Subj : Re: IOCP problem
To   : comp.programming.threads
From : SenderX
Date : Tue Jan 25 2005 04:19 am

> I've setup IOCP for a named pipe server that I have.
>
> I create the pipes and then create the completion port and assign it to 
> each
> pipe in turn.  Then I create a couple of worker threads and each worker 
> sits
> on GetQueuedCompletionStatus.

Ok. This sounds good so far.




> I then started stress testing the server and I've found that I get strange
> hangs if I use more than one worker thread.

Does the problem seem to only affect you if you're using more than one 
thread, while your stress testing?

If so, it's a probably a straight forward synchronization problem ( most 
likely ) that could be caused by a number of different things. One reason I 
say this, is because you say your "stress testing" with multi-threads. 
Cranking up an artificially heavy-load can and will expose most of your 
applications race-conditions ( if any ); Memory leaks, IOCP reordering 
side-effect, event-driven deadlock issues, and a host of other possible 
problems that could be directly related to your specific issue(s).




> I did some reading in this
> newsgroup yesterday and somebody mentioned that with IOCP it is possible 
> for
> one than one thread to wake up to handle the event

No.

One IOCP post == One Completion obtained through GQCS().

Any operation that is documented to generate an IOCP completion ( 
WSASend/Recv, PQCS(), ect. ) will generate only one completion event. This 
event could come in the form of GQCS() returning with a successful transfer 
of bytes, an error, a custom post, ect.. Could you post the relevant code 
that extracts and dispatches all of the events/errors returned from GQCS? 
Basically, a skeleton of your worker thread loop would help.




> so I handled this by
> using a "running" flag for the threads and let one thread actually handle
> the event, with the others doing nothing, but sending a completion packet
> back.

Ok. Yeah. Post the code that implements this, and any other code that links 
this logic to the dequeueing and dispatching of an event from GQCS().




> I though this was working, but it still seems to act strangely when
> more than one worker thread is created.

It sounds a bit strange. ;)




> I think I must be doing something wrong somewhere, but I'm not sure where.
> I thought I had it licked when I read about the problem that somebody else
> had and it was exactly the same.

Could you post a link to the information you read?


P.S.

Here are some relevant links that you might find useful:

http://www.jetbyte.com/portfolio-showarticle.asp?articleId=44&catId=1&subcatId=2

http://www.sysinternals.com/ntw2k/info/comport.shtml

http://support.microsoft.com/default.aspx?scid=kb;en-us;192800

http://msdn.microsoft.com/msdnmag/issues/1000/winsock/default.aspx

http://msdn.microsoft.com/msdnmag/issues/01/08/concur/default.aspx

.