Daniel Stenberg wrote:
> On Mon, 10 Mar 2008, Raf Nulens wrote:
>
>> "curl_socket_callback: fd=2792, index=0, event=2"
>>
>> ** we set the write flag for socket 2792
>>
>> "curl_multi_timer_callback: timeout=14798" // set the timeout
>> "curl_multi_socket: fd=2796, code=0 CURLM_OK, descr=no error, count=1"
>> "curl_multi_socket: done, code = 0" // end perform multi socket (2)
>>
>> now we dont get a write event, but instead we get a timeout after 15
>> seconds. The connection itself has been set up, but already timed out
>> after those 15 seconds wait.
>
> But what kind of event layer is this? libcurl said it wants to write on
> this socket so it wants to be called when the socket is writable. How
> come your event layer doesn't tell you the socket is writable if it is?

We're using winsock2. Problem is that we seem to set our read/write
event flags too late.

>
>> When we get a socket_callback is it possible that the socket in
>> question is already connected and write enabled so that we dont get a
>> write event?
>
> Yes of course. But if your event layer works that way, then you need to
> fix it or use it differently as libcurl will of course always use some
> CPU time before it returns back and there will be moments where it has
> told the callback that it wants to know when the socket is writable
> before it returns.
>
> Of course I may not understand your issue correctly.

If the following can happen, there is a problem:

1) connect
2) socket_callback ( write )
   *connect happens asynchronously*
   ::WSAEventSelect -> too late
3) ::WaitForMultipleObjects -> this should wait until the
"write/connect" event has happened, but the event was triggered before
::WSAEventSelect

This is wat we want to happen:

1) socket_callback ( write )
   ::WSAEventSelect
2) connect
   *connect happens asynchronously*
3) ::WaitForMultipleObjects -> this should wait until the
"write/connect" event happens

>
>> We suspect that when we set the write flag for socket 2792 in the
>> curl_socket_callback, we are too late to receive the event. Could
>> this be possible?
>
> I don't know how you "set the write flag" nor do I know how you receive
> events, so I really can't tell unless you tell us more. And of course I
> know very little windows details so I'm probably not the person to ask
> for windows- specific details, but I'm willing to offer insights and
> advice on how libcurl works and assumes the app to behave when using the
> multi_socket() API.
>
Received on 2008-03-11