Fix O(N^2) worker teardown with pending Atomics.waitAsync waiters

src/jsc/bindings/JSCTaskScheduler.cpp

@@ -35,17 +35,18 @@ class JSCDeferredWorkTask {
 
 // Drop `ticket` from whichever pending set holds it. Caller holds m_lock; the
 // event-loop ref is balanced after the caller releases the lock.
+//
+// Must stay O(1): ~VM -> WaiterListManager::unregister reaches this once per
+// still-pending Atomics.waitAsync ticket while holding the process-global
+// waiter-lists lock, so a per-call scan makes a worker exiting with N waiters
+// O(N^2) and stalls Atomics.notify/waitAsync on every other thread for the
+// whole window.
 static bool dropPendingTicketLocked(Bun::JSCTaskScheduler& scheduler, Ticket* ticket) WTF_REQUIRES_LOCK(scheduler.m_lock)
 {
-    bool isKeepingEventLoopAlive = scheduler.m_pendingTicketsKeepingEventLoopAlive.removeIf([ticket](auto pendingTicket) {
-        return pendingTicket.ptr() == ticket;
-    });
+    bool isKeepingEventLoopAlive = scheduler.m_pendingTicketsKeepingEventLoopAlive.remove(ticket);
     // -- At this point, ticket may be an invalid pointer.
-    if (!isKeepingEventLoopAlive) {
-        scheduler.m_pendingTicketsOther.removeIf([ticket](auto pendingTicket) {
-            return pendingTicket.ptr() == ticket;
-        });
-    }
+    if (!isKeepingEventLoopAlive)
+        scheduler.m_pendingTicketsOther.remove(ticket);
     return isKeepingEventLoopAlive;
 }
 

test/js/web/atomics.test.ts

@@ -1,4 +1,5 @@
 import { describe, expect, test } from "bun:test";
+import { bunEnv, bunExe, tempDir } from "harness";
 
 describe("Atomics", () => {
   describe("basic operations", () => {
@@ -307,3 +308,57 @@
     });
   });
 });
+
+describe("Atomics.waitAsync worker teardown", () => {
+  // A worker exiting with N pending Atomics.waitAsync waiters used to tear down in
+  // O(N^2): ~VM cancels each waiter's ticket, and each cancellation scanned the whole
+  // pending-ticket set. The scan also ran while JSC held the process-global
+  // waiter-lists lock, stalling Atomics.notify/waitAsync on every other thread.
+  // With N = 50000 the quadratic teardown took 38s+ in release; linear teardown is
+  // well under a second. The 15s bound leaves wide margin for slow debug/ASAN CI.
+  test("worker with many pending waiters exits quickly", async () => {
+    using dir = tempDir("atomics-waiter-teardown", {
+      "waiter-teardown-fixture.mjs": `
+        import { Worker, isMainThread, parentPort, workerData } from "node:worker_threads";
+        if (!isMainThread) {
+          const i32 = new Int32Array(workerData.sab);
+          for (let k = 0; k < 50000; k++) Atomics.waitAsync(i32, k % 1024, 0);
+          parentPort.postMessage("armed");
+          setInterval(() => {}, 1e6);
+        } else {
+          const sab = new SharedArrayBuffer(4096);
+          const w = new Worker(new URL(import.meta.url), { workerData: { sab } });
+          w.on("message", () => {
+            const start = Date.now();
+            w.terminate();
+            w.on("exit", () => {
+              console.log(JSON.stringify({ exitAfterMs: Date.now() - start }));
+              process.exit(0);
+            });
+          });
+        }
+      `,
+    });
+    await using proc = Bun.spawn({
+      cmd: [bunExe(), "waiter-teardown-fixture.mjs"],
+      env: bunEnv,
+      cwd: String(dir),
+      stdout: "pipe",
+    });
+
+    // Fail fast instead of waiting out a quadratic teardown (minutes in debug).
+    let timer: Timer | undefined;
+    const deadline = new Promise<"deadline">(resolve => {
+      timer = setTimeout(() => resolve("deadline"), 60_000);
+    });
+    const first = await Promise.race([proc.exited, deadline]);
+    clearTimeout(timer);
+    if (first === "deadline") proc.kill("SIGKILL");
+    expect(first).not.toBe("deadline");
+
+    const stdout = await proc.stdout.text();
+    const { exitAfterMs } = JSON.parse(stdout);
+    expect(exitAfterMs).toBeLessThan(15_000);
+    expect(await proc.exited).toBe(0);
+  }, 90_000);
+});