Async/await syntax for Rust and the futures
crate
The primary way of working with futures today in Rust is through the various
combinators on the Future
trait. This is not quite "callback hell" but can
sometimes feel like it as the rightward drift of code increases for each new
closure you tack on. The purpose of async/await is to provide a much more
"synchronous" feeling to code while retaining all the power of asynchronous
code!
Here's a small taste of what this crate does:
#[async]
fn fetch_rust_lang(client: hyper::Client) -> io::Result<String> {
let response = await!(client.get("https://www.rust-lang.org"))?;
if !response.status().is_success() {
return Err(io::Error::new(io::ErrorKind::Other, "request failed"))
}
let body = await!(response.body().concat())?;
let string = String::from_utf8(body)?;
Ok(string)
}
The notable points here are:
- Functions are tagged with
#[async]
, which means that they are asychronous and return a future when called instead of returning the result listed. - The
await!
macro allows blocking on a future to completion. This does not actually block the thread though, it just "blocks" the future returned fromfetch_rust_lang
from continuing. You can think of theawait!
macro as a function from a future to itsResult
, consuming the future along the way. - Error handling is much more natural here in
#[async]
functions. We get to use the?
operator as well as simplereturn Err
statements. With the combinators in thefutures
crate this is generally much more cumbersome. - The future returned from
fetch_rust_lang
is actually a state machine generated at compile time and does no implicit memory allocation. This is built on top of generators, described below.
You can also have async methods:
impl Foo {
#[async]
fn do_work(self) -> io::Result<u32> {
// ...
}
}
You can specify that you actually would prefer a trait object is returned
instead, e.g. Box<Future<Item = i32, Error = io::Error>>
#[async(boxed)]
fn foo() -> io::Result<i32> {
// ...
}
You can also have "async for
loops" which operate over the Stream
trait:
#[async]
for message in stream {
// ...
}
An async for
loop will propagate errors out of the function so message
has
the Item
type of the stream
passed in. Note that async for
loops can only
be used inside of an #[async]
function.
And finally, you can create a Stream
instead of a Future
via
#[async_stream(item = _)]
:
#[async]
fn fetch(client: hyper::Client, url: &'static str) -> io::Result<String> {
// ...
}
/// Fetch all provided urls one at a time
#[async_stream(item = String)]
fn fetch_all(client: hyper::Client, urls: Vec<&'static str>) -> io::Result<()> {
for url in urls {
let s = await!(fetch(client, url))?;
stream_yield!(s);
}
Ok(())
}
#[async_stream]
must have an item type specified via item = some::Path
and
the values output from the stream must be wrapped into a Result
and yielded
via the stream_yield!
macro. This macro also supports the same features as
#[async]
, an additional boxed
argument to return a Box<Stream>
, async
for
loops, etc.
This implementation is currently fundamentally built on generators/coroutines. This feature has just landed and will be in the nightly channel of Rust as of 2017-08-29. You can acquire the nightly channel via;
rustup update nightly
After doing this you'll want to ensure that Cargo's using the nightly toolchain by invoking it like:
cargo +nightly build
Next you'll add this to your Cargo.toml
[dependencies]
futures-await = "0.1"
and then...
#![feature(proc_macro, conservative_impl_trait, generators)]
extern crate futures_await as futures;
use futures::prelude::*;
#[async]
fn foo() -> Result<i32, i32> {
Ok(1 + await!(bar())?)
}
#[async]
fn bar() -> Result<i32, i32> {
Ok(2)
}
fn main() {
assert_eq!(foo().wait(), Ok(3));
}
This crate is intended to "masquerade" as the futures
crate, reexporting its
entire hierarchy and just augmenting it with the necessary runtime support, the
async
attribute, and the await!
macro. These imports are all contained in
futures::prelude::*
when you import it.
For a whole mess of examples in a whole mess of code, you can also check out the
async-await
branch of sccache
which is an in-progress transition
to using async/await syntax in many locations. You'll typically find that the
code is much more readable afterwards, especially these changes
As mentioned before this crate is fundamentally built on the feature of
generators in Rust. Generators, otherwise known in this case as stackless
coroutines, allow the compiler to generate an optimal implementation for a
Future
for a function to transform a synchronous-looking block of code into a
Future
that can execute asynchronously. The desugaring here is surprisingly
straightforward from code you write to code rustc compiles, and you can browse
the source of the futures-async-macro
crate for more details here.
Otherwise there's a few primary "APIs" provided by this crate:
-
#[async]
- this attribute can be applied to methods and functions to signify that it's an asynchronous function. The function's signature must return aResult
of some form (although it can return a typedef of results). Additionally, the function's arguments must all be owned values, or in other words must contain no references. This restriction may be lifted in the future!Some examples are:
// attribute on a bare function #[async] fn foo(a: i32) -> Result<u32, String> { // ... } // returning a typedef works too! #[async] fn foo() -> io::Result<u32> { // ... } impl Foo { // methods also work! #[async] fn foo(self) -> io::Result<u32> { // ... } } // For now, however, these do not work, they both have arguments which contain // references! This may work one day, but it does not work today. // // #[async] // fn foo(a: &i32) -> io::Result<u32> { /* ... */ } // // impl Foo { // #[async] // fn foo(&self) -> io::Result<u32> { /* ... */ } // }
Note that an
#[async]
function is intended to behave very similarly to that of its synchronous version. For example the?
operator works internally, you can use an earlyreturn
statement, etc.Under the hood an
#[async]
function is compiled to a state machine which represents a future for this function, and the state machine's suspension points will be where theawait!
macro is located. -
async_block!
- this macro is similar to#[async]
in that it creates a future, but unlike#[async]
it can be used in an expression context and doesn't need a dedicated function. This macro can be considered as "run this block of code asynchronously" where the block of code provided, like an async function, returns a result of some form. For example:let server = TcpListener::bind(..); let future = async_block! { #[async] for connection in server.incoming() { spawn(handle_connection(connection)); } Ok(()) }; core.run(future).unwrap();
or if you'd like to do some precomputation in a function:
fn hash_file(path: &Path, pool: &CpuPool) -> impl Future<Item = u32, Error = io::Error> { let abs_path = calculate_abs_path(path); async_block! { let contents = await!(read_file)?; Ok(hash(&contents)) } }
-
await!
- this is a macro provided in thefutures-await-macro
crate which allows waiting on a future to complete. Theawait!
macro can only be used inside of an#[async]
function or anasync_block!
and can be thought of as a function that looks like:fn await!<F: Future>(future: F) -> Result<F::Item, F::Error> { // magic }
Some examples of this macro are:
#[async] fn fetch_url(client: hyper::Client, url: String) -> io::Result<Vec<u8>> { // note the trailing `?` to propagate errors let response = await!(client.get(url))?; await!(response.body().concat()) }
-
#[async]
for loops - the ability to iterate asynchronously over aStream
. You can do this by attaching the#[async]
attribute to afor
loop where the object being iterated over implements theStream
trait.Errors from the stream will get propagated automatically, and otherwise, the for loop will exhauste the stream to completion, binding each element to the pattern provided.
Some examples are:
#[async] fn accept_connections(listener: TcpListener) -> io::Result<()> { #[async] for connection in server.incoming() { // `connection` here has type `TcpStream` } Ok(()) }
Note that an
#[async]
for loop, likeawait!
, can only be used in an async function or an async block. -
#[async_stream(item = ...)]
- defines a function which is an implementation ofStream
rather thanFuture
. This function uses thestream_yield!
macro to yield items, and otherwise works with theawait!
macro and#[async]
for loops.The declared function must return a
Result<(), E>
whereE
becomes the error of theStream
returned. The stream is terminated by returningOk(())
from the function or returning an error. Operations like?
work inside the function for propagating errors as well.An example is:
#[async_stream(item = u32)] fn accept_connections(listener: TcpListener) -> io::Result<()> { #[async] for object in fetch_all_objects() { let description = await!(fetch_object_description(object)); stream_yield!(description); } Ok(()) }
Right now this crate requires three nightly features to be used, and practically requires four features to be used to its fullest extent. These four features are:
-
#![feature(generators)]
- this is an experimental language feature that has yet to be stabilized but is the foundation for the implementation of async/await. The implementation landed a while ago and progress on stabilization can be found on its tracking issue. -
#![feature(pin)]
- this is another experimental language feature that works together withgenerators
to provide support for immovable self-referential generators, without thisasync
functions cannot have borrows that exist across the yield points used insideawait!
. The implementation landed very recently and progress on stabilization can be found on its tracking issue. -
#![feature(proc_macro)]
- this has also been dubbed "Macros 2.0" and is how the#[async]
attribute is defined in this crate and not the compiler itself. We then also take advantage of other macros 2.0 features like importing macros viause
instead of#[macro_use]
. Tracking issues for this feature include #38356 and #35896. -
#![feature(conservative_impl_trait)]
- this feature is not actually required to use the crate if you opt to always use trait objects via#[async(boxed)]
, but it's practically required to use the crate ergonomically. This feature is used because each function returnsimpl Future<...>
instead of a concrete type of future. This feature is tracked at #34511 and #42183.
The intention with this crate is that the newest feature, generators
, will
be the last to stabilize. The other two, proc_macro
and
conservative_impl_trait
, should hopefully stabilize ahead of generators!
This crate is still quite new and generators have only just landed on the nightly channel for Rust. While no major change is planned to this crate at this time, we'll have to see how this all pans out! One of the major motivational factors for landing generators so soon in the language was to transitively empower this implementation of async/await, and your help is needed in assessing that!
Notably, we're looking for feedback on async/await in this crate itself as well as generators as a language feature. We want to be sure that if we were to stabilize generators or procedural macros they're providing the optimal async/await experience!
At the moment you're encouraged to take caution when using this in production.
This crate itself has only been very lightly tested and the generators
language feature has also only been lightly tested so far. Caution is advised
along with a suggestion to keep your eyes peeled for bugs! If you run into any
questions or have any issues, you're more than welcome to open an
issue!
As can be expected with many nightly features, there are a number of caveats to be aware of when working with this project. Despite this bug reports or experience reports are more than welcome, always good to know what needs to be fixed regardless!
Borrowing doesn't really work so well today. The compiler will either reject many borrows or there may be some unsafety lurking as the generators feature is being developed. For example if you have a function such as:
#[async]
fn foo(s: &str) -> io::Result<()> {
// ..
}
This may not compile! The reason for this is that the returned future
typically needs to adhere to the 'static
bound. Async functions currently
execute no code when called, they only make progress when polled. This means
that when you call an async function what happens is it creates a future,
packages up the arguments, and then returns that to you. In this case it'd
have to return the &str
back up, which doesn't always have the lifetimes
work out.
An example of how to get the above function compiling would be to do:
#[async]
fn foo(s: String) -> io::Result<()> {
// ...
}
or somehow otherwise use an owned value instead of a borrowed reference.
Note that arguments are not the only point of pain with borrowing. For example code like this will not (or at least shouldn't) compile today:
for line in string.lines() {
await!(process(line));
println!("processed: {}", line);
}
The problem here is that line
is a borrowed value that is alive across a yield
point, namely the call to await!
. This means that when the future may return
back up the stack was part of the await!
process it'd have to restore the
line
variable when it reenters the future. This isn't really all implemented
and may be unsafe today. As a rule of thumb for now you'll need to only have
owned values (no borrowed internals) alive across calls to await!
or during
async for
loops.
Lots of thought is being put in to figure out how to alleviate this restriction! Borrowing is the crux of many ergonomic patterns in Rust, and we'd like this to work!
As one final point, a consequence of the "no borrowed arguments" today is that function signatures like:
#[async]
fn foo(&self) -> io::Result<()> {
// ...
}
unfortunately will not work. You'll either need to take self
by value or defer
to a different #[async]
function.
Let's say you've got a trait like so:
trait MyStuff {
fn do_async_task(??self) -> Box<Future<...>>;
}
We'll gloss over the self
details here for a bit, but in essence we've got a
function in a trait that wants to return a future. Unfortunately it's actually
quite difficult to use this! Right now there's a few of caveats:
- Ideally you want to tag this
#[async]
but this doesn't work because a trait function returningimpl Future
is not implemented in the compiler today. I'm told that this will eventually work, though! - Ok so then the next best thing is
#[async(boxed)]
to return a boxed trait object instead ofimpl Future
for the meantime. This may not be quite what we want runtime-wise but we also have problems with... - But now this brings us to the handling of
self
. Because of the limitations of#[async]
today we only have two options,self
andself: Box<Self>
. The former is unfortunately not object safe (now we can't use virtual dispatch with this trait) and the latter is typically wasteful (every invocation now requires a fresh allocation). Ideallyself: Rc<Self>
is exactly what we want here! But unfortunately this isn't implemented in the compiler 😦
So basically in summary you've got one of two options to return futures in traits today:
trait MyStuff {
// Trait is not object safe because of `self` so can't have virtual
// dispatch, and the allocation of `Box<..>` as a return value is required
// until the compiler implements returning `impl Future` from traits.
//
// Note that the upside of this approach, though, is that `self` could be
// something like `Rc` or have a bunch fo `Rc` inside of `self`, so this
// could be cheap to call.
#[async]
fn do_async_task(self) -> Box<Future<...>>;
}
or the alternative:
trait MyStuff {
// Like above we returned a trait object but here the trait is indeed object
// safe, allowing virtual dispatch. The downside is that we must have a
// `Box` on hand every time we call this function, which may be costly in
// some situations.
#[async]
fn do_async_task(self: Box<Self>) -> Box<Future<...>>;
}
The ideal end goal for futures-in-traits is this:
trait MyStuff {
#[async]
fn do_async_task(self: Rc<Self>) -> Result<i32, u32>;
}
but this needs two pieces to be implemented:
- The compiler must accept trait functions returning
impl Trait
- The compiler needs support for
self: Rc<Self>
, basically object-safe custom smart pointers in traits.
You can emulate this today with a non-object-safe implementation via:
trait Foo {
#[async]
fn do_async_task(me: Rc<Self>) -> Result<i32, u32>;
}
fn foo<T: Foo>(t: Rc<T>) {
let x = Foo::trait_fn(t);
// ...
}
But that's not exactly the most ergonomic!
Another limitation when using futures with traits is with associated types. Say you've got a trait like:
trait Service {
type Request;
type Error;
type Future: Future<Item = Self::Response, Error = Self::Error>;
fn call(&self) -> Self::Future;
}
If you want to implement call
with async_block!
, or by returning a future
from another function which was generated with #[async]
, you'd probably want
to use impl Future
. Unfortunately, it's not current possible to express an
associated constant like Service::Future
with an impl Trait
.
For now the best solution is to use Box<Future<...>>
:
impl Service for MyStruct {
type Request = ...;
type Error = ...;
type Future = Box<Future<Item = Self::Item, Error = Self::Error>>;
fn call(&self) -> Self::Future {
// ...
Box::new(future)
}
}
This project is licensed under either of
- Apache License, Version 2.0, (LICENSE-APACHE or http://www.apache.org/licenses/LICENSE-2.0)
- MIT license (LICENSE-MIT or http://opensource.org/licenses/MIT)
at your option.
Unless you explicitly state otherwise, any contribution intentionally submitted
for inclusion in futures-await
by you, as defined in the Apache-2.0 license, shall be
dual licensed as above, without any additional terms or conditions.