lost timer check in 2.6.7

lost timer check in 2.6.7

[Andi Kleen]

2.6.7 has 

+               /* ... but give the TSC a fair chance */
+               if (lost_count > 25)
+                       cpufreq_delayed_get();

While looking at porting this code to x86-64 I noticed that this only runs for 
the first lost timer event. In case of dynamic frequency which varies shouldn't this
be more like

		if ((lost_count % 25) == 0) 
			cpufreq_delayed_get();

? Otherwise this heuristic will only work once.

-Andi

Re: lost timer check in 2.6.7

[Dominik Brodowski]

On Wed, Jun 16, 2004 at 01:19:12PM +0200, Andi Kleen wrote:
> 
> 2.6.7 has 
> 
> +               /* ... but give the TSC a fair chance */
> +               if (lost_count > 25)
> +                       cpufreq_delayed_get();
> 
> While looking at porting this code to x86-64 I noticed that this only runs for 
> the first lost timer event.

Not exactly: lost_count is increased for every "tick" where lost ticks are
detected, but re-set anytime there is no lost tick.

> In case of dynamic frequency which varies shouldn't this
> be more like
> 
> 		if ((lost_count % 25) == 0) 
> 			cpufreq_delayed_get();
> 
> ? Otherwise this heuristic will only work once.

So if this heuristic works, no more ticks will be lost (at least in the near
future), so lost_count will be set to zero again. If, due to some new event,
ticks are lost again, lost_count will start at zero, reach 25 after some
time, and cpufreq_delayed_get() will be called again.

However: inaccuracies are only detected in _one_ direction: ticks being
missed, not "too many ticks" -- and only if it's a factor of two or
higher... Probably a better run-time check of the sanity of a timesource is
needed in future... on the other hand, frequencies shouldn't change behind
the kernel's back. That's what my other patch sent to the ACPI list a few
days ago (pstate_cnt) tries to address.

	Dominik