Re: [PATCH] Many groups patch.

[Tim Hockin <thockin@hockin.org>]

On Mon, Sep 29, 2003 at 06:29:32PM -0400, Pete Zaitcev wrote:
> >> This version drops the internal groups array (it's so often shared
> >> that it's not worth it, and the logic becomes a bit neater), and does
> >> vmalloc fallback in case someone has massive number of groups.
> > 
> > Why?
> 
> > The vmalloc space is limited, and the code just gets uglier.
> 
> Tim was going to write a version that segments groups into
> smaller arrays. I reckon it was too difficult?

I posted it once or twice then got busy.  It's an array of pages.  Rusty has
it, but didn't believe me when I said Linus wouldn't let vmalloc() fly.

> > Have you been looking at glibc sources lately, or why do you believe that 
> > we should encourage insane usage?
> 
> We have some customers who run insane number of groups,
> with their own patches. This practice is popular in the
> Beowulf crowd for some reason. I should note this is not
> very mainstream.

I'd be ok with the simplest "kmalloc or too bad" version, but our customers
wouldn't.

My version uses a struct group_info which has an array of pages.  The groups
are sorted and bsearched, instead of linear.  The perfomance is quite good.
An older version against 2.6.0-test1 or something is attached.  If this
method will fly, I'll take some of Rusty's good ideas and finish this
version of it..

Linus?