Poll event queue treats EINTR as fatal, tearing down the socket pool

[ianegordon]

Summary

a196c2e fixed SocketPool<ePoll> to treat epoll_wait returning -1/EINTR as "no events" instead of a fatal error. The Poll backend still has the pre-fix behavior, and fixing it turns out to be non-trivial on Windows — hence an issue rather than a PR, to agree on direction first. (The kQueue backend has the same gap; a separate PR for that is in preparation.)

The POSIX bug

Poll.getNotifications() (FlyingSocks/Sources/SocketPool+Poll.swift):

let status = Socket.poll(&buffer, UInt32(buffer.count), interval.milliseconds)
guard status > -1 else {
    throw SocketError.makeFailed("poll.getNotifications poll")
}

poll(2) (Apple manpage, Linux equivalent): "[EINTR] A signal is delivered before the time limit expires and before any of the selected events occurs." No events are lost — the correct response is to poll again. Instead the throw unwinds SocketPool.run(), whose defer { cancelAll() } cancels every waiting continuation: a single signal delivery (e.g. a debugger pause/resume) tears down the whole pool for any embedder using .poll.

Why the obvious fix is wrong on Windows

Mirroring a196c2e (if status == -1 && errno == EINTR { return [] }) misbehaves under WinSDK:

1. errno is shadowed. Socket+WinSock2.swift defines var errno: Int32 { WSAGetLastError() }, so the comparison would see Winsock codes (10000+), not CRT errno values.
2. EINTR has no Windows alias. The WinSock2 shim deliberately aliases EWOULDBLOCK/EINPROGRESS/EISCONN to their WSAE* equivalents but defines no EINTR, so a bare errno == EINTR may not even compile on Windows.
3. WSAEINTR means something different. Per Microsoft's Windows Sockets Error Codes, WSAEINTR (10004) is "Interrupted function call — a blocking operation was interrupted by a call to WSACancelBlockingCall" — an explicit cancellation, not a signal interruption — and the WSAPoll documentation does not list it as an expected result. Silently retrying it would be a behavior change, not a bug fix.

Proposed direction

Add the EINTR-retry branch under #if !canImport(WinSDK), matching a196c2e on POSIX platforms and leaving Windows behavior exactly as it is today:

guard status > -1 else {
    #if !canImport(WinSDK)
    // EINTR (signal) is not a failure: report no events so the caller
    // polls again rather than tearing down the server. See poll(2).
    if status == -1 && errno == EINTR {
        return []
    }
    #endif
    throw SocketError.makeFailed("poll.getNotifications poll")
}

Whether WSAEINTR should also be swallowed on Windows is left as a separate decision. Happy to send a PR for the POSIX-only version if this direction looks right.

🤖 Generated with Claude Code