use std::{
collections::HashMap,
panic::{AssertUnwindSafe, catch_unwind},
sync::{
Arc, Condvar, Mutex,
atomic::{AtomicBool, Ordering},
mpsc,
},
thread,
};
use rayon::{ThreadPool, ThreadPoolBuilder};
use crate::{Error, Result};
pub(crate) struct Job<T> { pub key: usize, pub reservation: usize, pub payload: T }
pub(crate) struct PipelineLimits { pub memory: usize, pub active: usize }
struct State {
next: usize,
reserved: usize,
active: usize,
cancelled: bool,
}
struct Budget {
limit: usize,
max_active: usize,
state: Mutex<State>,
wake: Condvar,
}
impl Budget {
fn new(limit: usize, max_active: usize) -> Self {
Self { limit, max_active, state: Mutex::new(State { next: 0, reserved: 0, active: 0, cancelled: false }), wake: Condvar::new() }
}
fn next<T>(&self, jobs: &[Job<T>]) -> Option<usize> {
let mut state = self.state.lock().unwrap_or_else(|error| error.into_inner());
loop {
if state.cancelled || state.next >= jobs.len() { return None; }
let reservation = jobs[state.next].reservation;
if state.active < self.max_active && reservation <= self.limit.saturating_sub(state.reserved) {
let next = state.next;
state.next += 1;
state.reserved += reservation;
state.active += 1;
return Some(next);
}
state = self.wake.wait(state).unwrap_or_else(|error| error.into_inner());
}
}
fn complete(&self, reservation: usize, retained: usize) {
let mut state = self.state.lock().unwrap_or_else(|error| error.into_inner());
debug_assert!(retained <= reservation);
state.reserved = state.reserved.saturating_sub(reservation).saturating_add(retained.min(reservation));
self.wake.notify_all();
}
fn retire(&self, retained: usize) {
let mut state = self.state.lock().unwrap_or_else(|error| error.into_inner());
state.reserved = state.reserved.saturating_sub(retained);
state.active = state.active.saturating_sub(1);
self.wake.notify_all();
}
fn cancel(&self) {
let mut state = self.state.lock().unwrap_or_else(|error| error.into_inner());
state.cancelled = true;
self.wake.notify_all();
}
}
type Message<T> = (usize, usize, T);
fn execute_job<T, O>(
index: usize,
jobs: &[Job<T>],
budget: &Budget,
sender: &mpsc::Sender<Message<O>>,
execute: &impl Fn(&T) -> O,
retained_size: &impl Fn(&O) -> usize,
) -> bool {
let job = &jobs[index];
let value = execute(&job.payload);
let retained = retained_size(&value).min(job.reservation);
budget.complete(job.reservation, retained);
if sender.send((job.key, retained, value)).is_ok() { true } else { budget.retire(retained); false }
}
pub(crate) struct OrderedResults<'a, T> { receiver: mpsc::Receiver<Message<T>>, ready: HashMap<usize, (usize, T)>, budget: &'a Budget }
impl<T> OrderedResults<'_, T> {
pub fn take(&mut self, key: usize) -> Result<T> {
while !self.ready.contains_key(&key) {
let (received_key, retained, value) = self.receiver.recv().map_err(|_| Error::InvalidConfiguration("parallel decoder stopped early".into()))?;
self.ready.insert(received_key, (retained, value));
}
let (retained, value) = self.ready.remove(&key).unwrap();
self.budget.retire(retained);
Ok(value)
}
pub fn discard_before(&mut self, key: usize) {
let stale: Vec<_> = self.ready.keys().copied().filter(|&candidate| candidate < key).collect();
for candidate in stale {
let (retained, _) = self.ready.remove(&candidate).unwrap();
self.budget.retire(retained);
}
}
}
pub(crate) fn run_ordered<T, O, R>(
pool: &ThreadPool,
jobs: &[Job<T>],
limits: PipelineLimits,
execute: impl Fn(&T) -> O + Sync,
retained_size: impl Fn(&O) -> usize + Sync,
consume: impl FnOnce(&mut OrderedResults<'_, O>) -> Result<R>,
) -> Result<R>
where
T: Sync,
O: Send,
{
if jobs.iter().any(|job| job.reservation > limits.memory) {
return Err(Error::InvalidConfiguration("a parallel job reservation exceeds the memory limit".into()));
}
let budget = Budget::new(limits.memory, limits.active);
let (sender, receiver) = mpsc::channel();
thread::scope(|scope| {
let worker = scope.spawn(|| {
pool.broadcast(|_| {
while let Some(index) = budget.next(jobs) { if !execute_job(index, jobs, &budget, &sender, &execute, &retained_size) { return; } }
});
});
let result = { let mut results = OrderedResults { receiver, ready: HashMap::new(), budget: &budget }; consume(&mut results) };
budget.cancel();
worker.join().map_err(|_| Error::InvalidConfiguration("parallel decoder worker panicked".into()))?;
result
})
}
enum TryNext { Ready(usize), Pending, Cancelled }
impl Budget {
fn try_next<T>(&self, jobs: &[Job<T>]) -> TryNext {
let mut state = self.state.lock().unwrap_or_else(|error| error.into_inner());
if state.cancelled { return TryNext::Cancelled; }
if state.next >= jobs.len() { return TryNext::Pending; }
let reservation = jobs[state.next].reservation;
if state.active >= self.max_active || reservation > self.limit.saturating_sub(state.reserved) { return TryNext::Pending; }
let next = state.next;
state.next += 1;
state.reserved += reservation;
state.active += 1;
TryNext::Ready(next)
}
fn wait_briefly(&self) {
let state = self.state.lock().unwrap_or_else(|error| error.into_inner());
let _ = self.wake.wait_timeout(state, std::time::Duration::from_millis(1));
}
fn notify(&self) { self.wake.notify_all(); }
}
pub(crate) struct Lease { retained: usize }
pub(crate) struct StagedResults<'a, O, S, Q> {
primary: OrderedResults<'a, O>,
stage_sender: mpsc::Sender<Message<S>>,
stage_receiver: mpsc::Receiver<(usize, Q)>,
stage_ready: HashMap<usize, Q>,
}
impl<'a, O, S, Q> StagedResults<'a, O, S, Q> {
pub fn take_primary(&mut self, key: usize) -> Result<(Lease, O)> {
while !self.primary.ready.contains_key(&key) {
let (received_key, retained, value) =
self.primary.receiver.recv().map_err(|_| Error::InvalidConfiguration("parallel decoder stopped early".into()))?;
self.primary.ready.insert(received_key, (retained, value));
}
let (retained, value) = self.primary.ready.remove(&key).unwrap();
Ok((Lease { retained }, value))
}
pub fn retire(&self, lease: Lease) { self.primary.budget.retire(lease.retained); }
pub fn submit(&self, key: usize, lease: Lease, payload: S) -> Result<()> {
if let Err(mpsc::SendError((_, retained, _))) = self.stage_sender.send((key, lease.retained, payload)) {
self.primary.budget.retire(retained);
return Err(Error::InvalidConfiguration("parallel decoder stopped before staged work was submitted".into()));
}
self.primary.budget.notify();
Ok(())
}
pub fn take_stage(&mut self, key: usize) -> Result<Q> {
while !self.stage_ready.contains_key(&key) {
let (received_key, value) = self.stage_receiver.recv().map_err(|_| Error::InvalidConfiguration("parallel staged worker stopped early".into()))?;
self.stage_ready.insert(received_key, value);
}
Ok(self.stage_ready.remove(&key).unwrap())
}
}
pub(crate) fn run_staged_ordered<T, O, S, Q, R>(
worker_count: usize,
jobs: &[Job<T>],
limits: PipelineLimits,
execute: impl Fn(&T) -> O + Sync,
retained_size: impl Fn(&O) -> usize + Sync,
execute_stage: impl Fn(S) -> Q + Sync,
consume: impl FnOnce(&mut StagedResults<'_, O, S, Q>) -> Result<R>,
) -> Result<R>
where
T: Sync,
O: Send,
S: Send,
Q: Send,
{
if jobs.iter().any(|job| job.reservation > limits.memory) {
return Err(Error::InvalidConfiguration("a parallel job reservation exceeds the memory limit".into()));
}
let budget = Budget::new(limits.memory, limits.active);
let (primary_sender, primary_receiver) = mpsc::channel();
let (stage_sender, stage_job_receiver) = mpsc::channel();
let stage_job_receiver = Mutex::new(stage_job_receiver);
let (stage_result_sender, stage_receiver) = mpsc::channel();
thread::scope(|scope| {
let workers: Vec<_> = (0..worker_count)
.map(|_| {
scope.spawn(|| {
loop {
let stage = stage_job_receiver.lock().unwrap_or_else(|error| error.into_inner()).try_recv();
match stage {
Ok((key, retained, payload)) => {
let value = execute_stage(payload);
budget.retire(retained);
if stage_result_sender.send((key, value)).is_err() { return; }
continue;
}
Err(mpsc::TryRecvError::Disconnected) => return,
Err(mpsc::TryRecvError::Empty) => {}
}
match budget.try_next(jobs) {
TryNext::Ready(index) => {
if !execute_job(index, jobs, &budget, &primary_sender, &execute, &retained_size) { return; }
}
TryNext::Pending => budget.wait_briefly(),
TryNext::Cancelled => return,
}
}
})
})
.collect();
let result = {
let primary = OrderedResults { receiver: primary_receiver, ready: HashMap::new(), budget: &budget };
let mut results = StagedResults { primary, stage_sender, stage_receiver, stage_ready: HashMap::new() };
consume(&mut results)
};
budget.cancel();
for worker in workers { worker.join().map_err(|_| Error::InvalidConfiguration("parallel decoder worker panicked".into()))?; }
result
})
}
type StreamingMessage<T> = (usize, usize, std::thread::Result<T>);
pub(crate) struct StreamingOrdered<T> {
pool: ThreadPool,
sender: mpsc::Sender<StreamingMessage<T>>,
receiver: mpsc::Receiver<StreamingMessage<T>>,
ready: HashMap<usize, (usize, std::thread::Result<T>)>,
next_submit: usize,
next_take: usize,
reserved: usize,
active: usize,
max_active: usize,
memory_limit: usize,
cancelled: Arc<AtomicBool>,
}
impl<T: Send + 'static> StreamingOrdered<T> {
pub fn new(threads: usize, memory_limit: usize, thread_name: &'static str) -> Result<Self> {
let max_active = threads.max(1);
let pool = ThreadPoolBuilder::new()
.num_threads(max_active)
.thread_name(move |index| format!("{thread_name}-{index}"))
.build()
.map_err(|error| Error::InvalidConfiguration(error.to_string()))?;
let (sender, receiver) = mpsc::channel();
Ok(Self {
pool,
sender,
receiver,
ready: HashMap::new(),
next_submit: 0,
next_take: 0,
reserved: 0,
active: 0,
max_active,
memory_limit,
cancelled: Arc::new(AtomicBool::new(false)),
})
}
pub fn can_submit(&self, reservation: usize) -> bool { self.active < self.max_active && reservation <= self.memory_limit.saturating_sub(self.reserved) }
pub fn submit(&mut self, reservation: usize, job: impl FnOnce() -> T + Send + 'static) -> Result<()> {
if reservation > self.memory_limit { return Err(Error::InvalidConfiguration("a streaming job reservation exceeds the memory limit".into())); }
if !self.can_submit(reservation) { return Err(Error::InvalidConfiguration("streaming pipeline is full".into())); }
let key = self.next_submit;
self.next_submit += 1;
self.active += 1;
self.reserved += reservation;
let sender = self.sender.clone();
let cancelled = Arc::clone(&self.cancelled);
self.pool.spawn(move || {
if cancelled.load(Ordering::Relaxed) { return; }
let result = catch_unwind(AssertUnwindSafe(job));
let _ = sender.send((key, reservation, result));
});
Ok(())
}
pub fn take_next(&mut self) -> Result<T> {
if self.next_take >= self.next_submit { return Err(Error::InvalidConfiguration("streaming pipeline has no pending result".into())); }
while !self.ready.contains_key(&self.next_take) {
let message = self.receiver.recv().map_err(|_| Error::InvalidConfiguration("streaming worker stopped early".into()))?;
self.ready.insert(message.0, (message.1, message.2));
}
let (reservation, result) = self.ready.remove(&self.next_take).unwrap();
self.next_take += 1;
self.active -= 1;
self.reserved -= reservation;
result.map_err(|_| Error::InvalidConfiguration("streaming worker panicked".into()))
}
pub fn has_pending(&self) -> bool { self.next_take < self.next_submit }
}
impl<T> Drop for StreamingOrdered<T> { fn drop(&mut self) { self.cancelled.store(true, Ordering::Relaxed); } }