Re: [PATCH] cpufreq: Processor Clocking Control interface driver

[Dominik Brodowski <linux@dominikbrodowski.net>]

Hey,

On Wed, Dec 16, 2009 at 03:31:48PM +0000, Chumbalkar, Nagananda wrote:
> >> >> +	if (target_freq <= 
> >> >(ioread32(&pcch_hdr->minimum_frequency) * 1000)) {
> >> >> +		target_freq = 
> >> >ioread32(&pcch_hdr->minimum_frequency) * 1000;
> >> >> +		dprintk("target: target_freq for cpu %d was 
> >> >below limit, "
> >> >> +			"converted it to %d\n", cpu, target_freq);
> >> >> +	}
> >> >
> >> >why not do this in the _verify() step? Does 
> >pcch_hdr->minimum_frequency
> >> >even change "on the fly"?
> >> 
> >> pcch_hdr->minimum_frequency does not change "on the fly". 
> >Also, there is no
> >> need for those IO accesses:
> >
> >target_freq cannot be below policy->min or above policy->max. 
> >If it were,
> >the whole cpufreq subsystem is broken. So there's no need for 
> >these checks,
> >AFAICS.
> >
> 
> I think the "ondemand" governor can ask for a target frequency that is
> below policy->min.
> 
> Let's say that at a given snapshot, the frequency of a CPU is policy->cur.
> At the next sampling, if the "load" on the CPU drops to a level that 
> satisifies the equation below, then freq_next requested by "ondemand" 
> will be below policy->min:
> 
> 	load < (threshold - differential) * policy->min / policy->cur
> 
> A patch such as below may be needed to sanitize the target frequency 
> requested by "ondemand". The "conservative" governor already has this check: 

If freq_next < policy->min is indeed possible in ondemand, that's a real bug
in ondemand.

Best,
	Dominik