FIX: Sockets Application Hangs or Exhibits Poor Performance

The information in this article applies to:

• Microsoft Visual C++, 32-bit Edition, versions 4.0, 4.1

SYMPTOMS

A Microsoft Foundation Classes (MFC) sockets application seems to hang until a message is generated for the application (for example, by moving the mouse over a window), or an application appears to take more than two seconds on single calls to Receive or Send.

In other cases the socket operations might execute more quickly if messages are being continuously generated (for example, by moving the mouse quickly over the application's window).

CAUSE

There is code in Sockcore.cpp that can prevent proper processing of socket notification messages if the messages arrive at a very specific time interval. The “More Information” section below describes this problem in further detail.

RESOLUTION

In your CSocket derived class, provide an override of the virtual function CSocket::OnMessagePending as follows:

#if _MFC_VER == 0x0400 || _MFC_VER == 0x0410

#define WM_SOCKET_NOTIFY  0x0373
#define WM_SOCKET_DEAD    0x0374
#define WM_KICKIDLE       0x036A

BOOL CMySocket::OnMessagePending() { #if _MFC_VER >= 0x0420

  return CSocket::OnMessagePending();

  if(CSocket::OnMessagePending())
    return TRUE;


  MSG msg;
  // Don't flood the queue with WM_KICKIDLE messages.
  while(::PeekMessage(&msg,NULL,WM_KICKIDLE,WM_KICKIDLE,PM_REMOVE));

  if(::PeekMessage( &msg,
                   NULL,
                   WM_SOCKET_NOTIFY,
                   WM_SOCKET_DEAD,
                   PM_NOREMOVE))
  {
    // Post another message so that we don't block on the
    // WaitMessage call in CSocket::OnMessagePending.
    ::PostMessage(NULL,WM_KICKIDLE,0,0);
  }
  return FALSE;

STATUS

Microsoft has confirmed this to be a bug in the Microsoft products listed at the beginning of this article. This bug was corrected in Microsoft Visual C++, version 4.2.

MORE INFORMATION

The following code can be found in CSocket::PumpMessages in Sockcore.cpp:

  ...

  if (::PeekMessage(&msg, NULL, 0, 0, PM_NOREMOVE) &&
      OnMessagePending())
  {
    // allow user-interface updates
    pThread->OnIdle(-1);
  }
  else
  {
    // no work to do -- allow CPU to sleep
    WaitMessage();
  }

Application Hangs Until Mouse Moves

The behavior of WaitMessage is such that if a message in the queue has been peeked at, then WaitMessage will not return until a new message arrives.

So if the socket notification we are looking for arrives just before the call to PeekMessage, then the PeekMessage call looks at it and WaitMessage won't return until a new message arrives. So, a new message must be generated before WaitMessage breaks out.

Socket Operations Are Very Slow

MFC has a built-in protection against cases where notifications are not arriving; it provides occasional time-outs so that the socket operation is attempted again even though the notification never arrived. The default time-out period is two seconds.

If the notification message was peeked at (as described above), then WaitMessage won't return even though the operation is ready to be completed. The application won't continue for at least two seconds. In some cases, the WM_TIMER message is also inadvertently peeked at, in which case the application exhibits the behavior shown above (hangs until a new message, such as a WM_MOUSEMOVE arrives).

In some situations, even though the application does not hang, it could take two seconds to perform the operation even though the notification message for the socket action has already arrived. This is what can cause the poor performance.

IMPORTANT NOTE: The CSocket class was written with the assumption that the socket has notifications enabled [that is, AsyncSelect(0) has NOT been called]. MFC provides protection against some cases where the notification posting has been disabled by using a time-out mechanism, which causes the operation to be re-tried. However, the workaround shown above does not take into account timer messages primarily because it can cause a CPU burden if another timer is active in the application. Thus, MFC's mechanism for protecting CSocket objects that have disabled their notifications won't necessarily be fixed when this workaround is implemented.

So you should ensure that notifications are always enabled for your CSocket objects. One of the reasons some developers disable notifications is to prevent multiple FD_READ notifications. If that is the case, then perhaps a better workaround might be to use IOCtl to determine whether there is data available on the socket.

For additional information on this approach, please see the following article in the Microsoft Knowledge Base:

   ARTICLE-ID: Q138694
   TITLE     : FIX: Using CSocket w/ CArchive Causes Block from Extra
               FD_READ