CooperativeExecutor.runUntil sleeps until the next timer while runnable jobs are queued

Comece em stdlib/public/Concurrency/CooperativeExecutor.swift, na região de CooperativeExecutor.runUntil(_:) nas linhas 286-331, e depois execute a reprodução WASI fornecida no runtime afetado. A correção estará completa quando jobs executáveis forem executados antes que o executor aguarde um timer e ambos os cenários terminarem em milissegundos com o watchdog cancelado; adicione um teste de regressão se a estrutura de testes ao redor oferecer suporte.

Description

On wasm32-unknown-wasip1, a task that becomes runnable while another task is inside Task.sleep does not run until that sleep's deadline. This applies to a task that called Task.yield() and to a task whose awaited child has finished.

CooperativeExecutor.runUntil(_:) takes the ready jobs with runQueue.take(), runs them, and then sleeps for the time to the next timer. Jobs enqueued while those ran are in runQueue at that point; runQueue.isEmpty is checked only when no timer is pending.

https://github.com/swiftlang/swift/blob/swift-6.4.0-RELEASE/stdlib/public/Concurrency/CooperativeExecutor.swift#L286-L331

Reproduction

// Package.swift: swift-tools-version 6.3, one executableTarget named repro.
// Sources/repro/main.swift
let sleepSeconds: UInt64 = 5
let clock = ContinuousClock()
let start = clock.now
func stamp(_ what: String) { print("[\(clock.now - start)] \(what)") }

func scenario(_ name: String, yields: Int) async {
    stamp("\(name): start")
    let watchdog = Task {
        try? await Task.sleep(nanoseconds: sleepSeconds * 1_000_000_000)
        stamp("\(name): watchdog woke, cancelled=\(Task.isCancelled)")
    }
    let worker = Task {
        for _ in 0 ..< yields { await Task.yield() }
        stamp("\(name): worker done")
    }
    await worker.value
    watchdog.cancel()
    stamp("\(name): end")
}

await scenario("no-yield", yields: 0)
await scenario("yield", yields: 50)
stamp("all done")

$ swift sdk install https://download.swift.org/swift-6.4.0-release/wasm-sdk/swift-6.4.0-RELEASE/swift-6.4.0-RELEASE_wasm.artifactbundle.tar.gz --checksum f07b7be3c586d92d7a07051fc6d303b87ebea67eadc40640ba59d5a8b79aa86d
$ swift build -c release --swift-sdk swift-6.4.0-RELEASE_wasm
$ wasmkit run .build/out/Products/Release-webassembly-wasm32/repro.wasm
[0.003 seconds] no-yield: start
[0.006 seconds] no-yield: worker done
[5.023 seconds] no-yield: end
[5.024 seconds] yield: start
[5.026 seconds] no-yield: watchdog woke, cancelled=true
[10.036 seconds] yield: watchdog woke, cancelled=false
[10.039 seconds] yield: worker done
[10.039 seconds] yield: end

With no yields the worker is done at 6 ms, yet the task awaiting it resumes at 5 s. With yields the worker does not advance at all until the watchdog's sleep has run its full length, by which time nobody has cancelled it.

Expected behavior

Jobs that became runnable during a pass run before the executor sleeps; it should sleep only when the run queue is empty. Each scenario above should end within milliseconds, with the watchdog cancelled before it wakes.

A possible fix is to guard the sleep: if let toWait, runQueue.isEmpty { _sleep(…) } (or continue when runQueue is not empty). Untested.

Environment

Swift version 6.4 (swift-6.4-RELEASE), Ubuntu 24.04 aarch64 host, Swift SDK swift-6.4.0-RELEASE_wasm (wasm32-unknown-wasip1), run with the WasmKit 0.3.1 the toolchain ships.

Also reproduced with a binary built by Swift 6.3.3 and its Wasm SDK. The same code is present in release/6.3, release/6.4.x, release/6.4.1, swift-6.4.0-RELEASE, and main (as of 2026-09-19).

Additional information

The Swift version makes no difference; whether the WASI runtime's poll_oneoff actually sleeps does. Time for a four-scenario variant of the program (5 s sleep each) to finish, same two binaries on three runtimes:

Impact: any program on these targets that keeps a timer pending while other tasks make progress — a timeout or watchdog around a body is the common case — runs at the pace of its timers. It surfaced as a test suite whose every deadline-bounded test took exactly its deadline (300 s) under the 6.4.0 toolchain while passing in seconds under 6.3.3: https://github.com/sinoru/swift-synchronization-kit/issues/4

#90096 (enabling Task.sleep in Embedded Swift) removes the #if !$Embedded guards around the same timer queues, so the same stall would presumably appear there once it lands.