[PATCH v2 3/6] cpufreq: intel_pstate: Set initial scaling_min_freq value

[Rafael J. Wysocki]

From: Rafael J. Wysocki <rafael.j.wysocki@xxxxxxxxx>

Currently, intel_pstate sets cpuinfo_min_freq to the minimum P-state
value retrieved from MSR_PLATFORM_INFO (bits 40:47) which prevents users
from setting scaling_min_freq below it.  However, there are systems
where CPUs may actually run at P-states below that limit and the power
of the processor is then lower than when they run at the limit P-state,
even though running them at the limit P-state may still be more
energy-efficient.

To allow users to utilize the low P-states in question, notice that when
HWP is enabled, the LOWEST_PERF value from MSR_HWP_CAPABILITIES can be
used for setting cpuinfo_min_freq instead of the minimum P-state value
mentioned above, but in order to preserve the default behavior, use the
observation that after commit 8c83947c5dbb ("cpufreq: Use policy->min/max
init as QoS request"), cpufreq drivers can request the initial
scaling_min_freq values above cpuinfo_min_freq by setting policy->min to
those values in their .init() callbacks.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@xxxxxxxxx>
---

v1 -> v2:
   * Rely on intel_pstate_update_freq_limits() to update the cached frequency
     values and modify it to take the min frequency into account (based on
     Sashiko feedback:
     https://sashiko.dev/#/patchset/2381464.iZASKD2KPV%40rafael.j.wysocki)
   * Minor tweaks in the changelog

---
 drivers/cpufreq/intel_pstate.c |   32 ++++++++++++++++++++------------
 1 file changed, 20 insertions(+), 12 deletions(-)

--- a/drivers/cpufreq/intel_pstate.c
+++ b/drivers/cpufreq/intel_pstate.c
@@ -589,14 +589,6 @@ static void intel_pstate_hybrid_hwp_adju
 
 	freq = perf_ctl_max_phys * perf_ctl_scaling;
 	cpu->pstate.max_pstate_physical = intel_pstate_freq_to_hwp(cpu, freq);
-
-	freq = cpu->pstate.min_pstate * perf_ctl_scaling;
-	cpu->pstate.min_freq = freq;
-	/*
-	 * Cast the min P-state value retrieved via pstate_funcs.get_min() to
-	 * the effective range of HWP performance levels.
-	 */
-	cpu->pstate.min_pstate = intel_pstate_freq_to_hwp(cpu, freq);
 }
 
 static bool turbo_is_disabled(void)
@@ -1181,15 +1173,21 @@ static void intel_pstate_update_freq_lim
 	int scaling = cpu->pstate.scaling;
 	unsigned int turbo_freq = cpu->pstate.turbo_pstate * scaling;
 	unsigned int max_freq = cpu->pstate.max_pstate * scaling;
+	unsigned int min_freq = cpu->pstate.min_pstate * scaling;
 	int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
 
 	if (scaling != perf_ctl_scaling) {
 		turbo_freq = rounddown(turbo_freq, perf_ctl_scaling);
 		max_freq = rounddown(max_freq, perf_ctl_scaling);
+		if (min_freq > perf_ctl_scaling)
+			min_freq = rounddown(min_freq, perf_ctl_scaling);
+		else
+			min_freq = perf_ctl_scaling;
 	}
 
 	cpu->pstate.turbo_freq = turbo_freq;
 	cpu->pstate.max_freq = max_freq;
+	cpu->pstate.min_freq = min_freq;
 }
 
 static void __intel_pstate_get_hwp_cap(struct cpudata *cpu)
@@ -1198,6 +1196,7 @@ static void __intel_pstate_get_hwp_cap(s
 
 	rdmsrq_on_cpu(cpu->cpu, MSR_HWP_CAPABILITIES, &cap);
 	WRITE_ONCE(cpu->hwp_cap_cached, cap);
+	cpu->pstate.min_pstate = HWP_LOWEST_PERF(cap);
 	cpu->pstate.max_pstate = HWP_GUARANTEED_PERF(cap);
 	cpu->pstate.turbo_pstate = HWP_HIGHEST_PERF(cap);
 }
@@ -2317,7 +2316,6 @@ static void intel_pstate_get_cpu_pstates
 	int perf_ctl_scaling = pstate_funcs.get_scaling();
 
 	cpu->pstate.max_pstate_physical = pstate_funcs.get_max_physical(cpu->cpu);
-	cpu->pstate.min_pstate = pstate_funcs.get_min(cpu->cpu);
 	cpu->pstate.perf_ctl_scaling = perf_ctl_scaling;
 
 	if (hwp_active && !hwp_mode_bdw) {
@@ -2337,6 +2335,7 @@ static void intel_pstate_get_cpu_pstates
 		hybrid_update_capacity(cpu);
 	} else {
 		cpu->pstate.scaling = perf_ctl_scaling;
+		cpu->pstate.min_pstate = pstate_funcs.get_min(cpu->cpu);
 		cpu->pstate.max_pstate = pstate_funcs.get_max(cpu->cpu);
 		cpu->pstate.turbo_pstate = pstate_funcs.get_turbo(cpu->cpu);
 	}
@@ -3047,6 +3046,13 @@ static int __intel_pstate_cpu_init(struc
 	policy->cpuinfo.max_freq = READ_ONCE(global.no_turbo) ?
 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
 
+	/*
+	 * If policy->min is greater than policy->cpuinfo.min_freq, the cpufreq
+	 * core will use this value for initializing scaling_min_freq.
+	 */
+	if (hwp_active)
+		policy->min = pstate_funcs.get_min(cpu->cpu) * cpu->pstate.perf_ctl_scaling;
+
 	intel_pstate_init_acpi_perf_limits(policy);
 
 	policy->fast_switch_possible = true;
@@ -3313,6 +3319,7 @@ static int intel_cpufreq_cpu_init(struct
 	cpu = all_cpu_data[policy->cpu];
 
 	if (hwp_active) {
+		unsigned int hwp;
 		u64 value;
 
 		policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY_HWP;
@@ -3324,14 +3331,15 @@ static int intel_cpufreq_cpu_init(struct
 
 		cpu->epp_cached = intel_pstate_get_epp(cpu, value);
 
+		hwp = intel_pstate_freq_to_hwp(cpu, policy->min);
 		intel_cpufreq_hwp_update(cpu, cpu->pstate.min_pstate,
-					 cpu->pstate.max_pstate,
-					 cpu->pstate.min_pstate, false);
+					 cpu->pstate.max_pstate, hwp, false);
+		policy->cur = policy->min;
 	} else {
 		policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY;
 		intel_pstate_set_min_pstate(cpu);
+		policy->cur = policy->cpuinfo.min_freq;
 	}
-	policy->cur = policy->cpuinfo.min_freq;
 
 	freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * global.min_perf_pct, 100);