[chenBright]

What problem does this PR solve?

Issue Number: resolve

Problem Summary:

On macOS, repeated futex_wake_private() calls can count an already notified waiter
again before it resumes. Callers such as TaskControl::signal_task() may consequently
stop notifying other parking lots prematurely, leaving idle workers asleep while tasks
remain queued

What is changed and the side effects?

Changed:

• Replace aggregate waiter counting with an intrusive queue of unnotified waiters, each
  owning a condition variable and notification flag.
• Remove each waiter from the queue when notified, preventing subsequent wake calls
  from counting it again.
• Remove timed-out waiters individually and retry spurious wakeups without extending
  the original deadline.
• Honor notifications that win the mutex race with a timeout.

Side effects:

• Performance effects:

• Breaking backward compatibility:

---

Check List:

• Please make sure your changes are compilable.
• When providing us with a new feature, it is best to add related tests.
• Please follow Contributor Covenant Code of Conduct.

[Copilot]

Warning

Copilot couldn't run its full agentic review because it didn't start before the timeout. Make sure your repository has a runner available, or add a copilot-code-review.yml file specifying one with the runs-on attribute. See the docs for more details.

Copilot review overview

5 open issues

• This test can become very slow/flaky under load: it may spin/yield for up to 5 seconds per… · New
• This test can become very slow/flaky under load: it may spin/yield for up to 5 seconds per… · New
• futex_wait_private sets errno = EAGAIN on mismatched value (per the current implementation),… · New
• The loop bound 256 and deadline offset 5000000L are magic numbers that encode test policy… · New
• The loop bound 256 and deadline offset 5000000L are magic numbers that encode test policy… · New

What changed in this PR

This PR fixes macOS simulated futex wake accounting so repeated futex_wake_private() calls cannot “re-count” a waiter that was already notified but hasn’t resumed yet, preventing premature cessation of wakeups in higher-level schedulers.

Changes:

• Replace aggregate waiter counting with a per-waiter intrusive queue and per-waiter condition variables.
• Ensure notified waiters are removed from the queue at wake time; timed-out waiters remove themselves.
• Add regression tests covering double-count prevention, timeout interactions, and non-queued mismatched waits.

File	Description
test/​bthread_futex_unittest.cpp	Adds regression tests validating one-time wake accounting and timeout/non-queue behaviors.
src/​bthread/​sys_futex.cpp	Reworks simulated futex internals to track unnotified waiters explicitly and avoid double-counting.

_{Review effort: Lite}

[Copilot]

Warning

Copilot couldn't run its full agentic review because it didn't start before the timeout. Make sure your repository has a runner available, or add a copilot-code-review.yml file specifying one with the runs-on attribute. See the docs for more details.

Copilot review overview

7 open issues

• This relies on RemoveFromList() being safe to call even if the node has already been removed in… · New
• This relies on RemoveFromList() being safe to call even if the node has already been removed in… · New
• INT_MAX is used in the new tests but this file doesn’t include a header that guarantees its… · New
• The “rescue” path can force the waiter to return -1/EAGAIN (by storing word=1 before the waiter… · New
• The “rescue” path can force the waiter to return -1/EAGAIN (by storing word=1 before the waiter… · New
• INT_MAX is used in the new tests but this file doesn’t include a header that guarantees its… · New
• INT_MAX is used in the new tests but this file doesn’t include a header that guarantees its… · New

5 resolved since last review

• futex_wait_private sets errno = EAGAIN on mismatched value (per the current implementation),…
• This test can become very slow/flaky under load: it may spin/yield for up to 5 seconds per…
• This test can become very slow/flaky under load: it may spin/yield for up to 5 seconds per…
• The loop bound 256 and deadline offset 5000000L are magic numbers that encode test policy…
• The loop bound 256 and deadline offset 5000000L are magic numbers that encode test policy…

_{Review effort: Lite}

[Copilot]

Warning

Copilot couldn't run its full agentic review because it didn't start before the timeout. Make sure your repository has a runner available, or add a copilot-code-review.yml file specifying one with the runs-on attribute. See the docs for more details.

Copilot review overview

3 open issues

• This change introduces pthread_cond_init/pthread_cond_destroy per wait operation (previously… · New
• The destructor has a non-obvious side effect (mutating the intrusive list) and depends on a… · New
• INT_MAX is used in the new tests but this file doesn’t include <climits>/<limits.h> in the… · New

7 resolved since last review

• INT_MAX is used in the new tests but this file doesn’t include a header that guarantees its…
• INT_MAX is used in the new tests but this file doesn’t include a header that guarantees its…
• The “rescue” path can force the waiter to return -1/EAGAIN (by storing word=1 before the waiter…
• The “rescue” path can force the waiter to return -1/EAGAIN (by storing word=1 before the waiter…
• INT_MAX is used in the new tests but this file doesn’t include a header that guarantees its…
• This relies on RemoveFromList() being safe to call even if the node has already been removed in…
• This relies on RemoveFromList() being safe to call even if the node has already been removed in…

_{Review effort: Lite}

[chenBright]

@wwbmmm PTAL.

The remaining Copilot review comments have been evaluated and are either false positives or non-blocking suggestions that we will not adopt in this bugfix. No further code changes are needed. All 65 relevant tests passed, and this PR is ready for maintainer review