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.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
42 #define ATOM_RATIOS 0x66a
43 #define ATOM_VIDS 0x66b
44 #define ATOM_TURBO_RATIOS 0x66c
45 #define ATOM_TURBO_VIDS 0x66d
48 #include <acpi/processor.h>
49 #include <acpi/cppc_acpi.h>
53 #define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
54 #define fp_toint(X) ((X) >> FRAC_BITS)
57 #define EXT_FRAC_BITS (EXT_BITS + FRAC_BITS)
58 #define fp_ext_toint(X) ((X) >> EXT_FRAC_BITS)
59 #define int_ext_tofp(X) ((int64_t)(X) << EXT_FRAC_BITS)
61 static inline int32_t mul_fp(int32_t x, int32_t y)
63 return ((int64_t)x * (int64_t)y) >> FRAC_BITS;
66 static inline int32_t div_fp(s64 x, s64 y)
68 return div64_s64((int64_t)x << FRAC_BITS, y);
71 static inline int ceiling_fp(int32_t x)
76 mask = (1 << FRAC_BITS) - 1;
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);
93 * struct sample - Store performance sample
94 * @core_avg_perf: Ratio of APERF/MPERF which is the actual average
95 * performance during last sample period
96 * @busy_scaled: Scaled busy value which is used to calculate next
97 * P state. This can be different than core_avg_perf
98 * to account for cpu idle period
99 * @aperf: Difference of actual performance frequency clock count
100 * read from APERF MSR between last and current sample
101 * @mperf: Difference of maximum performance frequency clock count
102 * read from MPERF MSR between last and current sample
103 * @tsc: Difference of time stamp counter between last and
105 * @time: Current time from scheduler
107 * This structure is used in the cpudata structure to store performance sample
108 * data for choosing next P State.
111 int32_t core_avg_perf;
120 * struct pstate_data - Store P state data
121 * @current_pstate: Current requested P state
122 * @min_pstate: Min P state possible for this platform
123 * @max_pstate: Max P state possible for this platform
124 * @max_pstate_physical:This is physical Max P state for a processor
125 * This can be higher than the max_pstate which can
126 * be limited by platform thermal design power limits
127 * @scaling: Scaling factor to convert frequency to cpufreq
129 * @turbo_pstate: Max Turbo P state possible for this platform
130 * @max_freq: @max_pstate frequency in cpufreq units
131 * @turbo_freq: @turbo_pstate frequency in cpufreq units
133 * Stores the per cpu model P state limits and current P state.
139 int max_pstate_physical;
142 unsigned int max_freq;
143 unsigned int turbo_freq;
147 * struct vid_data - Stores voltage information data
148 * @min: VID data for this platform corresponding to
150 * @max: VID data corresponding to the highest P State.
151 * @turbo: VID data for turbo P state
152 * @ratio: Ratio of (vid max - vid min) /
153 * (max P state - Min P State)
155 * Stores the voltage data for DVFS (Dynamic Voltage and Frequency Scaling)
156 * This data is used in Atom platforms, where in addition to target P state,
157 * the voltage data needs to be specified to select next P State.
167 * struct _pid - Stores PID data
168 * @setpoint: Target set point for busyness or performance
169 * @integral: Storage for accumulated error values
170 * @p_gain: PID proportional gain
171 * @i_gain: PID integral gain
172 * @d_gain: PID derivative gain
173 * @deadband: PID deadband
174 * @last_err: Last error storage for integral part of PID calculation
176 * Stores PID coefficients and last error for PID controller.
189 * struct perf_limits - Store user and policy limits
190 * @no_turbo: User requested turbo state from intel_pstate sysfs
191 * @turbo_disabled: Platform turbo status either from msr
192 * MSR_IA32_MISC_ENABLE or when maximum available pstate
193 * matches the maximum turbo pstate
194 * @max_perf_pct: Effective maximum performance limit in percentage, this
195 * is minimum of either limits enforced by cpufreq policy
196 * or limits from user set limits via intel_pstate sysfs
197 * @min_perf_pct: Effective minimum performance limit in percentage, this
198 * is maximum of either limits enforced by cpufreq policy
199 * or limits from user set limits via intel_pstate sysfs
200 * @max_perf: This is a scaled value between 0 to 255 for max_perf_pct
201 * This value is used to limit max pstate
202 * @min_perf: This is a scaled value between 0 to 255 for min_perf_pct
203 * This value is used to limit min pstate
204 * @max_policy_pct: The maximum performance in percentage enforced by
205 * cpufreq setpolicy interface
206 * @max_sysfs_pct: The maximum performance in percentage enforced by
207 * intel pstate sysfs interface, unused when per cpu
208 * controls are enforced
209 * @min_policy_pct: The minimum performance in percentage enforced by
210 * cpufreq setpolicy interface
211 * @min_sysfs_pct: The minimum performance in percentage enforced by
212 * intel pstate sysfs interface, unused when per cpu
213 * controls are enforced
215 * Storage for user and policy defined limits.
231 * struct cpudata - Per CPU instance data storage
232 * @cpu: CPU number for this instance data
233 * @policy: CPUFreq policy value
234 * @update_util: CPUFreq utility callback information
235 * @update_util_set: CPUFreq utility callback is set
236 * @iowait_boost: iowait-related boost fraction
237 * @last_update: Time of the last update.
238 * @pstate: Stores P state limits for this CPU
239 * @vid: Stores VID limits for this CPU
240 * @pid: Stores PID parameters for this CPU
241 * @last_sample_time: Last Sample time
242 * @prev_aperf: Last APERF value read from APERF MSR
243 * @prev_mperf: Last MPERF value read from MPERF MSR
244 * @prev_tsc: Last timestamp counter (TSC) value
245 * @prev_cummulative_iowait: IO Wait time difference from last and
247 * @sample: Storage for storing last Sample data
248 * @perf_limits: Pointer to perf_limit unique to this CPU
249 * Not all field in the structure are applicable
250 * when per cpu controls are enforced
251 * @acpi_perf_data: Stores ACPI perf information read from _PSS
252 * @valid_pss_table: Set to true for valid ACPI _PSS entries found
253 * @epp_powersave: Last saved HWP energy performance preference
254 * (EPP) or energy performance bias (EPB),
255 * when policy switched to performance
256 * @epp_policy: Last saved policy used to set EPP/EPB
257 * @epp_default: Power on default HWP energy performance
259 * @epp_saved: Saved EPP/EPB during system suspend or CPU offline
262 * This structure stores per CPU instance data for all CPUs.
268 struct update_util_data update_util;
269 bool update_util_set;
271 struct pstate_data pstate;
276 u64 last_sample_time;
280 u64 prev_cummulative_iowait;
281 struct sample sample;
282 struct perf_limits *perf_limits;
284 struct acpi_processor_performance acpi_perf_data;
285 bool valid_pss_table;
287 unsigned int iowait_boost;
294 static struct cpudata **all_cpu_data;
297 * struct pstate_adjust_policy - Stores static PID configuration data
298 * @sample_rate_ms: PID calculation sample rate in ms
299 * @sample_rate_ns: Sample rate calculation in ns
300 * @deadband: PID deadband
301 * @setpoint: PID Setpoint
302 * @p_gain_pct: PID proportional gain
303 * @i_gain_pct: PID integral gain
304 * @d_gain_pct: PID derivative gain
306 * Stores per CPU model static PID configuration data.
308 struct pstate_adjust_policy {
319 * struct pstate_funcs - Per CPU model specific callbacks
320 * @get_max: Callback to get maximum non turbo effective P state
321 * @get_max_physical: Callback to get maximum non turbo physical P state
322 * @get_min: Callback to get minimum P state
323 * @get_turbo: Callback to get turbo P state
324 * @get_scaling: Callback to get frequency scaling factor
325 * @get_val: Callback to convert P state to actual MSR write value
326 * @get_vid: Callback to get VID data for Atom platforms
327 * @get_target_pstate: Callback to a function to calculate next P state to use
329 * Core and Atom CPU models have different way to get P State limits. This
330 * structure is used to store those callbacks.
332 struct pstate_funcs {
333 int (*get_max)(void);
334 int (*get_max_physical)(void);
335 int (*get_min)(void);
336 int (*get_turbo)(void);
337 int (*get_scaling)(void);
338 u64 (*get_val)(struct cpudata*, int pstate);
339 void (*get_vid)(struct cpudata *);
340 int32_t (*get_target_pstate)(struct cpudata *);
344 * struct cpu_defaults- Per CPU model default config data
345 * @pid_policy: PID config data
346 * @funcs: Callback function data
348 struct cpu_defaults {
349 struct pstate_adjust_policy pid_policy;
350 struct pstate_funcs funcs;
353 static inline int32_t get_target_pstate_use_performance(struct cpudata *cpu);
354 static inline int32_t get_target_pstate_use_cpu_load(struct cpudata *cpu);
356 static struct pstate_adjust_policy pid_params __read_mostly;
357 static struct pstate_funcs pstate_funcs __read_mostly;
358 static int hwp_active __read_mostly;
359 static bool per_cpu_limits __read_mostly;
361 static bool driver_registered __read_mostly;
364 static bool acpi_ppc;
367 static struct perf_limits performance_limits = {
371 .max_perf = int_ext_tofp(1),
373 .min_perf = int_ext_tofp(1),
374 .max_policy_pct = 100,
375 .max_sysfs_pct = 100,
380 static struct perf_limits powersave_limits = {
384 .max_perf = int_ext_tofp(1),
387 .max_policy_pct = 100,
388 .max_sysfs_pct = 100,
393 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_PERFORMANCE
394 static struct perf_limits *limits = &performance_limits;
396 static struct perf_limits *limits = &powersave_limits;
399 static DEFINE_MUTEX(intel_pstate_driver_lock);
400 static DEFINE_MUTEX(intel_pstate_limits_lock);
404 static bool intel_pstate_get_ppc_enable_status(void)
406 if (acpi_gbl_FADT.preferred_profile == PM_ENTERPRISE_SERVER ||
407 acpi_gbl_FADT.preferred_profile == PM_PERFORMANCE_SERVER)
413 #ifdef CONFIG_ACPI_CPPC_LIB
415 /* The work item is needed to avoid CPU hotplug locking issues */
416 static void intel_pstste_sched_itmt_work_fn(struct work_struct *work)
418 sched_set_itmt_support();
421 static DECLARE_WORK(sched_itmt_work, intel_pstste_sched_itmt_work_fn);
423 static void intel_pstate_set_itmt_prio(int cpu)
425 struct cppc_perf_caps cppc_perf;
426 static u32 max_highest_perf = 0, min_highest_perf = U32_MAX;
429 ret = cppc_get_perf_caps(cpu, &cppc_perf);
434 * The priorities can be set regardless of whether or not
435 * sched_set_itmt_support(true) has been called and it is valid to
436 * update them at any time after it has been called.
438 sched_set_itmt_core_prio(cppc_perf.highest_perf, cpu);
440 if (max_highest_perf <= min_highest_perf) {
441 if (cppc_perf.highest_perf > max_highest_perf)
442 max_highest_perf = cppc_perf.highest_perf;
444 if (cppc_perf.highest_perf < min_highest_perf)
445 min_highest_perf = cppc_perf.highest_perf;
447 if (max_highest_perf > min_highest_perf) {
449 * This code can be run during CPU online under the
450 * CPU hotplug locks, so sched_set_itmt_support()
451 * cannot be called from here. Queue up a work item
454 schedule_work(&sched_itmt_work);
459 static void intel_pstate_set_itmt_prio(int cpu)
464 static void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
471 intel_pstate_set_itmt_prio(policy->cpu);
475 if (!intel_pstate_get_ppc_enable_status())
478 cpu = all_cpu_data[policy->cpu];
480 ret = acpi_processor_register_performance(&cpu->acpi_perf_data,
486 * Check if the control value in _PSS is for PERF_CTL MSR, which should
487 * guarantee that the states returned by it map to the states in our
490 if (cpu->acpi_perf_data.control_register.space_id !=
491 ACPI_ADR_SPACE_FIXED_HARDWARE)
495 * If there is only one entry _PSS, simply ignore _PSS and continue as
496 * usual without taking _PSS into account
498 if (cpu->acpi_perf_data.state_count < 2)
501 pr_debug("CPU%u - ACPI _PSS perf data\n", policy->cpu);
502 for (i = 0; i < cpu->acpi_perf_data.state_count; i++) {
503 pr_debug(" %cP%d: %u MHz, %u mW, 0x%x\n",
504 (i == cpu->acpi_perf_data.state ? '*' : ' '), i,
505 (u32) cpu->acpi_perf_data.states[i].core_frequency,
506 (u32) cpu->acpi_perf_data.states[i].power,
507 (u32) cpu->acpi_perf_data.states[i].control);
511 * The _PSS table doesn't contain whole turbo frequency range.
512 * This just contains +1 MHZ above the max non turbo frequency,
513 * with control value corresponding to max turbo ratio. But
514 * when cpufreq set policy is called, it will call with this
515 * max frequency, which will cause a reduced performance as
516 * this driver uses real max turbo frequency as the max
517 * frequency. So correct this frequency in _PSS table to
518 * correct max turbo frequency based on the turbo state.
519 * Also need to convert to MHz as _PSS freq is in MHz.
521 if (!limits->turbo_disabled)
522 cpu->acpi_perf_data.states[0].core_frequency =
523 policy->cpuinfo.max_freq / 1000;
524 cpu->valid_pss_table = true;
525 pr_debug("_PPC limits will be enforced\n");
530 cpu->valid_pss_table = false;
531 acpi_processor_unregister_performance(policy->cpu);
534 static void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
538 cpu = all_cpu_data[policy->cpu];
539 if (!cpu->valid_pss_table)
542 acpi_processor_unregister_performance(policy->cpu);
545 static inline void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
549 static inline void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
554 static inline void pid_reset(struct _pid *pid, int setpoint, int busy,
555 int deadband, int integral) {
556 pid->setpoint = int_tofp(setpoint);
557 pid->deadband = int_tofp(deadband);
558 pid->integral = int_tofp(integral);
559 pid->last_err = int_tofp(setpoint) - int_tofp(busy);
562 static inline void pid_p_gain_set(struct _pid *pid, int percent)
564 pid->p_gain = div_fp(percent, 100);
567 static inline void pid_i_gain_set(struct _pid *pid, int percent)
569 pid->i_gain = div_fp(percent, 100);
572 static inline void pid_d_gain_set(struct _pid *pid, int percent)
574 pid->d_gain = div_fp(percent, 100);
577 static signed int pid_calc(struct _pid *pid, int32_t busy)
580 int32_t pterm, dterm, fp_error;
581 int32_t integral_limit;
583 fp_error = pid->setpoint - busy;
585 if (abs(fp_error) <= pid->deadband)
588 pterm = mul_fp(pid->p_gain, fp_error);
590 pid->integral += fp_error;
593 * We limit the integral here so that it will never
594 * get higher than 30. This prevents it from becoming
595 * too large an input over long periods of time and allows
596 * it to get factored out sooner.
598 * The value of 30 was chosen through experimentation.
600 integral_limit = int_tofp(30);
601 if (pid->integral > integral_limit)
602 pid->integral = integral_limit;
603 if (pid->integral < -integral_limit)
604 pid->integral = -integral_limit;
606 dterm = mul_fp(pid->d_gain, fp_error - pid->last_err);
607 pid->last_err = fp_error;
609 result = pterm + mul_fp(pid->integral, pid->i_gain) + dterm;
610 result = result + (1 << (FRAC_BITS-1));
611 return (signed int)fp_toint(result);
614 static inline void intel_pstate_busy_pid_reset(struct cpudata *cpu)
616 pid_p_gain_set(&cpu->pid, pid_params.p_gain_pct);
617 pid_d_gain_set(&cpu->pid, pid_params.d_gain_pct);
618 pid_i_gain_set(&cpu->pid, pid_params.i_gain_pct);
620 pid_reset(&cpu->pid, pid_params.setpoint, 100, pid_params.deadband, 0);
623 static inline void intel_pstate_reset_all_pid(void)
627 for_each_online_cpu(cpu) {
628 if (all_cpu_data[cpu])
629 intel_pstate_busy_pid_reset(all_cpu_data[cpu]);
633 static inline void update_turbo_state(void)
638 cpu = all_cpu_data[0];
639 rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
640 limits->turbo_disabled =
641 (misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ||
642 cpu->pstate.max_pstate == cpu->pstate.turbo_pstate);
645 static s16 intel_pstate_get_epb(struct cpudata *cpu_data)
650 if (!static_cpu_has(X86_FEATURE_EPB))
653 ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
657 return (s16)(epb & 0x0f);
660 static s16 intel_pstate_get_epp(struct cpudata *cpu_data, u64 hwp_req_data)
664 if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
666 * When hwp_req_data is 0, means that caller didn't read
667 * MSR_HWP_REQUEST, so need to read and get EPP.
670 epp = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST,
675 epp = (hwp_req_data >> 24) & 0xff;
677 /* When there is no EPP present, HWP uses EPB settings */
678 epp = intel_pstate_get_epb(cpu_data);
684 static int intel_pstate_set_epb(int cpu, s16 pref)
689 if (!static_cpu_has(X86_FEATURE_EPB))
692 ret = rdmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
696 epb = (epb & ~0x0f) | pref;
697 wrmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, epb);
703 * EPP/EPB display strings corresponding to EPP index in the
704 * energy_perf_strings[]
706 *-------------------------------------
709 * 2 balance_performance
713 static const char * const energy_perf_strings[] = {
716 "balance_performance",
722 static int intel_pstate_get_energy_pref_index(struct cpudata *cpu_data)
727 epp = intel_pstate_get_epp(cpu_data, 0);
731 if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
734 * 0x00-0x3F : Performance
735 * 0x40-0x7F : Balance performance
736 * 0x80-0xBF : Balance power
738 * The EPP is a 8 bit value, but our ranges restrict the
739 * value which can be set. Here only using top two bits
742 index = (epp >> 6) + 1;
743 } else if (static_cpu_has(X86_FEATURE_EPB)) {
746 * 0x00-0x03 : Performance
747 * 0x04-0x07 : Balance performance
748 * 0x08-0x0B : Balance power
750 * The EPB is a 4 bit value, but our ranges restrict the
751 * value which can be set. Here only using top two bits
754 index = (epp >> 2) + 1;
760 static int intel_pstate_set_energy_pref_index(struct cpudata *cpu_data,
767 epp = cpu_data->epp_default;
769 mutex_lock(&intel_pstate_limits_lock);
771 if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
774 ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST, &value);
778 value &= ~GENMASK_ULL(31, 24);
781 * If epp is not default, convert from index into
782 * energy_perf_strings to epp value, by shifting 6
783 * bits left to use only top two bits in epp.
784 * The resultant epp need to shifted by 24 bits to
785 * epp position in MSR_HWP_REQUEST.
788 epp = (pref_index - 1) << 6;
790 value |= (u64)epp << 24;
791 ret = wrmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST, value);
794 epp = (pref_index - 1) << 2;
795 ret = intel_pstate_set_epb(cpu_data->cpu, epp);
798 mutex_unlock(&intel_pstate_limits_lock);
803 static ssize_t show_energy_performance_available_preferences(
804 struct cpufreq_policy *policy, char *buf)
809 while (energy_perf_strings[i] != NULL)
810 ret += sprintf(&buf[ret], "%s ", energy_perf_strings[i++]);
812 ret += sprintf(&buf[ret], "\n");
817 cpufreq_freq_attr_ro(energy_performance_available_preferences);
819 static ssize_t store_energy_performance_preference(
820 struct cpufreq_policy *policy, const char *buf, size_t count)
822 struct cpudata *cpu_data = all_cpu_data[policy->cpu];
823 char str_preference[21];
826 ret = sscanf(buf, "%20s", str_preference);
830 while (energy_perf_strings[i] != NULL) {
831 if (!strcmp(str_preference, energy_perf_strings[i])) {
832 intel_pstate_set_energy_pref_index(cpu_data, i);
841 static ssize_t show_energy_performance_preference(
842 struct cpufreq_policy *policy, char *buf)
844 struct cpudata *cpu_data = all_cpu_data[policy->cpu];
847 preference = intel_pstate_get_energy_pref_index(cpu_data);
851 return sprintf(buf, "%s\n", energy_perf_strings[preference]);
854 cpufreq_freq_attr_rw(energy_performance_preference);
856 static struct freq_attr *hwp_cpufreq_attrs[] = {
857 &energy_performance_preference,
858 &energy_performance_available_preferences,
862 static void intel_pstate_hwp_set(struct cpufreq_policy *policy)
864 int min, hw_min, max, hw_max, cpu, range, adj_range;
865 struct perf_limits *perf_limits = limits;
868 for_each_cpu(cpu, policy->cpus) {
869 int max_perf_pct, min_perf_pct;
870 struct cpudata *cpu_data = all_cpu_data[cpu];
874 perf_limits = all_cpu_data[cpu]->perf_limits;
876 rdmsrl_on_cpu(cpu, MSR_HWP_CAPABILITIES, &cap);
877 hw_min = HWP_LOWEST_PERF(cap);
878 hw_max = HWP_HIGHEST_PERF(cap);
879 range = hw_max - hw_min;
881 max_perf_pct = perf_limits->max_perf_pct;
882 min_perf_pct = perf_limits->min_perf_pct;
884 rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
885 adj_range = min_perf_pct * range / 100;
886 min = hw_min + adj_range;
887 value &= ~HWP_MIN_PERF(~0L);
888 value |= HWP_MIN_PERF(min);
890 adj_range = max_perf_pct * range / 100;
891 max = hw_min + adj_range;
892 if (limits->no_turbo) {
893 hw_max = HWP_GUARANTEED_PERF(cap);
898 value &= ~HWP_MAX_PERF(~0L);
899 value |= HWP_MAX_PERF(max);
901 if (cpu_data->epp_policy == cpu_data->policy)
904 cpu_data->epp_policy = cpu_data->policy;
906 if (cpu_data->epp_saved >= 0) {
907 epp = cpu_data->epp_saved;
908 cpu_data->epp_saved = -EINVAL;
912 if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE) {
913 epp = intel_pstate_get_epp(cpu_data, value);
914 cpu_data->epp_powersave = epp;
915 /* If EPP read was failed, then don't try to write */
922 /* skip setting EPP, when saved value is invalid */
923 if (cpu_data->epp_powersave < 0)
927 * No need to restore EPP when it is not zero. This
929 * - Policy is not changed
930 * - user has manually changed
931 * - Error reading EPB
933 epp = intel_pstate_get_epp(cpu_data, value);
937 epp = cpu_data->epp_powersave;
940 if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
941 value &= ~GENMASK_ULL(31, 24);
942 value |= (u64)epp << 24;
944 intel_pstate_set_epb(cpu, epp);
947 wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
951 static int intel_pstate_hwp_set_policy(struct cpufreq_policy *policy)
954 intel_pstate_hwp_set(policy);
959 static int intel_pstate_hwp_save_state(struct cpufreq_policy *policy)
961 struct cpudata *cpu_data = all_cpu_data[policy->cpu];
966 cpu_data->epp_saved = intel_pstate_get_epp(cpu_data, 0);
971 static int intel_pstate_resume(struct cpufreq_policy *policy)
978 mutex_lock(&intel_pstate_limits_lock);
980 all_cpu_data[policy->cpu]->epp_policy = 0;
982 ret = intel_pstate_hwp_set_policy(policy);
984 mutex_unlock(&intel_pstate_limits_lock);
989 static void intel_pstate_update_policies(void)
993 for_each_possible_cpu(cpu)
994 cpufreq_update_policy(cpu);
997 /************************** debugfs begin ************************/
998 static int pid_param_set(void *data, u64 val)
1001 intel_pstate_reset_all_pid();
1005 static int pid_param_get(void *data, u64 *val)
1007 *val = *(u32 *)data;
1010 DEFINE_SIMPLE_ATTRIBUTE(fops_pid_param, pid_param_get, pid_param_set, "%llu\n");
1012 static struct dentry *debugfs_parent;
1017 struct dentry *dentry;
1020 static struct pid_param pid_files[] = {
1021 {"sample_rate_ms", &pid_params.sample_rate_ms, },
1022 {"d_gain_pct", &pid_params.d_gain_pct, },
1023 {"i_gain_pct", &pid_params.i_gain_pct, },
1024 {"deadband", &pid_params.deadband, },
1025 {"setpoint", &pid_params.setpoint, },
1026 {"p_gain_pct", &pid_params.p_gain_pct, },
1030 static void intel_pstate_debug_expose_params(void)
1034 debugfs_parent = debugfs_create_dir("pstate_snb", NULL);
1035 if (IS_ERR_OR_NULL(debugfs_parent))
1038 for (i = 0; pid_files[i].name; i++) {
1039 struct dentry *dentry;
1041 dentry = debugfs_create_file(pid_files[i].name, 0660,
1042 debugfs_parent, pid_files[i].value,
1044 if (!IS_ERR(dentry))
1045 pid_files[i].dentry = dentry;
1049 static void intel_pstate_debug_hide_params(void)
1053 if (IS_ERR_OR_NULL(debugfs_parent))
1056 for (i = 0; pid_files[i].name; i++) {
1057 debugfs_remove(pid_files[i].dentry);
1058 pid_files[i].dentry = NULL;
1061 debugfs_remove(debugfs_parent);
1062 debugfs_parent = NULL;
1065 /************************** debugfs end ************************/
1067 /************************** sysfs begin ************************/
1068 #define show_one(file_name, object) \
1069 static ssize_t show_##file_name \
1070 (struct kobject *kobj, struct attribute *attr, char *buf) \
1072 return sprintf(buf, "%u\n", limits->object); \
1075 static ssize_t intel_pstate_show_status(char *buf);
1076 static int intel_pstate_update_status(const char *buf, size_t size);
1078 static ssize_t show_status(struct kobject *kobj,
1079 struct attribute *attr, char *buf)
1083 mutex_lock(&intel_pstate_driver_lock);
1084 ret = intel_pstate_show_status(buf);
1085 mutex_unlock(&intel_pstate_driver_lock);
1090 static ssize_t store_status(struct kobject *a, struct attribute *b,
1091 const char *buf, size_t count)
1093 char *p = memchr(buf, '\n', count);
1096 mutex_lock(&intel_pstate_driver_lock);
1097 ret = intel_pstate_update_status(buf, p ? p - buf : count);
1098 mutex_unlock(&intel_pstate_driver_lock);
1100 return ret < 0 ? ret : count;
1103 static ssize_t show_turbo_pct(struct kobject *kobj,
1104 struct attribute *attr, char *buf)
1106 struct cpudata *cpu;
1107 int total, no_turbo, turbo_pct;
1110 mutex_lock(&intel_pstate_driver_lock);
1112 if (!driver_registered) {
1113 mutex_unlock(&intel_pstate_driver_lock);
1117 cpu = all_cpu_data[0];
1119 total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1120 no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
1121 turbo_fp = div_fp(no_turbo, total);
1122 turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
1124 mutex_unlock(&intel_pstate_driver_lock);
1126 return sprintf(buf, "%u\n", turbo_pct);
1129 static ssize_t show_num_pstates(struct kobject *kobj,
1130 struct attribute *attr, char *buf)
1132 struct cpudata *cpu;
1135 mutex_lock(&intel_pstate_driver_lock);
1137 if (!driver_registered) {
1138 mutex_unlock(&intel_pstate_driver_lock);
1142 cpu = all_cpu_data[0];
1143 total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1145 mutex_unlock(&intel_pstate_driver_lock);
1147 return sprintf(buf, "%u\n", total);
1150 static ssize_t show_no_turbo(struct kobject *kobj,
1151 struct attribute *attr, char *buf)
1155 mutex_lock(&intel_pstate_driver_lock);
1157 if (!driver_registered) {
1158 mutex_unlock(&intel_pstate_driver_lock);
1162 update_turbo_state();
1163 if (limits->turbo_disabled)
1164 ret = sprintf(buf, "%u\n", limits->turbo_disabled);
1166 ret = sprintf(buf, "%u\n", limits->no_turbo);
1168 mutex_unlock(&intel_pstate_driver_lock);
1173 static ssize_t store_no_turbo(struct kobject *a, struct attribute *b,
1174 const char *buf, size_t count)
1179 ret = sscanf(buf, "%u", &input);
1183 mutex_lock(&intel_pstate_driver_lock);
1185 if (!driver_registered) {
1186 mutex_unlock(&intel_pstate_driver_lock);
1190 mutex_lock(&intel_pstate_limits_lock);
1192 update_turbo_state();
1193 if (limits->turbo_disabled) {
1194 pr_warn("Turbo disabled by BIOS or unavailable on processor\n");
1195 mutex_unlock(&intel_pstate_limits_lock);
1196 mutex_unlock(&intel_pstate_driver_lock);
1200 limits->no_turbo = clamp_t(int, input, 0, 1);
1202 mutex_unlock(&intel_pstate_limits_lock);
1204 intel_pstate_update_policies();
1206 mutex_unlock(&intel_pstate_driver_lock);
1211 static ssize_t store_max_perf_pct(struct kobject *a, struct attribute *b,
1212 const char *buf, size_t count)
1217 ret = sscanf(buf, "%u", &input);
1221 mutex_lock(&intel_pstate_driver_lock);
1223 if (!driver_registered) {
1224 mutex_unlock(&intel_pstate_driver_lock);
1228 mutex_lock(&intel_pstate_limits_lock);
1230 limits->max_sysfs_pct = clamp_t(int, input, 0 , 100);
1231 limits->max_perf_pct = min(limits->max_policy_pct,
1232 limits->max_sysfs_pct);
1233 limits->max_perf_pct = max(limits->min_policy_pct,
1234 limits->max_perf_pct);
1235 limits->max_perf_pct = max(limits->min_perf_pct,
1236 limits->max_perf_pct);
1237 limits->max_perf = div_ext_fp(limits->max_perf_pct, 100);
1239 mutex_unlock(&intel_pstate_limits_lock);
1241 intel_pstate_update_policies();
1243 mutex_unlock(&intel_pstate_driver_lock);
1248 static ssize_t store_min_perf_pct(struct kobject *a, struct attribute *b,
1249 const char *buf, size_t count)
1254 ret = sscanf(buf, "%u", &input);
1258 mutex_lock(&intel_pstate_driver_lock);
1260 if (!driver_registered) {
1261 mutex_unlock(&intel_pstate_driver_lock);
1265 mutex_lock(&intel_pstate_limits_lock);
1267 limits->min_sysfs_pct = clamp_t(int, input, 0 , 100);
1268 limits->min_perf_pct = max(limits->min_policy_pct,
1269 limits->min_sysfs_pct);
1270 limits->min_perf_pct = min(limits->max_policy_pct,
1271 limits->min_perf_pct);
1272 limits->min_perf_pct = min(limits->max_perf_pct,
1273 limits->min_perf_pct);
1274 limits->min_perf = div_ext_fp(limits->min_perf_pct, 100);
1276 mutex_unlock(&intel_pstate_limits_lock);
1278 intel_pstate_update_policies();
1280 mutex_unlock(&intel_pstate_driver_lock);
1285 show_one(max_perf_pct, max_perf_pct);
1286 show_one(min_perf_pct, min_perf_pct);
1288 define_one_global_rw(status);
1289 define_one_global_rw(no_turbo);
1290 define_one_global_rw(max_perf_pct);
1291 define_one_global_rw(min_perf_pct);
1292 define_one_global_ro(turbo_pct);
1293 define_one_global_ro(num_pstates);
1295 static struct attribute *intel_pstate_attributes[] = {
1303 static struct attribute_group intel_pstate_attr_group = {
1304 .attrs = intel_pstate_attributes,
1307 static void __init intel_pstate_sysfs_expose_params(void)
1309 struct kobject *intel_pstate_kobject;
1312 intel_pstate_kobject = kobject_create_and_add("intel_pstate",
1313 &cpu_subsys.dev_root->kobj);
1314 if (WARN_ON(!intel_pstate_kobject))
1317 rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
1322 * If per cpu limits are enforced there are no global limits, so
1323 * return without creating max/min_perf_pct attributes
1328 rc = sysfs_create_file(intel_pstate_kobject, &max_perf_pct.attr);
1331 rc = sysfs_create_file(intel_pstate_kobject, &min_perf_pct.attr);
1335 /************************** sysfs end ************************/
1337 static void intel_pstate_hwp_enable(struct cpudata *cpudata)
1339 /* First disable HWP notification interrupt as we don't process them */
1340 if (static_cpu_has(X86_FEATURE_HWP_NOTIFY))
1341 wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1343 wrmsrl_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1);
1344 cpudata->epp_policy = 0;
1345 if (cpudata->epp_default == -EINVAL)
1346 cpudata->epp_default = intel_pstate_get_epp(cpudata, 0);
1349 static int atom_get_min_pstate(void)
1353 rdmsrl(ATOM_RATIOS, value);
1354 return (value >> 8) & 0x7F;
1357 static int atom_get_max_pstate(void)
1361 rdmsrl(ATOM_RATIOS, value);
1362 return (value >> 16) & 0x7F;
1365 static int atom_get_turbo_pstate(void)
1369 rdmsrl(ATOM_TURBO_RATIOS, value);
1370 return value & 0x7F;
1373 static u64 atom_get_val(struct cpudata *cpudata, int pstate)
1379 val = (u64)pstate << 8;
1380 if (limits->no_turbo && !limits->turbo_disabled)
1381 val |= (u64)1 << 32;
1383 vid_fp = cpudata->vid.min + mul_fp(
1384 int_tofp(pstate - cpudata->pstate.min_pstate),
1385 cpudata->vid.ratio);
1387 vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
1388 vid = ceiling_fp(vid_fp);
1390 if (pstate > cpudata->pstate.max_pstate)
1391 vid = cpudata->vid.turbo;
1396 static int silvermont_get_scaling(void)
1400 /* Defined in Table 35-6 from SDM (Sept 2015) */
1401 static int silvermont_freq_table[] = {
1402 83300, 100000, 133300, 116700, 80000};
1404 rdmsrl(MSR_FSB_FREQ, value);
1408 return silvermont_freq_table[i];
1411 static int airmont_get_scaling(void)
1415 /* Defined in Table 35-10 from SDM (Sept 2015) */
1416 static int airmont_freq_table[] = {
1417 83300, 100000, 133300, 116700, 80000,
1418 93300, 90000, 88900, 87500};
1420 rdmsrl(MSR_FSB_FREQ, value);
1424 return airmont_freq_table[i];
1427 static void atom_get_vid(struct cpudata *cpudata)
1431 rdmsrl(ATOM_VIDS, value);
1432 cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
1433 cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
1434 cpudata->vid.ratio = div_fp(
1435 cpudata->vid.max - cpudata->vid.min,
1436 int_tofp(cpudata->pstate.max_pstate -
1437 cpudata->pstate.min_pstate));
1439 rdmsrl(ATOM_TURBO_VIDS, value);
1440 cpudata->vid.turbo = value & 0x7f;
1443 static int core_get_min_pstate(void)
1447 rdmsrl(MSR_PLATFORM_INFO, value);
1448 return (value >> 40) & 0xFF;
1451 static int core_get_max_pstate_physical(void)
1455 rdmsrl(MSR_PLATFORM_INFO, value);
1456 return (value >> 8) & 0xFF;
1459 static int core_get_max_pstate(void)
1466 rdmsrl(MSR_PLATFORM_INFO, plat_info);
1467 max_pstate = (plat_info >> 8) & 0xFF;
1469 err = rdmsrl_safe(MSR_TURBO_ACTIVATION_RATIO, &tar);
1471 /* Do some sanity checking for safety */
1472 if (plat_info & 0x600000000) {
1477 err = rdmsrl_safe(MSR_CONFIG_TDP_CONTROL, &tdp_ctrl);
1481 tdp_msr = MSR_CONFIG_TDP_NOMINAL + (tdp_ctrl & 0x3);
1482 err = rdmsrl_safe(tdp_msr, &tdp_ratio);
1486 /* For level 1 and 2, bits[23:16] contain the ratio */
1490 tdp_ratio &= 0xff; /* ratios are only 8 bits long */
1491 if (tdp_ratio - 1 == tar) {
1493 pr_debug("max_pstate=TAC %x\n", max_pstate);
1504 static int core_get_turbo_pstate(void)
1509 rdmsrl(MSR_TURBO_RATIO_LIMIT, value);
1510 nont = core_get_max_pstate();
1511 ret = (value) & 255;
1517 static inline int core_get_scaling(void)
1522 static u64 core_get_val(struct cpudata *cpudata, int pstate)
1526 val = (u64)pstate << 8;
1527 if (limits->no_turbo && !limits->turbo_disabled)
1528 val |= (u64)1 << 32;
1533 static int knl_get_turbo_pstate(void)
1538 rdmsrl(MSR_TURBO_RATIO_LIMIT, value);
1539 nont = core_get_max_pstate();
1540 ret = (((value) >> 8) & 0xFF);
1546 static struct cpu_defaults core_params = {
1548 .sample_rate_ms = 10,
1556 .get_max = core_get_max_pstate,
1557 .get_max_physical = core_get_max_pstate_physical,
1558 .get_min = core_get_min_pstate,
1559 .get_turbo = core_get_turbo_pstate,
1560 .get_scaling = core_get_scaling,
1561 .get_val = core_get_val,
1562 .get_target_pstate = get_target_pstate_use_performance,
1566 static const struct cpu_defaults silvermont_params = {
1568 .sample_rate_ms = 10,
1576 .get_max = atom_get_max_pstate,
1577 .get_max_physical = atom_get_max_pstate,
1578 .get_min = atom_get_min_pstate,
1579 .get_turbo = atom_get_turbo_pstate,
1580 .get_val = atom_get_val,
1581 .get_scaling = silvermont_get_scaling,
1582 .get_vid = atom_get_vid,
1583 .get_target_pstate = get_target_pstate_use_cpu_load,
1587 static const struct cpu_defaults airmont_params = {
1589 .sample_rate_ms = 10,
1597 .get_max = atom_get_max_pstate,
1598 .get_max_physical = atom_get_max_pstate,
1599 .get_min = atom_get_min_pstate,
1600 .get_turbo = atom_get_turbo_pstate,
1601 .get_val = atom_get_val,
1602 .get_scaling = airmont_get_scaling,
1603 .get_vid = atom_get_vid,
1604 .get_target_pstate = get_target_pstate_use_cpu_load,
1608 static const struct cpu_defaults knl_params = {
1610 .sample_rate_ms = 10,
1618 .get_max = core_get_max_pstate,
1619 .get_max_physical = core_get_max_pstate_physical,
1620 .get_min = core_get_min_pstate,
1621 .get_turbo = knl_get_turbo_pstate,
1622 .get_scaling = core_get_scaling,
1623 .get_val = core_get_val,
1624 .get_target_pstate = get_target_pstate_use_performance,
1628 static const struct cpu_defaults bxt_params = {
1630 .sample_rate_ms = 10,
1638 .get_max = core_get_max_pstate,
1639 .get_max_physical = core_get_max_pstate_physical,
1640 .get_min = core_get_min_pstate,
1641 .get_turbo = core_get_turbo_pstate,
1642 .get_scaling = core_get_scaling,
1643 .get_val = core_get_val,
1644 .get_target_pstate = get_target_pstate_use_cpu_load,
1648 static void intel_pstate_get_min_max(struct cpudata *cpu, int *min, int *max)
1650 int max_perf = cpu->pstate.turbo_pstate;
1653 struct perf_limits *perf_limits = limits;
1655 if (limits->no_turbo || limits->turbo_disabled)
1656 max_perf = cpu->pstate.max_pstate;
1659 perf_limits = cpu->perf_limits;
1662 * performance can be limited by user through sysfs, by cpufreq
1663 * policy, or by cpu specific default values determined through
1666 max_perf_adj = fp_ext_toint(max_perf * perf_limits->max_perf);
1667 *max = clamp_t(int, max_perf_adj,
1668 cpu->pstate.min_pstate, cpu->pstate.turbo_pstate);
1670 min_perf = fp_ext_toint(max_perf * perf_limits->min_perf);
1671 *min = clamp_t(int, min_perf, cpu->pstate.min_pstate, max_perf);
1674 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
1676 trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
1677 cpu->pstate.current_pstate = pstate;
1679 * Generally, there is no guarantee that this code will always run on
1680 * the CPU being updated, so force the register update to run on the
1683 wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
1684 pstate_funcs.get_val(cpu, pstate));
1687 static void intel_pstate_set_min_pstate(struct cpudata *cpu)
1689 intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
1692 static void intel_pstate_max_within_limits(struct cpudata *cpu)
1694 int min_pstate, max_pstate;
1696 update_turbo_state();
1697 intel_pstate_get_min_max(cpu, &min_pstate, &max_pstate);
1698 intel_pstate_set_pstate(cpu, max_pstate);
1701 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
1703 cpu->pstate.min_pstate = pstate_funcs.get_min();
1704 cpu->pstate.max_pstate = pstate_funcs.get_max();
1705 cpu->pstate.max_pstate_physical = pstate_funcs.get_max_physical();
1706 cpu->pstate.turbo_pstate = pstate_funcs.get_turbo();
1707 cpu->pstate.scaling = pstate_funcs.get_scaling();
1708 cpu->pstate.max_freq = cpu->pstate.max_pstate * cpu->pstate.scaling;
1709 cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
1711 if (pstate_funcs.get_vid)
1712 pstate_funcs.get_vid(cpu);
1714 intel_pstate_set_min_pstate(cpu);
1717 static inline void intel_pstate_calc_avg_perf(struct cpudata *cpu)
1719 struct sample *sample = &cpu->sample;
1721 sample->core_avg_perf = div_ext_fp(sample->aperf, sample->mperf);
1724 static inline bool intel_pstate_sample(struct cpudata *cpu, u64 time)
1727 unsigned long flags;
1730 local_irq_save(flags);
1731 rdmsrl(MSR_IA32_APERF, aperf);
1732 rdmsrl(MSR_IA32_MPERF, mperf);
1734 if (cpu->prev_mperf == mperf || cpu->prev_tsc == tsc) {
1735 local_irq_restore(flags);
1738 local_irq_restore(flags);
1740 cpu->last_sample_time = cpu->sample.time;
1741 cpu->sample.time = time;
1742 cpu->sample.aperf = aperf;
1743 cpu->sample.mperf = mperf;
1744 cpu->sample.tsc = tsc;
1745 cpu->sample.aperf -= cpu->prev_aperf;
1746 cpu->sample.mperf -= cpu->prev_mperf;
1747 cpu->sample.tsc -= cpu->prev_tsc;
1749 cpu->prev_aperf = aperf;
1750 cpu->prev_mperf = mperf;
1751 cpu->prev_tsc = tsc;
1753 * First time this function is invoked in a given cycle, all of the
1754 * previous sample data fields are equal to zero or stale and they must
1755 * be populated with meaningful numbers for things to work, so assume
1756 * that sample.time will always be reset before setting the utilization
1757 * update hook and make the caller skip the sample then.
1759 return !!cpu->last_sample_time;
1762 static inline int32_t get_avg_frequency(struct cpudata *cpu)
1764 return mul_ext_fp(cpu->sample.core_avg_perf,
1765 cpu->pstate.max_pstate_physical * cpu->pstate.scaling);
1768 static inline int32_t get_avg_pstate(struct cpudata *cpu)
1770 return mul_ext_fp(cpu->pstate.max_pstate_physical,
1771 cpu->sample.core_avg_perf);
1774 static inline int32_t get_target_pstate_use_cpu_load(struct cpudata *cpu)
1776 struct sample *sample = &cpu->sample;
1777 int32_t busy_frac, boost;
1778 int target, avg_pstate;
1780 busy_frac = div_fp(sample->mperf, sample->tsc);
1782 boost = cpu->iowait_boost;
1783 cpu->iowait_boost >>= 1;
1785 if (busy_frac < boost)
1788 sample->busy_scaled = busy_frac * 100;
1790 target = limits->no_turbo || limits->turbo_disabled ?
1791 cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
1792 target += target >> 2;
1793 target = mul_fp(target, busy_frac);
1794 if (target < cpu->pstate.min_pstate)
1795 target = cpu->pstate.min_pstate;
1798 * If the average P-state during the previous cycle was higher than the
1799 * current target, add 50% of the difference to the target to reduce
1800 * possible performance oscillations and offset possible performance
1801 * loss related to moving the workload from one CPU to another within
1804 avg_pstate = get_avg_pstate(cpu);
1805 if (avg_pstate > target)
1806 target += (avg_pstate - target) >> 1;
1811 static inline int32_t get_target_pstate_use_performance(struct cpudata *cpu)
1813 int32_t perf_scaled, max_pstate, current_pstate, sample_ratio;
1817 * perf_scaled is the ratio of the average P-state during the last
1818 * sampling period to the P-state requested last time (in percent).
1820 * That measures the system's response to the previous P-state
1823 max_pstate = cpu->pstate.max_pstate_physical;
1824 current_pstate = cpu->pstate.current_pstate;
1825 perf_scaled = mul_ext_fp(cpu->sample.core_avg_perf,
1826 div_fp(100 * max_pstate, current_pstate));
1829 * Since our utilization update callback will not run unless we are
1830 * in C0, check if the actual elapsed time is significantly greater (3x)
1831 * than our sample interval. If it is, then we were idle for a long
1832 * enough period of time to adjust our performance metric.
1834 duration_ns = cpu->sample.time - cpu->last_sample_time;
1835 if ((s64)duration_ns > pid_params.sample_rate_ns * 3) {
1836 sample_ratio = div_fp(pid_params.sample_rate_ns, duration_ns);
1837 perf_scaled = mul_fp(perf_scaled, sample_ratio);
1839 sample_ratio = div_fp(100 * cpu->sample.mperf, cpu->sample.tsc);
1840 if (sample_ratio < int_tofp(1))
1844 cpu->sample.busy_scaled = perf_scaled;
1845 return cpu->pstate.current_pstate - pid_calc(&cpu->pid, perf_scaled);
1848 static int intel_pstate_prepare_request(struct cpudata *cpu, int pstate)
1850 int max_perf, min_perf;
1852 intel_pstate_get_min_max(cpu, &min_perf, &max_perf);
1853 pstate = clamp_t(int, pstate, min_perf, max_perf);
1854 trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
1858 static void intel_pstate_update_pstate(struct cpudata *cpu, int pstate)
1860 pstate = intel_pstate_prepare_request(cpu, pstate);
1861 if (pstate == cpu->pstate.current_pstate)
1864 cpu->pstate.current_pstate = pstate;
1865 wrmsrl(MSR_IA32_PERF_CTL, pstate_funcs.get_val(cpu, pstate));
1868 static inline void intel_pstate_adjust_busy_pstate(struct cpudata *cpu)
1870 int from, target_pstate;
1871 struct sample *sample;
1873 from = cpu->pstate.current_pstate;
1875 target_pstate = cpu->policy == CPUFREQ_POLICY_PERFORMANCE ?
1876 cpu->pstate.turbo_pstate : pstate_funcs.get_target_pstate(cpu);
1878 update_turbo_state();
1880 intel_pstate_update_pstate(cpu, target_pstate);
1882 sample = &cpu->sample;
1883 trace_pstate_sample(mul_ext_fp(100, sample->core_avg_perf),
1884 fp_toint(sample->busy_scaled),
1886 cpu->pstate.current_pstate,
1890 get_avg_frequency(cpu),
1891 fp_toint(cpu->iowait_boost * 100));
1894 static void intel_pstate_update_util(struct update_util_data *data, u64 time,
1897 struct cpudata *cpu = container_of(data, struct cpudata, update_util);
1900 if (pstate_funcs.get_target_pstate == get_target_pstate_use_cpu_load) {
1901 if (flags & SCHED_CPUFREQ_IOWAIT) {
1902 cpu->iowait_boost = int_tofp(1);
1903 } else if (cpu->iowait_boost) {
1904 /* Clear iowait_boost if the CPU may have been idle. */
1905 delta_ns = time - cpu->last_update;
1906 if (delta_ns > TICK_NSEC)
1907 cpu->iowait_boost = 0;
1909 cpu->last_update = time;
1912 delta_ns = time - cpu->sample.time;
1913 if ((s64)delta_ns >= pid_params.sample_rate_ns) {
1914 bool sample_taken = intel_pstate_sample(cpu, time);
1917 intel_pstate_calc_avg_perf(cpu);
1919 intel_pstate_adjust_busy_pstate(cpu);
1924 #define ICPU(model, policy) \
1925 { X86_VENDOR_INTEL, 6, model, X86_FEATURE_APERFMPERF,\
1926 (unsigned long)&policy }
1928 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
1929 ICPU(INTEL_FAM6_SANDYBRIDGE, core_params),
1930 ICPU(INTEL_FAM6_SANDYBRIDGE_X, core_params),
1931 ICPU(INTEL_FAM6_ATOM_SILVERMONT1, silvermont_params),
1932 ICPU(INTEL_FAM6_IVYBRIDGE, core_params),
1933 ICPU(INTEL_FAM6_HASWELL_CORE, core_params),
1934 ICPU(INTEL_FAM6_BROADWELL_CORE, core_params),
1935 ICPU(INTEL_FAM6_IVYBRIDGE_X, core_params),
1936 ICPU(INTEL_FAM6_HASWELL_X, core_params),
1937 ICPU(INTEL_FAM6_HASWELL_ULT, core_params),
1938 ICPU(INTEL_FAM6_HASWELL_GT3E, core_params),
1939 ICPU(INTEL_FAM6_BROADWELL_GT3E, core_params),
1940 ICPU(INTEL_FAM6_ATOM_AIRMONT, airmont_params),
1941 ICPU(INTEL_FAM6_SKYLAKE_MOBILE, core_params),
1942 ICPU(INTEL_FAM6_BROADWELL_X, core_params),
1943 ICPU(INTEL_FAM6_SKYLAKE_DESKTOP, core_params),
1944 ICPU(INTEL_FAM6_BROADWELL_XEON_D, core_params),
1945 ICPU(INTEL_FAM6_XEON_PHI_KNL, knl_params),
1946 ICPU(INTEL_FAM6_XEON_PHI_KNM, knl_params),
1947 ICPU(INTEL_FAM6_ATOM_GOLDMONT, bxt_params),
1950 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
1952 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] __initconst = {
1953 ICPU(INTEL_FAM6_BROADWELL_XEON_D, core_params),
1954 ICPU(INTEL_FAM6_BROADWELL_X, core_params),
1955 ICPU(INTEL_FAM6_SKYLAKE_X, core_params),
1959 static int intel_pstate_init_cpu(unsigned int cpunum)
1961 struct cpudata *cpu;
1963 cpu = all_cpu_data[cpunum];
1966 unsigned int size = sizeof(struct cpudata);
1969 size += sizeof(struct perf_limits);
1971 cpu = kzalloc(size, GFP_KERNEL);
1975 all_cpu_data[cpunum] = cpu;
1977 cpu->perf_limits = (struct perf_limits *)(cpu + 1);
1979 cpu->epp_default = -EINVAL;
1980 cpu->epp_powersave = -EINVAL;
1981 cpu->epp_saved = -EINVAL;
1984 cpu = all_cpu_data[cpunum];
1989 intel_pstate_hwp_enable(cpu);
1990 pid_params.sample_rate_ms = 50;
1991 pid_params.sample_rate_ns = 50 * NSEC_PER_MSEC;
1994 intel_pstate_get_cpu_pstates(cpu);
1996 intel_pstate_busy_pid_reset(cpu);
1998 pr_debug("controlling: cpu %d\n", cpunum);
2003 static unsigned int intel_pstate_get(unsigned int cpu_num)
2005 struct cpudata *cpu = all_cpu_data[cpu_num];
2007 return cpu ? get_avg_frequency(cpu) : 0;
2010 static void intel_pstate_set_update_util_hook(unsigned int cpu_num)
2012 struct cpudata *cpu = all_cpu_data[cpu_num];
2014 if (cpu->update_util_set)
2017 /* Prevent intel_pstate_update_util() from using stale data. */
2018 cpu->sample.time = 0;
2019 cpufreq_add_update_util_hook(cpu_num, &cpu->update_util,
2020 intel_pstate_update_util);
2021 cpu->update_util_set = true;
2024 static void intel_pstate_clear_update_util_hook(unsigned int cpu)
2026 struct cpudata *cpu_data = all_cpu_data[cpu];
2028 if (!cpu_data->update_util_set)
2031 cpufreq_remove_update_util_hook(cpu);
2032 cpu_data->update_util_set = false;
2033 synchronize_sched();
2036 static void intel_pstate_set_performance_limits(struct perf_limits *limits)
2038 limits->no_turbo = 0;
2039 limits->turbo_disabled = 0;
2040 limits->max_perf_pct = 100;
2041 limits->max_perf = int_ext_tofp(1);
2042 limits->min_perf_pct = 100;
2043 limits->min_perf = int_ext_tofp(1);
2044 limits->max_policy_pct = 100;
2045 limits->max_sysfs_pct = 100;
2046 limits->min_policy_pct = 0;
2047 limits->min_sysfs_pct = 0;
2050 static void intel_pstate_update_perf_limits(struct cpufreq_policy *policy,
2051 struct perf_limits *limits)
2054 limits->max_policy_pct = DIV_ROUND_UP(policy->max * 100,
2055 policy->cpuinfo.max_freq);
2056 limits->max_policy_pct = clamp_t(int, limits->max_policy_pct, 0, 100);
2057 if (policy->max == policy->min) {
2058 limits->min_policy_pct = limits->max_policy_pct;
2060 limits->min_policy_pct = DIV_ROUND_UP(policy->min * 100,
2061 policy->cpuinfo.max_freq);
2062 limits->min_policy_pct = clamp_t(int, limits->min_policy_pct,
2066 /* Normalize user input to [min_policy_pct, max_policy_pct] */
2067 limits->min_perf_pct = max(limits->min_policy_pct,
2068 limits->min_sysfs_pct);
2069 limits->min_perf_pct = min(limits->max_policy_pct,
2070 limits->min_perf_pct);
2071 limits->max_perf_pct = min(limits->max_policy_pct,
2072 limits->max_sysfs_pct);
2073 limits->max_perf_pct = max(limits->min_policy_pct,
2074 limits->max_perf_pct);
2076 /* Make sure min_perf_pct <= max_perf_pct */
2077 limits->min_perf_pct = min(limits->max_perf_pct, limits->min_perf_pct);
2079 limits->min_perf = div_ext_fp(limits->min_perf_pct, 100);
2080 limits->max_perf = div_ext_fp(limits->max_perf_pct, 100);
2081 limits->max_perf = round_up(limits->max_perf, EXT_FRAC_BITS);
2082 limits->min_perf = round_up(limits->min_perf, EXT_FRAC_BITS);
2084 pr_debug("cpu:%d max_perf_pct:%d min_perf_pct:%d\n", policy->cpu,
2085 limits->max_perf_pct, limits->min_perf_pct);
2088 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
2090 struct cpudata *cpu;
2091 struct perf_limits *perf_limits = NULL;
2093 if (!policy->cpuinfo.max_freq)
2096 pr_debug("set_policy cpuinfo.max %u policy->max %u\n",
2097 policy->cpuinfo.max_freq, policy->max);
2099 cpu = all_cpu_data[policy->cpu];
2100 cpu->policy = policy->policy;
2102 if (cpu->pstate.max_pstate_physical > cpu->pstate.max_pstate &&
2103 policy->max < policy->cpuinfo.max_freq &&
2104 policy->max > cpu->pstate.max_pstate * cpu->pstate.scaling) {
2105 pr_debug("policy->max > max non turbo frequency\n");
2106 policy->max = policy->cpuinfo.max_freq;
2110 perf_limits = cpu->perf_limits;
2112 mutex_lock(&intel_pstate_limits_lock);
2114 if (policy->policy == CPUFREQ_POLICY_PERFORMANCE) {
2116 limits = &performance_limits;
2117 perf_limits = limits;
2119 if (policy->max >= policy->cpuinfo.max_freq &&
2120 !limits->no_turbo) {
2121 pr_debug("set performance\n");
2122 intel_pstate_set_performance_limits(perf_limits);
2126 pr_debug("set powersave\n");
2128 limits = &powersave_limits;
2129 perf_limits = limits;
2134 intel_pstate_update_perf_limits(policy, perf_limits);
2136 if (cpu->policy == CPUFREQ_POLICY_PERFORMANCE) {
2138 * NOHZ_FULL CPUs need this as the governor callback may not
2139 * be invoked on them.
2141 intel_pstate_clear_update_util_hook(policy->cpu);
2142 intel_pstate_max_within_limits(cpu);
2145 intel_pstate_set_update_util_hook(policy->cpu);
2147 intel_pstate_hwp_set_policy(policy);
2149 mutex_unlock(&intel_pstate_limits_lock);
2154 static int intel_pstate_verify_policy(struct cpufreq_policy *policy)
2156 cpufreq_verify_within_cpu_limits(policy);
2158 if (policy->policy != CPUFREQ_POLICY_POWERSAVE &&
2159 policy->policy != CPUFREQ_POLICY_PERFORMANCE)
2162 /* When per-CPU limits are used, sysfs limits are not used */
2163 if (!per_cpu_limits) {
2164 unsigned int max_freq, min_freq;
2166 max_freq = policy->cpuinfo.max_freq *
2167 limits->max_sysfs_pct / 100;
2168 min_freq = policy->cpuinfo.max_freq *
2169 limits->min_sysfs_pct / 100;
2170 cpufreq_verify_within_limits(policy, min_freq, max_freq);
2176 static void intel_cpufreq_stop_cpu(struct cpufreq_policy *policy)
2178 intel_pstate_set_min_pstate(all_cpu_data[policy->cpu]);
2181 static void intel_pstate_stop_cpu(struct cpufreq_policy *policy)
2183 pr_debug("CPU %d exiting\n", policy->cpu);
2185 intel_pstate_clear_update_util_hook(policy->cpu);
2187 intel_pstate_hwp_save_state(policy);
2189 intel_cpufreq_stop_cpu(policy);
2192 static int intel_pstate_cpu_exit(struct cpufreq_policy *policy)
2194 intel_pstate_exit_perf_limits(policy);
2196 policy->fast_switch_possible = false;
2201 static int __intel_pstate_cpu_init(struct cpufreq_policy *policy)
2203 struct cpudata *cpu;
2206 rc = intel_pstate_init_cpu(policy->cpu);
2210 cpu = all_cpu_data[policy->cpu];
2213 * We need sane value in the cpu->perf_limits, so inherit from global
2214 * perf_limits limits, which are seeded with values based on the
2215 * CONFIG_CPU_FREQ_DEFAULT_GOV_*, during boot up.
2218 memcpy(cpu->perf_limits, limits, sizeof(struct perf_limits));
2220 policy->min = cpu->pstate.min_pstate * cpu->pstate.scaling;
2221 policy->max = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
2223 /* cpuinfo and default policy values */
2224 policy->cpuinfo.min_freq = cpu->pstate.min_pstate * cpu->pstate.scaling;
2225 update_turbo_state();
2226 policy->cpuinfo.max_freq = limits->turbo_disabled ?
2227 cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
2228 policy->cpuinfo.max_freq *= cpu->pstate.scaling;
2230 intel_pstate_init_acpi_perf_limits(policy);
2231 cpumask_set_cpu(policy->cpu, policy->cpus);
2233 policy->fast_switch_possible = true;
2238 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
2240 int ret = __intel_pstate_cpu_init(policy);
2245 policy->cpuinfo.transition_latency = CPUFREQ_ETERNAL;
2246 if (limits->min_perf_pct == 100 && limits->max_perf_pct == 100)
2247 policy->policy = CPUFREQ_POLICY_PERFORMANCE;
2249 policy->policy = CPUFREQ_POLICY_POWERSAVE;
2254 static struct cpufreq_driver intel_pstate = {
2255 .flags = CPUFREQ_CONST_LOOPS,
2256 .verify = intel_pstate_verify_policy,
2257 .setpolicy = intel_pstate_set_policy,
2258 .suspend = intel_pstate_hwp_save_state,
2259 .resume = intel_pstate_resume,
2260 .get = intel_pstate_get,
2261 .init = intel_pstate_cpu_init,
2262 .exit = intel_pstate_cpu_exit,
2263 .stop_cpu = intel_pstate_stop_cpu,
2264 .name = "intel_pstate",
2267 static int intel_cpufreq_verify_policy(struct cpufreq_policy *policy)
2269 struct cpudata *cpu = all_cpu_data[policy->cpu];
2270 struct perf_limits *perf_limits = limits;
2272 update_turbo_state();
2273 policy->cpuinfo.max_freq = limits->turbo_disabled ?
2274 cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2276 cpufreq_verify_within_cpu_limits(policy);
2279 perf_limits = cpu->perf_limits;
2281 mutex_lock(&intel_pstate_limits_lock);
2283 intel_pstate_update_perf_limits(policy, perf_limits);
2285 mutex_unlock(&intel_pstate_limits_lock);
2290 static unsigned int intel_cpufreq_turbo_update(struct cpudata *cpu,
2291 struct cpufreq_policy *policy,
2292 unsigned int target_freq)
2294 unsigned int max_freq;
2296 update_turbo_state();
2298 max_freq = limits->no_turbo || limits->turbo_disabled ?
2299 cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2300 policy->cpuinfo.max_freq = max_freq;
2301 if (policy->max > max_freq)
2302 policy->max = max_freq;
2304 if (target_freq > max_freq)
2305 target_freq = max_freq;
2310 static int intel_cpufreq_target(struct cpufreq_policy *policy,
2311 unsigned int target_freq,
2312 unsigned int relation)
2314 struct cpudata *cpu = all_cpu_data[policy->cpu];
2315 struct cpufreq_freqs freqs;
2318 freqs.old = policy->cur;
2319 freqs.new = intel_cpufreq_turbo_update(cpu, policy, target_freq);
2321 cpufreq_freq_transition_begin(policy, &freqs);
2323 case CPUFREQ_RELATION_L:
2324 target_pstate = DIV_ROUND_UP(freqs.new, cpu->pstate.scaling);
2326 case CPUFREQ_RELATION_H:
2327 target_pstate = freqs.new / cpu->pstate.scaling;
2330 target_pstate = DIV_ROUND_CLOSEST(freqs.new, cpu->pstate.scaling);
2333 target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2334 if (target_pstate != cpu->pstate.current_pstate) {
2335 cpu->pstate.current_pstate = target_pstate;
2336 wrmsrl_on_cpu(policy->cpu, MSR_IA32_PERF_CTL,
2337 pstate_funcs.get_val(cpu, target_pstate));
2339 cpufreq_freq_transition_end(policy, &freqs, false);
2344 static unsigned int intel_cpufreq_fast_switch(struct cpufreq_policy *policy,
2345 unsigned int target_freq)
2347 struct cpudata *cpu = all_cpu_data[policy->cpu];
2350 target_freq = intel_cpufreq_turbo_update(cpu, policy, target_freq);
2351 target_pstate = DIV_ROUND_UP(target_freq, cpu->pstate.scaling);
2352 intel_pstate_update_pstate(cpu, target_pstate);
2356 static int intel_cpufreq_cpu_init(struct cpufreq_policy *policy)
2358 int ret = __intel_pstate_cpu_init(policy);
2363 policy->cpuinfo.transition_latency = INTEL_CPUFREQ_TRANSITION_LATENCY;
2364 /* This reflects the intel_pstate_get_cpu_pstates() setting. */
2365 policy->cur = policy->cpuinfo.min_freq;
2370 static struct cpufreq_driver intel_cpufreq = {
2371 .flags = CPUFREQ_CONST_LOOPS,
2372 .verify = intel_cpufreq_verify_policy,
2373 .target = intel_cpufreq_target,
2374 .fast_switch = intel_cpufreq_fast_switch,
2375 .init = intel_cpufreq_cpu_init,
2376 .exit = intel_pstate_cpu_exit,
2377 .stop_cpu = intel_cpufreq_stop_cpu,
2378 .name = "intel_cpufreq",
2381 static struct cpufreq_driver *intel_pstate_driver = &intel_pstate;
2383 static void intel_pstate_driver_cleanup(void)
2388 for_each_online_cpu(cpu) {
2389 if (all_cpu_data[cpu]) {
2390 if (intel_pstate_driver == &intel_pstate)
2391 intel_pstate_clear_update_util_hook(cpu);
2393 kfree(all_cpu_data[cpu]);
2394 all_cpu_data[cpu] = NULL;
2400 static int intel_pstate_register_driver(void)
2404 ret = cpufreq_register_driver(intel_pstate_driver);
2406 intel_pstate_driver_cleanup();
2410 mutex_lock(&intel_pstate_limits_lock);
2411 driver_registered = true;
2412 mutex_unlock(&intel_pstate_limits_lock);
2414 if (intel_pstate_driver == &intel_pstate && !hwp_active &&
2415 pstate_funcs.get_target_pstate != get_target_pstate_use_cpu_load)
2416 intel_pstate_debug_expose_params();
2421 static int intel_pstate_unregister_driver(void)
2426 if (intel_pstate_driver == &intel_pstate && !hwp_active &&
2427 pstate_funcs.get_target_pstate != get_target_pstate_use_cpu_load)
2428 intel_pstate_debug_hide_params();
2430 mutex_lock(&intel_pstate_limits_lock);
2431 driver_registered = false;
2432 mutex_unlock(&intel_pstate_limits_lock);
2434 cpufreq_unregister_driver(intel_pstate_driver);
2435 intel_pstate_driver_cleanup();
2440 static ssize_t intel_pstate_show_status(char *buf)
2442 if (!driver_registered)
2443 return sprintf(buf, "off\n");
2445 return sprintf(buf, "%s\n", intel_pstate_driver == &intel_pstate ?
2446 "active" : "passive");
2449 static int intel_pstate_update_status(const char *buf, size_t size)
2453 if (size == 3 && !strncmp(buf, "off", size))
2454 return driver_registered ?
2455 intel_pstate_unregister_driver() : -EINVAL;
2457 if (size == 6 && !strncmp(buf, "active", size)) {
2458 if (driver_registered) {
2459 if (intel_pstate_driver == &intel_pstate)
2462 ret = intel_pstate_unregister_driver();
2467 intel_pstate_driver = &intel_pstate;
2468 return intel_pstate_register_driver();
2471 if (size == 7 && !strncmp(buf, "passive", size)) {
2472 if (driver_registered) {
2473 if (intel_pstate_driver != &intel_pstate)
2476 ret = intel_pstate_unregister_driver();
2481 intel_pstate_driver = &intel_cpufreq;
2482 return intel_pstate_register_driver();
2488 static int no_load __initdata;
2489 static int no_hwp __initdata;
2490 static int hwp_only __initdata;
2491 static unsigned int force_load __initdata;
2493 static int __init intel_pstate_msrs_not_valid(void)
2495 if (!pstate_funcs.get_max() ||
2496 !pstate_funcs.get_min() ||
2497 !pstate_funcs.get_turbo())
2503 static void __init copy_pid_params(struct pstate_adjust_policy *policy)
2505 pid_params.sample_rate_ms = policy->sample_rate_ms;
2506 pid_params.sample_rate_ns = pid_params.sample_rate_ms * NSEC_PER_MSEC;
2507 pid_params.p_gain_pct = policy->p_gain_pct;
2508 pid_params.i_gain_pct = policy->i_gain_pct;
2509 pid_params.d_gain_pct = policy->d_gain_pct;
2510 pid_params.deadband = policy->deadband;
2511 pid_params.setpoint = policy->setpoint;
2515 static void intel_pstate_use_acpi_profile(void)
2517 if (acpi_gbl_FADT.preferred_profile == PM_MOBILE)
2518 pstate_funcs.get_target_pstate =
2519 get_target_pstate_use_cpu_load;
2522 static void intel_pstate_use_acpi_profile(void)
2527 static void __init copy_cpu_funcs(struct pstate_funcs *funcs)
2529 pstate_funcs.get_max = funcs->get_max;
2530 pstate_funcs.get_max_physical = funcs->get_max_physical;
2531 pstate_funcs.get_min = funcs->get_min;
2532 pstate_funcs.get_turbo = funcs->get_turbo;
2533 pstate_funcs.get_scaling = funcs->get_scaling;
2534 pstate_funcs.get_val = funcs->get_val;
2535 pstate_funcs.get_vid = funcs->get_vid;
2536 pstate_funcs.get_target_pstate = funcs->get_target_pstate;
2538 intel_pstate_use_acpi_profile();
2543 static bool __init intel_pstate_no_acpi_pss(void)
2547 for_each_possible_cpu(i) {
2549 union acpi_object *pss;
2550 struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
2551 struct acpi_processor *pr = per_cpu(processors, i);
2556 status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
2557 if (ACPI_FAILURE(status))
2560 pss = buffer.pointer;
2561 if (pss && pss->type == ACPI_TYPE_PACKAGE) {
2572 static bool __init intel_pstate_has_acpi_ppc(void)
2576 for_each_possible_cpu(i) {
2577 struct acpi_processor *pr = per_cpu(processors, i);
2581 if (acpi_has_method(pr->handle, "_PPC"))
2592 struct hw_vendor_info {
2594 char oem_id[ACPI_OEM_ID_SIZE];
2595 char oem_table_id[ACPI_OEM_TABLE_ID_SIZE];
2599 /* Hardware vendor-specific info that has its own power management modes */
2600 static struct hw_vendor_info vendor_info[] __initdata = {
2601 {1, "HP ", "ProLiant", PSS},
2602 {1, "ORACLE", "X4-2 ", PPC},
2603 {1, "ORACLE", "X4-2L ", PPC},
2604 {1, "ORACLE", "X4-2B ", PPC},
2605 {1, "ORACLE", "X3-2 ", PPC},
2606 {1, "ORACLE", "X3-2L ", PPC},
2607 {1, "ORACLE", "X3-2B ", PPC},
2608 {1, "ORACLE", "X4470M2 ", PPC},
2609 {1, "ORACLE", "X4270M3 ", PPC},
2610 {1, "ORACLE", "X4270M2 ", PPC},
2611 {1, "ORACLE", "X4170M2 ", PPC},
2612 {1, "ORACLE", "X4170 M3", PPC},
2613 {1, "ORACLE", "X4275 M3", PPC},
2614 {1, "ORACLE", "X6-2 ", PPC},
2615 {1, "ORACLE", "Sudbury ", PPC},
2619 static bool __init intel_pstate_platform_pwr_mgmt_exists(void)
2621 struct acpi_table_header hdr;
2622 struct hw_vendor_info *v_info;
2623 const struct x86_cpu_id *id;
2626 id = x86_match_cpu(intel_pstate_cpu_oob_ids);
2628 rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
2629 if ( misc_pwr & (1 << 8))
2633 if (acpi_disabled ||
2634 ACPI_FAILURE(acpi_get_table_header(ACPI_SIG_FADT, 0, &hdr)))
2637 for (v_info = vendor_info; v_info->valid; v_info++) {
2638 if (!strncmp(hdr.oem_id, v_info->oem_id, ACPI_OEM_ID_SIZE) &&
2639 !strncmp(hdr.oem_table_id, v_info->oem_table_id,
2640 ACPI_OEM_TABLE_ID_SIZE))
2641 switch (v_info->oem_pwr_table) {
2643 return intel_pstate_no_acpi_pss();
2645 return intel_pstate_has_acpi_ppc() &&
2653 static void intel_pstate_request_control_from_smm(void)
2656 * It may be unsafe to request P-states control from SMM if _PPC support
2657 * has not been enabled.
2660 acpi_processor_pstate_control();
2662 #else /* CONFIG_ACPI not enabled */
2663 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
2664 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
2665 static inline void intel_pstate_request_control_from_smm(void) {}
2666 #endif /* CONFIG_ACPI */
2668 static const struct x86_cpu_id hwp_support_ids[] __initconst = {
2669 { X86_VENDOR_INTEL, 6, X86_MODEL_ANY, X86_FEATURE_HWP },
2673 static int __init intel_pstate_init(void)
2675 const struct x86_cpu_id *id;
2676 struct cpu_defaults *cpu_def;
2682 if (x86_match_cpu(hwp_support_ids) && !no_hwp) {
2683 copy_cpu_funcs(&core_params.funcs);
2685 intel_pstate.attr = hwp_cpufreq_attrs;
2686 goto hwp_cpu_matched;
2689 id = x86_match_cpu(intel_pstate_cpu_ids);
2693 cpu_def = (struct cpu_defaults *)id->driver_data;
2695 copy_pid_params(&cpu_def->pid_policy);
2696 copy_cpu_funcs(&cpu_def->funcs);
2698 if (intel_pstate_msrs_not_valid())
2703 * The Intel pstate driver will be ignored if the platform
2704 * firmware has its own power management modes.
2706 if (intel_pstate_platform_pwr_mgmt_exists())
2709 if (!hwp_active && hwp_only)
2712 pr_info("Intel P-state driver initializing\n");
2714 all_cpu_data = vzalloc(sizeof(void *) * num_possible_cpus());
2718 intel_pstate_request_control_from_smm();
2720 intel_pstate_sysfs_expose_params();
2722 mutex_lock(&intel_pstate_driver_lock);
2723 rc = intel_pstate_register_driver();
2724 mutex_unlock(&intel_pstate_driver_lock);
2729 pr_info("HWP enabled\n");
2733 device_initcall(intel_pstate_init);
2735 static int __init intel_pstate_setup(char *str)
2740 if (!strcmp(str, "disable")) {
2742 } else if (!strcmp(str, "passive")) {
2743 pr_info("Passive mode enabled\n");
2744 intel_pstate_driver = &intel_cpufreq;
2747 if (!strcmp(str, "no_hwp")) {
2748 pr_info("HWP disabled\n");
2751 if (!strcmp(str, "force"))
2753 if (!strcmp(str, "hwp_only"))
2755 if (!strcmp(str, "per_cpu_perf_limits"))
2756 per_cpu_limits = true;
2759 if (!strcmp(str, "support_acpi_ppc"))
2765 early_param("intel_pstate", intel_pstate_setup);
2767 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
2768 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
2769 MODULE_LICENSE("GPL");