Improving async-await's "Future is not Send" diagnostic

Oct. 11, 2019 · David Wood on behalf of the Async Foundations WG

Async-await is due to hit stable in the 1.39 release (only a month away!), and as announced in the "Async Foundations Update: Time for polish!" post last month, the Async Foundations WG has shifted its focus to polish. This post will highlight one aspect of that focus, diagnostic improvements, and in particular, the improvements that the working group has been making to the once-unhelpful "future is not send" diagnostic.

Why doesn't my future implement Send?

One of the major places where async-await should be a pleasure to use is in multithreaded contexts, where having a future that can be sent to other threads is desirable. This might look something like the following (for brevity, there aren't any threads here, just a requirement that the future implement std::marker::Send):

use std::sync::{Mutex, MutexGuard};

fn is_send<T: Send>(t: T) { }

async fn foo() {
    bar(&Mutex::new(22)).await
}

async fn bar(x: &Mutex<u32>) {
    let g = x.lock().unwrap();
    baz().await
}

async fn baz() { }

fn main() {
    is_send(foo());
}

When we try to compile this, we'll get an unwieldly and hard-to-follow diagnostic:

error[E0277]: `std::sync::MutexGuard<'_, u32>` cannot be sent between threads safely
  --> src/main.rs:23:5
   |
23 |     is_send(foo());
   |     ^^^^^^^ `std::sync::MutexGuard<'_, u32>` cannot be sent between threads safely
   |
   = help: within `impl std::future::Future`, the trait `std::marker::Send` is not implemented for `std::sync::MutexGuard<'_, u32>`
   = note: required because it appears within the type `for<'r, 's> {&'r std::sync::Mutex<u32>, std::sync::MutexGuard<'s, u32>, impl std::future::Future, ()}`
   = note: required because it appears within the type `[static generator@src/main.rs:13:30: 16:2 x:&std::sync::Mutex<u32> for<'r, 's> {&'r std::sync::Mutex<u32>, std::sync::MutexGuard<'s, u32>, impl std::future::Future, ()}]`
   = note: required because it appears within the type `std::future::GenFuture<[static generator@src/main.rs:13:30: 16:2 x:&std::sync::Mutex<u32> for<'r, 's> {&'r std::sync::Mutex<u32>, std::sync::MutexGuard<'s, u32>, impl std::future::Future, ()}]>`
   = note: required because it appears within the type `impl std::future::Future`
   = note: required because it appears within the type `impl std::future::Future`
   = note: required because it appears within the type `for<'r> {impl std::future::Future, ()}`
   = note: required because it appears within the type `[static generator@src/main.rs:9:16: 11:2 for<'r> {impl std::future::Future, ()}]`
   = note: required because it appears within the type `std::future::GenFuture<[static generator@src/main.rs:9:16: 11:2 for<'r> {impl std::future::Future, ()}]>`
   = note: required because it appears within the type `impl std::future::Future`
   = note: required because it appears within the type `impl std::future::Future`
note: required by `is_send`
  --> src/main.rs:5:1
   |
5  | fn is_send<T: Send>(t: T) {
   | ^^^^^^^^^^^^^^^^^^^^^^^^^

That's.. not great. Let's break down what's happening and understand what this error is trying to tell us.

fn main() {
    is_send(foo());
}

In main, we are calling foo and passing the return value to is_send. foo is an async fn, so it doesn't return () (what you might expect for a function with no return type specified). Instead, it returns impl std::future::Future<Output = ()>, some unnamed type that implements std::future::Future:

async fn foo() {
    bar(&Mutex::new(22)).await
}

// becomes...

fn foo() -> impl std::future::Future<Output = ()> {
    async move {
        bar(&Mutex::new(22)).await
    }
}

If you want to learn more about the transformations that happen with async-await, consider reading the async/.await primer chapter of the async book.

fn is_send<T: Send>(t: T) { }

It looks like the error we're getting is because the future returned by foo doesn't satisfy the T: Send bound of is_send.

How are async functions implemented?

To explain why our future doesn't implement Send, we first need to understand a little bit more about what async-await is doing under the hood. rustc implements async fns using generators, an unstable language feature for resumable functions like the co-routines you might be familiar with from other languages. Generators are laid out like enums with variants containing all of the variables that are used across await points (which desugar to generator yields):

async fn bar(x: &Mutex<u32>) {
    let g = x.lock().unwrap();
    baz().await // <- await point (suspend #0), `g` and `x` are in use before await point
} // <- `g` and `x` dropped here, after await point

enum BarGenerator {
    // `bar`'s parameters.
    Unresumed { x: &Mutex<u32> },

    Suspend0 {
        // Future returned by `baz`, which is being polled.
        baz_future: BazGenerator,

        // Locals that are used across the await point.
        x: &Mutex<u32>,
        g: MutexGuard<'_, u32>,
    },

    Returned { value: () }
}

If you want to learn more about the async fn implementation details, then Tyler Mandry has written an excellent blog post diving into their work here in more depth which is definitely worth a read.

So, why doesn't my future implement Send?

We now know that an async fn is represented like an enum behind-the-scenes. In synchronous Rust, you'll be used to your types automatically implementing Send when the compiler determines it's appropriate - typically when all of the fields of your type also implement Send. It follows that the enum-like that represents our async fn would implement Send if all of the types in it do.

In other words, a future is safe to send across threads if all of the types that are held across .await points implement Send. This behaviour is useful because it lets us write generic code that interoperates smoothly with async-await, but without diagnostic support we get confusing error messages.

Well, which type is the problem in the example?

Returning to our example, the future must be holding a type across an .await point that doesn't implement Send, but where? This is the primary question that the diagnostic improvement aims to help answer. Let's start by looking at foo:

async fn foo() {
    bar(&Mutex::new(22)).await
}

foo invokes bar, passing a reference to a std::sync::Mutex and getting a future back, before awaiting it.

async fn bar(x: &Mutex<u32>) {
    let g: MutexGuard<u32> = x.lock().unwrap();
    baz().await
} // <- `g` is dropped here

bar locks the mutex before awaiting baz. std::sync::MutexGuard<u32> does not implement Send and lives across the baz().await point (because g is dropped at the end of the scope) which causes the entire future not to implement Send.

That wasn't obvious from the error: we had to know that futures might implement Send depending on the types they capture and find the type which lives across an await point ourselves.

Fortunately, the Async Foundations WG has been working to improve this error, and in nightly, we see the following diagnostic instead:

error[E0277]: `std::sync::MutexGuard<'_, u32>` cannot be sent between threads safely
  --> src/main.rs:23:5
   |
5  | fn is_send<T: Send>(t: T) {
   |    -------    ---- required by this bound in `is_send`
...
23 |     is_send(foo());
   |     ^^^^^^^ `std::sync::MutexGuard<'_, u32>` cannot be sent between threads safely
   |
   = help: within `impl std::future::Future`, the trait `std::marker::Send` is not implemented for `std::sync::MutexGuard<'_, u32>`
note: future does not implement `std::marker::Send` as this value is used across an await
  --> src/main.rs:15:3
   |
14 |   let g = x.lock().unwrap();
   |       - has type `std::sync::MutexGuard<'_, u32>`
15 |   baz().await;
   |   ^^^^^^^^^^^ await occurs here, with `g` maybe used later
16 | }
   | - `g` is later dropped here

Much better!

How does it work?

When rustc's trait system determines that a trait wasn't implemented, in this case std::marker::Send, it emits this error. The trait system produces a chain of "obligations". Obligations are types which denote where a bound (e.g T: Send in is_send) originated, or where a bound was propagated.

To improve this diagnostic, the chain of obligations is now treated like a stack frame, where each "frame" of obligations represents each function's contribution to the error. Let's make that more concrete with a very rough approximation:

Obligation {
    kind: DerivedObligation(/* generator that captured `g` */),
    source: /* `Span` type pointing at `bar`'s location in user code */,
    parent: Some(Obligation {
        kind: DerivedObligation(/* generator calling `bar` */),
        source: /* `Span` type pointing at `foo`'s location in user code */,
        parent: Some(Obligation {
            kind: ItemObligation(/* type representing `std::marker::Send` */),
            source: /* `Span` type pointing at `is_send`'s location in user code */,
            cause: None,
        }),
    }),
}

The compiler matches against the chain expecting an ItemObligation and some DerivedObligations containing generators, which identifies the error we want to improve. Using information from these obligations, rustc can construct the specialized error shown above - if you'd like to see what the actual implementation looks like, check out PR #64895.

If you're interested in improving diagnostics like this, or even just fixing bugs, consider contributing to the compiler! There are many working groups to join and resources to help you get started (like the rustc dev guide or the compiler team documentation).

What's next?

More improvements to this diagnostic are planned and being worked on, so that it is applicable in more cases, and has specialized messages for Send and Sync, like below:

error[E0277]: future cannot be sent between threads safely
  --> src/main.rs:23:5
   |
5  | fn is_send<T: Send>(t: T) {
   |    -------    ---- required by this bound in `is_send`
...
23 |     is_send(foo());
   |     ^^^^^^^ future returned by `foo` is not `Send`
   |
   = help:  future is not `Send` as this value is used across an await
note: future does not implement `std::marker::Send` as this value is used across an await
  --> src/main.rs:15:3
   |
14 |   let g = x.lock().unwrap();
   |       - has type `std::sync::MutexGuard<'_, u32>`
15 |   baz().await;
   |   ^^^^^^^^^^^ await occurs here, with `g` maybe used later
16 | }
   | - `g` is later dropped here