On Thu, 4 Sep 2008 10:21:11 -0600
Matthew Wilcox [off-list ref] wrote:

On Thu, Sep 04, 2008 at 06:14:24PM +0200, Ingo Molnar wrote:

quoted

quoted

If schedule() returned whether or not it had scheduled another
task, we could do something like:

		if (!schedule())
			udelay(10);

hm, i'm not really sure - this really just seems to be a higher
prio variant of yield() combined with some weird code. Do we really
want to promote such arguably broken behavior? If there's any
chance of any polling to take a material amount of CPU time it
should be event driven to begin with.

Oh, I'm not concerned about CPU utilisation, I'm concerned about PCI
bus utilisation.  Perhaps I'd like a yield_timeout() function instead
where I say that I'd like to not run for at least 10 microseconds?

Can we do that, or are we still jiffie-based there?

use schedule_hrtimerout() for this (hopefully will be in 2.6.28);
see this weeks LWN for an article describing it