Re: udelay function delays the wrong time interval in multiprocessor system, if ARCH_HAS_READ_CURRENT_TIMER is not defined and on current timer is used.

[Russell King - ARM Linux <linux@arm.linux.org.uk>]

On Sun, Mar 17, 2013 at 08:05:43PM +0000, Will Deacon wrote:
> On Sat, Mar 16, 2013 at 03:32:43AM +0000, chpoph wrote:
> > On Sat, Mar 16, 2013 at 2:14 AM, Russell King - ARM Linux
> > <linux@arm.linux.org.uk> wrote:
> > > We don't support different CPUs running at different frequencies with
> > > the delay loop.  Sorry.
> > 
> > Does it means that a timer-based delay implementation must be used to
> > get an accurate delay in SMP. I think it should print a warning
> > message if the CPU delay loop is used in SMP.  In my system, the wrong
> > delay interval fluctuated with CPU frequencies caused a control
> > problem.
> 
> I've been playing around with loops_per_jiffy recently, in an attempt to
> clean up the cpufreq scaling code so that the SMP-ness is in core code,
> rather than being duplicated by every architecture:
> 
>   git://git.kernel.org/pub/scm/linux/kernel/git/will/linux.git lpj
> 
> With those changes, it's pretty easy to get different delays depending on
> the current CPU, but it would require preempt_{enable,disable} calls around
> the delay, which I haven't convinced myself about.

Exactly, and that's why I said what I said.  If you start doing that,
then you might as well turn kernel preemption off altogether, because
the delays will impact your system latency.