Re: non-invariant tsc and cputicker

On Saturday 04 December 2010 06:12 am, Andriy Gapon wrote:
> on 04/12/2010 02:38 Jung-uk Kim said the following:
> > If my understanding is correct, your patch uses the dummy
> > timecounter until a real timecounter is chosen.
>
> Perhaps this is one way to look at it.
> But I look at it differently - the patch makes cpu_ticks refer to
> tc_cpu_ticks. That is, it make _the_ timecounter be used for cpu
> ticks.
> Exact timecounter backend is not important to me.
>
> > When a real timecounter is set,
> > tc_cpu_ticks() changes the frequency naturally.  How are you
> > going to solve this problem?
>
> Do we really care about cpu ticks accounting that early in the
> boot?
>
> >  What should we do when a user set a new
> > timecounter hardware via "sysctl kern.timecounter.hardware"?
>
> User can expect some instability (*if any*) when he does such a
> significant system reconfiguration.
> I put "if any", because I think that tc_cpu_ticks() should handle
> this. The same way as you don't see time returned by e.g.
> nanotime() going crazy at that moment.
>
> > I don't
> > think it is any better than current code.  Am I missing
> > something? :-(
>
> I think that it is much better.
> Handling of P-state changes for non-invariant TSC is just
> incorrect. kern.timecounter.hardware is not going to be changed as
> frequently as P-states, if ever.

Sigh...  Please see the history of calcru() in 
sys/kern/kern_resource.c.  Most important ones are:

http://svn.freebsd.org/viewvc/base?view=revision&revision=155444
http://svn.freebsd.org/viewvc/base?view=revision&revision=155534

Basically, we chose efficiency over accuracy and you are suggesting 
going backwards.

Jung-uk Kim