Amanieu / corosensei

A fast and safe implementation of stackful coroutines in Rust

corosensei

Documentation

Changelog

Overview

This crate provides a safe and efficient abstraction for context switching between different stacks, in the form of coroutines. A coroutine is a function that can be paused and resumed, yielding values to the caller. A coroutine can suspend itself from any point in its call stack. In addition to receiving yielded values from a coroutine, you can also pass data into the coroutine each time it is resumed.

Example

use corosensei::{Coroutine, CoroutineResult};

fn main() {
    println!("[main] creating coroutine");

    let mut coroutine = Coroutine::new(|yielder, input| {
        println!("[coroutine] coroutine started with input {}", input);
        for i in 0..5 {
            println!("[coroutine] yielding {}", i);
            let input = yielder.suspend(i);
            println!("[coroutine] got {} from parent", input)
        }
        println!("[coroutine] exiting coroutine");
    });

    let mut counter = 100;
    loop {
        println!("[main] resuming coroutine with argument {}", counter);
        match coroutine.resume(counter) {
            CoroutineResult::Yield(i) => println!("[main] got {:?} from coroutine", i),
            CoroutineResult::Return(()) => break,
        }

        counter += 1;
    }

    println!("[main] exiting");
}

Output

[main] creating coroutine
[main] resuming coroutine with argument 100
[coroutine] coroutine started with input 100
[coroutine] yielding 0
[main] got 0 from coroutine
[main] resuming coroutine with argument 101
[coroutine] got 101 from parent
[coroutine] yielding 1
[main] got 1 from coroutine
[main] resuming coroutine with argument 102
[coroutine] got 102 from parent
[coroutine] yielding 2
[main] got 2 from coroutine
[main] resuming coroutine with argument 103
[coroutine] got 103 from parent
[coroutine] yielding 3
[main] got 3 from coroutine
[main] resuming coroutine with argument 104
[coroutine] got 104 from parent
[coroutine] yielding 4
[main] got 4 from coroutine
[main] resuming coroutine with argument 105
[coroutine] got 105 from parent
[coroutine] exiting coroutine
[main] exiting

Supported targets

This crate currently supports the following targets:

	ELF (Linux, BSD, bare metal, etc)	Darwin (macOS, iOS, etc)	Windows
x86_64	✅	✅	✅
x86	✅	❌	⚠️*
AArch64	✅	✅	❌
ARM	✅	❌	❌
RISC-V	✅	❌	❌
LoongArch64	✅	❌	❌

* Linked backtraces are not supported on x86 Windows.

Feel free to open an issue if your target is not supported yet.

Features

Panic propagation

If a panic occurs in a coroutine then the panic will unwind through the coroutine stack and then continue to unwind out of the caller which last resumed it. Once this has happened, the coroutine is considered complete and can no longer be resumed.

use std::panic::{catch_unwind, AssertUnwindSafe};

use corosensei::Coroutine;

fn main() {
    println!("[main] creating coroutine");

    let mut coroutine = Coroutine::new(|yielder, ()| {
        println!("[coroutine] yielding 42");
        yielder.suspend(42);

        println!("[coroutine] panicking");
        panic!("foobar");
    });

    println!("[main] resuming coroutine");
    let result = catch_unwind(AssertUnwindSafe(|| coroutine.resume(())));
    println!(
        "[main] got value {} from coroutine",
        result.unwrap().as_yield().unwrap()
    );

    println!("[main] resuming coroutine");
    let result = catch_unwind(AssertUnwindSafe(|| coroutine.resume(())));
    println!(
        "[main] caught panic \"{}\" from coroutine",
        result.unwrap_err().downcast_ref::<&'static str>().unwrap()
    );

    println!("[main] exiting");
}

[main] creating coroutine
[main] resuming coroutine
[coroutine] yielding 42
[main] got value 42 from coroutine
[main] resuming coroutine
[coroutine] panicking
thread 'main' panicked at 'foobar', examples/panic.rs:13:9
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace
[main] caught panic "foobar" from coroutine
[main] exiting

Linked backtraces

Backtraces taken from within a coroutine will continue into the parent stack from the point where they were the coroutine was last resumed from. This is a significant help towards debugging issues with code in coroutines.

Notice how the backtrace in this example shows that the coroutine was last resumed from sub_function:

use backtrace::Backtrace;
use corosensei::Coroutine;

#[inline(never)]
fn sub_function(coroutine: &mut Coroutine<(), (), ()>) {
    coroutine.resume(());
}

fn main() {
    let mut coroutine = Coroutine::new(move |_, ()| {
        let trace = Backtrace::new();
        println!("{:?}", trace);
    });

    sub_function(&mut coroutine);
}

   0: backtrace::main::{{closure}}
             at examples/backtrace.rs:11:21
      corosensei::coroutine::ScopedCoroutine<Input,Yield,Return,Stack>::with_stack::coroutine_func::{{closure}}
             at src/coroutine.rs:174:62
      corosensei::unwind::catch_unwind_at_root
             at src/unwind.rs:43:16
   1: corosensei::coroutine::ScopedCoroutine<Input,Yield,Return,Stack>::with_stack::coroutine_func
             at src/coroutine.rs:174:30
   2: stack_init_trampoline
   3: corosensei::arch::x86_64::switch_and_link
             at src/arch/x86_64.rs:302:5
      corosensei::coroutine::ScopedCoroutine<Input,Yield,Return,Stack>::resume_inner
             at src/coroutine.rs:256:13
      corosensei::coroutine::ScopedCoroutine<Input,Yield,Return,Stack>::resume
             at src/coroutine.rs:235:19
      backtrace::sub_function
             at examples/backtrace.rs:6:5
   4: backtrace::main
             at examples/backtrace.rs:15:5

Linked backtraces use the platform's standard unwinding metadata and work with debuggers, profilers and many other tools.

Cleanup on drop

If a coroutine is dropped while it is suspended then the coroutine's stack will be safely unwound using the same mechanism as panics, which will drop any local variables on the stack. This is necessary to maintain a safe API: the coroutine's stack cannot be freed or reused while there are still live objects on it.

If this cleanup is not desired then a coroutine can forcibly be marked as completed using the unsafe force_reset function.

Trap handling

For advanced use cases, this crate provides the ability to recover from traps (e.g. segmentation fault, stack overflow) that occur in a coroutine. This API is unsafe and should be used from within a signal/exception handler: it will help set up the return from the signal/exception handler so that the coroutine will exit immediately with the given return value and return control to its parent.

This is particularly useful for running JIT-compiled code with a fixed stack limit: stack overflows can be caught by a signal/exception handler and easily recovered from.

Cargo features

This crate is compatible with #![no_std] when all Cargo features are disabled.

The following Cargo feature flags are available on this crate:

default-stack (Enabled by default)

This feature provides a DefaultStack implementation which can be used by the Coroutine type. This stack is allocated with a guard page using OS APIs and requires std.

If this feature is disabled then you must implement your own stack type which implements the Stack trait.

unwind (Enabled by default)

This feature adds support for:

• unwinding panics in a coroutine back out to its caller.
• forcibly unwinding a suspended coroutine via force_unwind or when the coroutine is dropped.

Note that if a coroutine is dropped while suspended (i.e. it has been resumed at least once but has not returned yet) when this feature is disabled then the program will abort.

Requires std.

asm-unwind

This feature uses the asm_unwind nightly feature to allow panics to unwind directly through the inline assembly used in this crate, which can improve performance since it doesn't need to be passed across stack boundaries as a Result.

Implies unwind.

Performance

Corosensei has been heavily optimized for performance, to the point where switching stacks only takes a handful of CPU cycles.

Two benchmarks are available via cargo bench:

• "Coroutine switch" measures the time taken to resume a coroutine and then for that coroutine to yield back to its caller.
• "Coroutine call" measures the time taken to create a new coroutine which returns immediately, call it and then drop the coroutine.

Benchmark results when running on Linux:

Arch	CPU	Frequency	Benchmark	Time	Cycles
AArch64	Apple M1 Max	2.06GHz - 3.22GHz	Coroutine switch	3.8665 ns	N/A
AArch64	Apple M1 Max	2.06GHz - 3.22GHz	Coroutine call	6.4813 ns	N/A
x86-64	AMD Ryzen 9 3950X	3.5GHz - 4.7GHz	Coroutine switch	4.2867 ns	14.8249
x86-64	AMD Ryzen 9 3950X	3.5GHz - 4.7GHz	Coroutine call	5.5082 ns	19.0069
AArch64	ARM Cortex-A72	1.6GHz	Coroutine switch	16.278 ns	N/A
AArch64	ARM Cortex-A72	1.6GHz	Coroutine call	18.769 ns	N/A
LoongArch64	Loongson 3A5000	2.3GHz	Coroutine switch	9.2831 ns	N/A
LoongArch64	Loongson 3A5000	2.3GHz	Coroutine call	16.101 ns	N/A

Credits

This crate was inspired by the following projects:

• libfringe
• Boost.Context
• wasmtime-fiber
• async-wormhole

License

Licensed under either of:

• Apache License, Version 2.0, (LICENSE-APACHE or https://www.apache.org/licenses/LICENSE-2.0)
• MIT license (LICENSE-MIT or https://opensource.org/licenses/MIT)

at your option.

Contribution

Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in the work by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.