RE: BUG in Winsock on P4 HT CPU
From: Glambert (Glambert_at_discussions.microsoft.com)
Date: 09/07/04
- Next message: Glambert: "Re: BUG in Winsock on P4 HT CPU"
- Previous message: Arkady Frenkel: "Re: how to transfer file between 2 computer"
- Messages sorted by: [ date ] [ thread ]
Date: Tue, 7 Sep 2004 08:05:03 -0700
If you read the help on WSAAsyncSelect it states
-------------------------------------------------------------
For FD_READ, FD_OOB, and FD_ACCEPT events, message posting is
level-triggered. This means that if the reenabling routine is called and the
relevant condition is still met after the call, a WSAAsyncSelect message is
posted to the application. This allows an application to be event-driven and
not be concerned with the amount of data that arrives at any one time.
Consider the following sequence:
Network transport stack receives 100 bytes of data on socket s and causes
Windows Sockets 2 to post an FD_READ message.
The application issues recv( s, buffptr, 50, 0) to read 50 bytes.
Another FD_READ message is posted since there is still data to be read.
With these semantics, an application need not read all available data in
response to an FD_READ message—a single recv in response to each FD_READ
message is appropriate. If an application issues multiple recv calls in
response to a single FD_READ, it can receive multiple FD_READ messages. Such
an application can need to disable FD_READ messages before starting the recv
calls by calling WSAAsyncSelect with the FD_READ event not set.
---------------------------------------------------------------
Based on the above, if you want to call recv multiple times within
OnReceive(), you should first disable FD_READ (via AsyncSelect()), issue all
of your recv (Receive()), and then reenable FD_READ (via
AsyncSelect(FD_READ)). This will work correctly. However, for CAsychSocket
derived classes you need to keep track of the events you have enabled (I wish
there was a way to query winsock for the events you are currently registered
for).
The tough part is I did not see any explicit documentation that states "You
cannot call Receive() multiple times from with OnReceive()". I read the
WSAAsyncSelect dcoumentation to eventually figure it out (and then I read it
about 10 times before realising what had to be done). I am not an MFC expert
by any means so perhaps the OnReceive() documentation is out there
somewhere...but it would have been nice to see it explicitly stated.
"Greg Ennis" wrote:
> I have confirmed a bug in Winsock API when running on fast Hyperthread P4
> CPU's. If you look in the newsgroups for posts by Glambert and myself you
> will see other references to it, but here is the basic facts of the matter.
>
> Winsock has a problem with intermittently not posting the socket
> notification message when data is received on a socket. This only happens on
> fast HT enabled P4 CPUs.
>
> Here is how you can reproduce it. Write a simple application that creates a
> loopback CAsyncSocket socket and connects to itself. In each socket's
> OnReceive handler, read 2 bytes and send 2 bytes back over the socket. When
> the connection is established, send 2 bytes from the 'client' socket to kick
> it off.
>
> This app should run forever, just sending bits back and forth to itself. And
> indeed it does... on CPU's that do not support HT, or CPUs with HT disabled.
> Run it on a HT CPU and guess what? After only exchanging a few packets, it
> stops. It stops dead because one of the sockets did NOT receive the socket
> notification message on its window handle. I have debugged it thoroughly and
> I have set it up using 2 computers over ethernet, one of which is a P4 HT,
> and watched using TCP packet sniffer. The P4 HT CPU is ALWAYS the guilty
> party, it is the one that stops sending!
>
> Here is further proof of the bug, as GLambert has discovered. If you call
> AsyncSelect() after reading all the available data, it works fine. The socket
> notification message appears to always be posted correctly. However, I can't
> guarantee this on all platforms, since it is a workaround at best.
>
> Waiting on an answer from MS from this.
>
- Next message: Glambert: "Re: BUG in Winsock on P4 HT CPU"
- Previous message: Arkady Frenkel: "Re: how to transfer file between 2 computer"
- Messages sorted by: [ date ] [ thread ]
Relevant Pages
|
