Async Closures Stabilization Report

This report proposes the stabilization of #![feature(async_closure)] (RFC 3668). This is a long-awaited feature that increases the expressiveness of the Rust language and fills a pressing gap in the async ecosystem.

Stabilization summary

• You can write async closures like async || {} which return futures that can borrow from their captures and can be higher-ranked in their argument lifetimes.
• You can express trait bounds for these async closures using the AsyncFn family of traits, analogous to the Fn family.

async fn takes_an_async_fn(f: impl AsyncFn(&str)) {
    futures::join(f("hello"), f("world")).await;
}

takes_an_async_fn(async |s| { other_fn(s).await }).await;

Motivation

Without this feature, users hit two major obstacles when writing async code that uses closures and Fn trait bounds:

• The inability to express higher-ranked async function signatures.
• That closures cannot return futures that borrow from the closure captures.

That is, for the first, we cannot write:

// We cannot express higher-ranked async function signatures.
async fn f<Fut>(_: impl for<'a> Fn(&'a u8) -> Fut)
where
    Fut: Future<Output = ()>,
{ todo!() }

async fn main() {
    async fn g(_: &u8) { todo!() }
    f(g).await;
    //~^ ERROR mismatched types
    //~| ERROR one type is more general than the other
}

And for the second, we cannot write:

// Closures cannot return futures that borrow closure captures.
async fn f<Fut: Future<Output = ()>>(_: impl FnMut() -> Fut)
{ todo!() }

async fn main() {
    let mut xs = vec![];
    f(|| async {
        async fn g() -> u8 { todo!() }
        xs.push(g().await);
    });
    //~^ ERROR captured variable cannot escape `FnMut` closure body
}

Async closures provide a first-class solution to these problems.

For further background, please refer to the motivation section of the RFC.

Major design decisions since RFC

The RFC had left open the question of whether we would spell the bounds syntax for async closures...

// ...as this...
fn f() -> impl AsyncFn() -> u8 { todo!() }
// ...or as this:
fn f() -> impl async Fn() -> u8 { todo!() }

We've decided to spell this as AsyncFn{,Mut,Once}.

The Fn family of traits is special in many ways. We had originally argued that, due to this specialness, that perhaps the async Fn syntax could be adopted without having to decide whether a general async Trait mechanism would ever be adopted. However, concerns have been raised that we may not want to use async Fn syntax unless we would pursue more general trait modifiers. Since there remain substantial open questions on those -- and we don't want to rush any design work there -- it makes sense to ship this needed feature using the AsyncFn-style bounds syntax.

Since we would, in no case, be shipping a generalized trait modifier system anytime soon, we'll be continuing to see AsyncFoo traits appear across the ecosystem regardless. If we were to ever later ship some general mechanism, we could at that time manage the migration from AsyncFn to async Fn, just as we'd be enabling and managing the migration of many other traits.

Note that, as specified in RFC 3668, the details of the AsyncFn* traits are not exposed and they can only be named via the "parentheses sugar". That is, we can write T: AsyncFn() -> u8 but not T: AsyncFn<Output = u8>.

Unlike the Fn traits, we cannot project to the Output associated type of the AsyncFn traits. That is, while we can write...

fn f<F: Fn() -> u8>(_: F::Output) {}

...we cannot write:

fn f<F: AsyncFn() -> u8>(_: F::Output) {}
//~^ ERROR

The choice of AsyncFn{,Mut,Once} bounds syntax obviates, for our purposes here, another question decided after that RFC, which was how to order bound modifiers such as for<'a> async Fn().

Other than answering the open question in the RFC on syntax, nothing has changed about the design of this feature between RFC 3668 and this stabilization.

What is stabilized

For those interested in the technical details, please see the dev guide section I authored.

Async closures

Other than in how they solve the problems described above, async closures act similarly to closures that return async blocks, and can have parts of their signatures specified:

// They can have arguments annotated with types:
let _ = async |_: u8| { todo!() };

// They can have their return types annotated:
let _ = async || -> u8 { todo!() };

// They can be higher-ranked:
let _ = async |_: &str| { todo!() };

// They can capture values by move:
let x = String::from("hello, world");
let _ = async move || do_something(&x).await };

When called, they return an anonymous future type corresponding to the (not-yet-executed) body of the closure. These can be awaited like any other future.

What distinguishes async closures is that, unlike closures that return async blocks, the futures returned from the async closure can capture state from the async closure. For example:

let vec: Vec<String> = vec![];

let closure = async || {
    vec.push(ready(String::from("")).await);
};

The async closure captures vec with some &'closure mut Vec<String> which lives until the closure is dropped. Every call to closure() returns a future which reborrows that mutable reference &'call mut Vec<String> which lives until the future is dropped (e.g. it is awaited).

As another example:

let string: String = "Hello, world".into();

let closure = async move || {
    ready(&string).await;
};

The closure is marked with move, which means it takes ownership of the string by value. The future that is returned by calling closure() returns a future which borrows a reference &'call String which lives until the future is dropped (e.g. it is awaited).

Async fn trait family

To support the lending capability of async closures, and to provide a first-class way to express higher-ranked async closures, we introduce the AsyncFn* family of traits. See the corresponding section of the RFC.

We stabilize naming AsyncFn* via the "parenthesized sugar" syntax that normal Fn* traits can be named. The AsyncFn* trait can be used anywhere a Fn* trait bound is allowed, such as:

/// In return-position impl trait:
fn closure() -> impl AsyncFn() { async || {} }

/// In trait bounds:
trait Foo<F>: Sized
where
    F: AsyncFn()
{
    fn new(f: F) -> Self;
}

/// in GATs:
trait Gat {
    type AsyncHasher<T>: AsyncFn(T) -> i32;
}

Other than using them in trait bounds, the definitions of these traits are not directly observable, but certain aspects of their behavior can be indirectly observed such as the fact that:

• AsyncFn::async_call and AsyncFnMut::async_call_mut return a future which is lending, and therefore borrows the &self lifetime of the callee.

fn by_ref_call(c: impl AsyncFn()) {
    let fut = c();
    drop(c);
    //   ^ Cannot drop `c` since it is borrowed by `fut`.
}

• AsyncFnOnce::async_call_once returns a future that takes ownership of the callee.

fn by_ref_call(c: impl AsyncFnOnce()) {
    let fut = c();
    let _ = c();
    //      ^ Cannot call `c` since calling it takes ownership the callee.
}

• All currently-stable callable types (i.e., closures, function items, function pointers, and dyn Fn* trait objects) automatically implement AsyncFn*() -> T if they implement Fn*() -> Fut for some output type Fut, and Fut implements Future<Output = T>.
  • This is to make sure that AsyncFn*() trait bounds have maximum compatibility with existing callable types which return futures, such as async function items and closures which return boxed futures.
  • For now, this only works currently for concrete callable types -- for example, a argument-position impl trait like impl Fn() -> impl Future<Output = ()> does not implement AsyncFn(), due to the fact that a AsyncFn-if-Fn blanket impl does not exist in reality. This may be relaxed in the future. Users can work around this by wrapping their type in an async closure and calling it. I expect this to not matter much in practice, as users are encouraged to write AsyncFn bounds directly.

fn is_async_fn(_: impl AsyncFn(&str)) {}

async fn async_fn_item(s: &str) { todo!() }
is_async_fn(s);
// ^^^ This works.

fn generic(f: impl Fn() -> impl Future<Output = ()>) {
    is_async_fn(f);
    // ^^^ This does not work (yet).
}

The by-move future

When async closures are called with AsyncFn/AsyncFnMut, they return a coroutine that borrows from the closure. However, when they are called via AsyncFnOnce, we consume that closure, and cannot return a coroutine that borrows from data that is now dropped.

To work around around this limitation, we synthesize a separate future type for calling the async closure via AsyncFnOnce.

This future executes identically to the by-ref future returned from calling the async closure, except for the fact that it has a different set of captures, since we must move the captures from the parent async into the child future.

Interactions between async closures and the Fn* family of traits

Async closures always implement FnOnce, since they always can be called once. They may also implement Fn or FnMut if their body is compatible with the calling mode (i.e. if they do not mutate their captures, or they do not capture their captures, respectively) and if the future returned by the async closure is not lending.

let id = String::new();

let mapped: Vec</* impl Future */> =
    [/* elements */]
    .into_iter()
    // `Iterator::map` takes an `impl FnMut`
    .map(async |element| {
        do_something(&id, element).await;
    })
    .collect();

See the dev guide for a detailed explanation for the situations where this may not be possible due to the lending nature of async closures.

Other notable features of async closures shared with synchronous closures

• Async closures are Copy and/or Clone if their captures are Copy/Clone.
• Async closures do closure signature inference: If an async closure is passed to a function with a AsyncFn or Fn trait bound, we can eagerly infer the argument types of the closure. More details are provided in the dev guide.

Lints

This PR also stabilizes the CLOSURE_RETURNING_ASYNC_BLOCK lint as an allow lint. This lints on "old-style" async closures:

#![warn(closure_returning_async_block)]
let c = |x: &str| async {};

We should encourage users to use async || {} where possible. This lint remains allow and may be refined in the future because it has a few false positives (namely, see: "Where do we expect rewriting || async {} into async || {} to fail?")

An alternative that could be made at the time of stabilization is to put this lint behind another gate, so we can decide to stabilize it later.

What isn't stabilized (aka, potential future work)

async Fn*() bound syntax

We decided to stabilize async closures without the async Fn*() bound modifier syntax. The general direction of this syntax and how it fits is still being considered by T-lang (e.g. in RFC 3710).

Naming the futures returned by async closures

This stabilization PR does not provide a way of naming the futures returned by calling AsyncFn*.

Exposing a stable way to refer to these futures is important for building async-closure-aware combinators, and will be an important future step.

Return type notation-style bounds for async closures

The RFC described an RTN-like syntax for putting bounds on the future returned by an async closure:

async fn foo(x: F) -> Result<()>
where
    F: AsyncFn(&str) -> Result<()>,
    // The future from calling `F` is `Send` and `'static`.
    F(..): Send + 'static,
{}

This stabilization PR does not stabilize that syntax yet, which remains unimplemented (though will be soon).

dyn AsyncFn*()

AsyncFn* are not dyn-compatible yet. This will likely be implemented in the future along with the dyn-compatibility of async fn in trait, since the same issue (dealing with the future returned by a call) applies there.

Tests

Tests exist for this feature in tests/ui/async-await/async-closures.

Remaining bugs and open issues

• Investigate getting rid of the AsyncFn* family of traits in favor of LendingFn* #120694 tracks moving onto more general LendingFn* traits. No action needed, since it's not observable.
• ICE: polymorphization does not support coroutines from async closures #124020 - Polymorphization ICE. Polymorphization needs to be heavily reworked. No action needed.
• rustdoc: support for async Fn* trait bounds #127227 - Tracking reworking the way that rustdoc re-sugars bounds.
  • The part relevant to to AsyncFn is fixed by Clean middle generics using paren sugar if trait has #[rustc_paren_sugar] #132697.

Where do we expect rewriting || async {} into async || {} to fail?

• Fn pointer coercions
  • Currently, it is not possible to coerce an async closure to an fn pointer like regular closures can be. This functionality may be implemented in the future.

let x: fn() -> _ = async || {};

• Argument capture
  • Like async functions, async closures always capture their input arguments. This is in contrast to something like |t: T| async {}, which doesn't capture t unless it is used in the async block. This may affect the Send-ness of the future or affect its outlives.

fn needs_send_future(_: impl Fn(NotSendArg) -> Fut)
where
    Fut: Future<Output = ()>,
{}

needs_send_future(async |_| {});

History

Important feature history

• async/await #51580
• Move async || ... closures into #![feature(async_closure)] #62292
• Rework support for async closures; allow them to return futures that borrow from the closure's captures #120361
• Harmonize AsyncFn implementations, make async closures conditionally impl Fn* traits #120712
• Implement async closure signature deduction #121857
• Make the computation of coroutine_captures_by_ref_ty more sophisticated #123660
• An async closure may implement FnMut/Fn if it has no self-borrows #125259
• Stop storing a special inner body for the coroutine by-move body for async closures #128506
• Infer async closure signature from (old-style) two-part Fn + Future bounds #127482

Acknowledgements

Thanks to @oli-obk for reviewing the bulk of the work for this feature. Thanks to @nikomatsakis for his design blog posts which generated interest for this feature, @traviscross for feedback and additions to this stabilization report. All errors are my own.

r? @ghost