2 * intel_pstate.c: Native P state management for Intel processors
4 * (C) Copyright 2012 Intel Corporation
5 * Author: Dirk Brandewie <dirk.j.brandewie@intel.com>
7 * This program is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU General Public License
9 * as published by the Free Software Foundation; version 2
13 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
15 #include <linux/kernel.h>
16 #include <linux/kernel_stat.h>
17 #include <linux/module.h>
18 #include <linux/ktime.h>
19 #include <linux/hrtimer.h>
20 #include <linux/tick.h>
21 #include <linux/slab.h>
22 #include <linux/sched/cpufreq.h>
23 #include <linux/list.h>
24 #include <linux/cpu.h>
25 #include <linux/cpufreq.h>
26 #include <linux/sysfs.h>
27 #include <linux/types.h>
29 #include <linux/acpi.h>
30 #include <linux/vmalloc.h>
31 #include <trace/events/power.h>
33 #include <asm/div64.h>
35 #include <asm/cpu_device_id.h>
36 #include <asm/cpufeature.h>
37 #include <asm/intel-family.h>
39 #define INTEL_PSTATE_SAMPLING_INTERVAL (10 * NSEC_PER_MSEC)
41 #define INTEL_CPUFREQ_TRANSITION_LATENCY 20000
42 #define INTEL_CPUFREQ_TRANSITION_DELAY 500
45 #include <acpi/processor.h>
46 #include <acpi/cppc_acpi.h>
50 #define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
51 #define fp_toint(X) ((X) >> FRAC_BITS)
54 #define EXT_FRAC_BITS (EXT_BITS + FRAC_BITS)
55 #define fp_ext_toint(X) ((X) >> EXT_FRAC_BITS)
56 #define int_ext_tofp(X) ((int64_t)(X) << EXT_FRAC_BITS)
58 static inline int32_t mul_fp(int32_t x, int32_t y)
60 return ((int64_t)x * (int64_t)y) >> FRAC_BITS;
63 static inline int32_t div_fp(s64 x, s64 y)
65 return div64_s64((int64_t)x << FRAC_BITS, y);
68 static inline int ceiling_fp(int32_t x)
73 mask = (1 << FRAC_BITS) - 1;
79 static inline int32_t percent_fp(int percent)
81 return div_fp(percent, 100);
84 static inline u64 mul_ext_fp(u64 x, u64 y)
86 return (x * y) >> EXT_FRAC_BITS;
89 static inline u64 div_ext_fp(u64 x, u64 y)
91 return div64_u64(x << EXT_FRAC_BITS, y);
94 static inline int32_t percent_ext_fp(int percent)
96 return div_ext_fp(percent, 100);
100 * struct sample - Store performance sample
101 * @core_avg_perf: Ratio of APERF/MPERF which is the actual average
102 * performance during last sample period
103 * @busy_scaled: Scaled busy value which is used to calculate next
104 * P state. This can be different than core_avg_perf
105 * to account for cpu idle period
106 * @aperf: Difference of actual performance frequency clock count
107 * read from APERF MSR between last and current sample
108 * @mperf: Difference of maximum performance frequency clock count
109 * read from MPERF MSR between last and current sample
110 * @tsc: Difference of time stamp counter between last and
112 * @time: Current time from scheduler
114 * This structure is used in the cpudata structure to store performance sample
115 * data for choosing next P State.
118 int32_t core_avg_perf;
127 * struct pstate_data - Store P state data
128 * @current_pstate: Current requested P state
129 * @min_pstate: Min P state possible for this platform
130 * @max_pstate: Max P state possible for this platform
131 * @max_pstate_physical:This is physical Max P state for a processor
132 * This can be higher than the max_pstate which can
133 * be limited by platform thermal design power limits
134 * @scaling: Scaling factor to convert frequency to cpufreq
136 * @turbo_pstate: Max Turbo P state possible for this platform
137 * @max_freq: @max_pstate frequency in cpufreq units
138 * @turbo_freq: @turbo_pstate frequency in cpufreq units
140 * Stores the per cpu model P state limits and current P state.
146 int max_pstate_physical;
149 unsigned int max_freq;
150 unsigned int turbo_freq;
154 * struct vid_data - Stores voltage information data
155 * @min: VID data for this platform corresponding to
157 * @max: VID data corresponding to the highest P State.
158 * @turbo: VID data for turbo P state
159 * @ratio: Ratio of (vid max - vid min) /
160 * (max P state - Min P State)
162 * Stores the voltage data for DVFS (Dynamic Voltage and Frequency Scaling)
163 * This data is used in Atom platforms, where in addition to target P state,
164 * the voltage data needs to be specified to select next P State.
174 * struct global_params - Global parameters, mostly tunable via sysfs.
175 * @no_turbo: Whether or not to use turbo P-states.
176 * @turbo_disabled: Whethet or not turbo P-states are available at all,
177 * based on the MSR_IA32_MISC_ENABLE value and whether or
178 * not the maximum reported turbo P-state is different from
179 * the maximum reported non-turbo one.
180 * @min_perf_pct: Minimum capacity limit in percent of the maximum turbo
182 * @max_perf_pct: Maximum capacity limit in percent of the maximum turbo
185 struct global_params {
193 * struct cpudata - Per CPU instance data storage
194 * @cpu: CPU number for this instance data
195 * @policy: CPUFreq policy value
196 * @update_util: CPUFreq utility callback information
197 * @update_util_set: CPUFreq utility callback is set
198 * @iowait_boost: iowait-related boost fraction
199 * @last_update: Time of the last update.
200 * @pstate: Stores P state limits for this CPU
201 * @vid: Stores VID limits for this CPU
202 * @last_sample_time: Last Sample time
203 * @aperf_mperf_shift: Number of clock cycles after aperf, merf is incremented
204 * This shift is a multiplier to mperf delta to
205 * calculate CPU busy.
206 * @prev_aperf: Last APERF value read from APERF MSR
207 * @prev_mperf: Last MPERF value read from MPERF MSR
208 * @prev_tsc: Last timestamp counter (TSC) value
209 * @prev_cummulative_iowait: IO Wait time difference from last and
211 * @sample: Storage for storing last Sample data
212 * @min_perf_ratio: Minimum capacity in terms of PERF or HWP ratios
213 * @max_perf_ratio: Maximum capacity in terms of PERF or HWP ratios
214 * @acpi_perf_data: Stores ACPI perf information read from _PSS
215 * @valid_pss_table: Set to true for valid ACPI _PSS entries found
216 * @epp_powersave: Last saved HWP energy performance preference
217 * (EPP) or energy performance bias (EPB),
218 * when policy switched to performance
219 * @epp_policy: Last saved policy used to set EPP/EPB
220 * @epp_default: Power on default HWP energy performance
222 * @epp_saved: Saved EPP/EPB during system suspend or CPU offline
224 * @hwp_req_cached: Cached value of the last HWP Request MSR
225 * @hwp_cap_cached: Cached value of the last HWP Capabilities MSR
226 * @last_io_update: Last time when IO wake flag was set
227 * @sched_flags: Store scheduler flags for possible cross CPU update
228 * @hwp_boost_min: Last HWP boosted min performance
230 * This structure stores per CPU instance data for all CPUs.
236 struct update_util_data update_util;
237 bool update_util_set;
239 struct pstate_data pstate;
243 u64 last_sample_time;
244 u64 aperf_mperf_shift;
248 u64 prev_cummulative_iowait;
249 struct sample sample;
250 int32_t min_perf_ratio;
251 int32_t max_perf_ratio;
253 struct acpi_processor_performance acpi_perf_data;
254 bool valid_pss_table;
256 unsigned int iowait_boost;
264 unsigned int sched_flags;
268 static struct cpudata **all_cpu_data;
271 * struct pstate_funcs - Per CPU model specific callbacks
272 * @get_max: Callback to get maximum non turbo effective P state
273 * @get_max_physical: Callback to get maximum non turbo physical P state
274 * @get_min: Callback to get minimum P state
275 * @get_turbo: Callback to get turbo P state
276 * @get_scaling: Callback to get frequency scaling factor
277 * @get_val: Callback to convert P state to actual MSR write value
278 * @get_vid: Callback to get VID data for Atom platforms
280 * Core and Atom CPU models have different way to get P State limits. This
281 * structure is used to store those callbacks.
283 struct pstate_funcs {
284 int (*get_max)(void);
285 int (*get_max_physical)(void);
286 int (*get_min)(void);
287 int (*get_turbo)(void);
288 int (*get_scaling)(void);
289 int (*get_aperf_mperf_shift)(void);
290 u64 (*get_val)(struct cpudata*, int pstate);
291 void (*get_vid)(struct cpudata *);
294 static struct pstate_funcs pstate_funcs __read_mostly;
296 static int hwp_active __read_mostly;
297 static bool per_cpu_limits __read_mostly;
298 static bool hwp_boost __read_mostly;
300 static struct cpufreq_driver *intel_pstate_driver __read_mostly;
303 static bool acpi_ppc;
306 static struct global_params global;
308 static DEFINE_MUTEX(intel_pstate_driver_lock);
309 static DEFINE_MUTEX(intel_pstate_limits_lock);
313 static bool intel_pstate_get_ppc_enable_status(void)
315 if (acpi_gbl_FADT.preferred_profile == PM_ENTERPRISE_SERVER ||
316 acpi_gbl_FADT.preferred_profile == PM_PERFORMANCE_SERVER)
322 #ifdef CONFIG_ACPI_CPPC_LIB
324 /* The work item is needed to avoid CPU hotplug locking issues */
325 static void intel_pstste_sched_itmt_work_fn(struct work_struct *work)
327 sched_set_itmt_support();
330 static DECLARE_WORK(sched_itmt_work, intel_pstste_sched_itmt_work_fn);
332 static void intel_pstate_set_itmt_prio(int cpu)
334 struct cppc_perf_caps cppc_perf;
335 static u32 max_highest_perf = 0, min_highest_perf = U32_MAX;
338 ret = cppc_get_perf_caps(cpu, &cppc_perf);
343 * The priorities can be set regardless of whether or not
344 * sched_set_itmt_support(true) has been called and it is valid to
345 * update them at any time after it has been called.
347 sched_set_itmt_core_prio(cppc_perf.highest_perf, cpu);
349 if (max_highest_perf <= min_highest_perf) {
350 if (cppc_perf.highest_perf > max_highest_perf)
351 max_highest_perf = cppc_perf.highest_perf;
353 if (cppc_perf.highest_perf < min_highest_perf)
354 min_highest_perf = cppc_perf.highest_perf;
356 if (max_highest_perf > min_highest_perf) {
358 * This code can be run during CPU online under the
359 * CPU hotplug locks, so sched_set_itmt_support()
360 * cannot be called from here. Queue up a work item
363 schedule_work(&sched_itmt_work);
368 static void intel_pstate_set_itmt_prio(int cpu)
373 static void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
380 intel_pstate_set_itmt_prio(policy->cpu);
384 if (!intel_pstate_get_ppc_enable_status())
387 cpu = all_cpu_data[policy->cpu];
389 ret = acpi_processor_register_performance(&cpu->acpi_perf_data,
395 * Check if the control value in _PSS is for PERF_CTL MSR, which should
396 * guarantee that the states returned by it map to the states in our
399 if (cpu->acpi_perf_data.control_register.space_id !=
400 ACPI_ADR_SPACE_FIXED_HARDWARE)
404 * If there is only one entry _PSS, simply ignore _PSS and continue as
405 * usual without taking _PSS into account
407 if (cpu->acpi_perf_data.state_count < 2)
410 pr_debug("CPU%u - ACPI _PSS perf data\n", policy->cpu);
411 for (i = 0; i < cpu->acpi_perf_data.state_count; i++) {
412 pr_debug(" %cP%d: %u MHz, %u mW, 0x%x\n",
413 (i == cpu->acpi_perf_data.state ? '*' : ' '), i,
414 (u32) cpu->acpi_perf_data.states[i].core_frequency,
415 (u32) cpu->acpi_perf_data.states[i].power,
416 (u32) cpu->acpi_perf_data.states[i].control);
420 * The _PSS table doesn't contain whole turbo frequency range.
421 * This just contains +1 MHZ above the max non turbo frequency,
422 * with control value corresponding to max turbo ratio. But
423 * when cpufreq set policy is called, it will call with this
424 * max frequency, which will cause a reduced performance as
425 * this driver uses real max turbo frequency as the max
426 * frequency. So correct this frequency in _PSS table to
427 * correct max turbo frequency based on the turbo state.
428 * Also need to convert to MHz as _PSS freq is in MHz.
430 if (!global.turbo_disabled)
431 cpu->acpi_perf_data.states[0].core_frequency =
432 policy->cpuinfo.max_freq / 1000;
433 cpu->valid_pss_table = true;
434 pr_debug("_PPC limits will be enforced\n");
439 cpu->valid_pss_table = false;
440 acpi_processor_unregister_performance(policy->cpu);
443 static void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
447 cpu = all_cpu_data[policy->cpu];
448 if (!cpu->valid_pss_table)
451 acpi_processor_unregister_performance(policy->cpu);
454 static inline void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
458 static inline void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
463 static inline void update_turbo_state(void)
468 cpu = all_cpu_data[0];
469 rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
470 global.turbo_disabled =
471 (misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ||
472 cpu->pstate.max_pstate == cpu->pstate.turbo_pstate);
475 static int min_perf_pct_min(void)
477 struct cpudata *cpu = all_cpu_data[0];
478 int turbo_pstate = cpu->pstate.turbo_pstate;
480 return turbo_pstate ?
481 (cpu->pstate.min_pstate * 100 / turbo_pstate) : 0;
484 static s16 intel_pstate_get_epb(struct cpudata *cpu_data)
489 if (!static_cpu_has(X86_FEATURE_EPB))
492 ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
496 return (s16)(epb & 0x0f);
499 static s16 intel_pstate_get_epp(struct cpudata *cpu_data, u64 hwp_req_data)
503 if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
505 * When hwp_req_data is 0, means that caller didn't read
506 * MSR_HWP_REQUEST, so need to read and get EPP.
509 epp = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST,
514 epp = (hwp_req_data >> 24) & 0xff;
516 /* When there is no EPP present, HWP uses EPB settings */
517 epp = intel_pstate_get_epb(cpu_data);
523 static int intel_pstate_set_epb(int cpu, s16 pref)
528 if (!static_cpu_has(X86_FEATURE_EPB))
531 ret = rdmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
535 epb = (epb & ~0x0f) | pref;
536 wrmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, epb);
542 * EPP/EPB display strings corresponding to EPP index in the
543 * energy_perf_strings[]
545 *-------------------------------------
548 * 2 balance_performance
552 static const char * const energy_perf_strings[] = {
555 "balance_performance",
560 static const unsigned int epp_values[] = {
562 HWP_EPP_BALANCE_PERFORMANCE,
563 HWP_EPP_BALANCE_POWERSAVE,
567 static int intel_pstate_get_energy_pref_index(struct cpudata *cpu_data)
572 epp = intel_pstate_get_epp(cpu_data, 0);
576 if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
577 if (epp == HWP_EPP_PERFORMANCE)
579 if (epp <= HWP_EPP_BALANCE_PERFORMANCE)
581 if (epp <= HWP_EPP_BALANCE_POWERSAVE)
585 } else if (static_cpu_has(X86_FEATURE_EPB)) {
588 * 0x00-0x03 : Performance
589 * 0x04-0x07 : Balance performance
590 * 0x08-0x0B : Balance power
592 * The EPB is a 4 bit value, but our ranges restrict the
593 * value which can be set. Here only using top two bits
596 index = (epp >> 2) + 1;
602 static int intel_pstate_set_energy_pref_index(struct cpudata *cpu_data,
609 epp = cpu_data->epp_default;
611 mutex_lock(&intel_pstate_limits_lock);
613 if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
616 ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST, &value);
620 value &= ~GENMASK_ULL(31, 24);
623 epp = epp_values[pref_index - 1];
625 value |= (u64)epp << 24;
626 ret = wrmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST, value);
629 epp = (pref_index - 1) << 2;
630 ret = intel_pstate_set_epb(cpu_data->cpu, epp);
633 mutex_unlock(&intel_pstate_limits_lock);
638 static ssize_t show_energy_performance_available_preferences(
639 struct cpufreq_policy *policy, char *buf)
644 while (energy_perf_strings[i] != NULL)
645 ret += sprintf(&buf[ret], "%s ", energy_perf_strings[i++]);
647 ret += sprintf(&buf[ret], "\n");
652 cpufreq_freq_attr_ro(energy_performance_available_preferences);
654 static ssize_t store_energy_performance_preference(
655 struct cpufreq_policy *policy, const char *buf, size_t count)
657 struct cpudata *cpu_data = all_cpu_data[policy->cpu];
658 char str_preference[21];
661 ret = sscanf(buf, "%20s", str_preference);
665 while (energy_perf_strings[i] != NULL) {
666 if (!strcmp(str_preference, energy_perf_strings[i])) {
667 intel_pstate_set_energy_pref_index(cpu_data, i);
676 static ssize_t show_energy_performance_preference(
677 struct cpufreq_policy *policy, char *buf)
679 struct cpudata *cpu_data = all_cpu_data[policy->cpu];
682 preference = intel_pstate_get_energy_pref_index(cpu_data);
686 return sprintf(buf, "%s\n", energy_perf_strings[preference]);
689 cpufreq_freq_attr_rw(energy_performance_preference);
691 static struct freq_attr *hwp_cpufreq_attrs[] = {
692 &energy_performance_preference,
693 &energy_performance_available_preferences,
697 static void intel_pstate_get_hwp_max(unsigned int cpu, int *phy_max,
702 rdmsrl_on_cpu(cpu, MSR_HWP_CAPABILITIES, &cap);
703 WRITE_ONCE(all_cpu_data[cpu]->hwp_cap_cached, cap);
705 *current_max = HWP_GUARANTEED_PERF(cap);
707 *current_max = HWP_HIGHEST_PERF(cap);
709 *phy_max = HWP_HIGHEST_PERF(cap);
712 static void intel_pstate_hwp_set(unsigned int cpu)
714 struct cpudata *cpu_data = all_cpu_data[cpu];
719 max = cpu_data->max_perf_ratio;
720 min = cpu_data->min_perf_ratio;
722 if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE)
725 rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
727 value &= ~HWP_MIN_PERF(~0L);
728 value |= HWP_MIN_PERF(min);
730 value &= ~HWP_MAX_PERF(~0L);
731 value |= HWP_MAX_PERF(max);
733 if (cpu_data->epp_policy == cpu_data->policy)
736 cpu_data->epp_policy = cpu_data->policy;
738 if (cpu_data->epp_saved >= 0) {
739 epp = cpu_data->epp_saved;
740 cpu_data->epp_saved = -EINVAL;
744 if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE) {
745 epp = intel_pstate_get_epp(cpu_data, value);
746 cpu_data->epp_powersave = epp;
747 /* If EPP read was failed, then don't try to write */
753 /* skip setting EPP, when saved value is invalid */
754 if (cpu_data->epp_powersave < 0)
758 * No need to restore EPP when it is not zero. This
760 * - Policy is not changed
761 * - user has manually changed
762 * - Error reading EPB
764 epp = intel_pstate_get_epp(cpu_data, value);
768 epp = cpu_data->epp_powersave;
771 if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
772 value &= ~GENMASK_ULL(31, 24);
773 value |= (u64)epp << 24;
775 intel_pstate_set_epb(cpu, epp);
778 WRITE_ONCE(cpu_data->hwp_req_cached, value);
779 wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
782 static int intel_pstate_hwp_save_state(struct cpufreq_policy *policy)
784 struct cpudata *cpu_data = all_cpu_data[policy->cpu];
789 cpu_data->epp_saved = intel_pstate_get_epp(cpu_data, 0);
794 static void intel_pstate_hwp_enable(struct cpudata *cpudata);
796 static int intel_pstate_resume(struct cpufreq_policy *policy)
801 mutex_lock(&intel_pstate_limits_lock);
803 if (policy->cpu == 0)
804 intel_pstate_hwp_enable(all_cpu_data[policy->cpu]);
806 all_cpu_data[policy->cpu]->epp_policy = 0;
807 intel_pstate_hwp_set(policy->cpu);
809 mutex_unlock(&intel_pstate_limits_lock);
814 static void intel_pstate_update_policies(void)
818 for_each_possible_cpu(cpu)
819 cpufreq_update_policy(cpu);
822 /************************** sysfs begin ************************/
823 #define show_one(file_name, object) \
824 static ssize_t show_##file_name \
825 (struct kobject *kobj, struct attribute *attr, char *buf) \
827 return sprintf(buf, "%u\n", global.object); \
830 static ssize_t intel_pstate_show_status(char *buf);
831 static int intel_pstate_update_status(const char *buf, size_t size);
833 static ssize_t show_status(struct kobject *kobj,
834 struct attribute *attr, char *buf)
838 mutex_lock(&intel_pstate_driver_lock);
839 ret = intel_pstate_show_status(buf);
840 mutex_unlock(&intel_pstate_driver_lock);
845 static ssize_t store_status(struct kobject *a, struct attribute *b,
846 const char *buf, size_t count)
848 char *p = memchr(buf, '\n', count);
851 mutex_lock(&intel_pstate_driver_lock);
852 ret = intel_pstate_update_status(buf, p ? p - buf : count);
853 mutex_unlock(&intel_pstate_driver_lock);
855 return ret < 0 ? ret : count;
858 static ssize_t show_turbo_pct(struct kobject *kobj,
859 struct attribute *attr, char *buf)
862 int total, no_turbo, turbo_pct;
865 mutex_lock(&intel_pstate_driver_lock);
867 if (!intel_pstate_driver) {
868 mutex_unlock(&intel_pstate_driver_lock);
872 cpu = all_cpu_data[0];
874 total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
875 no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
876 turbo_fp = div_fp(no_turbo, total);
877 turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
879 mutex_unlock(&intel_pstate_driver_lock);
881 return sprintf(buf, "%u\n", turbo_pct);
884 static ssize_t show_num_pstates(struct kobject *kobj,
885 struct attribute *attr, char *buf)
890 mutex_lock(&intel_pstate_driver_lock);
892 if (!intel_pstate_driver) {
893 mutex_unlock(&intel_pstate_driver_lock);
897 cpu = all_cpu_data[0];
898 total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
900 mutex_unlock(&intel_pstate_driver_lock);
902 return sprintf(buf, "%u\n", total);
905 static ssize_t show_no_turbo(struct kobject *kobj,
906 struct attribute *attr, char *buf)
910 mutex_lock(&intel_pstate_driver_lock);
912 if (!intel_pstate_driver) {
913 mutex_unlock(&intel_pstate_driver_lock);
917 update_turbo_state();
918 if (global.turbo_disabled)
919 ret = sprintf(buf, "%u\n", global.turbo_disabled);
921 ret = sprintf(buf, "%u\n", global.no_turbo);
923 mutex_unlock(&intel_pstate_driver_lock);
928 static ssize_t store_no_turbo(struct kobject *a, struct attribute *b,
929 const char *buf, size_t count)
934 ret = sscanf(buf, "%u", &input);
938 mutex_lock(&intel_pstate_driver_lock);
940 if (!intel_pstate_driver) {
941 mutex_unlock(&intel_pstate_driver_lock);
945 mutex_lock(&intel_pstate_limits_lock);
947 update_turbo_state();
948 if (global.turbo_disabled) {
949 pr_warn("Turbo disabled by BIOS or unavailable on processor\n");
950 mutex_unlock(&intel_pstate_limits_lock);
951 mutex_unlock(&intel_pstate_driver_lock);
955 global.no_turbo = clamp_t(int, input, 0, 1);
957 if (global.no_turbo) {
958 struct cpudata *cpu = all_cpu_data[0];
959 int pct = cpu->pstate.max_pstate * 100 / cpu->pstate.turbo_pstate;
961 /* Squash the global minimum into the permitted range. */
962 if (global.min_perf_pct > pct)
963 global.min_perf_pct = pct;
966 mutex_unlock(&intel_pstate_limits_lock);
968 intel_pstate_update_policies();
970 mutex_unlock(&intel_pstate_driver_lock);
975 static ssize_t store_max_perf_pct(struct kobject *a, struct attribute *b,
976 const char *buf, size_t count)
981 ret = sscanf(buf, "%u", &input);
985 mutex_lock(&intel_pstate_driver_lock);
987 if (!intel_pstate_driver) {
988 mutex_unlock(&intel_pstate_driver_lock);
992 mutex_lock(&intel_pstate_limits_lock);
994 global.max_perf_pct = clamp_t(int, input, global.min_perf_pct, 100);
996 mutex_unlock(&intel_pstate_limits_lock);
998 intel_pstate_update_policies();
1000 mutex_unlock(&intel_pstate_driver_lock);
1005 static ssize_t store_min_perf_pct(struct kobject *a, struct attribute *b,
1006 const char *buf, size_t count)
1011 ret = sscanf(buf, "%u", &input);
1015 mutex_lock(&intel_pstate_driver_lock);
1017 if (!intel_pstate_driver) {
1018 mutex_unlock(&intel_pstate_driver_lock);
1022 mutex_lock(&intel_pstate_limits_lock);
1024 global.min_perf_pct = clamp_t(int, input,
1025 min_perf_pct_min(), global.max_perf_pct);
1027 mutex_unlock(&intel_pstate_limits_lock);
1029 intel_pstate_update_policies();
1031 mutex_unlock(&intel_pstate_driver_lock);
1036 static ssize_t show_hwp_dynamic_boost(struct kobject *kobj,
1037 struct attribute *attr, char *buf)
1039 return sprintf(buf, "%u\n", hwp_boost);
1042 static ssize_t store_hwp_dynamic_boost(struct kobject *a, struct attribute *b,
1043 const char *buf, size_t count)
1048 ret = kstrtouint(buf, 10, &input);
1052 mutex_lock(&intel_pstate_driver_lock);
1053 hwp_boost = !!input;
1054 intel_pstate_update_policies();
1055 mutex_unlock(&intel_pstate_driver_lock);
1060 show_one(max_perf_pct, max_perf_pct);
1061 show_one(min_perf_pct, min_perf_pct);
1063 define_one_global_rw(status);
1064 define_one_global_rw(no_turbo);
1065 define_one_global_rw(max_perf_pct);
1066 define_one_global_rw(min_perf_pct);
1067 define_one_global_ro(turbo_pct);
1068 define_one_global_ro(num_pstates);
1069 define_one_global_rw(hwp_dynamic_boost);
1071 static struct attribute *intel_pstate_attributes[] = {
1079 static const struct attribute_group intel_pstate_attr_group = {
1080 .attrs = intel_pstate_attributes,
1083 static void __init intel_pstate_sysfs_expose_params(void)
1085 struct kobject *intel_pstate_kobject;
1088 intel_pstate_kobject = kobject_create_and_add("intel_pstate",
1089 &cpu_subsys.dev_root->kobj);
1090 if (WARN_ON(!intel_pstate_kobject))
1093 rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
1098 * If per cpu limits are enforced there are no global limits, so
1099 * return without creating max/min_perf_pct attributes
1104 rc = sysfs_create_file(intel_pstate_kobject, &max_perf_pct.attr);
1107 rc = sysfs_create_file(intel_pstate_kobject, &min_perf_pct.attr);
1111 rc = sysfs_create_file(intel_pstate_kobject,
1112 &hwp_dynamic_boost.attr);
1116 /************************** sysfs end ************************/
1118 static void intel_pstate_hwp_enable(struct cpudata *cpudata)
1120 /* First disable HWP notification interrupt as we don't process them */
1121 if (static_cpu_has(X86_FEATURE_HWP_NOTIFY))
1122 wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1124 wrmsrl_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1);
1125 cpudata->epp_policy = 0;
1126 if (cpudata->epp_default == -EINVAL)
1127 cpudata->epp_default = intel_pstate_get_epp(cpudata, 0);
1130 #define MSR_IA32_POWER_CTL_BIT_EE 19
1132 /* Disable energy efficiency optimization */
1133 static void intel_pstate_disable_ee(int cpu)
1138 ret = rdmsrl_on_cpu(cpu, MSR_IA32_POWER_CTL, &power_ctl);
1142 if (!(power_ctl & BIT(MSR_IA32_POWER_CTL_BIT_EE))) {
1143 pr_info("Disabling energy efficiency optimization\n");
1144 power_ctl |= BIT(MSR_IA32_POWER_CTL_BIT_EE);
1145 wrmsrl_on_cpu(cpu, MSR_IA32_POWER_CTL, power_ctl);
1149 static int atom_get_min_pstate(void)
1153 rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1154 return (value >> 8) & 0x7F;
1157 static int atom_get_max_pstate(void)
1161 rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1162 return (value >> 16) & 0x7F;
1165 static int atom_get_turbo_pstate(void)
1169 rdmsrl(MSR_ATOM_CORE_TURBO_RATIOS, value);
1170 return value & 0x7F;
1173 static u64 atom_get_val(struct cpudata *cpudata, int pstate)
1179 val = (u64)pstate << 8;
1180 if (global.no_turbo && !global.turbo_disabled)
1181 val |= (u64)1 << 32;
1183 vid_fp = cpudata->vid.min + mul_fp(
1184 int_tofp(pstate - cpudata->pstate.min_pstate),
1185 cpudata->vid.ratio);
1187 vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
1188 vid = ceiling_fp(vid_fp);
1190 if (pstate > cpudata->pstate.max_pstate)
1191 vid = cpudata->vid.turbo;
1196 static int silvermont_get_scaling(void)
1200 /* Defined in Table 35-6 from SDM (Sept 2015) */
1201 static int silvermont_freq_table[] = {
1202 83300, 100000, 133300, 116700, 80000};
1204 rdmsrl(MSR_FSB_FREQ, value);
1208 return silvermont_freq_table[i];
1211 static int airmont_get_scaling(void)
1215 /* Defined in Table 35-10 from SDM (Sept 2015) */
1216 static int airmont_freq_table[] = {
1217 83300, 100000, 133300, 116700, 80000,
1218 93300, 90000, 88900, 87500};
1220 rdmsrl(MSR_FSB_FREQ, value);
1224 return airmont_freq_table[i];
1227 static void atom_get_vid(struct cpudata *cpudata)
1231 rdmsrl(MSR_ATOM_CORE_VIDS, value);
1232 cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
1233 cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
1234 cpudata->vid.ratio = div_fp(
1235 cpudata->vid.max - cpudata->vid.min,
1236 int_tofp(cpudata->pstate.max_pstate -
1237 cpudata->pstate.min_pstate));
1239 rdmsrl(MSR_ATOM_CORE_TURBO_VIDS, value);
1240 cpudata->vid.turbo = value & 0x7f;
1243 static int core_get_min_pstate(void)
1247 rdmsrl(MSR_PLATFORM_INFO, value);
1248 return (value >> 40) & 0xFF;
1251 static int core_get_max_pstate_physical(void)
1255 rdmsrl(MSR_PLATFORM_INFO, value);
1256 return (value >> 8) & 0xFF;
1259 static int core_get_tdp_ratio(u64 plat_info)
1261 /* Check how many TDP levels present */
1262 if (plat_info & 0x600000000) {
1268 /* Get the TDP level (0, 1, 2) to get ratios */
1269 err = rdmsrl_safe(MSR_CONFIG_TDP_CONTROL, &tdp_ctrl);
1273 /* TDP MSR are continuous starting at 0x648 */
1274 tdp_msr = MSR_CONFIG_TDP_NOMINAL + (tdp_ctrl & 0x03);
1275 err = rdmsrl_safe(tdp_msr, &tdp_ratio);
1279 /* For level 1 and 2, bits[23:16] contain the ratio */
1280 if (tdp_ctrl & 0x03)
1283 tdp_ratio &= 0xff; /* ratios are only 8 bits long */
1284 pr_debug("tdp_ratio %x\n", (int)tdp_ratio);
1286 return (int)tdp_ratio;
1292 static int core_get_max_pstate(void)
1300 rdmsrl(MSR_PLATFORM_INFO, plat_info);
1301 max_pstate = (plat_info >> 8) & 0xFF;
1303 tdp_ratio = core_get_tdp_ratio(plat_info);
1308 /* Turbo activation ratio is not used on HWP platforms */
1312 err = rdmsrl_safe(MSR_TURBO_ACTIVATION_RATIO, &tar);
1316 /* Do some sanity checking for safety */
1317 tar_levels = tar & 0xff;
1318 if (tdp_ratio - 1 == tar_levels) {
1319 max_pstate = tar_levels;
1320 pr_debug("max_pstate=TAC %x\n", max_pstate);
1327 static int core_get_turbo_pstate(void)
1332 rdmsrl(MSR_TURBO_RATIO_LIMIT, value);
1333 nont = core_get_max_pstate();
1334 ret = (value) & 255;
1340 static inline int core_get_scaling(void)
1345 static u64 core_get_val(struct cpudata *cpudata, int pstate)
1349 val = (u64)pstate << 8;
1350 if (global.no_turbo && !global.turbo_disabled)
1351 val |= (u64)1 << 32;
1356 static int knl_get_aperf_mperf_shift(void)
1361 static int knl_get_turbo_pstate(void)
1366 rdmsrl(MSR_TURBO_RATIO_LIMIT, value);
1367 nont = core_get_max_pstate();
1368 ret = (((value) >> 8) & 0xFF);
1374 static int intel_pstate_get_base_pstate(struct cpudata *cpu)
1376 return global.no_turbo || global.turbo_disabled ?
1377 cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
1380 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
1382 trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
1383 cpu->pstate.current_pstate = pstate;
1385 * Generally, there is no guarantee that this code will always run on
1386 * the CPU being updated, so force the register update to run on the
1389 wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
1390 pstate_funcs.get_val(cpu, pstate));
1393 static void intel_pstate_set_min_pstate(struct cpudata *cpu)
1395 intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
1398 static void intel_pstate_max_within_limits(struct cpudata *cpu)
1402 update_turbo_state();
1403 pstate = intel_pstate_get_base_pstate(cpu);
1404 pstate = max(cpu->pstate.min_pstate, cpu->max_perf_ratio);
1405 intel_pstate_set_pstate(cpu, pstate);
1408 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
1410 cpu->pstate.min_pstate = pstate_funcs.get_min();
1411 cpu->pstate.max_pstate = pstate_funcs.get_max();
1412 cpu->pstate.max_pstate_physical = pstate_funcs.get_max_physical();
1413 cpu->pstate.turbo_pstate = pstate_funcs.get_turbo();
1414 cpu->pstate.scaling = pstate_funcs.get_scaling();
1415 cpu->pstate.max_freq = cpu->pstate.max_pstate * cpu->pstate.scaling;
1416 cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
1418 if (pstate_funcs.get_aperf_mperf_shift)
1419 cpu->aperf_mperf_shift = pstate_funcs.get_aperf_mperf_shift();
1421 if (pstate_funcs.get_vid)
1422 pstate_funcs.get_vid(cpu);
1424 intel_pstate_set_min_pstate(cpu);
1428 * Long hold time will keep high perf limits for long time,
1429 * which negatively impacts perf/watt for some workloads,
1430 * like specpower. 3ms is based on experiements on some
1433 static int hwp_boost_hold_time_ns = 3 * NSEC_PER_MSEC;
1435 static inline void intel_pstate_hwp_boost_up(struct cpudata *cpu)
1437 u64 hwp_req = READ_ONCE(cpu->hwp_req_cached);
1438 u32 max_limit = (hwp_req & 0xff00) >> 8;
1439 u32 min_limit = (hwp_req & 0xff);
1443 * Cases to consider (User changes via sysfs or boot time):
1444 * If, P0 (Turbo max) = P1 (Guaranteed max) = min:
1446 * If, P0 (Turbo max) > P1 (Guaranteed max) = min:
1447 * Should result in one level boost only for P0.
1448 * If, P0 (Turbo max) = P1 (Guaranteed max) > min:
1449 * Should result in two level boost:
1450 * (min + p1)/2 and P1.
1451 * If, P0 (Turbo max) > P1 (Guaranteed max) > min:
1452 * Should result in three level boost:
1453 * (min + p1)/2, P1 and P0.
1456 /* If max and min are equal or already at max, nothing to boost */
1457 if (max_limit == min_limit || cpu->hwp_boost_min >= max_limit)
1460 if (!cpu->hwp_boost_min)
1461 cpu->hwp_boost_min = min_limit;
1463 /* level at half way mark between min and guranteed */
1464 boost_level1 = (HWP_GUARANTEED_PERF(cpu->hwp_cap_cached) + min_limit) >> 1;
1466 if (cpu->hwp_boost_min < boost_level1)
1467 cpu->hwp_boost_min = boost_level1;
1468 else if (cpu->hwp_boost_min < HWP_GUARANTEED_PERF(cpu->hwp_cap_cached))
1469 cpu->hwp_boost_min = HWP_GUARANTEED_PERF(cpu->hwp_cap_cached);
1470 else if (cpu->hwp_boost_min == HWP_GUARANTEED_PERF(cpu->hwp_cap_cached) &&
1471 max_limit != HWP_GUARANTEED_PERF(cpu->hwp_cap_cached))
1472 cpu->hwp_boost_min = max_limit;
1476 hwp_req = (hwp_req & ~GENMASK_ULL(7, 0)) | cpu->hwp_boost_min;
1477 wrmsrl(MSR_HWP_REQUEST, hwp_req);
1478 cpu->last_update = cpu->sample.time;
1481 static inline void intel_pstate_hwp_boost_down(struct cpudata *cpu)
1483 if (cpu->hwp_boost_min) {
1486 /* Check if we are idle for hold time to boost down */
1487 expired = time_after64(cpu->sample.time, cpu->last_update +
1488 hwp_boost_hold_time_ns);
1490 wrmsrl(MSR_HWP_REQUEST, cpu->hwp_req_cached);
1491 cpu->hwp_boost_min = 0;
1494 cpu->last_update = cpu->sample.time;
1497 static inline void intel_pstate_update_util_hwp_local(struct cpudata *cpu,
1500 cpu->sample.time = time;
1502 if (cpu->sched_flags & SCHED_CPUFREQ_IOWAIT) {
1505 cpu->sched_flags = 0;
1507 * Set iowait_boost flag and update time. Since IO WAIT flag
1508 * is set all the time, we can't just conclude that there is
1509 * some IO bound activity is scheduled on this CPU with just
1510 * one occurrence. If we receive at least two in two
1511 * consecutive ticks, then we treat as boost candidate.
1513 if (time_before64(time, cpu->last_io_update + 2 * TICK_NSEC))
1516 cpu->last_io_update = time;
1519 intel_pstate_hwp_boost_up(cpu);
1522 intel_pstate_hwp_boost_down(cpu);
1526 static inline void intel_pstate_update_util_hwp(struct update_util_data *data,
1527 u64 time, unsigned int flags)
1529 struct cpudata *cpu = container_of(data, struct cpudata, update_util);
1531 cpu->sched_flags |= flags;
1533 if (smp_processor_id() == cpu->cpu)
1534 intel_pstate_update_util_hwp_local(cpu, time);
1537 static inline void intel_pstate_calc_avg_perf(struct cpudata *cpu)
1539 struct sample *sample = &cpu->sample;
1541 sample->core_avg_perf = div_ext_fp(sample->aperf, sample->mperf);
1544 static inline bool intel_pstate_sample(struct cpudata *cpu, u64 time)
1547 unsigned long flags;
1550 local_irq_save(flags);
1551 rdmsrl(MSR_IA32_APERF, aperf);
1552 rdmsrl(MSR_IA32_MPERF, mperf);
1554 if (cpu->prev_mperf == mperf || cpu->prev_tsc == tsc) {
1555 local_irq_restore(flags);
1558 local_irq_restore(flags);
1560 cpu->last_sample_time = cpu->sample.time;
1561 cpu->sample.time = time;
1562 cpu->sample.aperf = aperf;
1563 cpu->sample.mperf = mperf;
1564 cpu->sample.tsc = tsc;
1565 cpu->sample.aperf -= cpu->prev_aperf;
1566 cpu->sample.mperf -= cpu->prev_mperf;
1567 cpu->sample.tsc -= cpu->prev_tsc;
1569 cpu->prev_aperf = aperf;
1570 cpu->prev_mperf = mperf;
1571 cpu->prev_tsc = tsc;
1573 * First time this function is invoked in a given cycle, all of the
1574 * previous sample data fields are equal to zero or stale and they must
1575 * be populated with meaningful numbers for things to work, so assume
1576 * that sample.time will always be reset before setting the utilization
1577 * update hook and make the caller skip the sample then.
1579 if (cpu->last_sample_time) {
1580 intel_pstate_calc_avg_perf(cpu);
1586 static inline int32_t get_avg_frequency(struct cpudata *cpu)
1588 return mul_ext_fp(cpu->sample.core_avg_perf, cpu_khz);
1591 static inline int32_t get_avg_pstate(struct cpudata *cpu)
1593 return mul_ext_fp(cpu->pstate.max_pstate_physical,
1594 cpu->sample.core_avg_perf);
1597 static inline int32_t get_target_pstate(struct cpudata *cpu)
1599 struct sample *sample = &cpu->sample;
1600 int32_t busy_frac, boost;
1601 int target, avg_pstate;
1603 busy_frac = div_fp(sample->mperf << cpu->aperf_mperf_shift,
1606 boost = cpu->iowait_boost;
1607 cpu->iowait_boost >>= 1;
1609 if (busy_frac < boost)
1612 sample->busy_scaled = busy_frac * 100;
1614 target = global.no_turbo || global.turbo_disabled ?
1615 cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
1616 target += target >> 2;
1617 target = mul_fp(target, busy_frac);
1618 if (target < cpu->pstate.min_pstate)
1619 target = cpu->pstate.min_pstate;
1622 * If the average P-state during the previous cycle was higher than the
1623 * current target, add 50% of the difference to the target to reduce
1624 * possible performance oscillations and offset possible performance
1625 * loss related to moving the workload from one CPU to another within
1628 avg_pstate = get_avg_pstate(cpu);
1629 if (avg_pstate > target)
1630 target += (avg_pstate - target) >> 1;
1635 static int intel_pstate_prepare_request(struct cpudata *cpu, int pstate)
1637 int max_pstate = intel_pstate_get_base_pstate(cpu);
1640 min_pstate = max(cpu->pstate.min_pstate, cpu->min_perf_ratio);
1641 max_pstate = max(min_pstate, cpu->max_perf_ratio);
1642 return clamp_t(int, pstate, min_pstate, max_pstate);
1645 static void intel_pstate_update_pstate(struct cpudata *cpu, int pstate)
1647 if (pstate == cpu->pstate.current_pstate)
1650 cpu->pstate.current_pstate = pstate;
1651 wrmsrl(MSR_IA32_PERF_CTL, pstate_funcs.get_val(cpu, pstate));
1654 static void intel_pstate_adjust_pstate(struct cpudata *cpu)
1656 int from = cpu->pstate.current_pstate;
1657 struct sample *sample;
1660 update_turbo_state();
1662 target_pstate = get_target_pstate(cpu);
1663 target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
1664 trace_cpu_frequency(target_pstate * cpu->pstate.scaling, cpu->cpu);
1665 intel_pstate_update_pstate(cpu, target_pstate);
1667 sample = &cpu->sample;
1668 trace_pstate_sample(mul_ext_fp(100, sample->core_avg_perf),
1669 fp_toint(sample->busy_scaled),
1671 cpu->pstate.current_pstate,
1675 get_avg_frequency(cpu),
1676 fp_toint(cpu->iowait_boost * 100));
1679 static void intel_pstate_update_util(struct update_util_data *data, u64 time,
1682 struct cpudata *cpu = container_of(data, struct cpudata, update_util);
1685 /* Don't allow remote callbacks */
1686 if (smp_processor_id() != cpu->cpu)
1689 if (flags & SCHED_CPUFREQ_IOWAIT) {
1690 cpu->iowait_boost = int_tofp(1);
1691 cpu->last_update = time;
1693 * The last time the busy was 100% so P-state was max anyway
1694 * so avoid overhead of computation.
1696 if (fp_toint(cpu->sample.busy_scaled) == 100)
1700 } else if (cpu->iowait_boost) {
1701 /* Clear iowait_boost if the CPU may have been idle. */
1702 delta_ns = time - cpu->last_update;
1703 if (delta_ns > TICK_NSEC)
1704 cpu->iowait_boost = 0;
1706 cpu->last_update = time;
1707 delta_ns = time - cpu->sample.time;
1708 if ((s64)delta_ns < INTEL_PSTATE_SAMPLING_INTERVAL)
1712 if (intel_pstate_sample(cpu, time))
1713 intel_pstate_adjust_pstate(cpu);
1716 static struct pstate_funcs core_funcs = {
1717 .get_max = core_get_max_pstate,
1718 .get_max_physical = core_get_max_pstate_physical,
1719 .get_min = core_get_min_pstate,
1720 .get_turbo = core_get_turbo_pstate,
1721 .get_scaling = core_get_scaling,
1722 .get_val = core_get_val,
1725 static const struct pstate_funcs silvermont_funcs = {
1726 .get_max = atom_get_max_pstate,
1727 .get_max_physical = atom_get_max_pstate,
1728 .get_min = atom_get_min_pstate,
1729 .get_turbo = atom_get_turbo_pstate,
1730 .get_val = atom_get_val,
1731 .get_scaling = silvermont_get_scaling,
1732 .get_vid = atom_get_vid,
1735 static const struct pstate_funcs airmont_funcs = {
1736 .get_max = atom_get_max_pstate,
1737 .get_max_physical = atom_get_max_pstate,
1738 .get_min = atom_get_min_pstate,
1739 .get_turbo = atom_get_turbo_pstate,
1740 .get_val = atom_get_val,
1741 .get_scaling = airmont_get_scaling,
1742 .get_vid = atom_get_vid,
1745 static const struct pstate_funcs knl_funcs = {
1746 .get_max = core_get_max_pstate,
1747 .get_max_physical = core_get_max_pstate_physical,
1748 .get_min = core_get_min_pstate,
1749 .get_turbo = knl_get_turbo_pstate,
1750 .get_aperf_mperf_shift = knl_get_aperf_mperf_shift,
1751 .get_scaling = core_get_scaling,
1752 .get_val = core_get_val,
1755 #define ICPU(model, policy) \
1756 { X86_VENDOR_INTEL, 6, model, X86_FEATURE_APERFMPERF,\
1757 (unsigned long)&policy }
1759 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
1760 ICPU(INTEL_FAM6_SANDYBRIDGE, core_funcs),
1761 ICPU(INTEL_FAM6_SANDYBRIDGE_X, core_funcs),
1762 ICPU(INTEL_FAM6_ATOM_SILVERMONT1, silvermont_funcs),
1763 ICPU(INTEL_FAM6_IVYBRIDGE, core_funcs),
1764 ICPU(INTEL_FAM6_HASWELL_CORE, core_funcs),
1765 ICPU(INTEL_FAM6_BROADWELL_CORE, core_funcs),
1766 ICPU(INTEL_FAM6_IVYBRIDGE_X, core_funcs),
1767 ICPU(INTEL_FAM6_HASWELL_X, core_funcs),
1768 ICPU(INTEL_FAM6_HASWELL_ULT, core_funcs),
1769 ICPU(INTEL_FAM6_HASWELL_GT3E, core_funcs),
1770 ICPU(INTEL_FAM6_BROADWELL_GT3E, core_funcs),
1771 ICPU(INTEL_FAM6_ATOM_AIRMONT, airmont_funcs),
1772 ICPU(INTEL_FAM6_SKYLAKE_MOBILE, core_funcs),
1773 ICPU(INTEL_FAM6_BROADWELL_X, core_funcs),
1774 ICPU(INTEL_FAM6_SKYLAKE_DESKTOP, core_funcs),
1775 ICPU(INTEL_FAM6_BROADWELL_XEON_D, core_funcs),
1776 ICPU(INTEL_FAM6_XEON_PHI_KNL, knl_funcs),
1777 ICPU(INTEL_FAM6_XEON_PHI_KNM, knl_funcs),
1778 ICPU(INTEL_FAM6_ATOM_GOLDMONT, core_funcs),
1779 ICPU(INTEL_FAM6_ATOM_GEMINI_LAKE, core_funcs),
1780 ICPU(INTEL_FAM6_SKYLAKE_X, core_funcs),
1783 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
1785 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] __initconst = {
1786 ICPU(INTEL_FAM6_BROADWELL_XEON_D, core_funcs),
1787 ICPU(INTEL_FAM6_BROADWELL_X, core_funcs),
1788 ICPU(INTEL_FAM6_SKYLAKE_X, core_funcs),
1792 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[] = {
1793 ICPU(INTEL_FAM6_KABYLAKE_DESKTOP, core_funcs),
1797 static const struct x86_cpu_id intel_pstate_hwp_boost_ids[] = {
1798 ICPU(INTEL_FAM6_SKYLAKE_X, core_funcs),
1799 ICPU(INTEL_FAM6_SKYLAKE_DESKTOP, core_funcs),
1803 static int intel_pstate_init_cpu(unsigned int cpunum)
1805 struct cpudata *cpu;
1807 cpu = all_cpu_data[cpunum];
1810 cpu = kzalloc(sizeof(*cpu), GFP_KERNEL);
1814 all_cpu_data[cpunum] = cpu;
1816 cpu->epp_default = -EINVAL;
1817 cpu->epp_powersave = -EINVAL;
1818 cpu->epp_saved = -EINVAL;
1821 cpu = all_cpu_data[cpunum];
1826 const struct x86_cpu_id *id;
1828 id = x86_match_cpu(intel_pstate_cpu_ee_disable_ids);
1830 intel_pstate_disable_ee(cpunum);
1832 intel_pstate_hwp_enable(cpu);
1834 id = x86_match_cpu(intel_pstate_hwp_boost_ids);
1839 intel_pstate_get_cpu_pstates(cpu);
1841 pr_debug("controlling: cpu %d\n", cpunum);
1846 static void intel_pstate_set_update_util_hook(unsigned int cpu_num)
1848 struct cpudata *cpu = all_cpu_data[cpu_num];
1850 if (hwp_active && !hwp_boost)
1853 if (cpu->update_util_set)
1856 /* Prevent intel_pstate_update_util() from using stale data. */
1857 cpu->sample.time = 0;
1858 cpufreq_add_update_util_hook(cpu_num, &cpu->update_util,
1860 intel_pstate_update_util_hwp :
1861 intel_pstate_update_util));
1862 cpu->update_util_set = true;
1865 static void intel_pstate_clear_update_util_hook(unsigned int cpu)
1867 struct cpudata *cpu_data = all_cpu_data[cpu];
1869 if (!cpu_data->update_util_set)
1872 cpufreq_remove_update_util_hook(cpu);
1873 cpu_data->update_util_set = false;
1874 synchronize_sched();
1877 static int intel_pstate_get_max_freq(struct cpudata *cpu)
1879 return global.turbo_disabled || global.no_turbo ?
1880 cpu->pstate.max_freq : cpu->pstate.turbo_freq;
1883 static void intel_pstate_update_perf_limits(struct cpufreq_policy *policy,
1884 struct cpudata *cpu)
1886 int max_freq = intel_pstate_get_max_freq(cpu);
1887 int32_t max_policy_perf, min_policy_perf;
1888 int max_state, turbo_max;
1891 * HWP needs some special consideration, because on BDX the
1892 * HWP_REQUEST uses abstract value to represent performance
1893 * rather than pure ratios.
1896 intel_pstate_get_hwp_max(cpu->cpu, &turbo_max, &max_state);
1898 max_state = intel_pstate_get_base_pstate(cpu);
1899 turbo_max = cpu->pstate.turbo_pstate;
1902 max_policy_perf = max_state * policy->max / max_freq;
1903 if (policy->max == policy->min) {
1904 min_policy_perf = max_policy_perf;
1906 min_policy_perf = max_state * policy->min / max_freq;
1907 min_policy_perf = clamp_t(int32_t, min_policy_perf,
1908 0, max_policy_perf);
1911 pr_debug("cpu:%d max_state %d min_policy_perf:%d max_policy_perf:%d\n",
1912 policy->cpu, max_state,
1913 min_policy_perf, max_policy_perf);
1915 /* Normalize user input to [min_perf, max_perf] */
1916 if (per_cpu_limits) {
1917 cpu->min_perf_ratio = min_policy_perf;
1918 cpu->max_perf_ratio = max_policy_perf;
1920 int32_t global_min, global_max;
1922 /* Global limits are in percent of the maximum turbo P-state. */
1923 global_max = DIV_ROUND_UP(turbo_max * global.max_perf_pct, 100);
1924 global_min = DIV_ROUND_UP(turbo_max * global.min_perf_pct, 100);
1925 global_min = clamp_t(int32_t, global_min, 0, global_max);
1927 pr_debug("cpu:%d global_min:%d global_max:%d\n", policy->cpu,
1928 global_min, global_max);
1930 cpu->min_perf_ratio = max(min_policy_perf, global_min);
1931 cpu->min_perf_ratio = min(cpu->min_perf_ratio, max_policy_perf);
1932 cpu->max_perf_ratio = min(max_policy_perf, global_max);
1933 cpu->max_perf_ratio = max(min_policy_perf, cpu->max_perf_ratio);
1935 /* Make sure min_perf <= max_perf */
1936 cpu->min_perf_ratio = min(cpu->min_perf_ratio,
1937 cpu->max_perf_ratio);
1940 pr_debug("cpu:%d max_perf_ratio:%d min_perf_ratio:%d\n", policy->cpu,
1941 cpu->max_perf_ratio,
1942 cpu->min_perf_ratio);
1945 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
1947 struct cpudata *cpu;
1949 if (!policy->cpuinfo.max_freq)
1952 pr_debug("set_policy cpuinfo.max %u policy->max %u\n",
1953 policy->cpuinfo.max_freq, policy->max);
1955 cpu = all_cpu_data[policy->cpu];
1956 cpu->policy = policy->policy;
1958 mutex_lock(&intel_pstate_limits_lock);
1960 intel_pstate_update_perf_limits(policy, cpu);
1962 if (cpu->policy == CPUFREQ_POLICY_PERFORMANCE) {
1964 * NOHZ_FULL CPUs need this as the governor callback may not
1965 * be invoked on them.
1967 intel_pstate_clear_update_util_hook(policy->cpu);
1968 intel_pstate_max_within_limits(cpu);
1970 intel_pstate_set_update_util_hook(policy->cpu);
1975 * When hwp_boost was active before and dynamically it
1976 * was turned off, in that case we need to clear the
1980 intel_pstate_clear_update_util_hook(policy->cpu);
1981 intel_pstate_hwp_set(policy->cpu);
1984 mutex_unlock(&intel_pstate_limits_lock);
1989 static void intel_pstate_adjust_policy_max(struct cpufreq_policy *policy,
1990 struct cpudata *cpu)
1992 if (cpu->pstate.max_pstate_physical > cpu->pstate.max_pstate &&
1993 policy->max < policy->cpuinfo.max_freq &&
1994 policy->max > cpu->pstate.max_freq) {
1995 pr_debug("policy->max > max non turbo frequency\n");
1996 policy->max = policy->cpuinfo.max_freq;
2000 static int intel_pstate_verify_policy(struct cpufreq_policy *policy)
2002 struct cpudata *cpu = all_cpu_data[policy->cpu];
2004 update_turbo_state();
2005 cpufreq_verify_within_limits(policy, policy->cpuinfo.min_freq,
2006 intel_pstate_get_max_freq(cpu));
2008 if (policy->policy != CPUFREQ_POLICY_POWERSAVE &&
2009 policy->policy != CPUFREQ_POLICY_PERFORMANCE)
2012 intel_pstate_adjust_policy_max(policy, cpu);
2017 static void intel_cpufreq_stop_cpu(struct cpufreq_policy *policy)
2019 intel_pstate_set_min_pstate(all_cpu_data[policy->cpu]);
2022 static void intel_pstate_stop_cpu(struct cpufreq_policy *policy)
2024 pr_debug("CPU %d exiting\n", policy->cpu);
2026 intel_pstate_clear_update_util_hook(policy->cpu);
2028 intel_pstate_hwp_save_state(policy);
2030 intel_cpufreq_stop_cpu(policy);
2033 static int intel_pstate_cpu_exit(struct cpufreq_policy *policy)
2035 intel_pstate_exit_perf_limits(policy);
2037 policy->fast_switch_possible = false;
2042 static int __intel_pstate_cpu_init(struct cpufreq_policy *policy)
2044 struct cpudata *cpu;
2047 rc = intel_pstate_init_cpu(policy->cpu);
2051 cpu = all_cpu_data[policy->cpu];
2053 cpu->max_perf_ratio = 0xFF;
2054 cpu->min_perf_ratio = 0;
2056 policy->min = cpu->pstate.min_pstate * cpu->pstate.scaling;
2057 policy->max = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
2059 /* cpuinfo and default policy values */
2060 policy->cpuinfo.min_freq = cpu->pstate.min_pstate * cpu->pstate.scaling;
2061 update_turbo_state();
2062 policy->cpuinfo.max_freq = global.turbo_disabled ?
2063 cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
2064 policy->cpuinfo.max_freq *= cpu->pstate.scaling;
2066 intel_pstate_init_acpi_perf_limits(policy);
2068 policy->fast_switch_possible = true;
2073 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
2075 int ret = __intel_pstate_cpu_init(policy);
2080 if (IS_ENABLED(CONFIG_CPU_FREQ_DEFAULT_GOV_PERFORMANCE))
2081 policy->policy = CPUFREQ_POLICY_PERFORMANCE;
2083 policy->policy = CPUFREQ_POLICY_POWERSAVE;
2088 static struct cpufreq_driver intel_pstate = {
2089 .flags = CPUFREQ_CONST_LOOPS,
2090 .verify = intel_pstate_verify_policy,
2091 .setpolicy = intel_pstate_set_policy,
2092 .suspend = intel_pstate_hwp_save_state,
2093 .resume = intel_pstate_resume,
2094 .init = intel_pstate_cpu_init,
2095 .exit = intel_pstate_cpu_exit,
2096 .stop_cpu = intel_pstate_stop_cpu,
2097 .name = "intel_pstate",
2100 static int intel_cpufreq_verify_policy(struct cpufreq_policy *policy)
2102 struct cpudata *cpu = all_cpu_data[policy->cpu];
2104 update_turbo_state();
2105 cpufreq_verify_within_limits(policy, policy->cpuinfo.min_freq,
2106 intel_pstate_get_max_freq(cpu));
2108 intel_pstate_adjust_policy_max(policy, cpu);
2110 intel_pstate_update_perf_limits(policy, cpu);
2115 /* Use of trace in passive mode:
2117 * In passive mode the trace core_busy field (also known as the
2118 * performance field, and lablelled as such on the graphs; also known as
2119 * core_avg_perf) is not needed and so is re-assigned to indicate if the
2120 * driver call was via the normal or fast switch path. Various graphs
2121 * output from the intel_pstate_tracer.py utility that include core_busy
2122 * (or performance or core_avg_perf) have a fixed y-axis from 0 to 100%,
2123 * so we use 10 to indicate the the normal path through the driver, and
2124 * 90 to indicate the fast switch path through the driver.
2125 * The scaled_busy field is not used, and is set to 0.
2128 #define INTEL_PSTATE_TRACE_TARGET 10
2129 #define INTEL_PSTATE_TRACE_FAST_SWITCH 90
2131 static void intel_cpufreq_trace(struct cpudata *cpu, unsigned int trace_type, int old_pstate)
2133 struct sample *sample;
2135 if (!trace_pstate_sample_enabled())
2138 if (!intel_pstate_sample(cpu, ktime_get()))
2141 sample = &cpu->sample;
2142 trace_pstate_sample(trace_type,
2145 cpu->pstate.current_pstate,
2149 get_avg_frequency(cpu),
2150 fp_toint(cpu->iowait_boost * 100));
2153 static int intel_cpufreq_target(struct cpufreq_policy *policy,
2154 unsigned int target_freq,
2155 unsigned int relation)
2157 struct cpudata *cpu = all_cpu_data[policy->cpu];
2158 struct cpufreq_freqs freqs;
2159 int target_pstate, old_pstate;
2161 update_turbo_state();
2163 freqs.old = policy->cur;
2164 freqs.new = target_freq;
2166 cpufreq_freq_transition_begin(policy, &freqs);
2168 case CPUFREQ_RELATION_L:
2169 target_pstate = DIV_ROUND_UP(freqs.new, cpu->pstate.scaling);
2171 case CPUFREQ_RELATION_H:
2172 target_pstate = freqs.new / cpu->pstate.scaling;
2175 target_pstate = DIV_ROUND_CLOSEST(freqs.new, cpu->pstate.scaling);
2178 target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2179 old_pstate = cpu->pstate.current_pstate;
2180 if (target_pstate != cpu->pstate.current_pstate) {
2181 cpu->pstate.current_pstate = target_pstate;
2182 wrmsrl_on_cpu(policy->cpu, MSR_IA32_PERF_CTL,
2183 pstate_funcs.get_val(cpu, target_pstate));
2185 freqs.new = target_pstate * cpu->pstate.scaling;
2186 intel_cpufreq_trace(cpu, INTEL_PSTATE_TRACE_TARGET, old_pstate);
2187 cpufreq_freq_transition_end(policy, &freqs, false);
2192 static unsigned int intel_cpufreq_fast_switch(struct cpufreq_policy *policy,
2193 unsigned int target_freq)
2195 struct cpudata *cpu = all_cpu_data[policy->cpu];
2196 int target_pstate, old_pstate;
2198 update_turbo_state();
2200 target_pstate = DIV_ROUND_UP(target_freq, cpu->pstate.scaling);
2201 target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2202 old_pstate = cpu->pstate.current_pstate;
2203 intel_pstate_update_pstate(cpu, target_pstate);
2204 intel_cpufreq_trace(cpu, INTEL_PSTATE_TRACE_FAST_SWITCH, old_pstate);
2205 return target_pstate * cpu->pstate.scaling;
2208 static int intel_cpufreq_cpu_init(struct cpufreq_policy *policy)
2210 int ret = __intel_pstate_cpu_init(policy);
2215 policy->cpuinfo.transition_latency = INTEL_CPUFREQ_TRANSITION_LATENCY;
2216 policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY;
2217 /* This reflects the intel_pstate_get_cpu_pstates() setting. */
2218 policy->cur = policy->cpuinfo.min_freq;
2223 static struct cpufreq_driver intel_cpufreq = {
2224 .flags = CPUFREQ_CONST_LOOPS,
2225 .verify = intel_cpufreq_verify_policy,
2226 .target = intel_cpufreq_target,
2227 .fast_switch = intel_cpufreq_fast_switch,
2228 .init = intel_cpufreq_cpu_init,
2229 .exit = intel_pstate_cpu_exit,
2230 .stop_cpu = intel_cpufreq_stop_cpu,
2231 .name = "intel_cpufreq",
2234 static struct cpufreq_driver *default_driver = &intel_pstate;
2236 static void intel_pstate_driver_cleanup(void)
2241 for_each_online_cpu(cpu) {
2242 if (all_cpu_data[cpu]) {
2243 if (intel_pstate_driver == &intel_pstate)
2244 intel_pstate_clear_update_util_hook(cpu);
2246 kfree(all_cpu_data[cpu]);
2247 all_cpu_data[cpu] = NULL;
2251 intel_pstate_driver = NULL;
2254 static int intel_pstate_register_driver(struct cpufreq_driver *driver)
2258 memset(&global, 0, sizeof(global));
2259 global.max_perf_pct = 100;
2261 intel_pstate_driver = driver;
2262 ret = cpufreq_register_driver(intel_pstate_driver);
2264 intel_pstate_driver_cleanup();
2268 global.min_perf_pct = min_perf_pct_min();
2273 static int intel_pstate_unregister_driver(void)
2278 cpufreq_unregister_driver(intel_pstate_driver);
2279 intel_pstate_driver_cleanup();
2284 static ssize_t intel_pstate_show_status(char *buf)
2286 if (!intel_pstate_driver)
2287 return sprintf(buf, "off\n");
2289 return sprintf(buf, "%s\n", intel_pstate_driver == &intel_pstate ?
2290 "active" : "passive");
2293 static int intel_pstate_update_status(const char *buf, size_t size)
2297 if (size == 3 && !strncmp(buf, "off", size))
2298 return intel_pstate_driver ?
2299 intel_pstate_unregister_driver() : -EINVAL;
2301 if (size == 6 && !strncmp(buf, "active", size)) {
2302 if (intel_pstate_driver) {
2303 if (intel_pstate_driver == &intel_pstate)
2306 ret = intel_pstate_unregister_driver();
2311 return intel_pstate_register_driver(&intel_pstate);
2314 if (size == 7 && !strncmp(buf, "passive", size)) {
2315 if (intel_pstate_driver) {
2316 if (intel_pstate_driver == &intel_cpufreq)
2319 ret = intel_pstate_unregister_driver();
2324 return intel_pstate_register_driver(&intel_cpufreq);
2330 static int no_load __initdata;
2331 static int no_hwp __initdata;
2332 static int hwp_only __initdata;
2333 static unsigned int force_load __initdata;
2335 static int __init intel_pstate_msrs_not_valid(void)
2337 if (!pstate_funcs.get_max() ||
2338 !pstate_funcs.get_min() ||
2339 !pstate_funcs.get_turbo())
2345 static void __init copy_cpu_funcs(struct pstate_funcs *funcs)
2347 pstate_funcs.get_max = funcs->get_max;
2348 pstate_funcs.get_max_physical = funcs->get_max_physical;
2349 pstate_funcs.get_min = funcs->get_min;
2350 pstate_funcs.get_turbo = funcs->get_turbo;
2351 pstate_funcs.get_scaling = funcs->get_scaling;
2352 pstate_funcs.get_val = funcs->get_val;
2353 pstate_funcs.get_vid = funcs->get_vid;
2354 pstate_funcs.get_aperf_mperf_shift = funcs->get_aperf_mperf_shift;
2359 static bool __init intel_pstate_no_acpi_pss(void)
2363 for_each_possible_cpu(i) {
2365 union acpi_object *pss;
2366 struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
2367 struct acpi_processor *pr = per_cpu(processors, i);
2372 status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
2373 if (ACPI_FAILURE(status))
2376 pss = buffer.pointer;
2377 if (pss && pss->type == ACPI_TYPE_PACKAGE) {
2388 static bool __init intel_pstate_has_acpi_ppc(void)
2392 for_each_possible_cpu(i) {
2393 struct acpi_processor *pr = per_cpu(processors, i);
2397 if (acpi_has_method(pr->handle, "_PPC"))
2408 /* Hardware vendor-specific info that has its own power management modes */
2409 static struct acpi_platform_list plat_info[] __initdata = {
2410 {"HP ", "ProLiant", 0, ACPI_SIG_FADT, all_versions, 0, PSS},
2411 {"ORACLE", "X4-2 ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2412 {"ORACLE", "X4-2L ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2413 {"ORACLE", "X4-2B ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2414 {"ORACLE", "X3-2 ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2415 {"ORACLE", "X3-2L ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2416 {"ORACLE", "X3-2B ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2417 {"ORACLE", "X4470M2 ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2418 {"ORACLE", "X4270M3 ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2419 {"ORACLE", "X4270M2 ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2420 {"ORACLE", "X4170M2 ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2421 {"ORACLE", "X4170 M3", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2422 {"ORACLE", "X4275 M3", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2423 {"ORACLE", "X6-2 ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2424 {"ORACLE", "Sudbury ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2428 static bool __init intel_pstate_platform_pwr_mgmt_exists(void)
2430 const struct x86_cpu_id *id;
2434 id = x86_match_cpu(intel_pstate_cpu_oob_ids);
2436 rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
2437 if ( misc_pwr & (1 << 8))
2441 idx = acpi_match_platform_list(plat_info);
2445 switch (plat_info[idx].data) {
2447 return intel_pstate_no_acpi_pss();
2449 return intel_pstate_has_acpi_ppc() && !force_load;
2455 static void intel_pstate_request_control_from_smm(void)
2458 * It may be unsafe to request P-states control from SMM if _PPC support
2459 * has not been enabled.
2462 acpi_processor_pstate_control();
2464 #else /* CONFIG_ACPI not enabled */
2465 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
2466 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
2467 static inline void intel_pstate_request_control_from_smm(void) {}
2468 #endif /* CONFIG_ACPI */
2470 static const struct x86_cpu_id hwp_support_ids[] __initconst = {
2471 { X86_VENDOR_INTEL, 6, X86_MODEL_ANY, X86_FEATURE_HWP },
2475 static int __init intel_pstate_init(void)
2482 if (x86_match_cpu(hwp_support_ids)) {
2483 copy_cpu_funcs(&core_funcs);
2486 intel_pstate.attr = hwp_cpufreq_attrs;
2487 goto hwp_cpu_matched;
2490 const struct x86_cpu_id *id;
2492 id = x86_match_cpu(intel_pstate_cpu_ids);
2496 copy_cpu_funcs((struct pstate_funcs *)id->driver_data);
2499 if (intel_pstate_msrs_not_valid())
2504 * The Intel pstate driver will be ignored if the platform
2505 * firmware has its own power management modes.
2507 if (intel_pstate_platform_pwr_mgmt_exists())
2510 if (!hwp_active && hwp_only)
2513 pr_info("Intel P-state driver initializing\n");
2515 all_cpu_data = vzalloc(array_size(sizeof(void *), num_possible_cpus()));
2519 intel_pstate_request_control_from_smm();
2521 intel_pstate_sysfs_expose_params();
2523 mutex_lock(&intel_pstate_driver_lock);
2524 rc = intel_pstate_register_driver(default_driver);
2525 mutex_unlock(&intel_pstate_driver_lock);
2530 pr_info("HWP enabled\n");
2534 device_initcall(intel_pstate_init);
2536 static int __init intel_pstate_setup(char *str)
2541 if (!strcmp(str, "disable")) {
2543 } else if (!strcmp(str, "passive")) {
2544 pr_info("Passive mode enabled\n");
2545 default_driver = &intel_cpufreq;
2548 if (!strcmp(str, "no_hwp")) {
2549 pr_info("HWP disabled\n");
2552 if (!strcmp(str, "force"))
2554 if (!strcmp(str, "hwp_only"))
2556 if (!strcmp(str, "per_cpu_perf_limits"))
2557 per_cpu_limits = true;
2560 if (!strcmp(str, "support_acpi_ppc"))
2566 early_param("intel_pstate", intel_pstate_setup);
2568 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
2569 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
2570 MODULE_LICENSE("GPL");