cpufreq: intel_pstate: Proportional algorithm for Atom

The PID algorithm used by the intel_pstate driver tends to drive
performance to the minimum for workloads with utilization below the
setpoint, which is undesirable, so replace it with a modified
"proportional" algorithm on Atom.

The new algorithm will set the new P-state to be 1.25 times the
available maximum times the (frequency-invariant) utilization during
the previous sampling period except when the target P-state computed
this way is lower than the average P-state during the previous
sampling period.  In the latter case, it will increase the target by
50% of the difference between it and the average P-state to prevent
performance from dropping down too fast in some cases.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Tested-by: Srinivas Pandruvada <srinivas.pandruvada@linux.intel.com>

diff --git a/drivers/cpufreq/intel_pstate.c b/drivers/cpufreq/intel_pstate.c
index 1c7b91c5805bdfca5bbf8da83e870f466a55dea5..6c8c897d0a2dc421316879b84ce482c3b257ae40 100644 (file)
--- a/drivers/cpufreq/intel_pstate.c
+++ b/drivers/cpufreq/intel_pstate.c
@@ -1232,6 +1232,7 @@ static inline int32_t get_target_pstate_use_cpu_load(struct cpudata *cpu)
 {
        struct sample *sample = &cpu->sample;
        int32_t busy_frac, boost;
+       int target, avg_pstate;
 
        busy_frac = div_fp(sample->mperf, sample->tsc);
 
@@ -1242,7 +1243,26 @@ static inline int32_t get_target_pstate_use_cpu_load(struct cpudata *cpu)
                busy_frac = boost;
 
        sample->busy_scaled = busy_frac * 100;
-       return get_avg_pstate(cpu) - pid_calc(&cpu->pid, sample->busy_scaled);
+
+       target = limits->no_turbo || limits->turbo_disabled ?
+                       cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
+       target += target >> 2;
+       target = mul_fp(target, busy_frac);
+       if (target < cpu->pstate.min_pstate)
+               target = cpu->pstate.min_pstate;
+
+       /*
+        * If the average P-state during the previous cycle was higher than the
+        * current target, add 50% of the difference to the target to reduce
+        * possible performance oscillations and offset possible performance
+        * loss related to moving the workload from one CPU to another within
+        * a package/module.
+        */
+       avg_pstate = get_avg_pstate(cpu);
+       if (avg_pstate > target)
+               target += (avg_pstate - target) >> 1;
+
+       return target;
 }
 
 static inline int32_t get_target_pstate_use_performance(struct cpudata *cpu)