cpufreq: intel_pstate: Enforce _PPC limits
authorSrinivas Pandruvada <srinivas.pandruvada@linux.intel.com>
Wed, 27 Apr 2016 22:48:06 +0000 (15:48 -0700)
committerRafael J. Wysocki <rafael.j.wysocki@intel.com>
Wed, 27 Apr 2016 23:01:39 +0000 (01:01 +0200)
Use ACPI _PPC notification to limit max P state driver will request.
ACPI _PPC change notification is sent by BIOS to limit max P state
in several cases:
- Reduce impact of platform thermal condition
- When Config TDP feature is used, a changed _PPC is sent to
follow TDP change
- Remote node managers in server want to control platform power
via baseboard management controller (BMC)

This change registers with ACPI processor performance lib so that
_PPC changes are notified to cpufreq core, which in turns will
result in call to .setpolicy() callback. Also the way _PSS
table identifies a turbo frequency is not compatible to max turbo
frequency in intel_pstate, so the very first entry in _PSS needs
to be adjusted.

This feature can be turned on by using kernel parameters:
intel_pstate=support_acpi_ppc

Signed-off-by: Srinivas Pandruvada <srinivas.pandruvada@linux.intel.com>
[ rjw: Minor cleanups ]
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Documentation/kernel-parameters.txt
drivers/cpufreq/Kconfig.x86
drivers/cpufreq/intel_pstate.c

index ecc74fa4bfde8a55d9fee00d96ae5ae4123560e9..31c094fb1141e7048bddf7ff498af0a66b1903ca 100644 (file)
@@ -1661,6 +1661,8 @@ bytes respectively. Such letter suffixes can also be entirely omitted.
                hwp_only
                        Only load intel_pstate on systems which support
                        hardware P state control (HWP) if available.
+               support_acpi_ppc
+                       Enforce ACPI _PPC performance limits.
 
        intremap=       [X86-64, Intel-IOMMU]
                        on      enable Interrupt Remapping (default)
index c59bdcb83217071087cc0a90ef62e5cd01ce38d7..adbd1de1cea55cbe6777d3a1c4dab546a6a68adf 100644 (file)
@@ -5,6 +5,7 @@
 config X86_INTEL_PSTATE
        bool "Intel P state control"
        depends on X86
+       select ACPI_PROCESSOR if ACPI
        help
           This driver provides a P state for Intel core processors.
          The driver implements an internal governor and will become
index cfa6a6803e0e3e0161bf37b7260cfedd4bf1d3d4..c72a82a45872c43891d88bd47fa77a1e7e737244 100644 (file)
 #define ATOM_TURBO_RATIOS      0x66c
 #define ATOM_TURBO_VIDS                0x66d
 
+#ifdef CONFIG_ACPI
+#include <acpi/processor.h>
+#endif
+
 #define FRAC_BITS 8
 #define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
 #define fp_toint(X) ((X) >> FRAC_BITS)
@@ -174,6 +178,8 @@ struct _pid {
  * @prev_cummulative_iowait: IO Wait time difference from last and
  *                     current sample
  * @sample:            Storage for storing last Sample data
+ * @acpi_perf_data:    Stores ACPI perf information read from _PSS
+ * @valid_pss_table:   Set to true for valid ACPI _PSS entries found
  *
  * This structure stores per CPU instance data for all CPUs.
  */
@@ -192,6 +198,10 @@ struct cpudata {
        u64     prev_tsc;
        u64     prev_cummulative_iowait;
        struct sample sample;
+#ifdef CONFIG_ACPI
+       struct acpi_processor_performance acpi_perf_data;
+       bool valid_pss_table;
+#endif
 };
 
 static struct cpudata **all_cpu_data;
@@ -260,6 +270,9 @@ static struct pstate_adjust_policy pid_params;
 static struct pstate_funcs pstate_funcs;
 static int hwp_active;
 
+#ifdef CONFIG_ACPI
+static bool acpi_ppc;
+#endif
 
 /**
  * struct perf_limits - Store user and policy limits
@@ -333,6 +346,111 @@ static struct perf_limits *limits = &performance_limits;
 static struct perf_limits *limits = &powersave_limits;
 #endif
 
+#ifdef CONFIG_ACPI
+/*
+ * The max target pstate ratio is a 8 bit value in both PLATFORM_INFO MSR and
+ * in TURBO_RATIO_LIMIT MSR, which pstate driver stores in max_pstate and
+ * max_turbo_pstate fields. The PERF_CTL MSR contains 16 bit value for P state
+ * ratio, out of it only high 8 bits are used. For example 0x1700 is setting
+ * target ratio 0x17. The _PSS control value stores in a format which can be
+ * directly written to PERF_CTL MSR. But in intel_pstate driver this shift
+ * occurs during write to PERF_CTL (E.g. for cores core_set_pstate()).
+ * This function converts the _PSS control value to intel pstate driver format
+ * for comparison and assignment.
+ */
+static int convert_to_native_pstate_format(struct cpudata *cpu, int index)
+{
+       return cpu->acpi_perf_data.states[index].control >> 8;
+}
+
+static void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
+{
+       struct cpudata *cpu;
+       int turbo_pss_ctl;
+       int ret;
+       int i;
+
+       if (!acpi_ppc)
+               return;
+
+       cpu = all_cpu_data[policy->cpu];
+
+       ret = acpi_processor_register_performance(&cpu->acpi_perf_data,
+                                                 policy->cpu);
+       if (ret)
+               return;
+
+       /*
+        * Check if the control value in _PSS is for PERF_CTL MSR, which should
+        * guarantee that the states returned by it map to the states in our
+        * list directly.
+        */
+       if (cpu->acpi_perf_data.control_register.space_id !=
+                                               ACPI_ADR_SPACE_FIXED_HARDWARE)
+               goto err;
+
+       /*
+        * If there is only one entry _PSS, simply ignore _PSS and continue as
+        * usual without taking _PSS into account
+        */
+       if (cpu->acpi_perf_data.state_count < 2)
+               goto err;
+
+       pr_debug("CPU%u - ACPI _PSS perf data\n", policy->cpu);
+       for (i = 0; i < cpu->acpi_perf_data.state_count; i++) {
+               pr_debug("     %cP%d: %u MHz, %u mW, 0x%x\n",
+                        (i == cpu->acpi_perf_data.state ? '*' : ' '), i,
+                        (u32) cpu->acpi_perf_data.states[i].core_frequency,
+                        (u32) cpu->acpi_perf_data.states[i].power,
+                        (u32) cpu->acpi_perf_data.states[i].control);
+       }
+
+       /*
+        * The _PSS table doesn't contain whole turbo frequency range.
+        * This just contains +1 MHZ above the max non turbo frequency,
+        * with control value corresponding to max turbo ratio. But
+        * when cpufreq set policy is called, it will call with this
+        * max frequency, which will cause a reduced performance as
+        * this driver uses real max turbo frequency as the max
+        * frequency. So correct this frequency in _PSS table to
+        * correct max turbo frequency based on the turbo ratio.
+        * Also need to convert to MHz as _PSS freq is in MHz.
+        */
+       turbo_pss_ctl = convert_to_native_pstate_format(cpu, 0);
+       if (turbo_pss_ctl > cpu->pstate.max_pstate)
+               cpu->acpi_perf_data.states[0].core_frequency =
+                                       policy->cpuinfo.max_freq / 1000;
+       cpu->valid_pss_table = true;
+       pr_info("_PPC limits will be enforced\n");
+
+       return;
+
+ err:
+       cpu->valid_pss_table = false;
+       acpi_processor_unregister_performance(policy->cpu);
+}
+
+static void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
+{
+       struct cpudata *cpu;
+
+       cpu = all_cpu_data[policy->cpu];
+       if (!cpu->valid_pss_table)
+               return;
+
+       acpi_processor_unregister_performance(policy->cpu);
+}
+
+#else
+static void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
+{
+}
+
+static void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
+{
+}
+#endif
+
 static inline void pid_reset(struct _pid *pid, int setpoint, int busy,
                             int deadband, int integral) {
        pid->setpoint = int_tofp(setpoint);
@@ -1398,18 +1516,27 @@ static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
        policy->cpuinfo.min_freq = cpu->pstate.min_pstate * cpu->pstate.scaling;
        policy->cpuinfo.max_freq =
                cpu->pstate.turbo_pstate * cpu->pstate.scaling;
+       intel_pstate_init_acpi_perf_limits(policy);
        policy->cpuinfo.transition_latency = CPUFREQ_ETERNAL;
        cpumask_set_cpu(policy->cpu, policy->cpus);
 
        return 0;
 }
 
+static int intel_pstate_cpu_exit(struct cpufreq_policy *policy)
+{
+       intel_pstate_exit_perf_limits(policy);
+
+       return 0;
+}
+
 static struct cpufreq_driver intel_pstate_driver = {
        .flags          = CPUFREQ_CONST_LOOPS,
        .verify         = intel_pstate_verify_policy,
        .setpolicy      = intel_pstate_set_policy,
        .get            = intel_pstate_get,
        .init           = intel_pstate_cpu_init,
+       .exit           = intel_pstate_cpu_exit,
        .stop_cpu       = intel_pstate_stop_cpu,
        .name           = "intel_pstate",
 };
@@ -1453,8 +1580,7 @@ static void copy_cpu_funcs(struct pstate_funcs *funcs)
 
 }
 
-#if IS_ENABLED(CONFIG_ACPI)
-#include <acpi/processor.h>
+#ifdef CONFIG_ACPI
 
 static bool intel_pstate_no_acpi_pss(void)
 {
@@ -1660,6 +1786,12 @@ static int __init intel_pstate_setup(char *str)
                force_load = 1;
        if (!strcmp(str, "hwp_only"))
                hwp_only = 1;
+
+#ifdef CONFIG_ACPI
+       if (!strcmp(str, "support_acpi_ppc"))
+               acpi_ppc = true;
+#endif
+
        return 0;
 }
 early_param("intel_pstate", intel_pstate_setup);