Re: [PATCH 1/4] tickless idle cpu - Allow any CPU to update jiffies

[Srivatsa Vaddagiri <vatsa@in.ibm.com>]

On Sat, Apr 08, 2006 at 09:04:15AM +1000, Paul Mackerras wrote:
> Srivatsa Vaddagiri writes:
> 
> > Currently, only boot CPU calls do_timer to update jiffies. This prevents
> > idle boot CPU from skipping ticks. Patch below, against 2.6.17-rc1-mm1,
> > allows jiffies to be updated from any CPU.
> 
> We have to be very careful here.  The code that keeps xtime and
> gettimeofday in sync relies on xtime being incremented as close as
> possible in time to when the timebase passes specific values.  Since
> we currently stagger the timer interrupts for the cpus throughout a
> jiffy, having cpus other than the boot cpus calling do_timer will
> break this and introduce inaccuracies.  There are also implications
> for the stolen time accounting on shared-processor LPAR systems.
> 
> I think we need to remove the staggering, thus having all cpus take
> their timer interrupt at the same time.  That way, any of them can
> call do_timer.  However we then have to be much more careful about
> possible contention, e.g. on xtime_lock.  Your patch has every cpu
> taking xtime_lock for writing rather than just the boot cpu.  I'd like
> to see if there is some way to avoid that (while still having just one
> cpu call do_timer, of course).

Paul,
	Thanks for the feedback on the patches.

Avoiding contention on xtime_lock doesnt seem to be trivial. Any
solution to it is fraught with races. Anyway, I have attempted one
solution (in the followon Patch 2/2) which keeps the overhead in timer 
interrupt handler low. 

Let me know if you have other suggestions to avoid xtime_lock
contention!

Following patches are sent in separate mails:

Patch 1/2	-	Core patch to skip ticks - v2
Patch 2/2	- 	Allow boot CPU to skip ticks - v2

The sysctl control patch and decrementer statistics patch are as before
and hence I am not resending them this time.


-- 
Regards,
vatsa