Re: [PATCH] cpufreq: intel_pstate: Rework iowait boosting to be less aggressive

[Rafael J. Wysocki]

On Friday, February 1, 2019 5:54:37 PM CET Doug Smythies wrote:
> On 2019.01.30 16:05 Rafael J. Wysocki wrote:
> 
> > From: Rafael J. Wysocki <rafael.j.wysocki@xxxxxxxxx>
> >
> > The current iowait boosting mechanism in intel_pstate_update_util()
> > is quite aggressive, as it goes to the maximum P-state right away,
> > and may cause excessive amounts of energy to be used, which is not
> > desirable and arguably isn't necessary too.
> >
> > Follow commit a5a0809bc58e ("cpufreq: schedutil: Make iowait boost
> > more energy efficient") that reworked the analogous iowait boost
> > mechanism in the schedutil governor and make the iowait boosting
> > in intel_pstate_update_util() work along the same lines.
> > 
> > Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@xxxxxxxxx>
> > ---
> >  drivers/cpufreq/intel_pstate.c |   46 +++++++++++++++++++++++++----------------
> >  1 file changed, 29 insertions(+), 17 deletions(-)
> >
> > Index: linux-pm/drivers/cpufreq/intel_pstate.c
> > ===================================================================
> > --- linux-pm.orig/drivers/cpufreq/intel_pstate.c
> > +++ linux-pm/drivers/cpufreq/intel_pstate.c
> > @@ -50,6 +50,8 @@
> >  #define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
> >  #define fp_toint(X) ((X) >> FRAC_BITS)
> > 
> > +#define ONE_EIGHTH_FP ((int64_t)1 << (FRAC_BITS - 3))
> > +
> >  #define EXT_BITS 6
> >  #define EXT_FRAC_BITS (EXT_BITS + FRAC_BITS)
> >  #define fp_ext_toint(X) ((X) >> EXT_FRAC_BITS)
> > @@ -1678,17 +1680,14 @@ static inline int32_t get_avg_pstate(str
> >  static inline int32_t get_target_pstate(struct cpudata *cpu)
> >  {
> > 	struct sample *sample = &cpu->sample;
> > -	int32_t busy_frac, boost;
> > +	int32_t busy_frac;
> > 	int target, avg_pstate;
> > 
> > 	busy_frac = div_fp(sample->mperf << cpu->aperf_mperf_shift,
> > 			   sample->tsc);
> > 
> > -	boost = cpu->iowait_boost;
> > -	cpu->iowait_boost >>= 1;
> > -
> > -	if (busy_frac < boost)
> > -		busy_frac = boost;
> > +	if (busy_frac < cpu->iowait_boost)
> > +		busy_frac = cpu->iowait_boost;
> > 
> > 	sample->busy_scaled = busy_frac * 100;
> > 
> > @@ -1767,22 +1766,35 @@ static void intel_pstate_update_util(str
> > 	if (smp_processor_id() != cpu->cpu)
> >  		return;
> >
> > +	delta_ns = time - cpu->last_update;
> > 	if (flags & SCHED_CPUFREQ_IOWAIT) {
> > -		cpu->iowait_boost = int_tofp(1);
> > -		cpu->last_update = time;
> > -		/*
> > -		 * The last time the busy was 100% so P-state was max anyway
> > -		 * so avoid overhead of computation.
> > -		 */
> > -		if (fp_toint(cpu->sample.busy_scaled) == 100)
> > -			return;
> > -
> > -		goto set_pstate;
> > +		/* Start over if the CPU may have been idle. */
> > +		if (delta_ns > TICK_NSEC) {
> > +			cpu->iowait_boost = ONE_EIGHTH_FP;
> > +		} else if (cpu->iowait_boost) {
> > +			cpu->iowait_boost <<= 1;
> > +			if (cpu->iowait_boost >= int_tofp(1)) {
> > +				cpu->iowait_boost = int_tofp(1);
> > +				cpu->last_update = time;
> > +				/*
> > +				 * The last time the busy was 100% so P-state
> > +				 * was max anyway, so avoid the overhead of
> > +				 * computation.
> > +				 */
> > +				if (fp_toint(cpu->sample.busy_scaled) == 100)
> > +					return;
> 
> Hi Rafael,
> 
> By exiting here, the trace, if enabled, is also bypassed.
> We want the trace sample.

Fair enough, but the return is there regardless of this patch.

Maybe it should be fixed separately?

> Also, there is a generic:
> "If the target ptstate is the same as before, then don't set it"
> later on.
> Suggest to delete this test and exit condition. (I see that this early
> exit was done before also.)

Well, exactly.

It is not unreasonable to boost the frequency right away for an IO-waiter
without waiting for the next sample time IMO.