[Openpxa-users] Linux udelay() is way off

[Russell King - ARM Linux]

On Thu, Jan 20, 2011 at 09:25:36PM +0100, Bjørn Forsman wrote:
> 2011/1/20 Russell King - ARM Linux <linux at arm.linux.org.uk>:
> > On Thu, Jan 20, 2011 at 08:18:10PM +0100, Bjørn Forsman wrote:
> >> [   11.425363] cpufreq-core: saving 518144 as reference value for
> >> loops_per_jiffy; freq is 312000 kHz
> >> [   11.425523] cpufreq-core: scaling loops_per_jiffy to 1036288 for
> >> frequency 624000 kHz
> >
> > So despite scaling lpj by a factor of two for the doubling in clock
> > frequency, we see a 30% error in it?
> 
> No. udelay() is ~30 % of what it should be. My gpio + oscilloscope
> test showed that udelay(50) => 16 us in real time with cpufreq
> enabled. This gives 16/50 = 0.32.
> 
> > Is it possible to boot at 624MHz and report the resulting lpj?
> 
> I think I need some help doing that. I found some clock setup in
> <u-boot-pxa>/arch/arm/cpu/pxa/start.S (OpenPXA tree) but I'm not
> familiar with it.
> 
> > Alternatively, call calibrate_delay() after the cpufreq switch to
> > 624MHz, disabling the 'printed' stuff and see what it reports.
> 
> First attempts at moving calibrate_delay() after freq switch failed. I
> will try more. Just had to send this so stop the misunderstanding
> about how much udelay() is off...

Not a good idea to remove it from its current location.  I was suggesting
adding another call to calibrate_delay() after the cpufreq stuff has
initialized.

> What do you mean by disabling the 'printed' stuff? The cpufreq debug stuff?

See calibrate_delay() in init/calibrate.c.  'static bool printed;'  It's
set true after the first run, and there after it won't print the
information again.  So if you want to see it from subsequent runs, you
need to defeat it.