Re: [RFC PATCH 5/7] sched/schedutil: Add a new tunable to dictate response time

[Qais Yousef]

On 09/07/23 13:44, Peter Zijlstra wrote:
> On Mon, Aug 28, 2023 at 12:32:01AM +0100, Qais Yousef wrote:
> > +static inline unsigned long
> > +sugov_apply_response_time(struct sugov_policy *sg_policy, unsigned long util)
> > +{
> > +	unsigned long mult;
> > +
> > +	if (sg_policy->freq_response_time_ms == sg_policy->tunables->response_time_ms)
> > +		return util;
> > +
> > +	mult = sg_policy->freq_response_time_ms * SCHED_CAPACITY_SCALE;
> > +	mult /=	sg_policy->tunables->response_time_ms;
> > +	mult *= util;
> > +
> > +	return mult >> SCHED_CAPACITY_SHIFT;
> > +}
> > +
> >  static bool sugov_should_update_freq(struct sugov_policy *sg_policy, u64 time)
> >  {
> >  	s64 delta_ns;
> > @@ -143,6 +184,7 @@ static unsigned int get_next_freq(struct sugov_policy *sg_policy,
> >  	unsigned int freq = arch_scale_freq_invariant() ?
> >  				policy->cpuinfo.max_freq : policy->cur;
> >  
> > +	util = sugov_apply_response_time(sg_policy, util);
> >  	freq = map_util_freq(util, freq, max);
> >  
> >  	if (freq == sg_policy->cached_raw_freq && !sg_policy->need_freq_update)
> 
> Urgh, so instead of caching the multiplier you keep computing what is
> essentially a constant over and over and over and over again :/
> 
> That is, compute the whole 'freq_response_time_ms * SCHED_CAPACITY_SCALE
> / response_time_ms' thing *once*, when that file is written to, and then
> reduce the whole thing to:
> 
> 	return (freq_response_mult * util) >> SCHED_CAPACITY_SHIFT;
> 
> No need for that special case, no need for divisions, just go.

Yes! I was too focused am I doing the right thing, I didn't stop to think this
is actually a constant and can be done once too. I will fix it if this knobs
ends up hanging around.


Thanks!

--
Qais Yousef