in this manner.
Example: Passing borrowed Python references into the closure is not allowed
use pyo3::prelude::*;
use pyo3::types::PyString;
fn parallel_print(py: Python<'_>) {
let s = PyString::new(py, "This object cannot be accessed without holding the GIL >_<");
py.allow_threads(move || {
println!("{:?}", s); // This causes a compile error.
});
}
sourcepub fn eval(
self,
code: &str,
globals: Option<&PyDict>,
locals: Option<&PyDict>
) -> PyResult<&'py PyAny>
pub fn eval( self, code: &str, globals: Option<&PyDict>, locals: Option<&PyDict> ) -> PyResult<&'py PyAny>
Evaluates a Python expression in the given context and returns the result.
If globals
is None
, it defaults to Python module __main__
.
If locals
is None
, it defaults to the value of globals
.
Examples
let result = py.eval("[i * 10 for i in range(5)]", None, None).unwrap();
let res: Vec<i64> = result.extract().unwrap();
assert_eq!(res, vec![0, 10, 20, 30, 40])
sourcepub fn run(
self,
code: &str,
globals: Option<&PyDict>,
locals: Option<&PyDict>
) -> PyResult<()>
pub fn run( self, code: &str, globals: Option<&PyDict>, locals: Option<&PyDict> ) -> PyResult<()>
Executes one or more Python statements in the given context.
If globals
is None
, it defaults to Python module __main__
.
If locals
is None
, it defaults to the value of globals
.
Examples
use pyo3::{
prelude::*,
types::{PyBytes, PyDict},
};
Python::with_gil(|py| {
let locals = PyDict::new(py);
py.run(
r#"
import base64
s = 'Hello Rust!'
ret = base64.b64encode(s.encode('utf-8'))
"#,
None,
Some(locals),
)
.unwrap();
let ret = locals.get_item("ret").unwrap();
let b64: &PyBytes = ret.downcast().unwrap();
assert_eq!(b64.as_bytes(), b"SGVsbG8gUnVzdCE=");
});
You can use py_run!
for a handy alternative of run
if you don’t need globals
and unwrapping is OK.
sourcepub fn get_type<T>(self) -> &'py PyTypewhere
T: PyTypeInfo,
pub fn get_type<T>(self) -> &'py PyTypewhere T: PyTypeInfo,
Gets the Python type object for type T
.
sourcepub fn import<N>(self, name: N) -> PyResult<&'py PyModule>where
N: IntoPy<Py<PyString>>,
pub fn import<N>(self, name: N) -> PyResult<&'py PyModule>where N: IntoPy<Py<PyString>>,
Imports the Python module with the specified name.
sourcepub fn NotImplemented(self) -> PyObject
pub fn NotImplemented(self) -> PyObject
Gets the Python builtin value NotImplemented
.
sourcepub fn version(self) -> &'py str
pub fn version(self) -> &'py str
Gets the running Python interpreter version as a string.
Examples
Python::with_gil(|py| {
// The full string could be, for example:
// "3.10.0 (tags/v3.10.0:b494f59, Oct 4 2021, 19:00:18) [MSC v.1929 64 bit (AMD64)]"
assert!(py.version().starts_with("3."));
});
sourcepub fn version_info(self) -> PythonVersionInfo<'py>
pub fn version_info(self) -> PythonVersionInfo<'py>
Gets the running Python interpreter version as a struct similar to
sys.version_info
.
Examples
Python::with_gil(|py| {
// PyO3 supports Python 3.7 and up.
assert!(py.version_info() >= (3, 7));
assert!(py.version_info() >= (3, 7, 0));
});
sourcepub fn checked_cast_as<T>(
self,
obj: PyObject
) -> Result<&'py T, PyDowncastError<'py>>where
T: PyTryFrom<'py>,
pub fn checked_cast_as<T>( self, obj: PyObject ) -> Result<&'py T, PyDowncastError<'py>>where T: PyTryFrom<'py>,
Registers the object in the release pool, and tries to downcast to specific type.
sourcepub unsafe fn cast_as<T>(self, obj: PyObject) -> &'py Twhere
T: PyNativeType + PyTypeInfo,
pub unsafe fn cast_as<T>(self, obj: PyObject) -> &'py Twhere T: PyNativeType + PyTypeInfo,
Registers the object in the release pool, and does an unchecked downcast to the specific type.
Safety
Callers must ensure that ensure that the cast is valid.
sourcepub unsafe fn from_owned_ptr<T>(self, ptr: *mut PyObject) -> &'py Twhere
T: FromPyPointer<'py>,
pub unsafe fn from_owned_ptr<T>(self, ptr: *mut PyObject) -> &'py Twhere T: FromPyPointer<'py>,
Registers the object pointer in the release pool, and does an unchecked downcast to the specific type.
Safety
Callers must ensure that ensure that the cast is valid.
sourcepub unsafe fn from_owned_ptr_or_err<T>(
self,
ptr: *mut PyObject
) -> PyResult<&'py T>where
T: FromPyPointer<'py>,
pub unsafe fn from_owned_ptr_or_err<T>( self, ptr: *mut PyObject ) -> PyResult<&'py T>where T: FromPyPointer<'py>,
Registers the owned object pointer in the release pool.
Returns Err(PyErr)
if the pointer is NULL.
Does an unchecked downcast to the specific type.
Safety
Callers must ensure that ensure that the cast is valid.
sourcepub unsafe fn from_owned_ptr_or_opt<T>(
self,
ptr: *mut PyObject
) -> Option<&'py T>where
T: FromPyPointer<'py>,
pub unsafe fn from_owned_ptr_or_opt<T>( self, ptr: *mut PyObject ) -> Option<&'py T>where T: FromPyPointer<'py>,
Registers the owned object pointer in release pool.
Returns None
if the pointer is NULL.
Does an unchecked downcast to the specific type.
Safety
Callers must ensure that ensure that the cast is valid.
sourcepub unsafe fn from_borrowed_ptr<T>(self, ptr: *mut PyObject) -> &'py Twhere
T: FromPyPointer<'py>,
pub unsafe fn from_borrowed_ptr<T>(self, ptr: *mut PyObject) -> &'py Twhere T: FromPyPointer<'py>,
Does an unchecked downcast to the specific type.
Panics if the pointer is NULL.
Safety
Callers must ensure that ensure that the cast is valid.
sourcepub unsafe fn from_borrowed_ptr_or_err<T>(
self,
ptr: *mut PyObject
) -> PyResult<&'py T>where
T: FromPyPointer<'py>,
pub unsafe fn from_borrowed_ptr_or_err<T>( self, ptr: *mut PyObject ) -> PyResult<&'py T>where T: FromPyPointer<'py>,
Does an unchecked downcast to the specific type.
Returns Err(PyErr)
if the pointer is NULL.
Safety
Callers must ensure that ensure that the cast is valid.
sourcepub unsafe fn from_borrowed_ptr_or_opt<T>(
self,
ptr: *mut PyObject
) -> Option<&'py T>where
T: FromPyPointer<'py>,
pub unsafe fn from_borrowed_ptr_or_opt<T>( self, ptr: *mut PyObject ) -> Option<&'py T>where T: FromPyPointer<'py>,
Does an unchecked downcast to the specific type.
Returns None
if the pointer is NULL.
Safety
Callers must ensure that ensure that the cast is valid.
sourcepub fn check_signals(self) -> PyResult<()>
pub fn check_signals(self) -> PyResult<()>
Lets the Python interpreter check and handle any pending signals. This will invoke the corresponding signal handlers registered in Python (if any).
Returns Err(
PyErr
)
if any signal handler raises an exception.
These signals include SIGINT
(normally raised by CTRL + C), which by default raises
KeyboardInterrupt
. For this reason it is good practice to call this function regularly
as part of long-running Rust functions so that users can cancel it.
Example
use pyo3::prelude::*;
#[pyfunction]
fn loop_forever(py: Python<'_>) -> PyResult<()> {
loop {
// As this loop is infinite it should check for signals every once in a while.
// Using `?` causes any `PyErr` (potentially containing `KeyboardInterrupt`)
// to break out of the loop.
py.check_signals()?;
// do work here
}
}
Note
This function calls PyErr_CheckSignals()
which in turn may call signal handlers.
As Python’s signal
API allows users to define custom signal handlers, calling this
function allows arbitrary Python code inside signal handlers to run.
sourcepub unsafe fn new_pool(self) -> GILPool
pub unsafe fn new_pool(self) -> GILPool
Create a new pool for managing PyO3’s owned references.
When this GILPool
is dropped, all PyO3 owned references created after this GILPool
will
all have their Python reference counts decremented, potentially allowing Python to drop
the corresponding Python objects.
Typical usage of PyO3 will not need this API, as Python::with_gil
automatically creates
a GILPool
where appropriate.
Advanced uses of PyO3 which perform long-running tasks which never free the GIL may need to use this API to clear memory, as PyO3 usually does not clear memory until the GIL is released.
Examples
Python::with_gil(|py| {
// Some long-running process like a webserver, which never releases the GIL.
loop {
// Create a new pool, so that PyO3 can clear memory at the end of the loop.
let pool = unsafe { py.new_pool() };
// It is recommended to *always* immediately set py to the pool's Python, to help
// avoid creating references with invalid lifetimes.
let py = pool.python();
// do stuff...
}
});
Safety
Extreme care must be taken when using this API, as misuse can lead to accessing invalid
memory. In addition, the caller is responsible for guaranteeing that the GIL remains held
for the entire lifetime of the returned GILPool
.
Two best practices are required when using this API:
- From the moment
new_pool()
is called, only thePython
token from the returnedGILPool
(accessible using.python()
) should be used in PyO3 APIs. All other olderPython
tokens with longer lifetimes are unsafe to use until theGILPool
is dropped, because they can be used to create PyO3 owned references which have lifetimes which outlive theGILPool
. - Similarly, methods on existing owned references will implicitly refer back to the
Python
token which that reference was originally created with. If the returned values from these methods are owned references they will inherit the same lifetime. As a result, Rust’s lifetime rules may allow them to outlive theGILPool
, even though this is not safe for reasons discussed above. Care must be taken to never access these return values after theGILPool
is dropped, unless they are converted toPy<T>
before the pool is dropped.
source§impl Python<'_>
impl Python<'_>
sourcepub fn with_pool<F, R>(&self, f: F) -> Rwhere
F: for<'py> FnOnce(Python<'py>) -> R + Ungil,
pub fn with_pool<F, R>(&self, f: F) -> Rwhere F: for<'py> FnOnce(Python<'py>) -> R + Ungil,
Creates a scope using a new pool for managing PyO3’s owned references.
This is a safe alterantive to new_pool
as
it limits the closure to using the new GIL token at the cost of
being unable to capture existing GIL-bound references.
Note that on stable Rust, this API suffers from the same the SendWrapper
loophole
as allow_threads
, c.f. the documentation of the Ungil
trait,
Examples
Python::with_gil(|py| {
// Some long-running process like a webserver, which never releases the GIL.
loop {
// Create a new scope, so that PyO3 can clear memory at the end of the loop.
py.with_pool(|py| {
// do stuff...
});
}
});
The Ungil
bound on the closure does prevent hanging on to existing GIL-bound references
Python::with_gil(|py| {
let old_str = PyString::new(py, "a message from the past");
py.with_pool(|_py| {
print!("{:?}", old_str);
});
});
or continuing to use the old GIL token
Python::with_gil(|old_py| {
old_py.with_pool(|_new_py| {
let _none = old_py.None();
});
});
source§impl<'unbound> Python<'unbound>
impl<'unbound> Python<'unbound>
sourcepub unsafe fn assume_gil_acquired() -> Python<'unbound>
pub unsafe fn assume_gil_acquired() -> Python<'unbound>
Unsafely creates a Python token with an unbounded lifetime.
Many of PyO3 APIs use Python<'_>
as proof that the GIL is held, but this function can be
used to call them unsafely.
Safety
- This token and any borrowed Python references derived from it can only be safely used whilst the currently executing thread is actually holding the GIL.
- This function creates a token with an unbounded lifetime. Safe code can assume that
holding a
Python<'py>
token means the GIL is and stays acquired for the lifetime'py
. If you let it or borrowed Python references escape to safe code you are responsible for bounding the lifetime'unbound
appropriately. For more on unbounded lifetimes, see the nomicon.