turbopack_node/worker_pool/

mod.rs

use std::{
    path::PathBuf,
    sync::{
        Arc,
        atomic::{AtomicU32, Ordering},
    },
};

use anyhow::{Context, Result};
use rustc_hash::FxHashMap;
use tokio::{
    select,
    sync::{Semaphore, oneshot},
    time::sleep,
};
use turbo_rcstr::{RcStr, rcstr};
use turbo_tasks::{ResolvedVc, duration_span};
use turbo_tasks_fs::FileSystemPath;

use crate::{
    AssetsForSourceMapping,
    backend::{CreatePoolFuture, CreatePoolOptions, NodeBackend},
    evaluate::{EvaluateOperation, EvaluatePool, Operation},
    pool_stats::{AcquiredPermits, PoolStatsSnapshot},
    worker_pool::{
        operation::{
            PoolState, TaskChannels, WORKER_POOL_OPERATION, WorkerOperation, WorkerOptions,
            get_pool_state,
        },
        worker_thread::create_worker,
    },
};

mod operation;
mod worker_thread;

static OPERATION_TASK_ID: AtomicU32 = AtomicU32::new(1);

#[turbo_tasks::value(cell = "new", serialization = "none", eq = "manual", shared)]
pub(crate) struct WorkerThreadPool {
    worker_options: Arc<WorkerOptions>,
    concurrency: usize,
    pub(crate) assets_for_source_mapping: ResolvedVc<AssetsForSourceMapping>,
    pub(crate) assets_root: FileSystemPath,
    pub(crate) project_dir: FileSystemPath,
    #[turbo_tasks(trace_ignore, debug_ignore)]
    state: Arc<PoolState>,
    #[turbo_tasks(trace_ignore, debug_ignore)]
    concurrency_semaphore: Arc<Semaphore>,
    #[turbo_tasks(trace_ignore, debug_ignore)]
    bootup_semaphore: Arc<Semaphore>,
}

impl WorkerThreadPool {
    pub(crate) async fn create(
        cwd: PathBuf,
        entrypoint: PathBuf,
        // The worker thread will inherit env from parent process, so it's not needed
        _env: FxHashMap<RcStr, RcStr>,
        assets_for_source_mapping: ResolvedVc<AssetsForSourceMapping>,
        assets_root: FileSystemPath,
        project_dir: FileSystemPath,
        concurrency: usize,
        debug: bool,
    ) -> EvaluatePool {
        let cwd: RcStr = cwd.to_string_lossy().into();
        let filename: RcStr = entrypoint.to_string_lossy().into();
        let worker_options = Arc::new(WorkerOptions { cwd, filename });
        let state = get_pool_state(worker_options.clone()).await;
        EvaluatePool::new(
            Box::new(Self {
                worker_options,
                concurrency: (if debug { 1 } else { concurrency }),
                assets_for_source_mapping,
                assets_root: assets_root.clone(),
                project_dir: project_dir.clone(),
                state,
                concurrency_semaphore: Arc::new(Semaphore::new(if debug {
                    1
                } else {
                    concurrency
                })),
                bootup_semaphore: Arc::new(Semaphore::new(1)),
            }) as Box<dyn EvaluateOperation>,
            assets_for_source_mapping,
            assets_root,
            project_dir,
        )
    }

    async fn acquire_worker(&self) -> Result<(u32, AcquiredPermits)> {
        let concurrency_permit = self.concurrency_semaphore.clone().acquire_owned().await?;

        {
            let mut idle = self.state.idle_workers.lock();
            if let Some(worker_id) = idle.pop() {
                return Ok((
                    worker_id,
                    AcquiredPermits::Idle {
                        _concurrency_permit: concurrency_permit,
                    },
                ));
            }
        }

        let (tx, rx) = oneshot::channel();
        {
            let mut waiters = self.state.waiters.lock();
            let mut idle = self.state.idle_workers.lock();
            if let Some(worker_id) = idle.pop() {
                return Ok((
                    worker_id,
                    AcquiredPermits::Idle {
                        _concurrency_permit: concurrency_permit,
                    },
                ));
            }
            waiters.push(tx);
        }

        let bootup = async {
            let permit = self.bootup_semaphore.clone().acquire_owned().await;
            let wait_time = self.state.stats.lock().wait_time_before_bootup();
            sleep(wait_time).await;
            permit
        };

        select! {
            worker_id = rx => {
                let worker_id = worker_id?;
                Ok((worker_id, AcquiredPermits::Idle { _concurrency_permit: concurrency_permit }))
            }
            bootup_permit = bootup => {
                let bootup_permit = bootup_permit.context("acquiring bootup permit")?;
                {
                    self.state.stats.lock().add_booting_worker();
                }
                let worker_id = create_worker(self.worker_options.clone()).await?;

                {
                    let mut stats = self.state.stats.lock();
                    stats.finished_booting_worker();
                }

                self.bootup_semaphore.add_permits(1);
                Ok((worker_id, AcquiredPermits::Fresh { _concurrency_permit: concurrency_permit, _bootup_permit: bootup_permit }))
            }
        }
    }
}

#[turbo_tasks::value(shared)]
pub(crate) struct WorkerThreadsBackend;

#[turbo_tasks::value_impl]
impl NodeBackend for WorkerThreadsBackend {
    fn runtime_module_path(&self) -> RcStr {
        rcstr!("worker_thread/evaluate.ts")
    }

    fn globals_module_path(&self) -> RcStr {
        rcstr!("worker_thread/globals.ts")
    }

    fn create_pool(&self, options: CreatePoolOptions) -> CreatePoolFuture {
        Box::pin(async move {
            let CreatePoolOptions {
                cwd,
                entrypoint,
                env,
                assets_for_source_mapping,
                assets_root,
                project_dir,
                concurrency,
                debug,
            } = options;

            Ok(WorkerThreadPool::create(
                cwd,
                entrypoint,
                env,
                assets_for_source_mapping,
                assets_root,
                project_dir,
                concurrency,
                debug,
            )
            .await)
        })
    }

    fn scale_down(&self) -> Result<()> {
        WorkerThreadPool::scale_down();
        Ok(())
    }

    fn scale_zero(&self) -> Result<()> {
        WorkerThreadPool::scale_zero();
        Ok(())
    }
}

impl WorkerThreadPool {
    pub fn scale_down() {
        let _ = WORKER_POOL_OPERATION.scale_down();
    }

    pub fn scale_zero() {
        let _ = WORKER_POOL_OPERATION.scale_zero();
    }
}

#[async_trait::async_trait]
impl EvaluateOperation for WorkerThreadPool {
    async fn operation(&self) -> Result<Box<dyn Operation>> {
        let operation = {
            let _guard = duration_span!("Node.js operation");
            let worker_options = self.worker_options.clone();

            let task_id = OPERATION_TASK_ID.fetch_add(1, Ordering::Release);

            if task_id == 0 {
                panic!("Node.js operation task id overflow")
            }

            let (worker_id, permits) = self.acquire_worker().await?;

            let state = self.state.clone();

            // Pre-allocate channels for this task to avoid HashMap lookups during communication
            let channels = TaskChannels::new(task_id, worker_id);

            WorkerOperation {
                worker_options,
                worker_id,
                state: state.clone(),
                on_drop: Some(Box::new(move |worker_id| {
                    let mut waiters = state.waiters.lock();
                    loop {
                        if let Some(tx) = waiters.pop() {
                            if tx.send(worker_id).is_ok() {
                                break;
                            }
                        } else {
                            state.idle_workers.lock().push(worker_id);
                            break;
                        }
                    }
                })),
                _permits: permits,
                channels,
            }
        };

        Ok(Box::new(operation))
    }

    /// Returns a snapshot of the pool's internal statistics.
    fn stats(&self) -> PoolStatsSnapshot {
        self.state.stats.lock().snapshot()
    }

    fn pre_warm(&self) {
        // TODO: This is a no-op for worker_pool right now, only process_pool implements it
    }
}