Fix CI: Windows errno clobbering, event loop timeout, socket test race, NIO connect verification

[Copilot]

CI run 23856315454 fails across Windows socket tests and emulated Linux targets. Four distinct root causes identified.

Windows impl_facade! errno clobbering

The facade macro logs via tracing::info!() after the NIO layer returns, which triggers CreateFileW through tracing-appender. This clobbers GetLastError(), so callers see stale errors (e.g., error 6 "The handle is invalid" from a CreateFileW failure) instead of the actual socket error.

Fix: capture GetLastError() immediately after the inner syscall, use the saved value for logging, restore before returning.

let r = self.inner.$syscall(fn_ptr, $($arg, )*);
let saved_errno = std::io::Error::last_os_error();
// ... logging and coroutine bookkeeping (clobbers GetLastError) ...
$crate::info!("exit syscall {} {:?} {}", syscall, r, saved_errno);
if let Some(e) = saved_errno.raw_os_error() {
    $crate::syscall::set_errno(u32::try_from(e).unwrap_or_default());
}
r

Event loop timeout ("did not completed within the specified time")

PR #432 renamed OrderedLocalQueue::len() → local_len() and added a new len() that iterates all queues (local + shared + siblings). is_empty() now means "entire work-steal system is empty" rather than "this pool's local queue is empty." The event loop exit condition !consumer.is_empty() and do_stop_sync() check block until all sibling queues drain — causing timeouts on Windows and slower emulated targets.

Fix: add a new CoroutinePool::is_local_empty() method that delegates to task_queue.is_local_empty(), and use it in the two event loop exit conditions (event_loop.rs lines 394 and 447). CoroutinePool::is_empty(), try_grow(), and the Drop impl remain unchanged with global semantics.

Socket test race condition and port conflicts

All four socket test examples (socket_not_co, socket_co, socket_co_server, socket_co_client) and the operator tests (core/src/net/operator/linux/tests.rs, core/src/net/operator/windows/tests.rs) had two issues:

1. Race condition: Server and client threads spawned simultaneously without ensuring the server is listening before the client connects. On Windows, the first syscall through the tracing facade takes ~1s (tracing-appender's CreateFileW initialization), making this race nearly 100% reproducible.

2. Port conflicts: Hardcoded ports (7060, 7061, 8888, 8889, 8899, 8999) can collide with other services on CI runners. The NIO connect handler may report a false success (connect returns 0) against a foreign listener, but the subsequent send fails with WSAENOTCONN (10057) because the foreign process resets the connection. This also interacts with EventLoops::wait_write_event() returning on any event from the shared selector (not specifically the write event for the connecting fd), which can produce false positives from getsockopt(SO_ERROR) during a pending connect.

Fix: use dynamic port allocation (TcpListener::bind("127.0.0.1:0")) and pass the OS-assigned SocketAddr from server to client through a Mutex<Option<SocketAddr>> condvar barrier. The server signals after TcpListener::bind() returns with the actual address, and the client waits for this signal before attempting to connect. This eliminates both the race condition and port conflict issues. Applied to all socket examples and both Linux and Windows operator tests.

NIO connect false success prevention

NioConnectSyscall on both Unix and Windows could report a false connect success. When EventLoops::wait_write_event() returns due to an unrelated event on the shared selector while the connect is still pending, getsockopt(SO_ERROR) reads 0 (no error yet) and getpeername() may transiently succeed, causing the handler to report success. The subsequent send then fails with ENOTCONN/WSAENOTCONN (10057).

Fix: swap the order of getsockopt(SO_ERROR) and getpeername() in the NIO connect handler on both Unix and Windows. getpeername() is now checked first as a quick connectivity test, and getsockopt(SO_ERROR) is checked last as the final authority. If getpeername() fails but SO_ERROR is 0, the connect is still pending — errno is set to EINPROGRESS/WSAEINPROGRESS and the wait loop continues. This ensures SO_ERROR always has the last word on connection status.

[codecov]

Codecov Report

❌ Patch coverage is 88.23529% with 2 lines in your changes missing coverage. Please review.
✅ Project coverage is 75.74%. Comparing base (a60d251) to head (e52143b).
⚠️ Report is 11 commits behind head on master.

Files with missing lines	Patch %	Lines
core/src/syscall/unix/connect.rs	77.77%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #434      +/-   ##
==========================================
+ Coverage   75.32%   75.74%   +0.41%     
==========================================
  Files          44       44              
  Lines        4698     4709      +11     
==========================================
+ Hits         3539     3567      +28     
+ Misses       1159     1142      -17     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[loongs-zhang]

Focus on did not completed within the specified time and fix them.

[Copilot]

Focus on did not completed within the specified time and fix them.

Fixed in 5722944. Added CoroutinePool::is_local_empty() and use it in the event loop thread exit condition (line 394) and do_stop_sync() (line 447). After PR #432, is_empty() checks all queues globally — the event loop was blocking until sibling queues drained, causing the 60s timeout. Now each event loop exits when its own local work is done. CoroutinePool::is_empty(), try_grow(), and Drop remain unchanged with global semantics.

[loongs-zhang]

Do not modify the code I have changed, continue to investigate and fix the CI errors.

[Copilot]

Do not modify the code I have changed, continue to investigate and fix the CI errors.

Understood — co_pool/mod.rs and ordered_work_steal.rs are untouched. Investigated the CI timeouts: the root cause is a race condition in the socket test examples — the client thread can connect_timeout() before the server's TcpListener::bind() + listen() completes (on Windows the first syscall through the tracing facade takes ~1s). Added a server_started condvar barrier to all four socket examples so the client waits until the server is listening. Fix in 0013de0.

[loongs-zhang]

I think your change is effective, but strangely, according to https://github.com/acl-dev/open-coroutine/actions/runs/23886892084/job/69651487944, we see Thu, 02 Apr 2026 14:38:35 +0800 INFO ThreadId(04) open_coroutine_core::syscall::windows::connect: 22: exit syscall connect 0 The operation completed successfully. (os error 0), then why Thu, 02 Apr 2026 14:38:35 +0800 INFO ThreadId(04) open_coroutine_core::syscall::windows::send: 20: exit syscall send -1 A request to send or receive data was disallowed because the socket is not connected and (when sending on a datagram socket using a sendto call) no address was supplied. (os error 10057)?

[Copilot]

I think your change is effective, but strangely, according to https://github.com/acl-dev/open-coroutine/actions/runs/23886892084/job/69651487944, we see `Thu, 02 Apr 2026 14:38:35 +0800 INFO Thread...

Analyzed the CI log from that run. The NIO connect handler (connect.rs) reports success (r=0) based on getsockopt(SO_ERROR) returning 0 and getpeername() returning 0. But the subsequent raw send() fails with WSAENOTCONN (10057), meaning the socket was never truly connected.

Two issues contribute:

1. Port conflict on CI runner — the hardcoded port 8888 may collide with another service on the Windows runner. The client's connect could succeed against a foreign listener, then fail on send because the foreign process resets the connection or doesn't accept it properly. The WSAENOTCONN (not WSAECONNRESET) suggests the TCP handshake appeared complete to getpeername but the socket state was invalidated before send.

2. EventLoops::wait_write_event() in the NIO connect handler returns when any event fires on the shared selector (not specifically the write event for the connect fd). If the selector returns due to an unrelated event while the connect is still pending, getsockopt(SO_ERROR) reads 0 (no error yet) and getpeername() may transiently succeed on Windows, giving a false positive.

Fixed in 1853064: all four socket examples now bind to port 0 (OS-assigned dynamic port) and pass the actual SocketAddr from server to client through the condvar barrier. This eliminates port conflicts entirely — the client always connects to the exact port the server is listening on.

[loongs-zhang]

Can NioConnectSyscall add code to ensure success and avoid fake success? Both check them on unix and windows.

[Copilot]

Can NioConnectSyscall add code to ensure success and avoid fake success? Both check them on unix and windows.

Added in 1e2a7f5. After getpeername() returns 0, NioConnectSyscall now rechecks getsockopt(SO_ERROR) on both Unix and Windows. If SO_ERROR is non-zero, the connection failed despite getpeername() appearing to succeed — errno is set to the actual error and r is set to -1, which retries the wait loop. This catches the false positive where wait_write_event() returns due to an unrelated selector event while the connect is still pending.

[loongs-zhang]

How about simply swapping the order of getsockopt and getpeername in the original code? Both on unix and windows.

[Copilot]

How about simply swapping the order of getsockopt and getpeername in the original code? Both on unix and windows.

Done in ea2ce66. Swapped the order on both Unix and Windows — getpeername() first, getsockopt(SO_ERROR) last as the final authority. If getpeername fails but SO_ERROR is 0, the connect is still pending so errno is set to EINPROGRESS/WSAEINPROGRESS and the wait loop continues. Removed the triple-check verification block.

[loongs-zhang]

Good, now I need you to remove all let port = 7060; and let port = 7061; on mod operator, just like your modifications to socket_co.

[Copilot]

Good, now I need you to remove all let port = 7060; and let port = 7061; on mod operator, just like your modifications to socket_co.

Done in e52143b. Removed all hardcoded let port = 7060; and let port = 7061; from both core/src/net/operator/linux/tests.rs and core/src/net/operator/windows/tests.rs. Now using TcpListener::bind("127.0.0.1:0") with Arc<(Mutex<Option<SocketAddr>>, Condvar)> barrier, same pattern as the socket examples.