Crate corona

A library combining futures and coroutines.

This library brings stack-full coroutines. Each coroutine can asynchronously wait on futures and provides a future of its result.

Motivation

The current aim of Rust in regards to asynchronous programming is on futures. They have some good properties, but some tasks are more conveniently done in a more imperative way.

There's the work in progress of async-await. But it requires nightly (for now), provides stack-less coroutines (which means the asynchronous waiting can be done in a top-level function only) and there are too many 'static bounds. Something might improve over time.

This library brings a more convenient interface. However, it comes with a run-time cost, so you might want to consider if you prefer ease of development or memory efficiency. Often, the asynchronous communication isn't the bottleneck and you won't be handling millions of concurrent connections, only tens of thousands, so this might be OK.

Pros

• Easier to use than futures.
• Can integrate with futures.
• Allows working with borrowed futures.
• Provides safe interface.

Cons

• Each coroutine needs its own stack, which is at least few memory pages large. This makes the library unsuitable when there are many concurrent coroutines.
• The coroutines can't move between threads once created safely. This library is tied into tokio-current-threaded executor (it might be possible to remove this tie-in, but not the single-threadiness).

How to use

By bringing the corona::prelude::* into scope, all Futures, Streams and Sinks get new methods for asynchronous waiting on their completion or progress.

The Coroutine is used to start a new coroutine. It acts a bit like std::thread::spawn. However, all the coroutines run on the current thread and switch to other coroutines whenever they wait on something. This must be done from the context of current-thread executor, which is easiest by wrapping the whole application into tokio::runtime::current_thread::block_on_all and future::lazy.

There's also the Coroutine::run that does the same (but is available only in case the convenient-run feature is not turned off).

Each new coroutine returns a future. It resolves whenever the coroutine terminates. However, the coroutine is eager ‒ it doesn't wait with the execution for the future to be polled. The future can be dropped and the coroutine will still execute.

extern crate corona;
extern crate tokio;

use std::time::Duration;

use corona::prelude::*;
use tokio::clock;
use tokio::prelude::*;
use tokio::runtime::current_thread;
use tokio::timer::Delay;

fn main() {
    let result = current_thread::block_on_all(future::lazy(|| {
        Coroutine::with_defaults(|| {
            let timeout = Delay::new(clock::now() + Duration::from_millis(50));
            timeout.coro_wait().unwrap(); // Timeouts don't error
            42
        })
    })).unwrap();
    assert_eq!(42, result);
}

extern crate corona;
extern crate tokio;

use std::time::Duration;

use corona::prelude::*;
use tokio::clock;
use tokio::prelude::*;
use tokio::runtime::current_thread;
use tokio::timer::Delay;

fn main() {
    let result = Coroutine::new()
        .run(|| {
            let timeout = Delay::new(clock::now() + Duration::from_millis(50));
            timeout.coro_wait().unwrap(); // Timeouts don't error
            42
        }).unwrap();
    assert_eq!(42, result);
}

Few things of note:

• All the coroutine-aware methods panic outside of a coroutine.
• Many of them panic outside of a current-thread executor.
• You can freely mix future and coroutine approach. Therefore, you can use combinators to build a future and then coro_wait on it.
• A coroutine spawned by spawn or with_defaults will propagate panics outside. One spawned with spawn_catch_panic captures the panic and passes it on through its result.
• Panicking outside of the coroutine where the executor runs may lead to ugly things, like aborting the program (this'd usually lead to a double panic).
• Any of the waiting methods may switch to a different coroutine. Therefore it is not a good idea to hold a RefCell borrowed or a Mutex locked around that if another coroutine could also borrow it.

The new methods are here:

• Futures
• Streams
• Sinks

Coroutine-blocking IO

Furthermore, if the blocking-wrappers feature is enabled (it is by default), all AsyncRead and AsyncWrite objects can be wrapped in corona::io::BlockingWrapper. This implements Read and Write in a way that mimics blocking, but it blocks only the coroutine, not the whole thread. This allows it to be used with usual blocking routines, like serde_json::from_reader.

The API is still a bit rough (it exposes just the Read and Write traits, all other methods need to be accessed through .inner() or .inner_mut), this will be improved in future versions.

use std::io::{Read, Result as IoResult};
use corona::io::BlockingWrapper;
use tokio::net::TcpStream;

fn blocking_read(connection: &mut TcpStream) -> IoResult<()> {
    let mut connection = BlockingWrapper::new(connection);
    let mut buf = [0u8; 64];
    // This will block the coroutine, but not the thread
    connection.read_exact(&mut buf)
}

Cleaning up

If the executor is dropped while a coroutine waits on something, the waiting method will panic. That way the coroutine's stack is unwinded, releasing resources on its stack (there doesn't seem to be a better way to drop the whole stack).

However, if the executor is dropped because of a panic, Rust abort the whole program because of a double-panic. Ideas how to overcome this (since the second panic is on a different stack, but Rust doesn't know that) are welcome.

There are waiting methods that return an error instead of panicking, but they are less convenient to use.

It also can be configured to leak the stack in such case instead of double-panicking.

Pitfalls

If the coroutine is created with default configuration, it gets really small stack. If you overflow it, you get a segfault (it happens more often on debug builds than release ones) with really useless backtrace. Try making the stack bigger in that case.

API Stability

The API is likely to get stabilized soon (I hope it won't change much any more). But I still want to do more experimentation before making it official.

There are two areas where I expect some changes will still be needed:

• I want to support scoped coroutines (similar to some libraries that provide scoped threads).

Other experiments from consumers are also welcome.

Known problems

These are the problems I'm aware of and which I want to find a solution some day.

• Many handy abstractions are still missing, like waiting for a future with a timeout, or conveniently waiting for a first of a set of futures or streams.
• The coroutines can't move between threads. This is likely impossible, since Rust's type system doesn't expect whole stacks with all local data to move.
• It relies on Tokio. This might change in the future.
• The API doesn't prevent some footguns ‒ leaving a RefCell borrowed across coroutine switch, deadlocking, calling the waiting methods outside of a coroutine or using .wait() by a mistake and blocking the whole thread. These manifest during runtime.
• The cleaning up of coroutines when the executor is dropped is done through panics.

Contribution

All kinds of contributions are welcome, including reporting bugs, improving the documentation, submitting code, etc. However, the most valuable contribution for now would be trying it out and providing some feedback ‒ if the thing works, where the API needs improvements, etc.

Examples

One that shows the API.

use std::time::Duration;
use futures::unsync::mpsc;
use corona::prelude::*;
use tokio::clock;
use tokio::prelude::*;
use tokio::runtime::current_thread;
use tokio::timer::Delay;

let result = Coroutine::new().run(|| {
    let (sender, receiver) = mpsc::channel(1);
    corona::spawn(|| {
        let mut sender = sender;
        sender = sender.send(1).coro_wait().unwrap();
        sender = sender.send(2).coro_wait().unwrap();
    });

    for item in receiver.iter_ok() {
        println!("{}", item);
    }

    let timeout = Delay::new(clock::now() + Duration::from_millis(100));
    timeout.coro_wait().unwrap();

    42
}).unwrap();
assert_eq!(42, result);

Further examples can be found in the repository.

Behind the scenes

There are few things that might help understanding how the library works inside.

First, there's some thread-local state. This state is used for caching currently unused stacks as well as the state that is used when waiting for something and switching coroutines.

Whenever one of the waiting methods is used, a wrapper future is created. After the original future resolves, it resumes the execution to the current stack. This future is spawned onto the executor and a switch is made to the parent coroutine (it's the coroutine that started or resumed the current one). This way, the „outside“ coroutine is reached eventually. It is expected this outside coroutine will run the executor, waking up the ready to proceed coroutines and then switching to them.

That's about it, the rest of the library are just implementation details about what is stored where and how to pass the information around without breaking any lifetime bounds.

Re-exports

pub use errors::Dropped;

pub use errors::TaskFailed;

pub use coroutine::spawn;

pub use coroutine::Coroutine;

pub use coroutine::CoroutineResult;

Modules

coroutine	The Coroutine and related things.
errors	Various errors.
io	Primitives to turn AsyncRead and AsyncWrite into (coroutine) blocking Read and Write.
prelude	A module for wildcard import.
wrappers	Various wrappers and helper structs.