PollSet::poll on Windows always returns immediately after first connection, due to bad event socket implementation

[obiltschnig]

PollSet uses and event fd to interrupt an ongoing poll() in wakeUp(). On POSIX platforms this uses eventfd() and works fine. On Windows, a ServerSocket is created (either as Unix domain socket or TCP socket) instead. The problem is that when that event fd is signaled, the implementation attempts to receive from it. Which is not going to work as it's a ServerSocket. So this eventfd stays signaled forever, causing high CPU load as poll() always returns immediately.

[lanthora]

The commit that fixes this problem should also solve the problem that the
Poco::Net::ServiceNotFoundException exception is thrown when creating a PollSet with a Chinese username in Windows.

// TemporaryFile path contains Chinese characters
_pSockFile.reset(new TemporaryFile);
// SocketAddress throw ServiceNotFoundException
_pSocket.reset(new ServerSocket(SocketAddress(_pSockFile->path())));

[micheleselea]

I think @obiltschnig that the best way to deal with this behavior in windows, is not to create a tcp or udp listening on port, but is to use

void wakeUp()
	{		
#ifdef WEPOLL_H_
		if (_port > 0) {
			StreamSocket ss(SocketAddress("127.0.0.1", _port));
		}
		else if (_epollfd > 0) {
			PostQueuedCompletionStatus(_epollfd, 0, 0, NULL);
		}
#else
		uint64_t val = 1;
		// This is guaranteed to write into a valid fd,
		// or 0 (meaning PollSet is being destroyed).
		// Errors are ignored.
		write(_eventfd, &val, sizeof(val));
#endif
}

Than we should add to the SocketImpl::poll something like

#ifdef WEPOLL_H_
		setWEPoll(epollfd);
#endif
		rc = epoll_wait(epollfd, &evout, 1, static_cast<int>(remainingTime.totalMilliseconds()));
#ifdef WEPOLL_H_
		setWEPoll(0);
#endif

to let the ::close function do something like

void SocketImpl::close()
{
	if (_sockfd != POCO_INVALID_SOCKET)
	{
#ifdef WEPOLL_H_
		HANDLE epollfd = getWEPoll();
		if (epollfd > 0) {
			PostQueuedCompletionStatus(epollfd, 0, 0, NULL);
		}
#endif
		poco_closesocket(_sockfd);
		_sockfd = POCO_INVALID_SOCKET;
	}
}

[micheleselea]

in wepoll.c I found that we miss one check in function port__feed_events that cause me some problem
The check before doing the line
epoll_event_count += sock_feed_event(port_state, io_status_block, ev);

static inline int port__feed_events(port_state_t* port_state,
                                    struct epoll_event* epoll_events,
                                    OVERLAPPED_ENTRY* iocp_events,
                                    DWORD iocp_event_count) {
  int epoll_event_count = 0;
  DWORD i;

  for (i = 0; i < iocp_event_count; i++) {
    IO_STATUS_BLOCK* io_status_block =
        (IO_STATUS_BLOCK*) iocp_events[i].lpOverlapped;
    struct epoll_event* ev = &epoll_events[epoll_event_count];
    
    if (io_status_block)
        epoll_event_count += sock_feed_event(port_state, io_status_block, ev);
    else {
        ev->events = 0;
        ev->data.u64 = 0;
        epoll_event_count += 1;
    }
  }

  return epoll_event_count;
}