Keir,

In my work on layering hpet on get_s_time, I found a problem
in get_s_time and related code. Because of the problem I was getting
large jumps in the offset between local time and ntp time.
These jumps were on the order of many seconds.

The issue is the race between local_time_calibration() executing
on one processor and platform_time_calibration() on another.

I have included a patch which addresses the race in
local_time_calibration(), cpu_frequency_change(), and
init_percpu_time().

I'm giving you this ahead of the hpet work as it affects all users
of get_s_time().

I'm confident of the fix in local_time_calibration() as I had failures there
before the fix and no failures after. The other two I'm less confident 
in, so check
my work closely there.

On the hpet over get_s_time() front, this fix allows me to get .0014% error.
This is very close to the error going to the hardware hpet each time.

Regards,
Dave