diff mbox series

[2/2] cpufreq: intel_pstate: hybrid: CPU-specific scaling factor

Message ID 2790305.e9J7NaK4W3@kreacher
State New
Headers show
Series cpufreq: intel_pstate: HWP support on hybrid processors | expand

Commit Message

Rafael J. Wysocki May 12, 2021, 2:19 p.m. UTC
From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

The scaling factor between HWP performance levels and CPU frequency
may be different for different types of CPUs in a hybrid processor
and in general the HWP performance levels need not correspond to
"P-states" representing values that would be written to
MSR_IA32_PERF_CTL if HWP was disabled.

However, the policy limits control in cpufreq is defined in terms
of CPU frequency, so it is necessary to map the frequency limits set
through that interface to HWP performance levels with reasonable
accuracy and the behavior of that interface on hybrid processors
has to be compatible with its behavior on non-hybrid ones.

To address this problem, use the observations that (1) on hybrid
processors the sysfs interface can operate by mapping frequency
to "P-states" and translating those "P-states" to specific HWP
performance levels of the given CPU and (2) the scaling factor
between the MSR_IA32_PERF_CTL "P-states" and CPU frequency can be
regarded as a known value.  Moreover, the mapping between the
HWP performance levels and CPU frequency can be assumed to be
linear and such that HWP performance level 0 correspond to the
frequency value of 0, so it is only necessary to know the
frequency corresponding to one specific HWP performance level
to compute the scaling factor applicable to all of them.

One possibility is to take the nominal performance value from CPPC,
if available, and use cpu_khz as the corresponding frequency.  If
the CPPC capabilities interface is not there or the nominal
performance value provided by it is out of range, though, something
else needs to be done.

Namely, the guaranteed performance level either from CPPC or from
MSR_HWP_CAPABILITIES can be used instead, but the corresponding
frequency needs to be determined.  That can be done by computing the
product of the (known) scaling factor between the MSR_IA32_PERF_CTL
P-states and CPU frequency (the PERF_CTL scaling factor) and the
P-state value referred to as the "TDP ratio".

If the HWP-to-frequency scaling factor value obtained in one of the
ways above turns out to be euqal to the PERF_CTL scaling factor, it
can be assumed that the number of HWP performance levels is equal to
the number of P-states and the given CPU can be handled as though
this was not a hybrid processor.

Otherwise, one more adjustment may still need to be made, because the
HWP-to-frequency scaling factor computed so far may not be accurate
enough (e.g. because the CPPC information does not match the exact
behavior of the processor).  Specifically, in that case the frequency
corresponding to the highest HWP performance value from
MSR_HWP_CAPABILITIES (computed as the product of that value and the
HWP-to-frequency scaling factor) cannot exceed the frequency that
corresponds to the maximum 1-core turbo P-state value from
MSR_TURBO_RATIO_LIMIT (computed as the procuct of that value and the
PERF_CTL scaling factor) and the HWP-to-frequency scaling factor may
need to be adjusted accordingly.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
---
 drivers/cpufreq/intel_pstate.c |  233 ++++++++++++++++++++++++++++++++++++-----
 1 file changed, 210 insertions(+), 23 deletions(-)
diff mbox series

Patch

Index: linux-pm/drivers/cpufreq/intel_pstate.c
===================================================================
--- linux-pm.orig/drivers/cpufreq/intel_pstate.c
+++ linux-pm/drivers/cpufreq/intel_pstate.c
@@ -121,9 +121,10 @@  struct sample {
  * @max_pstate_physical:This is physical Max P state for a processor
  *			This can be higher than the max_pstate which can
  *			be limited by platform thermal design power limits
- * @scaling:		Scaling factor to  convert frequency to cpufreq
- *			frequency units
+ * @perf_ctl_scaling:	PERF_CTL P-state to frequency scaling factor
+ * @scaling:		Scaling factor between performance and frequency
  * @turbo_pstate:	Max Turbo P state possible for this platform
+ * @min_freq:		@min_pstate frequency in cpufreq units
  * @max_freq:		@max_pstate frequency in cpufreq units
  * @turbo_freq:		@turbo_pstate frequency in cpufreq units
  *
@@ -134,8 +135,10 @@  struct pstate_data {
 	int	min_pstate;
 	int	max_pstate;
 	int	max_pstate_physical;
+	int	perf_ctl_scaling;
 	int	scaling;
 	int	turbo_pstate;
+	unsigned int min_freq;
 	unsigned int max_freq;
 	unsigned int turbo_freq;
 };
@@ -489,6 +492,149 @@  static int intel_pstate_get_cppc_gurante
 }
 #endif /* CONFIG_ACPI_CPPC_LIB */
 
+static bool intel_pstate_cppc_perf_valid(u32 perf, struct cppc_perf_caps *caps)
+{
+	return perf && perf <= caps->highest_perf && perf >= caps->lowest_perf;
+}
+
+static bool intel_pstate_cppc_perf_caps(struct cpudata *cpu,
+					struct cppc_perf_caps *caps)
+{
+	if (cppc_get_perf_caps(cpu->cpu, caps))
+		return false;
+
+	return caps->highest_perf && caps->lowest_perf <= caps->highest_perf;
+}
+
+static void intel_pstate_hybrid_hwp_perf_ctl_parity(struct cpudata *cpu)
+{
+	pr_debug("CPU%d: Using PERF_CTL scaling for HWP\n", cpu->cpu);
+
+	cpu->pstate.scaling = cpu->pstate.perf_ctl_scaling;
+}
+
+/**
+ * intel_pstate_hybrid_hwp_calibrate - Calibrate HWP performance levels.
+ * @cpu: Target CPU.
+ *
+ * On hybrid processors, HWP may expose more performance levels than there are
+ * P-states accessible through the PERF_CTL interface.  If that happens, the
+ * scaling factor between HWP performance levels and CPU frequency will be less
+ * than the scaling factor between P-state values and CPU frequency.
+ *
+ * In that case, the scaling factor between HWP performance levels and CPU
+ * frequency needs to be determined which can be done with the help of the
+ * observation that certain HWP performance levels should correspond to certain
+ * P-states, like for example the HWP highest performance should correspond
+ * to the maximum turbo P-state of the CPU.
+ */
+static void intel_pstate_hybrid_hwp_calibrate(struct cpudata *cpu)
+{
+	struct cppc_perf_caps caps;
+	int perf_ctl_max_phys = cpu->pstate.max_pstate_physical;
+	int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
+	int perf_ctl_turbo = pstate_funcs.get_turbo();
+	int turbo_freq = perf_ctl_turbo * perf_ctl_scaling;
+	int perf_ctl_max = pstate_funcs.get_max();
+	int max_freq = perf_ctl_max * perf_ctl_scaling;
+	int scaling = INT_MAX;
+	int freq;
+
+	pr_debug("CPU%d: perf_ctl_max_phys = %d\n", cpu->cpu, perf_ctl_max_phys);
+	pr_debug("CPU%d: perf_ctl_max = %d\n", cpu->cpu, perf_ctl_max);
+	pr_debug("CPU%d: perf_ctl_turbo = %d\n", cpu->cpu, perf_ctl_turbo);
+	pr_debug("CPU%d: perf_ctl_scaling = %d\n", cpu->cpu, perf_ctl_scaling);
+
+	pr_debug("CPU%d: HWP_CAP guaranteed = %d\n", cpu->cpu, cpu->pstate.max_pstate);
+	pr_debug("CPU%d: HWP_CAP highest = %d\n", cpu->cpu, cpu->pstate.turbo_pstate);
+
+	if (intel_pstate_cppc_perf_caps(cpu, &caps)) {
+		if (intel_pstate_cppc_perf_valid(caps.nominal_perf, &caps)) {
+			pr_debug("CPU%d: Using CPPC nominal\n", cpu->cpu);
+
+			/*
+			 * If the CPPC nominal performance is valid, it can be
+			 * assumed to correspond to cpu_khz.
+			 */
+			if (caps.nominal_perf == perf_ctl_max_phys) {
+				intel_pstate_hybrid_hwp_perf_ctl_parity(cpu);
+				return;
+			}
+			scaling = DIV_ROUND_UP(cpu_khz, caps.nominal_perf);
+		} else if (intel_pstate_cppc_perf_valid(caps.guaranteed_perf, &caps)) {
+			pr_debug("CPU%d: Using CPPC guaranteed\n", cpu->cpu);
+
+			/*
+			 * If the CPPC guaranteed performance is valid, it can
+			 * be assumed to correspond to max_freq.
+			 */
+			if (caps.guaranteed_perf == perf_ctl_max) {
+				intel_pstate_hybrid_hwp_perf_ctl_parity(cpu);
+				return;
+			}
+			scaling = DIV_ROUND_UP(max_freq, caps.guaranteed_perf);
+		}
+	}
+	/*
+	 * If using the CPPC data to compute the HWP-to-frequency scaling factor
+	 * doesn't work, use the HWP_CAP gauranteed perf for this purpose with
+	 * the assumption that it corresponds to max_freq.
+	 */
+	if (scaling > perf_ctl_scaling) {
+		pr_debug("CPU%d: Using HWP_CAP guaranteed\n", cpu->cpu);
+
+		if (cpu->pstate.max_pstate == perf_ctl_max) {
+			intel_pstate_hybrid_hwp_perf_ctl_parity(cpu);
+			return;
+		}
+		scaling = DIV_ROUND_UP(max_freq, cpu->pstate.max_pstate);
+		if (scaling > perf_ctl_scaling) {
+			/*
+			 * This should not happen, because it would mean that
+			 * the number of HWP perf levels was less than the
+			 * number of P-states, so use the PERF_CTL scaling in
+			 * that case.
+			 */
+			pr_debug("CPU%d: scaling (%d) out of range\n", cpu->cpu,
+				scaling);
+
+			intel_pstate_hybrid_hwp_perf_ctl_parity(cpu);
+			return;
+		}
+	}
+
+	/*
+	 * If the product of the HWP performance scaling factor obtained above
+	 * and the HWP_CAP highest performance is greater than the maximum turbo
+	 * frequency corresponding to the pstate_funcs.get_turbo() return value,
+	 * the scaling factor is too high, so recompute it so that the HWP_CAP
+	 * highest performance corresponds to the maximum turbo frequency.
+	 */
+	if (turbo_freq < cpu->pstate.turbo_pstate * scaling) {
+		pr_debug("CPU%d: scaling too high (%d)\n", cpu->cpu, scaling);
+
+		cpu->pstate.turbo_freq = turbo_freq;
+		scaling = DIV_ROUND_UP(turbo_freq, cpu->pstate.turbo_pstate);
+	}
+
+	cpu->pstate.scaling = scaling;
+
+	pr_debug("CPU%d: HWP-to-frequency scaling factor: %d\n", cpu->cpu, scaling);
+
+	cpu->pstate.max_freq = rounddown(cpu->pstate.max_pstate * scaling,
+					 perf_ctl_scaling);
+
+	freq = perf_ctl_max_phys * perf_ctl_scaling;
+	cpu->pstate.max_pstate_physical = DIV_ROUND_UP(freq, scaling);
+
+	cpu->pstate.min_freq = cpu->pstate.min_pstate * perf_ctl_scaling;
+	/*
+	 * Cast the min P-state value retrieved via pstate_funcs.get_min() to
+	 * the effective range of HWP performance levels.
+	 */
+	cpu->pstate.min_pstate = DIV_ROUND_UP(cpu->pstate.min_freq, scaling);
+}
+
 static inline void update_turbo_state(void)
 {
 	u64 misc_en;
@@ -795,19 +941,22 @@  cpufreq_freq_attr_rw(energy_performance_
 
 static ssize_t show_base_frequency(struct cpufreq_policy *policy, char *buf)
 {
-	struct cpudata *cpu;
-	u64 cap;
-	int ratio;
+	struct cpudata *cpu = all_cpu_data[policy->cpu];
+	int ratio, freq;
 
 	ratio = intel_pstate_get_cppc_guranteed(policy->cpu);
 	if (ratio <= 0) {
+		u64 cap;
+
 		rdmsrl_on_cpu(policy->cpu, MSR_HWP_CAPABILITIES, &cap);
 		ratio = HWP_GUARANTEED_PERF(cap);
 	}
 
-	cpu = all_cpu_data[policy->cpu];
+	freq = ratio * cpu->pstate.scaling;
+	if (cpu->pstate.scaling != cpu->pstate.perf_ctl_scaling)
+		freq = rounddown(freq, cpu->pstate.perf_ctl_scaling);
 
-	return sprintf(buf, "%d\n", ratio * cpu->pstate.scaling);
+	return sprintf(buf, "%d\n", freq);
 }
 
 cpufreq_freq_attr_ro(base_frequency);
@@ -831,9 +980,20 @@  static void __intel_pstate_get_hwp_cap(s
 
 static void intel_pstate_get_hwp_cap(struct cpudata *cpu)
 {
+	int scaling = cpu->pstate.scaling;
+
 	__intel_pstate_get_hwp_cap(cpu);
-	cpu->pstate.max_freq = cpu->pstate.max_pstate * cpu->pstate.scaling;
-	cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
+
+	cpu->pstate.max_freq = cpu->pstate.max_pstate * scaling;
+	cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * scaling;
+	if (scaling != cpu->pstate.perf_ctl_scaling) {
+		int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
+
+		cpu->pstate.max_freq = rounddown(cpu->pstate.max_freq,
+						 perf_ctl_scaling);
+		cpu->pstate.turbo_freq = rounddown(cpu->pstate.turbo_freq,
+						   perf_ctl_scaling);
+	}
 }
 
 static void intel_pstate_hwp_set(unsigned int cpu)
@@ -1724,19 +1884,33 @@  static void intel_pstate_max_within_limi
 
 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
 {
+	bool hybrid_cpu = boot_cpu_has(X86_FEATURE_HYBRID_CPU);
+	int perf_ctl_max_phys = pstate_funcs.get_max_physical();
+	int perf_ctl_scaling = hybrid_cpu ? cpu_khz / perf_ctl_max_phys :
+					    pstate_funcs.get_scaling();
+
 	cpu->pstate.min_pstate = pstate_funcs.get_min();
-	cpu->pstate.max_pstate_physical = pstate_funcs.get_max_physical();
-	cpu->pstate.scaling = pstate_funcs.get_scaling();
+	cpu->pstate.max_pstate_physical = perf_ctl_max_phys;
+	cpu->pstate.perf_ctl_scaling = perf_ctl_scaling;
 
 	if (hwp_active && !hwp_mode_bdw) {
 		__intel_pstate_get_hwp_cap(cpu);
+
+		if (hybrid_cpu)
+			intel_pstate_hybrid_hwp_calibrate(cpu);
+		else
+			cpu->pstate.scaling = perf_ctl_scaling;
 	} else {
+		cpu->pstate.scaling = perf_ctl_scaling;
 		cpu->pstate.max_pstate = pstate_funcs.get_max();
 		cpu->pstate.turbo_pstate = pstate_funcs.get_turbo();
 	}
 
-	cpu->pstate.max_freq = cpu->pstate.max_pstate * cpu->pstate.scaling;
-	cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
+	if (cpu->pstate.scaling == perf_ctl_scaling) {
+		cpu->pstate.min_freq = cpu->pstate.min_pstate * perf_ctl_scaling;
+		cpu->pstate.max_freq = cpu->pstate.max_pstate * perf_ctl_scaling;
+		cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * perf_ctl_scaling;
+	}
 
 	if (pstate_funcs.get_aperf_mperf_shift)
 		cpu->aperf_mperf_shift = pstate_funcs.get_aperf_mperf_shift();
@@ -2206,23 +2380,34 @@  static void intel_pstate_update_perf_lim
 					    unsigned int policy_min,
 					    unsigned int policy_max)
 {
-	int scaling = cpu->pstate.scaling;
+	int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
 	int32_t max_policy_perf, min_policy_perf;
 
+	max_policy_perf = policy_max / perf_ctl_scaling;
+	if (policy_max == policy_min) {
+		min_policy_perf = max_policy_perf;
+	} else {
+		min_policy_perf = policy_min / perf_ctl_scaling;
+		min_policy_perf = clamp_t(int32_t, min_policy_perf,
+					  0, max_policy_perf);
+	}
+
 	/*
 	 * HWP needs some special consideration, because HWP_REQUEST uses
 	 * abstract values to represent performance rather than pure ratios.
 	 */
-	if (hwp_active)
+	if (hwp_active) {
 		intel_pstate_get_hwp_cap(cpu);
 
-	max_policy_perf = policy_max / scaling;
-	if (policy_max == policy_min) {
-		min_policy_perf = max_policy_perf;
-	} else {
-		min_policy_perf = policy_min / scaling;
-		min_policy_perf = clamp_t(int32_t, min_policy_perf,
-					  0, max_policy_perf);
+		if (cpu->pstate.scaling != perf_ctl_scaling) {
+			int scaling = cpu->pstate.scaling;
+			int freq;
+
+			freq = max_policy_perf * perf_ctl_scaling;
+			max_policy_perf = DIV_ROUND_UP(freq, scaling);
+			freq = min_policy_perf * perf_ctl_scaling;
+			min_policy_perf = DIV_ROUND_UP(freq, scaling);
+		}
 	}
 
 	pr_debug("cpu:%d min_policy_perf:%d max_policy_perf:%d\n",
@@ -2416,7 +2601,7 @@  static int __intel_pstate_cpu_init(struc
 	cpu->min_perf_ratio = 0;
 
 	/* cpuinfo and default policy values */
-	policy->cpuinfo.min_freq = cpu->pstate.min_pstate * cpu->pstate.scaling;
+	policy->cpuinfo.min_freq = cpu->pstate.min_freq;
 	update_turbo_state();
 	global.turbo_disabled_mf = global.turbo_disabled;
 	policy->cpuinfo.max_freq = global.turbo_disabled ?
@@ -3146,6 +3331,8 @@  hwp_cpu_matched:
 		}
 
 		pr_info("HWP enabled\n");
+	} else if (boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
+		pr_warn("Problematic setup: Hybrid processor with disabled HWP\n");
 	}
 
 	return 0;