Are you sure about this?

Yes. Pretty sure.

I've never programmed under Windows,

Don't write this off as a "windows problem".

but under Linux I haven't seen yet any single case when the file descriptor would be closed by the kernel.

You may not have, but others have:

Platform:
 osname=linux, osvers=2.6.24-27-server, archname=i486-linux-gnu-thread
+-multi
 uname='linux vernadsky 2.6.24-27-server #1 smp fri mar 12 01:45:06 ut
+c 2010 i686 gnulinux '
[download]

As to example, I've never used IO::Select module, but many years ago I've created more or less simple module that uses "select" directly. This module still works for many of my application, even though I consider it to be very poor. You can get it at http://vandal.sdf-eu.org/NetHandling.pm if you want.

Thanks for teh offer, but the way I eventually stopped writing off my long term lack of success with IO::Select as "user error" was by re-implementing (most) of the API without reference to the module source:

I guess the main reason why one can't find good examples of using "select" is just the fact, that it is not that simple.

Maybe, but wrapping over some of the low-level nitty-gritty in an IO:Select-type module certainly makes it easier.

But if there has been a latent bug in that module for the last 12 or more years that prevented people's attempts from working, it would have had a significant effect upon the availability of good examples.

And given that there are plenty of examples of many other things that are at least as complex, I find that a pretty persuasive argument for that inhibitory effect.

I've never really needed to use select for serious work -- I find threading far simpler, more intuitive and more powerful -- on windows at least -- so my previous attempts have been half-hearted -- usually with a view to demonstrating the virtues of threading over the select model -- and so I never had any real incentive to work past the problems.

Now I have understood the problem that has dogged my attempts for years, I feel the need to write a (pair of) full-featured servers. In part so I know I have done so. In part to allow me to make the real-world comparisons and measurements that will test my gut feel hypothesis about the relative merits of the two approaches.

Watch this space. (But don't hold your breath :)