Re: [Libevent-users] libevent + bufferevent + IOCP = confused

[Nick Mathewson <nickm@xxxxxxxxxxxxx>]

On Fri, Jul 1, 2011 at 11:26 AM, Juan Pablo Fernandez
<this.juan@xxxxxxxxx> wrote:
> Hi,
>
> I've been experimenting with libevent on windows. I want to use IOCP,
> worker threads filling a queue layer in a server application.
> Bufferevents is a nice fit! And I've written a small server using
> bufferevents.  However, I'm confused reading the documentation,
> saying:
>
> 1, It's experimental (online book)
> 2, Back-end is still select-based.
> 3, Can be used for sending (no mention of receive)
>
> ==> What is the "true" status?

It is indeed experimental.

The select-based backend is still used for the event interface
provided by event_add, event_del, etc.  It turns out that you can't
emulate that using IOCP, as far as I can tell.[*]  If you want to use
IOCP, you need to use bufferevents: the bufferevent interface _can_ be
implemented with IOCP.

And so there's an implementation of it.

To use it, initialize Libevent passing the
EVENT_BASE_FLAG_STARTUP_IOCP flag in the event_base_config.  This will
make subsequent calls to bufferevent_socket_new() return an IOCP-based
bufferevent, and evconnlistener_*new() return IOCP-based listeners.

[*] Some people have in the past tried to implement the event backend
using IOCP, using a hack where "completing" a zero-length read or
write indicated that the socket was readable/writable.  When we
investigated, it turned out that this works for reading, but not for
writing: sockets whose buffers are too full to write anything will
still accept a zero-length write and report it as "completed."  If
there's a way around this, or somebody knows differently, I'd love to
know about it.

yrs,
-- 
Nick
***********************************************************************
To unsubscribe, send an e-mail to majordomo@xxxxxxxxxxxxx with
unsubscribe libevent-users    in the body.