Re: NSRunLoop limitations

[Richard Frith-Macdonald]

On Wednesday, January 23, 2002, at 02:06 AM, Kaelin Colclasure wrote:

> I have run squarely into a limitation of the NSRunLoop
> implementation. For the distributed benchmarking project I am working
> on, I have found it necessary to increase the number of file
> descriptors a single process can have open. Unfortunately, this has
> illuminated the fact that the bowels of NSRunLoop are based on
> select(2) rather than poll(2). This means that NSRunLoop suffers from
> the FD_SETSIZE limitation that years ago convinced me to abandon
> select in favor of poll and never look back. [Yes, I know about
> -DFD_SETSIZE=nnn and I also know how non-portable it is and what a
> PITA it is even when it works!]
>
> There are (at least) two ways I can solve my problem:
>
>  1) I can replace NSRunLoop and the classes which intimately depend on
>     it within my programs.
>
>  2) I can rework GNUstep's NSRunLoop and intimates to use poll instead
>     of select.
>
> Solution 1 has the virtue of being straightforward to code, but the
> vice of being intimately coupled to other parts of the GNUstep code
> base while being separately maintained. Solution 2 is probably more
> work in the short term, but (assuming for the moment I got it right)
> would solve this problem `once and for all'.
>
> I'd like to hear from (at least) one of the GNUstep maintainers on
> this point... Is this a limitation that bears removing? Or am I the
> only developer who's ever (ab)used NSRunLoop in this particular
> fashion?

The use of select() is a historical limitation ... yes the base library
really *has* been around so long that when it was written a lot of 
systems
didn't have poll().

I've never found it to be a real difficulty - as the default set size 
under
GNU/Linux is 1024 and I've never needed more than that many descriptors,
but I know it can be a problem occasionally and it would be nice to have
a poll() implementation.
We could use 'if HAVE_POLL'  to control fallback to select() in case 
there
are still select only operating systems in use.

The work of implementing a poll() solution is really almost trivial - it
would require a new implementation of a single method 
-acceptInputForMode:beforeDate:
or some conditional compilation within that method, but I think a 
separate
version of the whole method might be more readable.
I can't imagine how your solution 1 could be even close to being as
simple as solution 2, since what you are really talking about in 1 is 
rewriting
the entire class, but in 2 you would just rewrite one method.