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/debugfs.h>
30 #include <linux/acpi.h>
31 #include <linux/vmalloc.h>
32 #include <trace/events/power.h>
34 #include <asm/div64.h>
36 #include <asm/cpu_device_id.h>
37 #include <asm/cpufeature.h>
38 #include <asm/intel-family.h>
40 #define INTEL_CPUFREQ_TRANSITION_LATENCY 20000
43 #include <acpi/processor.h>
44 #include <acpi/cppc_acpi.h>
48 #define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
49 #define fp_toint(X) ((X) >> FRAC_BITS)
52 #define EXT_FRAC_BITS (EXT_BITS + FRAC_BITS)
53 #define fp_ext_toint(X) ((X) >> EXT_FRAC_BITS)
54 #define int_ext_tofp(X) ((int64_t)(X) << EXT_FRAC_BITS)
56 static inline int32_t mul_fp(int32_t x, int32_t y)
58 return ((int64_t)x * (int64_t)y) >> FRAC_BITS;
61 static inline int32_t div_fp(s64 x, s64 y)
63 return div64_s64((int64_t)x << FRAC_BITS, y);
66 static inline int ceiling_fp(int32_t x)
71 mask = (1 << FRAC_BITS) - 1;
77 static inline int32_t percent_fp(int percent)
79 return div_fp(percent, 100);
82 static inline u64 mul_ext_fp(u64 x, u64 y)
84 return (x * y) >> EXT_FRAC_BITS;
87 static inline u64 div_ext_fp(u64 x, u64 y)
89 return div64_u64(x << EXT_FRAC_BITS, y);
92 static inline int32_t percent_ext_fp(int percent)
94 return div_ext_fp(percent, 100);
98 * struct sample - Store performance sample
99 * @core_avg_perf: Ratio of APERF/MPERF which is the actual average
100 * performance during last sample period
101 * @busy_scaled: Scaled busy value which is used to calculate next
102 * P state. This can be different than core_avg_perf
103 * to account for cpu idle period
104 * @aperf: Difference of actual performance frequency clock count
105 * read from APERF MSR between last and current sample
106 * @mperf: Difference of maximum performance frequency clock count
107 * read from MPERF MSR between last and current sample
108 * @tsc: Difference of time stamp counter between last and
110 * @time: Current time from scheduler
112 * This structure is used in the cpudata structure to store performance sample
113 * data for choosing next P State.
116 int32_t core_avg_perf;
125 * struct pstate_data - Store P state data
126 * @current_pstate: Current requested P state
127 * @min_pstate: Min P state possible for this platform
128 * @max_pstate: Max P state possible for this platform
129 * @max_pstate_physical:This is physical Max P state for a processor
130 * This can be higher than the max_pstate which can
131 * be limited by platform thermal design power limits
132 * @scaling: Scaling factor to convert frequency to cpufreq
134 * @turbo_pstate: Max Turbo P state possible for this platform
135 * @max_freq: @max_pstate frequency in cpufreq units
136 * @turbo_freq: @turbo_pstate frequency in cpufreq units
138 * Stores the per cpu model P state limits and current P state.
144 int max_pstate_physical;
147 unsigned int max_freq;
148 unsigned int turbo_freq;
152 * struct vid_data - Stores voltage information data
153 * @min: VID data for this platform corresponding to
155 * @max: VID data corresponding to the highest P State.
156 * @turbo: VID data for turbo P state
157 * @ratio: Ratio of (vid max - vid min) /
158 * (max P state - Min P State)
160 * Stores the voltage data for DVFS (Dynamic Voltage and Frequency Scaling)
161 * This data is used in Atom platforms, where in addition to target P state,
162 * the voltage data needs to be specified to select next P State.
172 * struct _pid - Stores PID data
173 * @setpoint: Target set point for busyness or performance
174 * @integral: Storage for accumulated error values
175 * @p_gain: PID proportional gain
176 * @i_gain: PID integral gain
177 * @d_gain: PID derivative gain
178 * @deadband: PID deadband
179 * @last_err: Last error storage for integral part of PID calculation
181 * Stores PID coefficients and last error for PID controller.
194 * struct global_params - Global parameters, mostly tunable via sysfs.
195 * @no_turbo: Whether or not to use turbo P-states.
196 * @turbo_disabled: Whethet or not turbo P-states are available at all,
197 * based on the MSR_IA32_MISC_ENABLE value and whether or
198 * not the maximum reported turbo P-state is different from
199 * the maximum reported non-turbo one.
200 * @min_perf_pct: Minimum capacity limit in percent of the maximum turbo
202 * @max_perf_pct: Maximum capacity limit in percent of the maximum turbo
205 struct global_params {
213 * struct cpudata - Per CPU instance data storage
214 * @cpu: CPU number for this instance data
215 * @policy: CPUFreq policy value
216 * @update_util: CPUFreq utility callback information
217 * @update_util_set: CPUFreq utility callback is set
218 * @iowait_boost: iowait-related boost fraction
219 * @last_update: Time of the last update.
220 * @pstate: Stores P state limits for this CPU
221 * @vid: Stores VID limits for this CPU
222 * @pid: Stores PID parameters for this CPU
223 * @last_sample_time: Last Sample time
224 * @prev_aperf: Last APERF value read from APERF MSR
225 * @prev_mperf: Last MPERF value read from MPERF MSR
226 * @prev_tsc: Last timestamp counter (TSC) value
227 * @prev_cummulative_iowait: IO Wait time difference from last and
229 * @sample: Storage for storing last Sample data
230 * @min_perf: Minimum capacity limit as a fraction of the maximum
231 * turbo P-state capacity.
232 * @max_perf: Maximum capacity limit as a fraction of the maximum
233 * turbo P-state capacity.
234 * @acpi_perf_data: Stores ACPI perf information read from _PSS
235 * @valid_pss_table: Set to true for valid ACPI _PSS entries found
236 * @epp_powersave: Last saved HWP energy performance preference
237 * (EPP) or energy performance bias (EPB),
238 * when policy switched to performance
239 * @epp_policy: Last saved policy used to set EPP/EPB
240 * @epp_default: Power on default HWP energy performance
242 * @epp_saved: Saved EPP/EPB during system suspend or CPU offline
245 * This structure stores per CPU instance data for all CPUs.
251 struct update_util_data update_util;
252 bool update_util_set;
254 struct pstate_data pstate;
259 u64 last_sample_time;
263 u64 prev_cummulative_iowait;
264 struct sample sample;
268 struct acpi_processor_performance acpi_perf_data;
269 bool valid_pss_table;
271 unsigned int iowait_boost;
278 static struct cpudata **all_cpu_data;
281 * struct pstate_adjust_policy - Stores static PID configuration data
282 * @sample_rate_ms: PID calculation sample rate in ms
283 * @sample_rate_ns: Sample rate calculation in ns
284 * @deadband: PID deadband
285 * @setpoint: PID Setpoint
286 * @p_gain_pct: PID proportional gain
287 * @i_gain_pct: PID integral gain
288 * @d_gain_pct: PID derivative gain
290 * Stores per CPU model static PID configuration data.
292 struct pstate_adjust_policy {
303 * struct pstate_funcs - Per CPU model specific callbacks
304 * @get_max: Callback to get maximum non turbo effective P state
305 * @get_max_physical: Callback to get maximum non turbo physical P state
306 * @get_min: Callback to get minimum P state
307 * @get_turbo: Callback to get turbo P state
308 * @get_scaling: Callback to get frequency scaling factor
309 * @get_val: Callback to convert P state to actual MSR write value
310 * @get_vid: Callback to get VID data for Atom platforms
311 * @get_target_pstate: Callback to a function to calculate next P state to use
313 * Core and Atom CPU models have different way to get P State limits. This
314 * structure is used to store those callbacks.
316 struct pstate_funcs {
317 int (*get_max)(void);
318 int (*get_max_physical)(void);
319 int (*get_min)(void);
320 int (*get_turbo)(void);
321 int (*get_scaling)(void);
322 u64 (*get_val)(struct cpudata*, int pstate);
323 void (*get_vid)(struct cpudata *);
324 int32_t (*get_target_pstate)(struct cpudata *);
328 * struct cpu_defaults- Per CPU model default config data
329 * @funcs: Callback function data
331 struct cpu_defaults {
332 struct pstate_funcs funcs;
335 static inline int32_t get_target_pstate_use_performance(struct cpudata *cpu);
336 static inline int32_t get_target_pstate_use_cpu_load(struct cpudata *cpu);
338 static struct pstate_funcs pstate_funcs __read_mostly;
339 static struct pstate_adjust_policy pid_params __read_mostly = {
340 .sample_rate_ms = 10,
341 .sample_rate_ns = 10 * NSEC_PER_MSEC,
349 static int hwp_active __read_mostly;
350 static bool per_cpu_limits __read_mostly;
352 static struct cpufreq_driver *intel_pstate_driver __read_mostly;
355 static bool acpi_ppc;
358 static struct global_params global;
360 static DEFINE_MUTEX(intel_pstate_driver_lock);
361 static DEFINE_MUTEX(intel_pstate_limits_lock);
365 static bool intel_pstate_get_ppc_enable_status(void)
367 if (acpi_gbl_FADT.preferred_profile == PM_ENTERPRISE_SERVER ||
368 acpi_gbl_FADT.preferred_profile == PM_PERFORMANCE_SERVER)
374 #ifdef CONFIG_ACPI_CPPC_LIB
376 /* The work item is needed to avoid CPU hotplug locking issues */
377 static void intel_pstste_sched_itmt_work_fn(struct work_struct *work)
379 sched_set_itmt_support();
382 static DECLARE_WORK(sched_itmt_work, intel_pstste_sched_itmt_work_fn);
384 static void intel_pstate_set_itmt_prio(int cpu)
386 struct cppc_perf_caps cppc_perf;
387 static u32 max_highest_perf = 0, min_highest_perf = U32_MAX;
390 ret = cppc_get_perf_caps(cpu, &cppc_perf);
395 * The priorities can be set regardless of whether or not
396 * sched_set_itmt_support(true) has been called and it is valid to
397 * update them at any time after it has been called.
399 sched_set_itmt_core_prio(cppc_perf.highest_perf, cpu);
401 if (max_highest_perf <= min_highest_perf) {
402 if (cppc_perf.highest_perf > max_highest_perf)
403 max_highest_perf = cppc_perf.highest_perf;
405 if (cppc_perf.highest_perf < min_highest_perf)
406 min_highest_perf = cppc_perf.highest_perf;
408 if (max_highest_perf > min_highest_perf) {
410 * This code can be run during CPU online under the
411 * CPU hotplug locks, so sched_set_itmt_support()
412 * cannot be called from here. Queue up a work item
415 schedule_work(&sched_itmt_work);
420 static void intel_pstate_set_itmt_prio(int cpu)
425 static void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
432 intel_pstate_set_itmt_prio(policy->cpu);
436 if (!intel_pstate_get_ppc_enable_status())
439 cpu = all_cpu_data[policy->cpu];
441 ret = acpi_processor_register_performance(&cpu->acpi_perf_data,
447 * Check if the control value in _PSS is for PERF_CTL MSR, which should
448 * guarantee that the states returned by it map to the states in our
451 if (cpu->acpi_perf_data.control_register.space_id !=
452 ACPI_ADR_SPACE_FIXED_HARDWARE)
456 * If there is only one entry _PSS, simply ignore _PSS and continue as
457 * usual without taking _PSS into account
459 if (cpu->acpi_perf_data.state_count < 2)
462 pr_debug("CPU%u - ACPI _PSS perf data\n", policy->cpu);
463 for (i = 0; i < cpu->acpi_perf_data.state_count; i++) {
464 pr_debug(" %cP%d: %u MHz, %u mW, 0x%x\n",
465 (i == cpu->acpi_perf_data.state ? '*' : ' '), i,
466 (u32) cpu->acpi_perf_data.states[i].core_frequency,
467 (u32) cpu->acpi_perf_data.states[i].power,
468 (u32) cpu->acpi_perf_data.states[i].control);
472 * The _PSS table doesn't contain whole turbo frequency range.
473 * This just contains +1 MHZ above the max non turbo frequency,
474 * with control value corresponding to max turbo ratio. But
475 * when cpufreq set policy is called, it will call with this
476 * max frequency, which will cause a reduced performance as
477 * this driver uses real max turbo frequency as the max
478 * frequency. So correct this frequency in _PSS table to
479 * correct max turbo frequency based on the turbo state.
480 * Also need to convert to MHz as _PSS freq is in MHz.
482 if (!global.turbo_disabled)
483 cpu->acpi_perf_data.states[0].core_frequency =
484 policy->cpuinfo.max_freq / 1000;
485 cpu->valid_pss_table = true;
486 pr_debug("_PPC limits will be enforced\n");
491 cpu->valid_pss_table = false;
492 acpi_processor_unregister_performance(policy->cpu);
495 static void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
499 cpu = all_cpu_data[policy->cpu];
500 if (!cpu->valid_pss_table)
503 acpi_processor_unregister_performance(policy->cpu);
506 static inline void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
510 static inline void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
515 static signed int pid_calc(struct _pid *pid, int32_t busy)
518 int32_t pterm, dterm, fp_error;
519 int32_t integral_limit;
521 fp_error = pid->setpoint - busy;
523 if (abs(fp_error) <= pid->deadband)
526 pterm = mul_fp(pid->p_gain, fp_error);
528 pid->integral += fp_error;
531 * We limit the integral here so that it will never
532 * get higher than 30. This prevents it from becoming
533 * too large an input over long periods of time and allows
534 * it to get factored out sooner.
536 * The value of 30 was chosen through experimentation.
538 integral_limit = int_tofp(30);
539 if (pid->integral > integral_limit)
540 pid->integral = integral_limit;
541 if (pid->integral < -integral_limit)
542 pid->integral = -integral_limit;
544 dterm = mul_fp(pid->d_gain, fp_error - pid->last_err);
545 pid->last_err = fp_error;
547 result = pterm + mul_fp(pid->integral, pid->i_gain) + dterm;
548 result = result + (1 << (FRAC_BITS-1));
549 return (signed int)fp_toint(result);
552 static inline void intel_pstate_pid_reset(struct cpudata *cpu)
554 struct _pid *pid = &cpu->pid;
556 pid->p_gain = percent_fp(pid_params.p_gain_pct);
557 pid->d_gain = percent_fp(pid_params.d_gain_pct);
558 pid->i_gain = percent_fp(pid_params.i_gain_pct);
559 pid->setpoint = int_tofp(pid_params.setpoint);
560 pid->last_err = pid->setpoint - int_tofp(100);
561 pid->deadband = int_tofp(pid_params.deadband);
565 static inline void update_turbo_state(void)
570 cpu = all_cpu_data[0];
571 rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
572 global.turbo_disabled =
573 (misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ||
574 cpu->pstate.max_pstate == cpu->pstate.turbo_pstate);
577 static int min_perf_pct_min(void)
579 struct cpudata *cpu = all_cpu_data[0];
581 return DIV_ROUND_UP(cpu->pstate.min_pstate * 100,
582 cpu->pstate.turbo_pstate);
585 static s16 intel_pstate_get_epb(struct cpudata *cpu_data)
590 if (!static_cpu_has(X86_FEATURE_EPB))
593 ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
597 return (s16)(epb & 0x0f);
600 static s16 intel_pstate_get_epp(struct cpudata *cpu_data, u64 hwp_req_data)
604 if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
606 * When hwp_req_data is 0, means that caller didn't read
607 * MSR_HWP_REQUEST, so need to read and get EPP.
610 epp = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST,
615 epp = (hwp_req_data >> 24) & 0xff;
617 /* When there is no EPP present, HWP uses EPB settings */
618 epp = intel_pstate_get_epb(cpu_data);
624 static int intel_pstate_set_epb(int cpu, s16 pref)
629 if (!static_cpu_has(X86_FEATURE_EPB))
632 ret = rdmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
636 epb = (epb & ~0x0f) | pref;
637 wrmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, epb);
643 * EPP/EPB display strings corresponding to EPP index in the
644 * energy_perf_strings[]
646 *-------------------------------------
649 * 2 balance_performance
653 static const char * const energy_perf_strings[] = {
656 "balance_performance",
662 static int intel_pstate_get_energy_pref_index(struct cpudata *cpu_data)
667 epp = intel_pstate_get_epp(cpu_data, 0);
671 if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
674 * 0x00-0x3F : Performance
675 * 0x40-0x7F : Balance performance
676 * 0x80-0xBF : Balance power
678 * The EPP is a 8 bit value, but our ranges restrict the
679 * value which can be set. Here only using top two bits
682 index = (epp >> 6) + 1;
683 } else if (static_cpu_has(X86_FEATURE_EPB)) {
686 * 0x00-0x03 : Performance
687 * 0x04-0x07 : Balance performance
688 * 0x08-0x0B : Balance power
690 * The EPB is a 4 bit value, but our ranges restrict the
691 * value which can be set. Here only using top two bits
694 index = (epp >> 2) + 1;
700 static int intel_pstate_set_energy_pref_index(struct cpudata *cpu_data,
707 epp = cpu_data->epp_default;
709 mutex_lock(&intel_pstate_limits_lock);
711 if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
714 ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST, &value);
718 value &= ~GENMASK_ULL(31, 24);
721 * If epp is not default, convert from index into
722 * energy_perf_strings to epp value, by shifting 6
723 * bits left to use only top two bits in epp.
724 * The resultant epp need to shifted by 24 bits to
725 * epp position in MSR_HWP_REQUEST.
728 epp = (pref_index - 1) << 6;
730 value |= (u64)epp << 24;
731 ret = wrmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST, value);
734 epp = (pref_index - 1) << 2;
735 ret = intel_pstate_set_epb(cpu_data->cpu, epp);
738 mutex_unlock(&intel_pstate_limits_lock);
743 static ssize_t show_energy_performance_available_preferences(
744 struct cpufreq_policy *policy, char *buf)
749 while (energy_perf_strings[i] != NULL)
750 ret += sprintf(&buf[ret], "%s ", energy_perf_strings[i++]);
752 ret += sprintf(&buf[ret], "\n");
757 cpufreq_freq_attr_ro(energy_performance_available_preferences);
759 static ssize_t store_energy_performance_preference(
760 struct cpufreq_policy *policy, const char *buf, size_t count)
762 struct cpudata *cpu_data = all_cpu_data[policy->cpu];
763 char str_preference[21];
766 ret = sscanf(buf, "%20s", str_preference);
770 while (energy_perf_strings[i] != NULL) {
771 if (!strcmp(str_preference, energy_perf_strings[i])) {
772 intel_pstate_set_energy_pref_index(cpu_data, i);
781 static ssize_t show_energy_performance_preference(
782 struct cpufreq_policy *policy, char *buf)
784 struct cpudata *cpu_data = all_cpu_data[policy->cpu];
787 preference = intel_pstate_get_energy_pref_index(cpu_data);
791 return sprintf(buf, "%s\n", energy_perf_strings[preference]);
794 cpufreq_freq_attr_rw(energy_performance_preference);
796 static struct freq_attr *hwp_cpufreq_attrs[] = {
797 &energy_performance_preference,
798 &energy_performance_available_preferences,
802 static void intel_pstate_hwp_set(struct cpufreq_policy *policy)
804 int min, hw_min, max, hw_max, cpu;
807 for_each_cpu(cpu, policy->cpus) {
808 struct cpudata *cpu_data = all_cpu_data[cpu];
811 rdmsrl_on_cpu(cpu, MSR_HWP_CAPABILITIES, &cap);
812 hw_min = HWP_LOWEST_PERF(cap);
814 hw_max = HWP_GUARANTEED_PERF(cap);
816 hw_max = HWP_HIGHEST_PERF(cap);
818 max = fp_ext_toint(hw_max * cpu_data->max_perf);
819 if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE)
822 min = fp_ext_toint(hw_max * cpu_data->min_perf);
824 rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
826 value &= ~HWP_MIN_PERF(~0L);
827 value |= HWP_MIN_PERF(min);
829 value &= ~HWP_MAX_PERF(~0L);
830 value |= HWP_MAX_PERF(max);
832 if (cpu_data->epp_policy == cpu_data->policy)
835 cpu_data->epp_policy = cpu_data->policy;
837 if (cpu_data->epp_saved >= 0) {
838 epp = cpu_data->epp_saved;
839 cpu_data->epp_saved = -EINVAL;
843 if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE) {
844 epp = intel_pstate_get_epp(cpu_data, value);
845 cpu_data->epp_powersave = epp;
846 /* If EPP read was failed, then don't try to write */
853 /* skip setting EPP, when saved value is invalid */
854 if (cpu_data->epp_powersave < 0)
858 * No need to restore EPP when it is not zero. This
860 * - Policy is not changed
861 * - user has manually changed
862 * - Error reading EPB
864 epp = intel_pstate_get_epp(cpu_data, value);
868 epp = cpu_data->epp_powersave;
871 if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
872 value &= ~GENMASK_ULL(31, 24);
873 value |= (u64)epp << 24;
875 intel_pstate_set_epb(cpu, epp);
878 wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
882 static int intel_pstate_hwp_save_state(struct cpufreq_policy *policy)
884 struct cpudata *cpu_data = all_cpu_data[policy->cpu];
889 cpu_data->epp_saved = intel_pstate_get_epp(cpu_data, 0);
894 static int intel_pstate_resume(struct cpufreq_policy *policy)
899 mutex_lock(&intel_pstate_limits_lock);
901 all_cpu_data[policy->cpu]->epp_policy = 0;
902 intel_pstate_hwp_set(policy);
904 mutex_unlock(&intel_pstate_limits_lock);
909 static void intel_pstate_update_policies(void)
913 for_each_possible_cpu(cpu)
914 cpufreq_update_policy(cpu);
917 /************************** debugfs begin ************************/
918 static int pid_param_set(void *data, u64 val)
923 pid_params.sample_rate_ns = pid_params.sample_rate_ms * NSEC_PER_MSEC;
924 for_each_possible_cpu(cpu)
925 if (all_cpu_data[cpu])
926 intel_pstate_pid_reset(all_cpu_data[cpu]);
931 static int pid_param_get(void *data, u64 *val)
936 DEFINE_SIMPLE_ATTRIBUTE(fops_pid_param, pid_param_get, pid_param_set, "%llu\n");
938 static struct dentry *debugfs_parent;
943 struct dentry *dentry;
946 static struct pid_param pid_files[] = {
947 {"sample_rate_ms", &pid_params.sample_rate_ms, },
948 {"d_gain_pct", &pid_params.d_gain_pct, },
949 {"i_gain_pct", &pid_params.i_gain_pct, },
950 {"deadband", &pid_params.deadband, },
951 {"setpoint", &pid_params.setpoint, },
952 {"p_gain_pct", &pid_params.p_gain_pct, },
956 static void intel_pstate_debug_expose_params(void)
960 debugfs_parent = debugfs_create_dir("pstate_snb", NULL);
961 if (IS_ERR_OR_NULL(debugfs_parent))
964 for (i = 0; pid_files[i].name; i++) {
965 struct dentry *dentry;
967 dentry = debugfs_create_file(pid_files[i].name, 0660,
968 debugfs_parent, pid_files[i].value,
971 pid_files[i].dentry = dentry;
975 static void intel_pstate_debug_hide_params(void)
979 if (IS_ERR_OR_NULL(debugfs_parent))
982 for (i = 0; pid_files[i].name; i++) {
983 debugfs_remove(pid_files[i].dentry);
984 pid_files[i].dentry = NULL;
987 debugfs_remove(debugfs_parent);
988 debugfs_parent = NULL;
991 /************************** debugfs end ************************/
993 /************************** sysfs begin ************************/
994 #define show_one(file_name, object) \
995 static ssize_t show_##file_name \
996 (struct kobject *kobj, struct attribute *attr, char *buf) \
998 return sprintf(buf, "%u\n", global.object); \
1001 static ssize_t intel_pstate_show_status(char *buf);
1002 static int intel_pstate_update_status(const char *buf, size_t size);
1004 static ssize_t show_status(struct kobject *kobj,
1005 struct attribute *attr, char *buf)
1009 mutex_lock(&intel_pstate_driver_lock);
1010 ret = intel_pstate_show_status(buf);
1011 mutex_unlock(&intel_pstate_driver_lock);
1016 static ssize_t store_status(struct kobject *a, struct attribute *b,
1017 const char *buf, size_t count)
1019 char *p = memchr(buf, '\n', count);
1022 mutex_lock(&intel_pstate_driver_lock);
1023 ret = intel_pstate_update_status(buf, p ? p - buf : count);
1024 mutex_unlock(&intel_pstate_driver_lock);
1026 return ret < 0 ? ret : count;
1029 static ssize_t show_turbo_pct(struct kobject *kobj,
1030 struct attribute *attr, char *buf)
1032 struct cpudata *cpu;
1033 int total, no_turbo, turbo_pct;
1036 mutex_lock(&intel_pstate_driver_lock);
1038 if (!intel_pstate_driver) {
1039 mutex_unlock(&intel_pstate_driver_lock);
1043 cpu = all_cpu_data[0];
1045 total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1046 no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
1047 turbo_fp = div_fp(no_turbo, total);
1048 turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
1050 mutex_unlock(&intel_pstate_driver_lock);
1052 return sprintf(buf, "%u\n", turbo_pct);
1055 static ssize_t show_num_pstates(struct kobject *kobj,
1056 struct attribute *attr, char *buf)
1058 struct cpudata *cpu;
1061 mutex_lock(&intel_pstate_driver_lock);
1063 if (!intel_pstate_driver) {
1064 mutex_unlock(&intel_pstate_driver_lock);
1068 cpu = all_cpu_data[0];
1069 total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1071 mutex_unlock(&intel_pstate_driver_lock);
1073 return sprintf(buf, "%u\n", total);
1076 static ssize_t show_no_turbo(struct kobject *kobj,
1077 struct attribute *attr, char *buf)
1081 mutex_lock(&intel_pstate_driver_lock);
1083 if (!intel_pstate_driver) {
1084 mutex_unlock(&intel_pstate_driver_lock);
1088 update_turbo_state();
1089 if (global.turbo_disabled)
1090 ret = sprintf(buf, "%u\n", global.turbo_disabled);
1092 ret = sprintf(buf, "%u\n", global.no_turbo);
1094 mutex_unlock(&intel_pstate_driver_lock);
1099 static ssize_t store_no_turbo(struct kobject *a, struct attribute *b,
1100 const char *buf, size_t count)
1105 ret = sscanf(buf, "%u", &input);
1109 mutex_lock(&intel_pstate_driver_lock);
1111 if (!intel_pstate_driver) {
1112 mutex_unlock(&intel_pstate_driver_lock);
1116 mutex_lock(&intel_pstate_limits_lock);
1118 update_turbo_state();
1119 if (global.turbo_disabled) {
1120 pr_warn("Turbo disabled by BIOS or unavailable on processor\n");
1121 mutex_unlock(&intel_pstate_limits_lock);
1122 mutex_unlock(&intel_pstate_driver_lock);
1126 global.no_turbo = clamp_t(int, input, 0, 1);
1128 if (global.no_turbo) {
1129 struct cpudata *cpu = all_cpu_data[0];
1130 int pct = cpu->pstate.max_pstate * 100 / cpu->pstate.turbo_pstate;
1132 /* Squash the global minimum into the permitted range. */
1133 if (global.min_perf_pct > pct)
1134 global.min_perf_pct = pct;
1137 mutex_unlock(&intel_pstate_limits_lock);
1139 intel_pstate_update_policies();
1141 mutex_unlock(&intel_pstate_driver_lock);
1146 static ssize_t store_max_perf_pct(struct kobject *a, struct attribute *b,
1147 const char *buf, size_t count)
1152 ret = sscanf(buf, "%u", &input);
1156 mutex_lock(&intel_pstate_driver_lock);
1158 if (!intel_pstate_driver) {
1159 mutex_unlock(&intel_pstate_driver_lock);
1163 mutex_lock(&intel_pstate_limits_lock);
1165 global.max_perf_pct = clamp_t(int, input, global.min_perf_pct, 100);
1167 mutex_unlock(&intel_pstate_limits_lock);
1169 intel_pstate_update_policies();
1171 mutex_unlock(&intel_pstate_driver_lock);
1176 static ssize_t store_min_perf_pct(struct kobject *a, struct attribute *b,
1177 const char *buf, size_t count)
1182 ret = sscanf(buf, "%u", &input);
1186 mutex_lock(&intel_pstate_driver_lock);
1188 if (!intel_pstate_driver) {
1189 mutex_unlock(&intel_pstate_driver_lock);
1193 mutex_lock(&intel_pstate_limits_lock);
1195 global.min_perf_pct = clamp_t(int, input,
1196 min_perf_pct_min(), global.max_perf_pct);
1198 mutex_unlock(&intel_pstate_limits_lock);
1200 intel_pstate_update_policies();
1202 mutex_unlock(&intel_pstate_driver_lock);
1207 show_one(max_perf_pct, max_perf_pct);
1208 show_one(min_perf_pct, min_perf_pct);
1210 define_one_global_rw(status);
1211 define_one_global_rw(no_turbo);
1212 define_one_global_rw(max_perf_pct);
1213 define_one_global_rw(min_perf_pct);
1214 define_one_global_ro(turbo_pct);
1215 define_one_global_ro(num_pstates);
1217 static struct attribute *intel_pstate_attributes[] = {
1225 static struct attribute_group intel_pstate_attr_group = {
1226 .attrs = intel_pstate_attributes,
1229 static void __init intel_pstate_sysfs_expose_params(void)
1231 struct kobject *intel_pstate_kobject;
1234 intel_pstate_kobject = kobject_create_and_add("intel_pstate",
1235 &cpu_subsys.dev_root->kobj);
1236 if (WARN_ON(!intel_pstate_kobject))
1239 rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
1244 * If per cpu limits are enforced there are no global limits, so
1245 * return without creating max/min_perf_pct attributes
1250 rc = sysfs_create_file(intel_pstate_kobject, &max_perf_pct.attr);
1253 rc = sysfs_create_file(intel_pstate_kobject, &min_perf_pct.attr);
1257 /************************** sysfs end ************************/
1259 static void intel_pstate_hwp_enable(struct cpudata *cpudata)
1261 /* First disable HWP notification interrupt as we don't process them */
1262 if (static_cpu_has(X86_FEATURE_HWP_NOTIFY))
1263 wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1265 wrmsrl_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1);
1266 cpudata->epp_policy = 0;
1267 if (cpudata->epp_default == -EINVAL)
1268 cpudata->epp_default = intel_pstate_get_epp(cpudata, 0);
1271 #define MSR_IA32_POWER_CTL_BIT_EE 19
1273 /* Disable energy efficiency optimization */
1274 static void intel_pstate_disable_ee(int cpu)
1279 ret = rdmsrl_on_cpu(cpu, MSR_IA32_POWER_CTL, &power_ctl);
1283 if (!(power_ctl & BIT(MSR_IA32_POWER_CTL_BIT_EE))) {
1284 pr_info("Disabling energy efficiency optimization\n");
1285 power_ctl |= BIT(MSR_IA32_POWER_CTL_BIT_EE);
1286 wrmsrl_on_cpu(cpu, MSR_IA32_POWER_CTL, power_ctl);
1290 static int atom_get_min_pstate(void)
1294 rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1295 return (value >> 8) & 0x7F;
1298 static int atom_get_max_pstate(void)
1302 rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1303 return (value >> 16) & 0x7F;
1306 static int atom_get_turbo_pstate(void)
1310 rdmsrl(MSR_ATOM_CORE_TURBO_RATIOS, value);
1311 return value & 0x7F;
1314 static u64 atom_get_val(struct cpudata *cpudata, int pstate)
1320 val = (u64)pstate << 8;
1321 if (global.no_turbo && !global.turbo_disabled)
1322 val |= (u64)1 << 32;
1324 vid_fp = cpudata->vid.min + mul_fp(
1325 int_tofp(pstate - cpudata->pstate.min_pstate),
1326 cpudata->vid.ratio);
1328 vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
1329 vid = ceiling_fp(vid_fp);
1331 if (pstate > cpudata->pstate.max_pstate)
1332 vid = cpudata->vid.turbo;
1337 static int silvermont_get_scaling(void)
1341 /* Defined in Table 35-6 from SDM (Sept 2015) */
1342 static int silvermont_freq_table[] = {
1343 83300, 100000, 133300, 116700, 80000};
1345 rdmsrl(MSR_FSB_FREQ, value);
1349 return silvermont_freq_table[i];
1352 static int airmont_get_scaling(void)
1356 /* Defined in Table 35-10 from SDM (Sept 2015) */
1357 static int airmont_freq_table[] = {
1358 83300, 100000, 133300, 116700, 80000,
1359 93300, 90000, 88900, 87500};
1361 rdmsrl(MSR_FSB_FREQ, value);
1365 return airmont_freq_table[i];
1368 static void atom_get_vid(struct cpudata *cpudata)
1372 rdmsrl(MSR_ATOM_CORE_VIDS, value);
1373 cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
1374 cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
1375 cpudata->vid.ratio = div_fp(
1376 cpudata->vid.max - cpudata->vid.min,
1377 int_tofp(cpudata->pstate.max_pstate -
1378 cpudata->pstate.min_pstate));
1380 rdmsrl(MSR_ATOM_CORE_TURBO_VIDS, value);
1381 cpudata->vid.turbo = value & 0x7f;
1384 static int core_get_min_pstate(void)
1388 rdmsrl(MSR_PLATFORM_INFO, value);
1389 return (value >> 40) & 0xFF;
1392 static int core_get_max_pstate_physical(void)
1396 rdmsrl(MSR_PLATFORM_INFO, value);
1397 return (value >> 8) & 0xFF;
1400 static int core_get_tdp_ratio(u64 plat_info)
1402 /* Check how many TDP levels present */
1403 if (plat_info & 0x600000000) {
1409 /* Get the TDP level (0, 1, 2) to get ratios */
1410 err = rdmsrl_safe(MSR_CONFIG_TDP_CONTROL, &tdp_ctrl);
1414 /* TDP MSR are continuous starting at 0x648 */
1415 tdp_msr = MSR_CONFIG_TDP_NOMINAL + (tdp_ctrl & 0x03);
1416 err = rdmsrl_safe(tdp_msr, &tdp_ratio);
1420 /* For level 1 and 2, bits[23:16] contain the ratio */
1421 if (tdp_ctrl & 0x03)
1424 tdp_ratio &= 0xff; /* ratios are only 8 bits long */
1425 pr_debug("tdp_ratio %x\n", (int)tdp_ratio);
1427 return (int)tdp_ratio;
1433 static int core_get_max_pstate(void)
1441 rdmsrl(MSR_PLATFORM_INFO, plat_info);
1442 max_pstate = (plat_info >> 8) & 0xFF;
1444 tdp_ratio = core_get_tdp_ratio(plat_info);
1449 /* Turbo activation ratio is not used on HWP platforms */
1453 err = rdmsrl_safe(MSR_TURBO_ACTIVATION_RATIO, &tar);
1457 /* Do some sanity checking for safety */
1458 tar_levels = tar & 0xff;
1459 if (tdp_ratio - 1 == tar_levels) {
1460 max_pstate = tar_levels;
1461 pr_debug("max_pstate=TAC %x\n", max_pstate);
1468 static int core_get_turbo_pstate(void)
1473 rdmsrl(MSR_TURBO_RATIO_LIMIT, value);
1474 nont = core_get_max_pstate();
1475 ret = (value) & 255;
1481 static inline int core_get_scaling(void)
1486 static u64 core_get_val(struct cpudata *cpudata, int pstate)
1490 val = (u64)pstate << 8;
1491 if (global.no_turbo && !global.turbo_disabled)
1492 val |= (u64)1 << 32;
1497 static int knl_get_turbo_pstate(void)
1502 rdmsrl(MSR_TURBO_RATIO_LIMIT, value);
1503 nont = core_get_max_pstate();
1504 ret = (((value) >> 8) & 0xFF);
1510 static struct cpu_defaults core_params = {
1512 .get_max = core_get_max_pstate,
1513 .get_max_physical = core_get_max_pstate_physical,
1514 .get_min = core_get_min_pstate,
1515 .get_turbo = core_get_turbo_pstate,
1516 .get_scaling = core_get_scaling,
1517 .get_val = core_get_val,
1518 .get_target_pstate = get_target_pstate_use_performance,
1522 static const struct cpu_defaults silvermont_params = {
1524 .get_max = atom_get_max_pstate,
1525 .get_max_physical = atom_get_max_pstate,
1526 .get_min = atom_get_min_pstate,
1527 .get_turbo = atom_get_turbo_pstate,
1528 .get_val = atom_get_val,
1529 .get_scaling = silvermont_get_scaling,
1530 .get_vid = atom_get_vid,
1531 .get_target_pstate = get_target_pstate_use_cpu_load,
1535 static const struct cpu_defaults airmont_params = {
1537 .get_max = atom_get_max_pstate,
1538 .get_max_physical = atom_get_max_pstate,
1539 .get_min = atom_get_min_pstate,
1540 .get_turbo = atom_get_turbo_pstate,
1541 .get_val = atom_get_val,
1542 .get_scaling = airmont_get_scaling,
1543 .get_vid = atom_get_vid,
1544 .get_target_pstate = get_target_pstate_use_cpu_load,
1548 static const struct cpu_defaults knl_params = {
1550 .get_max = core_get_max_pstate,
1551 .get_max_physical = core_get_max_pstate_physical,
1552 .get_min = core_get_min_pstate,
1553 .get_turbo = knl_get_turbo_pstate,
1554 .get_scaling = core_get_scaling,
1555 .get_val = core_get_val,
1556 .get_target_pstate = get_target_pstate_use_performance,
1560 static const struct cpu_defaults bxt_params = {
1562 .get_max = core_get_max_pstate,
1563 .get_max_physical = core_get_max_pstate_physical,
1564 .get_min = core_get_min_pstate,
1565 .get_turbo = core_get_turbo_pstate,
1566 .get_scaling = core_get_scaling,
1567 .get_val = core_get_val,
1568 .get_target_pstate = get_target_pstate_use_cpu_load,
1572 static void intel_pstate_get_min_max(struct cpudata *cpu, int *min, int *max)
1574 int max_perf = cpu->pstate.turbo_pstate;
1578 if (global.no_turbo || global.turbo_disabled)
1579 max_perf = cpu->pstate.max_pstate;
1582 * performance can be limited by user through sysfs, by cpufreq
1583 * policy, or by cpu specific default values determined through
1586 max_perf_adj = fp_ext_toint(max_perf * cpu->max_perf);
1587 *max = clamp_t(int, max_perf_adj,
1588 cpu->pstate.min_pstate, cpu->pstate.turbo_pstate);
1590 min_perf = fp_ext_toint(max_perf * cpu->min_perf);
1591 *min = clamp_t(int, min_perf, cpu->pstate.min_pstate, max_perf);
1594 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
1596 trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
1597 cpu->pstate.current_pstate = pstate;
1599 * Generally, there is no guarantee that this code will always run on
1600 * the CPU being updated, so force the register update to run on the
1603 wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
1604 pstate_funcs.get_val(cpu, pstate));
1607 static void intel_pstate_set_min_pstate(struct cpudata *cpu)
1609 intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
1612 static void intel_pstate_max_within_limits(struct cpudata *cpu)
1614 int min_pstate, max_pstate;
1616 update_turbo_state();
1617 intel_pstate_get_min_max(cpu, &min_pstate, &max_pstate);
1618 intel_pstate_set_pstate(cpu, max_pstate);
1621 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
1623 cpu->pstate.min_pstate = pstate_funcs.get_min();
1624 cpu->pstate.max_pstate = pstate_funcs.get_max();
1625 cpu->pstate.max_pstate_physical = pstate_funcs.get_max_physical();
1626 cpu->pstate.turbo_pstate = pstate_funcs.get_turbo();
1627 cpu->pstate.scaling = pstate_funcs.get_scaling();
1628 cpu->pstate.max_freq = cpu->pstate.max_pstate * cpu->pstate.scaling;
1629 cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
1631 if (pstate_funcs.get_vid)
1632 pstate_funcs.get_vid(cpu);
1634 intel_pstate_set_min_pstate(cpu);
1637 static inline void intel_pstate_calc_avg_perf(struct cpudata *cpu)
1639 struct sample *sample = &cpu->sample;
1641 sample->core_avg_perf = div_ext_fp(sample->aperf, sample->mperf);
1644 static inline bool intel_pstate_sample(struct cpudata *cpu, u64 time)
1647 unsigned long flags;
1650 local_irq_save(flags);
1651 rdmsrl(MSR_IA32_APERF, aperf);
1652 rdmsrl(MSR_IA32_MPERF, mperf);
1654 if (cpu->prev_mperf == mperf || cpu->prev_tsc == tsc) {
1655 local_irq_restore(flags);
1658 local_irq_restore(flags);
1660 cpu->last_sample_time = cpu->sample.time;
1661 cpu->sample.time = time;
1662 cpu->sample.aperf = aperf;
1663 cpu->sample.mperf = mperf;
1664 cpu->sample.tsc = tsc;
1665 cpu->sample.aperf -= cpu->prev_aperf;
1666 cpu->sample.mperf -= cpu->prev_mperf;
1667 cpu->sample.tsc -= cpu->prev_tsc;
1669 cpu->prev_aperf = aperf;
1670 cpu->prev_mperf = mperf;
1671 cpu->prev_tsc = tsc;
1673 * First time this function is invoked in a given cycle, all of the
1674 * previous sample data fields are equal to zero or stale and they must
1675 * be populated with meaningful numbers for things to work, so assume
1676 * that sample.time will always be reset before setting the utilization
1677 * update hook and make the caller skip the sample then.
1679 return !!cpu->last_sample_time;
1682 static inline int32_t get_avg_frequency(struct cpudata *cpu)
1684 return mul_ext_fp(cpu->sample.core_avg_perf,
1685 cpu->pstate.max_pstate_physical * cpu->pstate.scaling);
1688 static inline int32_t get_avg_pstate(struct cpudata *cpu)
1690 return mul_ext_fp(cpu->pstate.max_pstate_physical,
1691 cpu->sample.core_avg_perf);
1694 static inline int32_t get_target_pstate_use_cpu_load(struct cpudata *cpu)
1696 struct sample *sample = &cpu->sample;
1697 int32_t busy_frac, boost;
1698 int target, avg_pstate;
1700 busy_frac = div_fp(sample->mperf, sample->tsc);
1702 boost = cpu->iowait_boost;
1703 cpu->iowait_boost >>= 1;
1705 if (busy_frac < boost)
1708 sample->busy_scaled = busy_frac * 100;
1710 target = global.no_turbo || global.turbo_disabled ?
1711 cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
1712 target += target >> 2;
1713 target = mul_fp(target, busy_frac);
1714 if (target < cpu->pstate.min_pstate)
1715 target = cpu->pstate.min_pstate;
1718 * If the average P-state during the previous cycle was higher than the
1719 * current target, add 50% of the difference to the target to reduce
1720 * possible performance oscillations and offset possible performance
1721 * loss related to moving the workload from one CPU to another within
1724 avg_pstate = get_avg_pstate(cpu);
1725 if (avg_pstate > target)
1726 target += (avg_pstate - target) >> 1;
1731 static inline int32_t get_target_pstate_use_performance(struct cpudata *cpu)
1733 int32_t perf_scaled, max_pstate, current_pstate, sample_ratio;
1737 * perf_scaled is the ratio of the average P-state during the last
1738 * sampling period to the P-state requested last time (in percent).
1740 * That measures the system's response to the previous P-state
1743 max_pstate = cpu->pstate.max_pstate_physical;
1744 current_pstate = cpu->pstate.current_pstate;
1745 perf_scaled = mul_ext_fp(cpu->sample.core_avg_perf,
1746 div_fp(100 * max_pstate, current_pstate));
1749 * Since our utilization update callback will not run unless we are
1750 * in C0, check if the actual elapsed time is significantly greater (3x)
1751 * than our sample interval. If it is, then we were idle for a long
1752 * enough period of time to adjust our performance metric.
1754 duration_ns = cpu->sample.time - cpu->last_sample_time;
1755 if ((s64)duration_ns > pid_params.sample_rate_ns * 3) {
1756 sample_ratio = div_fp(pid_params.sample_rate_ns, duration_ns);
1757 perf_scaled = mul_fp(perf_scaled, sample_ratio);
1759 sample_ratio = div_fp(100 * cpu->sample.mperf, cpu->sample.tsc);
1760 if (sample_ratio < int_tofp(1))
1764 cpu->sample.busy_scaled = perf_scaled;
1765 return cpu->pstate.current_pstate - pid_calc(&cpu->pid, perf_scaled);
1768 static int intel_pstate_prepare_request(struct cpudata *cpu, int pstate)
1770 int max_perf, min_perf;
1772 intel_pstate_get_min_max(cpu, &min_perf, &max_perf);
1773 pstate = clamp_t(int, pstate, min_perf, max_perf);
1777 static void intel_pstate_update_pstate(struct cpudata *cpu, int pstate)
1779 if (pstate == cpu->pstate.current_pstate)
1782 cpu->pstate.current_pstate = pstate;
1783 wrmsrl(MSR_IA32_PERF_CTL, pstate_funcs.get_val(cpu, pstate));
1786 static inline void intel_pstate_adjust_busy_pstate(struct cpudata *cpu)
1788 int from, target_pstate;
1789 struct sample *sample;
1791 from = cpu->pstate.current_pstate;
1793 target_pstate = cpu->policy == CPUFREQ_POLICY_PERFORMANCE ?
1794 cpu->pstate.turbo_pstate : pstate_funcs.get_target_pstate(cpu);
1796 update_turbo_state();
1798 target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
1799 trace_cpu_frequency(target_pstate * cpu->pstate.scaling, cpu->cpu);
1800 intel_pstate_update_pstate(cpu, target_pstate);
1802 sample = &cpu->sample;
1803 trace_pstate_sample(mul_ext_fp(100, sample->core_avg_perf),
1804 fp_toint(sample->busy_scaled),
1806 cpu->pstate.current_pstate,
1810 get_avg_frequency(cpu),
1811 fp_toint(cpu->iowait_boost * 100));
1814 static void intel_pstate_update_util(struct update_util_data *data, u64 time,
1817 struct cpudata *cpu = container_of(data, struct cpudata, update_util);
1820 if (pstate_funcs.get_target_pstate == get_target_pstate_use_cpu_load) {
1821 if (flags & SCHED_CPUFREQ_IOWAIT) {
1822 cpu->iowait_boost = int_tofp(1);
1823 } else if (cpu->iowait_boost) {
1824 /* Clear iowait_boost if the CPU may have been idle. */
1825 delta_ns = time - cpu->last_update;
1826 if (delta_ns > TICK_NSEC)
1827 cpu->iowait_boost = 0;
1829 cpu->last_update = time;
1832 delta_ns = time - cpu->sample.time;
1833 if ((s64)delta_ns >= pid_params.sample_rate_ns) {
1834 bool sample_taken = intel_pstate_sample(cpu, time);
1837 intel_pstate_calc_avg_perf(cpu);
1839 intel_pstate_adjust_busy_pstate(cpu);
1844 #define ICPU(model, policy) \
1845 { X86_VENDOR_INTEL, 6, model, X86_FEATURE_APERFMPERF,\
1846 (unsigned long)&policy }
1848 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
1849 ICPU(INTEL_FAM6_SANDYBRIDGE, core_params),
1850 ICPU(INTEL_FAM6_SANDYBRIDGE_X, core_params),
1851 ICPU(INTEL_FAM6_ATOM_SILVERMONT1, silvermont_params),
1852 ICPU(INTEL_FAM6_IVYBRIDGE, core_params),
1853 ICPU(INTEL_FAM6_HASWELL_CORE, core_params),
1854 ICPU(INTEL_FAM6_BROADWELL_CORE, core_params),
1855 ICPU(INTEL_FAM6_IVYBRIDGE_X, core_params),
1856 ICPU(INTEL_FAM6_HASWELL_X, core_params),
1857 ICPU(INTEL_FAM6_HASWELL_ULT, core_params),
1858 ICPU(INTEL_FAM6_HASWELL_GT3E, core_params),
1859 ICPU(INTEL_FAM6_BROADWELL_GT3E, core_params),
1860 ICPU(INTEL_FAM6_ATOM_AIRMONT, airmont_params),
1861 ICPU(INTEL_FAM6_SKYLAKE_MOBILE, core_params),
1862 ICPU(INTEL_FAM6_BROADWELL_X, core_params),
1863 ICPU(INTEL_FAM6_SKYLAKE_DESKTOP, core_params),
1864 ICPU(INTEL_FAM6_BROADWELL_XEON_D, core_params),
1865 ICPU(INTEL_FAM6_XEON_PHI_KNL, knl_params),
1866 ICPU(INTEL_FAM6_XEON_PHI_KNM, knl_params),
1867 ICPU(INTEL_FAM6_ATOM_GOLDMONT, bxt_params),
1870 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
1872 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] __initconst = {
1873 ICPU(INTEL_FAM6_BROADWELL_XEON_D, core_params),
1874 ICPU(INTEL_FAM6_BROADWELL_X, core_params),
1875 ICPU(INTEL_FAM6_SKYLAKE_X, core_params),
1879 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[] = {
1880 ICPU(INTEL_FAM6_KABYLAKE_DESKTOP, core_params),
1884 static int intel_pstate_init_cpu(unsigned int cpunum)
1886 struct cpudata *cpu;
1888 cpu = all_cpu_data[cpunum];
1891 cpu = kzalloc(sizeof(*cpu), GFP_KERNEL);
1895 all_cpu_data[cpunum] = cpu;
1897 cpu->epp_default = -EINVAL;
1898 cpu->epp_powersave = -EINVAL;
1899 cpu->epp_saved = -EINVAL;
1902 cpu = all_cpu_data[cpunum];
1907 const struct x86_cpu_id *id;
1909 id = x86_match_cpu(intel_pstate_cpu_ee_disable_ids);
1911 intel_pstate_disable_ee(cpunum);
1913 intel_pstate_hwp_enable(cpu);
1914 } else if (pstate_funcs.get_target_pstate == get_target_pstate_use_performance) {
1915 intel_pstate_pid_reset(cpu);
1918 intel_pstate_get_cpu_pstates(cpu);
1920 pr_debug("controlling: cpu %d\n", cpunum);
1925 static unsigned int intel_pstate_get(unsigned int cpu_num)
1927 struct cpudata *cpu = all_cpu_data[cpu_num];
1929 return cpu ? get_avg_frequency(cpu) : 0;
1932 static void intel_pstate_set_update_util_hook(unsigned int cpu_num)
1934 struct cpudata *cpu = all_cpu_data[cpu_num];
1936 if (cpu->update_util_set)
1939 /* Prevent intel_pstate_update_util() from using stale data. */
1940 cpu->sample.time = 0;
1941 cpufreq_add_update_util_hook(cpu_num, &cpu->update_util,
1942 intel_pstate_update_util);
1943 cpu->update_util_set = true;
1946 static void intel_pstate_clear_update_util_hook(unsigned int cpu)
1948 struct cpudata *cpu_data = all_cpu_data[cpu];
1950 if (!cpu_data->update_util_set)
1953 cpufreq_remove_update_util_hook(cpu);
1954 cpu_data->update_util_set = false;
1955 synchronize_sched();
1958 static int intel_pstate_get_max_freq(struct cpudata *cpu)
1960 return global.turbo_disabled || global.no_turbo ?
1961 cpu->pstate.max_freq : cpu->pstate.turbo_freq;
1964 static void intel_pstate_update_perf_limits(struct cpufreq_policy *policy,
1965 struct cpudata *cpu)
1967 int max_freq = intel_pstate_get_max_freq(cpu);
1968 int32_t max_policy_perf, min_policy_perf;
1970 max_policy_perf = div_ext_fp(policy->max, max_freq);
1971 max_policy_perf = clamp_t(int32_t, max_policy_perf, 0, int_ext_tofp(1));
1972 if (policy->max == policy->min) {
1973 min_policy_perf = max_policy_perf;
1975 min_policy_perf = div_ext_fp(policy->min, max_freq);
1976 min_policy_perf = clamp_t(int32_t, min_policy_perf,
1977 0, max_policy_perf);
1980 /* Normalize user input to [min_perf, max_perf] */
1981 if (per_cpu_limits) {
1982 cpu->min_perf = min_policy_perf;
1983 cpu->max_perf = max_policy_perf;
1985 int32_t global_min, global_max;
1987 /* Global limits are in percent of the maximum turbo P-state. */
1988 global_max = percent_ext_fp(global.max_perf_pct);
1989 global_min = percent_ext_fp(global.min_perf_pct);
1990 if (max_freq != cpu->pstate.turbo_freq) {
1991 int32_t turbo_factor;
1993 turbo_factor = div_ext_fp(cpu->pstate.turbo_pstate,
1994 cpu->pstate.max_pstate);
1995 global_min = mul_ext_fp(global_min, turbo_factor);
1996 global_max = mul_ext_fp(global_max, turbo_factor);
1998 global_min = clamp_t(int32_t, global_min, 0, global_max);
2000 cpu->min_perf = max(min_policy_perf, global_min);
2001 cpu->min_perf = min(cpu->min_perf, max_policy_perf);
2002 cpu->max_perf = min(max_policy_perf, global_max);
2003 cpu->max_perf = max(min_policy_perf, cpu->max_perf);
2005 /* Make sure min_perf <= max_perf */
2006 cpu->min_perf = min(cpu->min_perf, cpu->max_perf);
2009 cpu->max_perf = round_up(cpu->max_perf, EXT_FRAC_BITS);
2010 cpu->min_perf = round_up(cpu->min_perf, EXT_FRAC_BITS);
2012 pr_debug("cpu:%d max_perf_pct:%d min_perf_pct:%d\n", policy->cpu,
2013 fp_ext_toint(cpu->max_perf * 100),
2014 fp_ext_toint(cpu->min_perf * 100));
2017 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
2019 struct cpudata *cpu;
2021 if (!policy->cpuinfo.max_freq)
2024 pr_debug("set_policy cpuinfo.max %u policy->max %u\n",
2025 policy->cpuinfo.max_freq, policy->max);
2027 cpu = all_cpu_data[policy->cpu];
2028 cpu->policy = policy->policy;
2030 mutex_lock(&intel_pstate_limits_lock);
2032 intel_pstate_update_perf_limits(policy, cpu);
2034 if (cpu->policy == CPUFREQ_POLICY_PERFORMANCE) {
2036 * NOHZ_FULL CPUs need this as the governor callback may not
2037 * be invoked on them.
2039 intel_pstate_clear_update_util_hook(policy->cpu);
2040 intel_pstate_max_within_limits(cpu);
2043 intel_pstate_set_update_util_hook(policy->cpu);
2046 intel_pstate_hwp_set(policy);
2048 mutex_unlock(&intel_pstate_limits_lock);
2053 static void intel_pstate_adjust_policy_max(struct cpufreq_policy *policy,
2054 struct cpudata *cpu)
2056 if (cpu->pstate.max_pstate_physical > cpu->pstate.max_pstate &&
2057 policy->max < policy->cpuinfo.max_freq &&
2058 policy->max > cpu->pstate.max_freq) {
2059 pr_debug("policy->max > max non turbo frequency\n");
2060 policy->max = policy->cpuinfo.max_freq;
2064 static int intel_pstate_verify_policy(struct cpufreq_policy *policy)
2066 struct cpudata *cpu = all_cpu_data[policy->cpu];
2068 update_turbo_state();
2069 cpufreq_verify_within_limits(policy, policy->cpuinfo.min_freq,
2070 intel_pstate_get_max_freq(cpu));
2072 if (policy->policy != CPUFREQ_POLICY_POWERSAVE &&
2073 policy->policy != CPUFREQ_POLICY_PERFORMANCE)
2076 intel_pstate_adjust_policy_max(policy, cpu);
2081 static void intel_cpufreq_stop_cpu(struct cpufreq_policy *policy)
2083 intel_pstate_set_min_pstate(all_cpu_data[policy->cpu]);
2086 static void intel_pstate_stop_cpu(struct cpufreq_policy *policy)
2088 pr_debug("CPU %d exiting\n", policy->cpu);
2090 intel_pstate_clear_update_util_hook(policy->cpu);
2092 intel_pstate_hwp_save_state(policy);
2094 intel_cpufreq_stop_cpu(policy);
2097 static int intel_pstate_cpu_exit(struct cpufreq_policy *policy)
2099 intel_pstate_exit_perf_limits(policy);
2101 policy->fast_switch_possible = false;
2106 static int __intel_pstate_cpu_init(struct cpufreq_policy *policy)
2108 struct cpudata *cpu;
2111 rc = intel_pstate_init_cpu(policy->cpu);
2115 cpu = all_cpu_data[policy->cpu];
2117 cpu->max_perf = int_ext_tofp(1);
2120 policy->min = cpu->pstate.min_pstate * cpu->pstate.scaling;
2121 policy->max = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
2123 /* cpuinfo and default policy values */
2124 policy->cpuinfo.min_freq = cpu->pstate.min_pstate * cpu->pstate.scaling;
2125 update_turbo_state();
2126 policy->cpuinfo.max_freq = global.turbo_disabled ?
2127 cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
2128 policy->cpuinfo.max_freq *= cpu->pstate.scaling;
2130 intel_pstate_init_acpi_perf_limits(policy);
2131 cpumask_set_cpu(policy->cpu, policy->cpus);
2133 policy->fast_switch_possible = true;
2138 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
2140 int ret = __intel_pstate_cpu_init(policy);
2145 policy->cpuinfo.transition_latency = CPUFREQ_ETERNAL;
2146 if (IS_ENABLED(CONFIG_CPU_FREQ_DEFAULT_GOV_PERFORMANCE))
2147 policy->policy = CPUFREQ_POLICY_PERFORMANCE;
2149 policy->policy = CPUFREQ_POLICY_POWERSAVE;
2154 static struct cpufreq_driver intel_pstate = {
2155 .flags = CPUFREQ_CONST_LOOPS,
2156 .verify = intel_pstate_verify_policy,
2157 .setpolicy = intel_pstate_set_policy,
2158 .suspend = intel_pstate_hwp_save_state,
2159 .resume = intel_pstate_resume,
2160 .get = intel_pstate_get,
2161 .init = intel_pstate_cpu_init,
2162 .exit = intel_pstate_cpu_exit,
2163 .stop_cpu = intel_pstate_stop_cpu,
2164 .name = "intel_pstate",
2167 static int intel_cpufreq_verify_policy(struct cpufreq_policy *policy)
2169 struct cpudata *cpu = all_cpu_data[policy->cpu];
2171 update_turbo_state();
2172 cpufreq_verify_within_limits(policy, policy->cpuinfo.min_freq,
2173 intel_pstate_get_max_freq(cpu));
2175 intel_pstate_adjust_policy_max(policy, cpu);
2177 intel_pstate_update_perf_limits(policy, cpu);
2182 static int intel_cpufreq_target(struct cpufreq_policy *policy,
2183 unsigned int target_freq,
2184 unsigned int relation)
2186 struct cpudata *cpu = all_cpu_data[policy->cpu];
2187 struct cpufreq_freqs freqs;
2190 update_turbo_state();
2192 freqs.old = policy->cur;
2193 freqs.new = target_freq;
2195 cpufreq_freq_transition_begin(policy, &freqs);
2197 case CPUFREQ_RELATION_L:
2198 target_pstate = DIV_ROUND_UP(freqs.new, cpu->pstate.scaling);
2200 case CPUFREQ_RELATION_H:
2201 target_pstate = freqs.new / cpu->pstate.scaling;
2204 target_pstate = DIV_ROUND_CLOSEST(freqs.new, cpu->pstate.scaling);
2207 target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2208 if (target_pstate != cpu->pstate.current_pstate) {
2209 cpu->pstate.current_pstate = target_pstate;
2210 wrmsrl_on_cpu(policy->cpu, MSR_IA32_PERF_CTL,
2211 pstate_funcs.get_val(cpu, target_pstate));
2213 freqs.new = target_pstate * cpu->pstate.scaling;
2214 cpufreq_freq_transition_end(policy, &freqs, false);
2219 static unsigned int intel_cpufreq_fast_switch(struct cpufreq_policy *policy,
2220 unsigned int target_freq)
2222 struct cpudata *cpu = all_cpu_data[policy->cpu];
2225 update_turbo_state();
2227 target_pstate = DIV_ROUND_UP(target_freq, cpu->pstate.scaling);
2228 target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2229 intel_pstate_update_pstate(cpu, target_pstate);
2230 return target_pstate * cpu->pstate.scaling;
2233 static int intel_cpufreq_cpu_init(struct cpufreq_policy *policy)
2235 int ret = __intel_pstate_cpu_init(policy);
2240 policy->cpuinfo.transition_latency = INTEL_CPUFREQ_TRANSITION_LATENCY;
2241 /* This reflects the intel_pstate_get_cpu_pstates() setting. */
2242 policy->cur = policy->cpuinfo.min_freq;
2247 static struct cpufreq_driver intel_cpufreq = {
2248 .flags = CPUFREQ_CONST_LOOPS,
2249 .verify = intel_cpufreq_verify_policy,
2250 .target = intel_cpufreq_target,
2251 .fast_switch = intel_cpufreq_fast_switch,
2252 .init = intel_cpufreq_cpu_init,
2253 .exit = intel_pstate_cpu_exit,
2254 .stop_cpu = intel_cpufreq_stop_cpu,
2255 .name = "intel_cpufreq",
2258 static struct cpufreq_driver *default_driver = &intel_pstate;
2260 static void intel_pstate_driver_cleanup(void)
2265 for_each_online_cpu(cpu) {
2266 if (all_cpu_data[cpu]) {
2267 if (intel_pstate_driver == &intel_pstate)
2268 intel_pstate_clear_update_util_hook(cpu);
2270 kfree(all_cpu_data[cpu]);
2271 all_cpu_data[cpu] = NULL;
2275 intel_pstate_driver = NULL;
2278 static int intel_pstate_register_driver(struct cpufreq_driver *driver)
2282 memset(&global, 0, sizeof(global));
2283 global.max_perf_pct = 100;
2285 intel_pstate_driver = driver;
2286 ret = cpufreq_register_driver(intel_pstate_driver);
2288 intel_pstate_driver_cleanup();
2292 global.min_perf_pct = min_perf_pct_min();
2294 if (intel_pstate_driver == &intel_pstate && !hwp_active &&
2295 pstate_funcs.get_target_pstate != get_target_pstate_use_cpu_load)
2296 intel_pstate_debug_expose_params();
2301 static int intel_pstate_unregister_driver(void)
2306 if (intel_pstate_driver == &intel_pstate && !hwp_active &&
2307 pstate_funcs.get_target_pstate != get_target_pstate_use_cpu_load)
2308 intel_pstate_debug_hide_params();
2310 cpufreq_unregister_driver(intel_pstate_driver);
2311 intel_pstate_driver_cleanup();
2316 static ssize_t intel_pstate_show_status(char *buf)
2318 if (!intel_pstate_driver)
2319 return sprintf(buf, "off\n");
2321 return sprintf(buf, "%s\n", intel_pstate_driver == &intel_pstate ?
2322 "active" : "passive");
2325 static int intel_pstate_update_status(const char *buf, size_t size)
2329 if (size == 3 && !strncmp(buf, "off", size))
2330 return intel_pstate_driver ?
2331 intel_pstate_unregister_driver() : -EINVAL;
2333 if (size == 6 && !strncmp(buf, "active", size)) {
2334 if (intel_pstate_driver) {
2335 if (intel_pstate_driver == &intel_pstate)
2338 ret = intel_pstate_unregister_driver();
2343 return intel_pstate_register_driver(&intel_pstate);
2346 if (size == 7 && !strncmp(buf, "passive", size)) {
2347 if (intel_pstate_driver) {
2348 if (intel_pstate_driver != &intel_pstate)
2351 ret = intel_pstate_unregister_driver();
2356 return intel_pstate_register_driver(&intel_cpufreq);
2362 static int no_load __initdata;
2363 static int no_hwp __initdata;
2364 static int hwp_only __initdata;
2365 static unsigned int force_load __initdata;
2367 static int __init intel_pstate_msrs_not_valid(void)
2369 if (!pstate_funcs.get_max() ||
2370 !pstate_funcs.get_min() ||
2371 !pstate_funcs.get_turbo())
2378 static void intel_pstate_use_acpi_profile(void)
2380 switch (acpi_gbl_FADT.preferred_profile) {
2383 case PM_APPLIANCE_PC:
2385 case PM_WORKSTATION:
2386 pstate_funcs.get_target_pstate =
2387 get_target_pstate_use_cpu_load;
2391 static void intel_pstate_use_acpi_profile(void)
2396 static void __init copy_cpu_funcs(struct pstate_funcs *funcs)
2398 pstate_funcs.get_max = funcs->get_max;
2399 pstate_funcs.get_max_physical = funcs->get_max_physical;
2400 pstate_funcs.get_min = funcs->get_min;
2401 pstate_funcs.get_turbo = funcs->get_turbo;
2402 pstate_funcs.get_scaling = funcs->get_scaling;
2403 pstate_funcs.get_val = funcs->get_val;
2404 pstate_funcs.get_vid = funcs->get_vid;
2405 pstate_funcs.get_target_pstate = funcs->get_target_pstate;
2407 intel_pstate_use_acpi_profile();
2412 static bool __init intel_pstate_no_acpi_pss(void)
2416 for_each_possible_cpu(i) {
2418 union acpi_object *pss;
2419 struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
2420 struct acpi_processor *pr = per_cpu(processors, i);
2425 status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
2426 if (ACPI_FAILURE(status))
2429 pss = buffer.pointer;
2430 if (pss && pss->type == ACPI_TYPE_PACKAGE) {
2441 static bool __init intel_pstate_has_acpi_ppc(void)
2445 for_each_possible_cpu(i) {
2446 struct acpi_processor *pr = per_cpu(processors, i);
2450 if (acpi_has_method(pr->handle, "_PPC"))
2461 struct hw_vendor_info {
2463 char oem_id[ACPI_OEM_ID_SIZE];
2464 char oem_table_id[ACPI_OEM_TABLE_ID_SIZE];
2468 /* Hardware vendor-specific info that has its own power management modes */
2469 static struct hw_vendor_info vendor_info[] __initdata = {
2470 {1, "HP ", "ProLiant", PSS},
2471 {1, "ORACLE", "X4-2 ", PPC},
2472 {1, "ORACLE", "X4-2L ", PPC},
2473 {1, "ORACLE", "X4-2B ", PPC},
2474 {1, "ORACLE", "X3-2 ", PPC},
2475 {1, "ORACLE", "X3-2L ", PPC},
2476 {1, "ORACLE", "X3-2B ", PPC},
2477 {1, "ORACLE", "X4470M2 ", PPC},
2478 {1, "ORACLE", "X4270M3 ", PPC},
2479 {1, "ORACLE", "X4270M2 ", PPC},
2480 {1, "ORACLE", "X4170M2 ", PPC},
2481 {1, "ORACLE", "X4170 M3", PPC},
2482 {1, "ORACLE", "X4275 M3", PPC},
2483 {1, "ORACLE", "X6-2 ", PPC},
2484 {1, "ORACLE", "Sudbury ", PPC},
2488 static bool __init intel_pstate_platform_pwr_mgmt_exists(void)
2490 struct acpi_table_header hdr;
2491 struct hw_vendor_info *v_info;
2492 const struct x86_cpu_id *id;
2495 id = x86_match_cpu(intel_pstate_cpu_oob_ids);
2497 rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
2498 if ( misc_pwr & (1 << 8))
2502 if (acpi_disabled ||
2503 ACPI_FAILURE(acpi_get_table_header(ACPI_SIG_FADT, 0, &hdr)))
2506 for (v_info = vendor_info; v_info->valid; v_info++) {
2507 if (!strncmp(hdr.oem_id, v_info->oem_id, ACPI_OEM_ID_SIZE) &&
2508 !strncmp(hdr.oem_table_id, v_info->oem_table_id,
2509 ACPI_OEM_TABLE_ID_SIZE))
2510 switch (v_info->oem_pwr_table) {
2512 return intel_pstate_no_acpi_pss();
2514 return intel_pstate_has_acpi_ppc() &&
2522 static void intel_pstate_request_control_from_smm(void)
2525 * It may be unsafe to request P-states control from SMM if _PPC support
2526 * has not been enabled.
2529 acpi_processor_pstate_control();
2531 #else /* CONFIG_ACPI not enabled */
2532 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
2533 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
2534 static inline void intel_pstate_request_control_from_smm(void) {}
2535 #endif /* CONFIG_ACPI */
2537 static const struct x86_cpu_id hwp_support_ids[] __initconst = {
2538 { X86_VENDOR_INTEL, 6, X86_MODEL_ANY, X86_FEATURE_HWP },
2542 static int __init intel_pstate_init(void)
2549 if (x86_match_cpu(hwp_support_ids)) {
2550 copy_cpu_funcs(&core_params.funcs);
2552 pstate_funcs.get_target_pstate = get_target_pstate_use_cpu_load;
2555 intel_pstate.attr = hwp_cpufreq_attrs;
2556 pid_params.sample_rate_ns = 50 * NSEC_PER_MSEC;
2557 goto hwp_cpu_matched;
2560 const struct x86_cpu_id *id;
2561 struct cpu_defaults *cpu_def;
2563 id = x86_match_cpu(intel_pstate_cpu_ids);
2567 cpu_def = (struct cpu_defaults *)id->driver_data;
2568 copy_cpu_funcs(&cpu_def->funcs);
2571 if (intel_pstate_msrs_not_valid())
2576 * The Intel pstate driver will be ignored if the platform
2577 * firmware has its own power management modes.
2579 if (intel_pstate_platform_pwr_mgmt_exists())
2582 if (!hwp_active && hwp_only)
2585 pr_info("Intel P-state driver initializing\n");
2587 all_cpu_data = vzalloc(sizeof(void *) * num_possible_cpus());
2591 intel_pstate_request_control_from_smm();
2593 intel_pstate_sysfs_expose_params();
2595 mutex_lock(&intel_pstate_driver_lock);
2596 rc = intel_pstate_register_driver(default_driver);
2597 mutex_unlock(&intel_pstate_driver_lock);
2602 pr_info("HWP enabled\n");
2606 device_initcall(intel_pstate_init);
2608 static int __init intel_pstate_setup(char *str)
2613 if (!strcmp(str, "disable")) {
2615 } else if (!strcmp(str, "passive")) {
2616 pr_info("Passive mode enabled\n");
2617 default_driver = &intel_cpufreq;
2620 if (!strcmp(str, "no_hwp")) {
2621 pr_info("HWP disabled\n");
2624 if (!strcmp(str, "force"))
2626 if (!strcmp(str, "hwp_only"))
2628 if (!strcmp(str, "per_cpu_perf_limits"))
2629 per_cpu_limits = true;
2632 if (!strcmp(str, "support_acpi_ppc"))
2638 early_param("intel_pstate", intel_pstate_setup);
2640 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
2641 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
2642 MODULE_LICENSE("GPL");