speice.io/_posts/2020-07-08-on-implementing-futures.md

layout	title	description	category	tags
post	Tips for Implementing `Future`			python

When support for async/await launched in Rust, it came with a couple of technical caveats; it was deemed more beneficial to release a minimum viable product than wait for a feature-complete release. So far, this model seems to have worked out well. Asynchronous code originally required thread-local storage for context tracking which mean that it could only be used in projects that included the Rust standard library. It wasn't a hard requirement; nothing about the async design mandated context tracking in this way. But given that most users of asynchronous code relied on the standard library anyway, supporting asynchronous no_std projects was left as something to be addressed later. After some fantastic work, thread-local storage is no longer used and there's some incredibly cool work being done to enable Rust on no_std and embedded systems. While asynchronous programming is very frequently used to model disk or network I/O, the same principles can be applied to monitoring signals received from GPIO pins.

NOTE: Should I mention something about how cool it is that we can have async without needing heap allocations or type erasure like in every other async implementation?

One other missing feature in the initial async support was being able to write traits that contained async fn methods Normally, when an async fn function is declared, the compiler does some magic to the function signature:

struct R;

// When you write a function like this:
async fn read_bytes(s: TcpStream) -> R { /* ... */ }

// ...the compiler effectively transforms it into this:
fn read_bytes(s: TcpStream) -> impl Future<Output = R> { /* ... */ }

This special return type (the impl Future thing) tells the compiler "I have no idea what the exact return type will be, but it will be something that implements Future, just figure it out for me." If you're writing static or struct functions, that's no issue, the compiler can figure everything out for you.

However, this "figure it out for me" mentality doesn't work when used with traits. The reasons are varied and complex and out of scope for this discussion. But if we want to mix traits and asynchronous code, we simply need to make sure the trait method returns a type that implements the Future trait:

use std::future::Future;
use std::pin::Pin;
use std::task::{Context, Poll};

pub struct Byte(u8);

// Implementing this section of the code is what we'll be talking about.
// vvvvvvvvvvvvvvvvvvv
impl Future for Byte {
    type Output = u8;

    fn poll(self: Pin<&mut Self>, _cx: &mut Context<'_>) -> Poll<Self::Output> {
        return Poll::Ready(self.0);
    }
}
// ^^^^^^^^^^^^^^^^^^^

pub trait ByteReader {
    fn get_byte(&self) -> Byte;
}

pub async fn my_function(b: impl ByteReader) -> u8 {
    b.get_byte().await
}

Because of some Rust-specific issues (the Pin/Unpin system, unhelpful compiler messages), implementing Future directly can be rather difficult. It's possible to use crates like async_trait to work around the limitation, but if you're interested in building your own futures, these techniques should make the process at least a bit easier.

Implement functionality before structure

Don't impl Future right away; use a separate method and pass eevrything in. It's helpful to de-couple "what you need in order to function" from "how you get those things"; are you supposed to use #[pin_project] or let Self { } = &mut *self or maybe just &mut self.value? Self-pinning makes things weird, and it's typically safe to deal with those questions later. Two guidelines:

1. Everything that needs to be poll-ed should be passed as Pin<&mut T>
2. Everything else passed by reference.

Don't call this function before it's ready; errors elsewhere in the code can make it difficult to understand if the problem is in your "inner" function implementation, or the impl Future implementation.

Dealing with unfulfilled trait bounds

Should also add something about how AsyncBufRead isn't implemented for &R3, but is after deref (R3). The errors become a lot more obvious if you try to deref self.reader:

use futures_io::AsyncBufRead;
use std::future::Future;
use std::pin::Pin;
use std::task::{Context, Poll};

fn poll_once<R1: AsyncBufRead + ?Sized>(mut reader: Pin<&mut R1>, cx: &mut Context<'_>) -> Poll<()> {
    reader.as_mut().poll_fill_buf(cx);
    return Poll::Ready(());
}

struct MyStruct<'a, R2: ?Sized> {
    reader: &'a R2,
}

impl<R3: AsyncBufRead + ?Sized + Unpin> Future for MyStruct<'_, R3> {
    type Output = ();

    fn poll(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Self::Output> {
        // Important bit is the `*self.reader` here
        poll_once(Pin::new(&mut *self.reader), cx)
    }
}

error[E0596]: cannot borrow data in a dereference of `std::pin::Pin<&mut MyStruct<'_, R3>>` as mutable
  --> src/lib.rs:19:28
   |
12 |     reader: &'a R2,
   |             ------ help: consider changing this to be mutable: `&'a mut R2`
...
19 |         poll_once(Pin::new(&mut *self.reader), cx)
   |                            ^^^^^^^^^^^^^^^^^ cannot borrow as mutable

error[E0596]: cannot borrow `self` as mutable, as it is not declared as mutable
  --> src/lib.rs:19:34
   |
18 |     fn poll(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Self::Output> {
   |             ---- help: consider changing this to be mutable: `mut self`
19 |         poll_once(Pin::new(&mut *self.reader), cx)
   |                                  ^^^^ cannot borrow as mutable

Now, we can see that self can't be borrowed as mutable (it needs to be marked mut self: Pin<&mut Self>) and that the reader can't be borrowed as mutable (the struct definition needs &'a mut R2). After those are fixed, we're good to go.

Don't feel bad about requiring Unpin

For trait bounds, don't require it unless you need to, but don't hesitate to add it if the compiler thinks you should.

use futures_io::AsyncBufRead;
use std::future::Future;
use std::pin::Pin;
use std::task::{Context, Poll};

fn poll_once<R1: AsyncBufRead + ?Sized>(mut reader: Pin<&mut R1>, cx: &mut Context<'_>) -> Poll<()> {
    reader.as_mut().poll_fill_buf(cx);
    return Poll::Ready(());
}

struct MyStruct<'a, R2: ?Sized> {
    reader: &'a R2,
}

impl<R3: AsyncBufRead + ?Sized> Future for MyStruct<'_, R3> {
    type Output = ();

    fn poll(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Self::Output> {
        poll_once(Pin::new(&mut self.reader), cx)
    }
}

The type bounds for R1 and R3 seem to be identical, but are actually slightly different:

error[E0277]: `R3` cannot be unpinned
  --> src/lib.rs:19:9
   |
6  | fn poll_once<R1: AsyncBufRead + ?Sized>(mut reader: Pin<&mut R1>, cx: &mut Context<'_>) -> Poll<()> {
   |                  ------------ required by this bound in `poll_once`
...
19 |         poll_once(Pin::new(&mut self.reader), cx)
   |         ^^^^^^^^^ the trait `std::marker::Unpin` is not implemented for `R3`
   |
   = note: required because of the requirements on the impl of `futures_io::if_std::AsyncBufRead` for `&mut R3`
help: consider further restricting this bound
   |
15 | impl<R3: AsyncBufRead + ?Sized + std::marker::Unpin> Future for MyStruct<'_, R3> {
   |                                ^^^^^^^^^^^^^^^^^^^^

For struct, if they have no Pin elements, Unpin is automatically implemented. Just need to make sure that type bounds contain Unpin, or weird things happen when trying to use them:

#![allow(unused_mut)]
use std::future::Future;
use std::pin::Pin;
use std::task::{Context, Poll};

struct CantUnpin<T> {
    items: Vec<T>
}

impl<T: Default> Future for CantUnpin<T> {
    type Output = ();

    fn poll(mut self: Pin<&mut Self>, _cx: &mut Context<'_>) -> Poll<Self::Output> {
        self.items.push(T::default());
        Poll::Ready(())
    }
}

struct CanUnpin<T> {
    items: Vec<T>
}

impl<T: Default + Unpin> Future for CanUnpin<T> {
    type Output = ();

    fn poll(mut self: Pin<&mut Self>, _cx: &mut Context<'_>) -> Poll<Self::Output> {
        self.items.push(T::default());
        Poll::Ready(())
    }
}

error[E0596]: cannot borrow data in a dereference of `std::pin::Pin<&mut CantUnpin<T>>` as mutable
  --> src/lib.rs:14:9
   |
14 |         self.items.push(T::default());
   |         ^^^^^^^^^^ cannot borrow as mutable
   |
   = help: trait `DerefMut` is required to modify through a dereference, but it is not implemented for `std::pin::Pin<&mut CantUnpin<T>>`

Rule of thumb: If you don't know whether it's safe to require Unpin, it almost certainly is. Worst case, can write a "compile test"; put code in a closure that's never called.

Know what the escape hatches are

When used sparingly, either #[async_trait] or BoxFuture can enable async functionality in code that will later not need the allocations. Use the escape hatch when you need to such that you can continue making incremental improvements later.

Specific trick: use BoxFuture for type erasure:

use std::future::Future;
use std::pin::Pin;
use std::task::{Context, Poll};
use futures::future::BoxFuture;

async fn function1() {}

async fn function2() -> u8 { 0 }

pub struct MyStruct<T> {
    f: BoxFuture<'static, T>
}

impl<T> Future for MyStruct<T> {
    type Output = T;

    fn poll(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<T> {
        self.f.as_mut().poll(cx)
    }
}

pub fn another_function() -> MyStruct<u8> {
    MyStruct { f: Box::pin(async {
        function1().await;
        function2().await
    }) }
}

There's one allocation because of Box::pin(), but that's it. We're allowed to use an opaque impl Future and still return values from it.

---

Other thoughts that may be helpful in writing:

Plenty of reasons to write low level futures code; if you feel guilty about every heap allocation because you wonder if it's really necessary, if you have to write traits, no_std or no alloc environments

These are tools to help the mortals who don't really understand the Pin system yet. It's like fighting the borrow checker; you'll probably figure it out eventually, but eventually doesn't help you right now.

Unpin seems to mostly mean "safe to move", so everything that doesn't interact with the pin system normally is probably fine.

Also need a note about enum type parameters and pinning

Principle: pinning is needed so Rust can desugar references across await points. Practically? No idea how to meaningfully use it, what the purpose of pin project is, or how to actually create a struct with internal reference.