Rust Async Internals

Rust Async Internals

Purpose

Guide agents through Rust async/await internals: the Future trait and poll loop, Pin/Unpin for self-referential types, tokio's task model, diagnosing async stack traces with tokio-console, finding waker leaks, and common select!/join! pitfalls.

Triggers

• "How does async/await actually work in Rust?"
• "What is Pin and Unpin in async Rust?"
• "My async code is slow — how do I profile it?"
• "How do I use tokio-console to debug async tasks?"
• "I have a blocking call in async — what do I do?"
• "How does select! work and what are the pitfalls?"

Workflow

1. The Future trait — poll model

// std::future::Future (simplified)
pub trait Future {
    type Output;
    fn poll(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Self::Output>;
}

pub enum Poll<T> {
    Ready(T),    // computation done, T is the result
    Pending,     // not ready yet, waker registered, will be polled again
}

Execution model:

1. Calling .await calls poll() on the future
2. If Pending: current task registers its waker and yields to the runtime
3. When the waker is triggered (I/O ready, timer fired), the runtime re-polls
4. If Ready(val): the .await expression evaluates to val

2. Implementing a simple Future

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

struct Delay { deadline: Instant }

impl Delay {
    fn new(dur: Duration) -> Self {
        Delay { deadline: Instant::now() + dur }
    }
}

impl Future for Delay {
    type Output = ();

    fn poll(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<()> {
        if Instant::now() >= self.deadline {
            Poll::Ready(())
        } else {
            // Register the waker — runtime calls waker.wake() to re-poll
            // In production: register with I/O reactor or timer wheel
            let waker = cx.waker().clone();
            let deadline = self.deadline;
            std::thread::spawn(move || {
                let now = Instant::now();
                if deadline > now {
                    std::thread::sleep(deadline - now);
                }
                waker.wake();  // notify runtime to re-poll
            });
            Poll::Pending
        }
    }
}

// Usage
async fn main() {
    Delay::new(Duration::from_secs(1)).await;
    println!("Done");
}

3. Pin and Unpin

Pin<P> prevents moving the value behind pointer P. This matters because async state machines contain self-referential pointers (a reference into the same struct where the future lives):

// Why Pin is needed: async fn compiles to a state machine struct
// that may have self-references across await points

async fn example() {
    let data = vec![1, 2, 3];
    let ref_to_data = &data;              // reference into same stack frame
    some_async_op().await;                // suspension point
    println!("{:?}", ref_to_data);       // reference still used after suspend
}
// The state machine stores both `data` and `ref_to_data`.
// If the struct were moved, `ref_to_data` would dangle.
// Pin<&mut State> prevents moving the state machine.

// Unpin: a marker trait for types that are safe to move even when pinned
// Most types implement Unpin automatically
// Futures generated by async/await do NOT implement Unpin

// Creating a Pin from Box (heap allocation → safe)
let boxed: Pin<Box<dyn Future<Output = ()>>> = Box::pin(my_future);

// Pinning to stack (unsafe, use pin! macro)
use std::pin::pin;
let fut = pin!(my_future);
fut.await;   // or poll it directly

4. tokio task model

use tokio::task;

// Spawn a task (runs concurrently on the runtime thread pool)
let handle = tokio::spawn(async {
    // ... async work ...
    42
});
let result = handle.await.unwrap();  // wait for completion

// spawn_blocking — for CPU-bound or blocking I/O
let result = task::spawn_blocking(|| {
    // runs on a dedicated blocking thread pool
    std::fs::read_to_string("big_file.txt")
}).await.unwrap();

// yield to runtime (cooperative multitasking)
tokio::task::yield_now().await;

// LocalSet — for !Send futures (single-threaded)
let local = task::LocalSet::new();
local.run_until(async {
    task::spawn_local(async { /* !Send future */ }).await.unwrap();
}).await;

5. tokio-console — async task inspector

# Cargo.toml
[dependencies]
console-subscriber = "0.3"
tokio = { version = "1", features = ["full", "tracing"] }

// main.rs
fn main() {
    console_subscriber::init();  // must be called before tokio runtime
    tokio::runtime::Builder::new_multi_thread()
        .enable_all()
        .build()
        .unwrap()
        .block_on(async_main());
}

# Install tokio-console CLI
cargo install --locked tokio-console

# Run your app with tracing enabled
RUSTFLAGS="--cfg tokio_unstable" cargo run

# In another terminal, connect tokio-console
tokio-console

# tokio-console shows:
# - Running tasks with their names, poll times, and wakeup counts
# - Slow tasks (high poll duration = blocking in async!)
# - Tasks that have been pending for a long time (stuck?)
# - Resource contention (mutex/semaphore wait times)

6. Blocking in async — common mistake

// WRONG: blocking call in async context blocks entire thread
async fn bad() {
    std::thread::sleep(Duration::from_secs(1));  // blocks runtime thread!
    std::fs::read_to_string("file.txt").unwrap(); // blocking I/O blocks runtime!
}

// CORRECT: use async equivalents
async fn good() {
    tokio::time::sleep(Duration::from_secs(1)).await;   // async sleep
    tokio::fs::read_to_string("file.txt").await.unwrap(); // async I/O
}

// CORRECT: if you must block, use spawn_blocking
async fn with_blocking() {
    let content = tokio::task::spawn_blocking(|| {
        heavy_cpu_computation()   // runs on blocking thread pool
    }).await.unwrap();
}

7. select! and join! pitfalls

use tokio::select;

// select! — complete when FIRST branch completes, cancels others
select! {
    result = fetch_a() => println!("A: {:?}", result),
    result = fetch_b() => println!("B: {:?}", result),
    // Pitfall: the LOSING branches are DROPPED immediately
    // If fetch_a wins, fetch_b's future is dropped (and its state machine cleaned up)
    // This is correct and safe — but can be surprising
}

// join! — wait for ALL to complete
let (a, b) = tokio::join!(fetch_a(), fetch_b());

// Biased select (always check first branch first)
loop {
    select! {
        biased;                        // prevents fairness, checks in order
        _ = shutdown_signal.recv() => break,
        msg = queue.recv() => process(msg),
    }
}

// select! with values from loop (use fuse)
let mut fut = some_future().fuse();   // FusedFuture: safe to poll after completion
loop {
    select! {
        val = &mut fut => { /* ... */ break; }
        _ = interval.tick() => { /* periodic work */ }
    }
}

Related skills

• Use skills/rust/rust-debugging for GDB/LLDB debugging of async Rust programs
• Use skills/rust/rust-profiling for cargo-flamegraph with async stack frames
• Use skills/low-level-programming/cpp-coroutines for C++20 coroutine comparison
• Use skills/low-level-programming/memory-model for memory ordering in async contexts