cranpose_services/

async_io.rs

//! Waking a future from another thread.
//!
//! A platform's own I/O is blocking — a synchronous HTTP read, a provider call,
//! a file descriptor a system API hands over — and the framework has no thread
//! pool to hide that behind. What it has is the waker: work runs on a thread of
//! its own and, when it has something, wakes whoever was awaiting it.
//!
//! Two shapes cover everything here. A [`Signal`] carries one value, which is
//! what a request's status line is. A [`ChunkChannel`] carries a stream of byte
//! chunks with the consumer's progress bounding the producer, which is what a
//! response body is — without the bound, a slow reader and a fast server put the
//! whole download in memory, which is the thing streaming exists to avoid.

use std::collections::VecDeque;
use std::future::Future;
use std::pin::Pin;
use std::sync::{Arc, Condvar, Mutex};
use std::task::{Context, Poll, Waker};

/// How many chunks may wait ahead of the consumer before the producer stops
/// reading. Enough to keep a socket busy across one scheduling gap, and far
/// short of holding a download in memory.
pub const MAX_PENDING_CHUNKS: usize = 8;

struct SignalState<T> {
    value: Option<T>,
    waker: Option<Waker>,
    closed: bool,
}

/// A one-value hand-off between a worker and a future.
///
/// The worker calls [`Signal::set`]; whoever awaits [`Signal::wait`] is woken
/// with it. A worker that dies without setting anything closes the signal, and
/// the wait resolves to `None` rather than hanging for ever.
pub struct Signal<T> {
    state: Arc<Mutex<SignalState<T>>>,
}

impl<T> Clone for Signal<T> {
    fn clone(&self) -> Self {
        Self {
            state: Arc::clone(&self.state),
        }
    }
}

impl<T> Default for Signal<T> {
    fn default() -> Self {
        Self::new()
    }
}

impl<T> Signal<T> {
    pub fn new() -> Self {
        Self {
            state: Arc::new(Mutex::new(SignalState {
                value: None,
                waker: None,
                closed: false,
            })),
        }
    }

    /// Delivers the value and wakes the waiter. A second call is ignored: one
    /// signal carries one value.
    pub fn set(&self, value: T) {
        let waker = {
            let mut state = lock(&self.state);
            if state.closed {
                return;
            }
            state.value = Some(value);
            state.closed = true;
            state.waker.take()
        };
        if let Some(waker) = waker {
            waker.wake();
        }
    }

    /// Ends the signal with no value, so a waiter stops waiting.
    pub fn close(&self) {
        let waker = {
            let mut state = lock(&self.state);
            if state.closed {
                return;
            }
            state.closed = true;
            state.waker.take()
        };
        if let Some(waker) = waker {
            waker.wake();
        }
    }

    /// Resolves with the value, or `None` when the signal was closed empty.
    pub fn wait(&self) -> SignalWait<T> {
        SignalWait {
            state: Arc::clone(&self.state),
        }
    }
}

/// The future [`Signal::wait`] returns.
pub struct SignalWait<T> {
    state: Arc<Mutex<SignalState<T>>>,
}

impl<T> Future for SignalWait<T> {
    type Output = Option<T>;

    fn poll(self: Pin<&mut Self>, context: &mut Context<'_>) -> Poll<Option<T>> {
        let mut state = lock(&self.state);
        if let Some(value) = state.value.take() {
            return Poll::Ready(Some(value));
        }
        if state.closed {
            return Poll::Ready(None);
        }
        state.waker = Some(context.waker().clone());
        Poll::Pending
    }
}

struct ChunkState<E> {
    ready: VecDeque<Vec<u8>>,
    error: Option<E>,
    finished: bool,
    /// The consumer has gone, or asked to stop. The producer notices at its
    /// next push and stops reading rather than filling a queue nobody drains.
    abandoned: bool,
    waker: Option<Waker>,
}

struct ChunkShared<E> {
    state: Mutex<ChunkState<E>>,
    /// The producer parks here while the consumer is behind.
    room: Condvar,
}

/// The producing half of a chunked byte stream.
///
/// Held by whatever is doing the blocking read. Dropping it without calling
/// [`ChunkChannel::finish`] or [`ChunkChannel::fail`] ends the stream, so a
/// worker that panics does not leave a reader waiting for ever.
pub struct ChunkChannel<E> {
    shared: Arc<ChunkShared<E>>,
}

impl<E> ChunkChannel<E> {
    /// Creates a channel and its reading half.
    pub fn new() -> (Self, ChunkStream<E>) {
        let shared = Arc::new(ChunkShared {
            state: Mutex::new(ChunkState {
                ready: VecDeque::new(),
                error: None,
                finished: false,
                abandoned: false,
                waker: None,
            }),
            room: Condvar::new(),
        });
        (
            Self {
                shared: Arc::clone(&shared),
            },
            ChunkStream { shared },
        )
    }

    /// Publishes one chunk, waiting while the consumer is more than
    /// [`MAX_PENDING_CHUNKS`] behind.
    ///
    /// Returns `false` once the consumer has gone, which is the producer's
    /// signal to stop reading.
    pub fn push(&self, chunk: Vec<u8>) -> bool {
        let waker = {
            let mut state = lock(&self.shared.state);
            // The browser has one thread: parking it would stop the very task
            // that drains the queue, so there the bound is advisory.
            #[cfg(not(target_arch = "wasm32"))]
            while state.ready.len() >= MAX_PENDING_CHUNKS && !state.abandoned {
                state = self
                    .shared
                    .room
                    .wait(state)
                    .unwrap_or_else(|error| error.into_inner());
            }
            if state.abandoned || state.finished {
                return false;
            }
            state.ready.push_back(chunk);
            state.waker.take()
        };
        if let Some(waker) = waker {
            waker.wake();
        }
        true
    }

    /// Ends the stream with an error.
    pub fn fail(&self, error: E) {
        let waker = {
            let mut state = lock(&self.shared.state);
            if state.finished {
                return;
            }
            state.error = Some(error);
            state.finished = true;
            state.waker.take()
        };
        if let Some(waker) = waker {
            waker.wake();
        }
    }

    /// Ends the stream normally.
    pub fn finish(&self) {
        let waker = {
            let mut state = lock(&self.shared.state);
            if state.finished {
                return;
            }
            state.finished = true;
            state.waker.take()
        };
        if let Some(waker) = waker {
            waker.wake();
        }
    }

    /// Whether the consumer has gone.
    pub fn is_abandoned(&self) -> bool {
        lock(&self.shared.state).abandoned
    }
}

impl<E> Drop for ChunkChannel<E> {
    fn drop(&mut self) {
        self.finish();
    }
}

/// The consuming half of a chunked byte stream.
pub struct ChunkStream<E> {
    shared: Arc<ChunkShared<E>>,
}

impl<E> ChunkStream<E> {
    /// Resolves with the next chunk, `Ok(None)` at the end of the stream, or
    /// the error the producer ended with.
    pub fn next(&self) -> ChunkNext<'_, E> {
        ChunkNext { stream: self }
    }
}

impl<E> Drop for ChunkStream<E> {
    fn drop(&mut self) {
        let mut state = lock(&self.shared.state);
        state.abandoned = true;
        drop(state);
        // A producer parked on the bound has to learn nobody is reading.
        self.shared.room.notify_all();
    }
}

/// The future [`ChunkStream::next`] returns.
pub struct ChunkNext<'a, E> {
    stream: &'a ChunkStream<E>,
}

impl<E> Future for ChunkNext<'_, E> {
    type Output = Result<Option<Vec<u8>>, E>;

    fn poll(self: Pin<&mut Self>, context: &mut Context<'_>) -> Poll<Self::Output> {
        let shared = Arc::clone(&self.stream.shared);
        let mut state = lock(&shared.state);
        if let Some(chunk) = state.ready.pop_front() {
            drop(state);
            shared.room.notify_one();
            return Poll::Ready(Ok(Some(chunk)));
        }
        if let Some(error) = state.error.take() {
            return Poll::Ready(Err(error));
        }
        if state.finished {
            return Poll::Ready(Ok(None));
        }
        state.waker = Some(context.waker().clone());
        Poll::Pending
    }
}

fn lock<T>(mutex: &Mutex<T>) -> std::sync::MutexGuard<'_, T> {
    mutex.lock().unwrap_or_else(|error| error.into_inner())
}

#[cfg(test)]
mod tests {
    use super::*;

    #[derive(Debug, PartialEq, Eq)]
    struct Failed(&'static str);

    #[test]
    fn a_signal_carries_one_value_to_whoever_waits() {
        let signal = Signal::new();
        signal.set(7u32);
        assert_eq!(pollster::block_on(signal.wait()), Some(7));
    }

    /// A worker that dies without answering must not leave a reader hanging.
    #[test]
    fn a_closed_signal_resolves_to_nothing_rather_than_waiting_for_ever() {
        let signal = Signal::<u32>::new();
        signal.close();
        assert_eq!(pollster::block_on(signal.wait()), None);
    }

    #[test]
    fn a_signal_set_from_another_thread_wakes_the_waiter() {
        let signal = Signal::new();
        let worker = signal.clone();
        let handle = std::thread::spawn(move || {
            std::thread::sleep(std::time::Duration::from_millis(20));
            worker.set(11u32);
        });
        assert_eq!(pollster::block_on(signal.wait()), Some(11));
        handle.join().expect("the worker finishes");
    }

    #[test]
    fn chunks_arrive_in_the_order_they_were_produced() {
        let (channel, stream) = ChunkChannel::<Failed>::new();
        assert!(channel.push(b"one".to_vec()));
        assert!(channel.push(b"two".to_vec()));
        channel.finish();

        assert_eq!(pollster::block_on(stream.next()), Ok(Some(b"one".to_vec())));
        assert_eq!(pollster::block_on(stream.next()), Ok(Some(b"two".to_vec())));
        assert_eq!(pollster::block_on(stream.next()), Ok(None));
    }

    #[test]
    fn a_failed_stream_reports_the_error_after_what_it_already_produced() {
        let (channel, stream) = ChunkChannel::new();
        assert!(channel.push(b"partial".to_vec()));
        channel.fail(Failed("the connection dropped"));

        assert_eq!(
            pollster::block_on(stream.next()),
            Ok(Some(b"partial".to_vec()))
        );
        assert_eq!(
            pollster::block_on(stream.next()),
            Err(Failed("the connection dropped"))
        );
    }

    /// A worker that panics must end the stream rather than leaving a reader
    /// waiting on a chunk nobody will produce.
    #[test]
    fn dropping_the_producer_ends_the_stream() {
        let (channel, stream) = ChunkChannel::<Failed>::new();
        drop(channel);
        assert_eq!(pollster::block_on(stream.next()), Ok(None));
    }

    /// Without a bound a fast server and a slow reader put the whole download in
    /// memory, which is the thing streaming exists to avoid.
    #[test]
    fn the_producer_waits_while_the_consumer_is_behind() {
        let (channel, stream) = ChunkChannel::<Failed>::new();
        let pushed = Arc::new(std::sync::atomic::AtomicUsize::new(0));
        let counter = Arc::clone(&pushed);
        let worker = std::thread::spawn(move || {
            for index in 0..MAX_PENDING_CHUNKS + 4 {
                if !channel.push(vec![index as u8]) {
                    break;
                }
                counter.fetch_add(1, std::sync::atomic::Ordering::Release);
            }
            channel.finish();
        });

        // Give the producer a chance to fill the queue and park on the bound.
        std::thread::sleep(std::time::Duration::from_millis(50));
        assert!(
            pushed.load(std::sync::atomic::Ordering::Acquire) <= MAX_PENDING_CHUNKS,
            "the producer must stop at the bound rather than reading ahead without limit"
        );

        let mut received = 0usize;
        while let Ok(Some(_)) = pollster::block_on(stream.next()) {
            received += 1;
        }
        assert_eq!(received, MAX_PENDING_CHUNKS + 4);
        worker.join().expect("the worker finishes");
    }

    /// A reader that stops early — a cancelled download, a dropped screen —
    /// must stop the producer rather than leaving it reading into a queue
    /// nobody drains.
    #[test]
    fn abandoning_the_stream_stops_the_producer() {
        let (channel, stream) = ChunkChannel::<Failed>::new();
        assert!(channel.push(b"first".to_vec()));
        drop(stream);
        assert!(channel.is_abandoned());
        assert!(
            !channel.push(b"second".to_vec()),
            "a push after the consumer has gone must report that nobody is reading"
        );
    }
}