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>
52 #define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
53 #define fp_toint(X) ((X) >> FRAC_BITS)
56 #define EXT_FRAC_BITS (EXT_BITS + FRAC_BITS)
57 #define fp_ext_toint(X) ((X) >> EXT_FRAC_BITS)
58 #define int_ext_tofp(X) ((int64_t)(X) << EXT_FRAC_BITS)
60 static inline int32_t mul_fp(int32_t x, int32_t y)
62 return ((int64_t)x * (int64_t)y) >> FRAC_BITS;
65 static inline int32_t div_fp(s64 x, s64 y)
67 return div64_s64((int64_t)x << FRAC_BITS, y);
70 static inline int ceiling_fp(int32_t x)
75 mask = (1 << FRAC_BITS) - 1;
81 static inline u64 mul_ext_fp(u64 x, u64 y)
83 return (x * y) >> EXT_FRAC_BITS;
86 static inline u64 div_ext_fp(u64 x, u64 y)
88 return div64_u64(x << EXT_FRAC_BITS, y);
92 * struct sample - Store performance sample
93 * @core_avg_perf: Ratio of APERF/MPERF which is the actual average
94 * performance during last sample period
95 * @busy_scaled: Scaled busy value which is used to calculate next
96 * P state. This can be different than core_avg_perf
97 * to account for cpu idle period
98 * @aperf: Difference of actual performance frequency clock count
99 * read from APERF MSR between last and current sample
100 * @mperf: Difference of maximum performance frequency clock count
101 * read from MPERF MSR between last and current sample
102 * @tsc: Difference of time stamp counter between last and
104 * @time: Current time from scheduler
106 * This structure is used in the cpudata structure to store performance sample
107 * data for choosing next P State.
110 int32_t core_avg_perf;
119 * struct pstate_data - Store P state data
120 * @current_pstate: Current requested P state
121 * @min_pstate: Min P state possible for this platform
122 * @max_pstate: Max P state possible for this platform
123 * @max_pstate_physical:This is physical Max P state for a processor
124 * This can be higher than the max_pstate which can
125 * be limited by platform thermal design power limits
126 * @scaling: Scaling factor to convert frequency to cpufreq
128 * @turbo_pstate: Max Turbo P state possible for this platform
129 * @max_freq: @max_pstate frequency in cpufreq units
130 * @turbo_freq: @turbo_pstate frequency in cpufreq units
132 * Stores the per cpu model P state limits and current P state.
138 int max_pstate_physical;
141 unsigned int max_freq;
142 unsigned int turbo_freq;
146 * struct vid_data - Stores voltage information data
147 * @min: VID data for this platform corresponding to
149 * @max: VID data corresponding to the highest P State.
150 * @turbo: VID data for turbo P state
151 * @ratio: Ratio of (vid max - vid min) /
152 * (max P state - Min P State)
154 * Stores the voltage data for DVFS (Dynamic Voltage and Frequency Scaling)
155 * This data is used in Atom platforms, where in addition to target P state,
156 * the voltage data needs to be specified to select next P State.
166 * struct _pid - Stores PID data
167 * @setpoint: Target set point for busyness or performance
168 * @integral: Storage for accumulated error values
169 * @p_gain: PID proportional gain
170 * @i_gain: PID integral gain
171 * @d_gain: PID derivative gain
172 * @deadband: PID deadband
173 * @last_err: Last error storage for integral part of PID calculation
175 * Stores PID coefficients and last error for PID controller.
188 * struct perf_limits - Store user and policy limits
189 * @no_turbo: User requested turbo state from intel_pstate sysfs
190 * @turbo_disabled: Platform turbo status either from msr
191 * MSR_IA32_MISC_ENABLE or when maximum available pstate
192 * matches the maximum turbo pstate
193 * @max_perf_pct: Effective maximum performance limit in percentage, this
194 * is minimum of either limits enforced by cpufreq policy
195 * or limits from user set limits via intel_pstate sysfs
196 * @min_perf_pct: Effective minimum performance limit in percentage, this
197 * is maximum of either limits enforced by cpufreq policy
198 * or limits from user set limits via intel_pstate sysfs
199 * @max_perf: This is a scaled value between 0 to 255 for max_perf_pct
200 * This value is used to limit max pstate
201 * @min_perf: This is a scaled value between 0 to 255 for min_perf_pct
202 * This value is used to limit min pstate
203 * @max_policy_pct: The maximum performance in percentage enforced by
204 * cpufreq setpolicy interface
205 * @max_sysfs_pct: The maximum performance in percentage enforced by
206 * intel pstate sysfs interface, unused when per cpu
207 * controls are enforced
208 * @min_policy_pct: The minimum performance in percentage enforced by
209 * cpufreq setpolicy interface
210 * @min_sysfs_pct: The minimum performance in percentage enforced by
211 * intel pstate sysfs interface, unused when per cpu
212 * controls are enforced
214 * Storage for user and policy defined limits.
230 * struct cpudata - Per CPU instance data storage
231 * @cpu: CPU number for this instance data
232 * @policy: CPUFreq policy value
233 * @update_util: CPUFreq utility callback information
234 * @update_util_set: CPUFreq utility callback is set
235 * @iowait_boost: iowait-related boost fraction
236 * @last_update: Time of the last update.
237 * @pstate: Stores P state limits for this CPU
238 * @vid: Stores VID limits for this CPU
239 * @pid: Stores PID parameters for this CPU
240 * @last_sample_time: Last Sample time
241 * @prev_aperf: Last APERF value read from APERF MSR
242 * @prev_mperf: Last MPERF value read from MPERF MSR
243 * @prev_tsc: Last timestamp counter (TSC) value
244 * @prev_cummulative_iowait: IO Wait time difference from last and
246 * @sample: Storage for storing last Sample data
247 * @perf_limits: Pointer to perf_limit unique to this CPU
248 * Not all field in the structure are applicable
249 * when per cpu controls are enforced
250 * @acpi_perf_data: Stores ACPI perf information read from _PSS
251 * @valid_pss_table: Set to true for valid ACPI _PSS entries found
252 * @epp_saved: Last saved HWP energy performance preference
253 * (EPP) or energy performance bias (EPB)
254 * @epp_policy: Last saved policy used to set EPP/EPB
256 * This structure stores per CPU instance data for all CPUs.
262 struct update_util_data update_util;
263 bool update_util_set;
265 struct pstate_data pstate;
270 u64 last_sample_time;
274 u64 prev_cummulative_iowait;
275 struct sample sample;
276 struct perf_limits *perf_limits;
278 struct acpi_processor_performance acpi_perf_data;
279 bool valid_pss_table;
281 unsigned int iowait_boost;
286 static struct cpudata **all_cpu_data;
289 * struct pstate_adjust_policy - Stores static PID configuration data
290 * @sample_rate_ms: PID calculation sample rate in ms
291 * @sample_rate_ns: Sample rate calculation in ns
292 * @deadband: PID deadband
293 * @setpoint: PID Setpoint
294 * @p_gain_pct: PID proportional gain
295 * @i_gain_pct: PID integral gain
296 * @d_gain_pct: PID derivative gain
298 * Stores per CPU model static PID configuration data.
300 struct pstate_adjust_policy {
311 * struct pstate_funcs - Per CPU model specific callbacks
312 * @get_max: Callback to get maximum non turbo effective P state
313 * @get_max_physical: Callback to get maximum non turbo physical P state
314 * @get_min: Callback to get minimum P state
315 * @get_turbo: Callback to get turbo P state
316 * @get_scaling: Callback to get frequency scaling factor
317 * @get_val: Callback to convert P state to actual MSR write value
318 * @get_vid: Callback to get VID data for Atom platforms
319 * @get_target_pstate: Callback to a function to calculate next P state to use
321 * Core and Atom CPU models have different way to get P State limits. This
322 * structure is used to store those callbacks.
324 struct pstate_funcs {
325 int (*get_max)(void);
326 int (*get_max_physical)(void);
327 int (*get_min)(void);
328 int (*get_turbo)(void);
329 int (*get_scaling)(void);
330 u64 (*get_val)(struct cpudata*, int pstate);
331 void (*get_vid)(struct cpudata *);
332 int32_t (*get_target_pstate)(struct cpudata *);
336 * struct cpu_defaults- Per CPU model default config data
337 * @pid_policy: PID config data
338 * @funcs: Callback function data
340 struct cpu_defaults {
341 struct pstate_adjust_policy pid_policy;
342 struct pstate_funcs funcs;
345 static inline int32_t get_target_pstate_use_performance(struct cpudata *cpu);
346 static inline int32_t get_target_pstate_use_cpu_load(struct cpudata *cpu);
348 static struct pstate_adjust_policy pid_params __read_mostly;
349 static struct pstate_funcs pstate_funcs __read_mostly;
350 static int hwp_active __read_mostly;
351 static bool per_cpu_limits __read_mostly;
354 static bool acpi_ppc;
357 static struct perf_limits performance_limits = {
361 .max_perf = int_ext_tofp(1),
363 .min_perf = int_ext_tofp(1),
364 .max_policy_pct = 100,
365 .max_sysfs_pct = 100,
370 static struct perf_limits powersave_limits = {
374 .max_perf = int_ext_tofp(1),
377 .max_policy_pct = 100,
378 .max_sysfs_pct = 100,
383 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_PERFORMANCE
384 static struct perf_limits *limits = &performance_limits;
386 static struct perf_limits *limits = &powersave_limits;
389 static DEFINE_MUTEX(intel_pstate_limits_lock);
393 static bool intel_pstate_get_ppc_enable_status(void)
395 if (acpi_gbl_FADT.preferred_profile == PM_ENTERPRISE_SERVER ||
396 acpi_gbl_FADT.preferred_profile == PM_PERFORMANCE_SERVER)
402 static void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
411 if (!intel_pstate_get_ppc_enable_status())
414 cpu = all_cpu_data[policy->cpu];
416 ret = acpi_processor_register_performance(&cpu->acpi_perf_data,
422 * Check if the control value in _PSS is for PERF_CTL MSR, which should
423 * guarantee that the states returned by it map to the states in our
426 if (cpu->acpi_perf_data.control_register.space_id !=
427 ACPI_ADR_SPACE_FIXED_HARDWARE)
431 * If there is only one entry _PSS, simply ignore _PSS and continue as
432 * usual without taking _PSS into account
434 if (cpu->acpi_perf_data.state_count < 2)
437 pr_debug("CPU%u - ACPI _PSS perf data\n", policy->cpu);
438 for (i = 0; i < cpu->acpi_perf_data.state_count; i++) {
439 pr_debug(" %cP%d: %u MHz, %u mW, 0x%x\n",
440 (i == cpu->acpi_perf_data.state ? '*' : ' '), i,
441 (u32) cpu->acpi_perf_data.states[i].core_frequency,
442 (u32) cpu->acpi_perf_data.states[i].power,
443 (u32) cpu->acpi_perf_data.states[i].control);
447 * The _PSS table doesn't contain whole turbo frequency range.
448 * This just contains +1 MHZ above the max non turbo frequency,
449 * with control value corresponding to max turbo ratio. But
450 * when cpufreq set policy is called, it will call with this
451 * max frequency, which will cause a reduced performance as
452 * this driver uses real max turbo frequency as the max
453 * frequency. So correct this frequency in _PSS table to
454 * correct max turbo frequency based on the turbo state.
455 * Also need to convert to MHz as _PSS freq is in MHz.
457 if (!limits->turbo_disabled)
458 cpu->acpi_perf_data.states[0].core_frequency =
459 policy->cpuinfo.max_freq / 1000;
460 cpu->valid_pss_table = true;
461 pr_debug("_PPC limits will be enforced\n");
466 cpu->valid_pss_table = false;
467 acpi_processor_unregister_performance(policy->cpu);
470 static void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
474 cpu = all_cpu_data[policy->cpu];
475 if (!cpu->valid_pss_table)
478 acpi_processor_unregister_performance(policy->cpu);
482 static inline void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
486 static inline void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
491 static inline void pid_reset(struct _pid *pid, int setpoint, int busy,
492 int deadband, int integral) {
493 pid->setpoint = int_tofp(setpoint);
494 pid->deadband = int_tofp(deadband);
495 pid->integral = int_tofp(integral);
496 pid->last_err = int_tofp(setpoint) - int_tofp(busy);
499 static inline void pid_p_gain_set(struct _pid *pid, int percent)
501 pid->p_gain = div_fp(percent, 100);
504 static inline void pid_i_gain_set(struct _pid *pid, int percent)
506 pid->i_gain = div_fp(percent, 100);
509 static inline void pid_d_gain_set(struct _pid *pid, int percent)
511 pid->d_gain = div_fp(percent, 100);
514 static signed int pid_calc(struct _pid *pid, int32_t busy)
517 int32_t pterm, dterm, fp_error;
518 int32_t integral_limit;
520 fp_error = pid->setpoint - busy;
522 if (abs(fp_error) <= pid->deadband)
525 pterm = mul_fp(pid->p_gain, fp_error);
527 pid->integral += fp_error;
530 * We limit the integral here so that it will never
531 * get higher than 30. This prevents it from becoming
532 * too large an input over long periods of time and allows
533 * it to get factored out sooner.
535 * The value of 30 was chosen through experimentation.
537 integral_limit = int_tofp(30);
538 if (pid->integral > integral_limit)
539 pid->integral = integral_limit;
540 if (pid->integral < -integral_limit)
541 pid->integral = -integral_limit;
543 dterm = mul_fp(pid->d_gain, fp_error - pid->last_err);
544 pid->last_err = fp_error;
546 result = pterm + mul_fp(pid->integral, pid->i_gain) + dterm;
547 result = result + (1 << (FRAC_BITS-1));
548 return (signed int)fp_toint(result);
551 static inline void intel_pstate_busy_pid_reset(struct cpudata *cpu)
553 pid_p_gain_set(&cpu->pid, pid_params.p_gain_pct);
554 pid_d_gain_set(&cpu->pid, pid_params.d_gain_pct);
555 pid_i_gain_set(&cpu->pid, pid_params.i_gain_pct);
557 pid_reset(&cpu->pid, pid_params.setpoint, 100, pid_params.deadband, 0);
560 static inline void intel_pstate_reset_all_pid(void)
564 for_each_online_cpu(cpu) {
565 if (all_cpu_data[cpu])
566 intel_pstate_busy_pid_reset(all_cpu_data[cpu]);
570 static inline void update_turbo_state(void)
575 cpu = all_cpu_data[0];
576 rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
577 limits->turbo_disabled =
578 (misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ||
579 cpu->pstate.max_pstate == cpu->pstate.turbo_pstate);
582 static s16 intel_pstate_get_epb(struct cpudata *cpu_data)
587 if (!static_cpu_has(X86_FEATURE_EPB))
590 ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
594 return (s16)(epb & 0x0f);
597 static s16 intel_pstate_get_epp(struct cpudata *cpu_data, u64 hwp_req_data)
601 if (static_cpu_has(X86_FEATURE_HWP_EPP))
602 epp = (hwp_req_data >> 24) & 0xff;
604 /* When there is no EPP present, HWP uses EPB settings */
605 epp = intel_pstate_get_epb(cpu_data);
610 static void intel_pstate_set_epb(int cpu, s16 pref)
614 if (!static_cpu_has(X86_FEATURE_EPB))
617 if (rdmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb))
620 epb = (epb & ~0x0f) | pref;
621 wrmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, epb);
624 static void intel_pstate_hwp_set(const struct cpumask *cpumask)
626 int min, hw_min, max, hw_max, cpu, range, adj_range;
627 struct perf_limits *perf_limits = limits;
630 for_each_cpu(cpu, cpumask) {
631 int max_perf_pct, min_perf_pct;
632 struct cpudata *cpu_data = all_cpu_data[cpu];
636 perf_limits = all_cpu_data[cpu]->perf_limits;
638 rdmsrl_on_cpu(cpu, MSR_HWP_CAPABILITIES, &cap);
639 hw_min = HWP_LOWEST_PERF(cap);
640 hw_max = HWP_HIGHEST_PERF(cap);
641 range = hw_max - hw_min;
643 max_perf_pct = perf_limits->max_perf_pct;
644 min_perf_pct = perf_limits->min_perf_pct;
646 rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
647 adj_range = min_perf_pct * range / 100;
648 min = hw_min + adj_range;
649 value &= ~HWP_MIN_PERF(~0L);
650 value |= HWP_MIN_PERF(min);
652 adj_range = max_perf_pct * range / 100;
653 max = hw_min + adj_range;
654 if (limits->no_turbo) {
655 hw_max = HWP_GUARANTEED_PERF(cap);
660 value &= ~HWP_MAX_PERF(~0L);
661 value |= HWP_MAX_PERF(max);
663 if (cpu_data->epp_policy == cpu_data->policy)
666 cpu_data->epp_policy = cpu_data->policy;
668 if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE) {
669 epp = intel_pstate_get_epp(cpu_data, value);
670 /* If EPP read was failed, then don't try to write */
672 cpu_data->epp_saved = epp;
676 cpu_data->epp_saved = epp;
680 /* skip setting EPP, when saved value is invalid */
681 if (cpu_data->epp_saved < 0)
685 * No need to restore EPP when it is not zero. This
687 * - Policy is not changed
688 * - user has manually changed
689 * - Error reading EPB
691 epp = intel_pstate_get_epp(cpu_data, value);
695 epp = cpu_data->epp_saved;
697 if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
698 value &= ~GENMASK_ULL(31, 24);
699 value |= (u64)epp << 24;
701 intel_pstate_set_epb(cpu, epp);
704 wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
708 static int intel_pstate_hwp_set_policy(struct cpufreq_policy *policy)
711 intel_pstate_hwp_set(policy->cpus);
716 static int intel_pstate_resume(struct cpufreq_policy *policy)
721 all_cpu_data[policy->cpu]->epp_policy = 0;
722 all_cpu_data[policy->cpu]->epp_saved = -EINVAL;
724 return intel_pstate_hwp_set_policy(policy);
727 static void intel_pstate_hwp_set_online_cpus(void)
730 intel_pstate_hwp_set(cpu_online_mask);
734 /************************** debugfs begin ************************/
735 static int pid_param_set(void *data, u64 val)
738 intel_pstate_reset_all_pid();
742 static int pid_param_get(void *data, u64 *val)
747 DEFINE_SIMPLE_ATTRIBUTE(fops_pid_param, pid_param_get, pid_param_set, "%llu\n");
754 static struct pid_param pid_files[] = {
755 {"sample_rate_ms", &pid_params.sample_rate_ms},
756 {"d_gain_pct", &pid_params.d_gain_pct},
757 {"i_gain_pct", &pid_params.i_gain_pct},
758 {"deadband", &pid_params.deadband},
759 {"setpoint", &pid_params.setpoint},
760 {"p_gain_pct", &pid_params.p_gain_pct},
764 static void __init intel_pstate_debug_expose_params(void)
766 struct dentry *debugfs_parent;
770 pstate_funcs.get_target_pstate == get_target_pstate_use_cpu_load)
773 debugfs_parent = debugfs_create_dir("pstate_snb", NULL);
774 if (IS_ERR_OR_NULL(debugfs_parent))
776 while (pid_files[i].name) {
777 debugfs_create_file(pid_files[i].name, 0660,
778 debugfs_parent, pid_files[i].value,
784 /************************** debugfs end ************************/
786 /************************** sysfs begin ************************/
787 #define show_one(file_name, object) \
788 static ssize_t show_##file_name \
789 (struct kobject *kobj, struct attribute *attr, char *buf) \
791 return sprintf(buf, "%u\n", limits->object); \
794 static ssize_t show_turbo_pct(struct kobject *kobj,
795 struct attribute *attr, char *buf)
798 int total, no_turbo, turbo_pct;
801 cpu = all_cpu_data[0];
803 total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
804 no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
805 turbo_fp = div_fp(no_turbo, total);
806 turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
807 return sprintf(buf, "%u\n", turbo_pct);
810 static ssize_t show_num_pstates(struct kobject *kobj,
811 struct attribute *attr, char *buf)
816 cpu = all_cpu_data[0];
817 total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
818 return sprintf(buf, "%u\n", total);
821 static ssize_t show_no_turbo(struct kobject *kobj,
822 struct attribute *attr, char *buf)
826 update_turbo_state();
827 if (limits->turbo_disabled)
828 ret = sprintf(buf, "%u\n", limits->turbo_disabled);
830 ret = sprintf(buf, "%u\n", limits->no_turbo);
835 static ssize_t store_no_turbo(struct kobject *a, struct attribute *b,
836 const char *buf, size_t count)
841 ret = sscanf(buf, "%u", &input);
845 mutex_lock(&intel_pstate_limits_lock);
847 update_turbo_state();
848 if (limits->turbo_disabled) {
849 pr_warn("Turbo disabled by BIOS or unavailable on processor\n");
850 mutex_unlock(&intel_pstate_limits_lock);
854 limits->no_turbo = clamp_t(int, input, 0, 1);
856 mutex_unlock(&intel_pstate_limits_lock);
859 intel_pstate_hwp_set_online_cpus();
864 static ssize_t store_max_perf_pct(struct kobject *a, struct attribute *b,
865 const char *buf, size_t count)
870 ret = sscanf(buf, "%u", &input);
874 mutex_lock(&intel_pstate_limits_lock);
876 limits->max_sysfs_pct = clamp_t(int, input, 0 , 100);
877 limits->max_perf_pct = min(limits->max_policy_pct,
878 limits->max_sysfs_pct);
879 limits->max_perf_pct = max(limits->min_policy_pct,
880 limits->max_perf_pct);
881 limits->max_perf_pct = max(limits->min_perf_pct,
882 limits->max_perf_pct);
883 limits->max_perf = div_ext_fp(limits->max_perf_pct, 100);
885 mutex_unlock(&intel_pstate_limits_lock);
888 intel_pstate_hwp_set_online_cpus();
892 static ssize_t store_min_perf_pct(struct kobject *a, struct attribute *b,
893 const char *buf, size_t count)
898 ret = sscanf(buf, "%u", &input);
902 mutex_lock(&intel_pstate_limits_lock);
904 limits->min_sysfs_pct = clamp_t(int, input, 0 , 100);
905 limits->min_perf_pct = max(limits->min_policy_pct,
906 limits->min_sysfs_pct);
907 limits->min_perf_pct = min(limits->max_policy_pct,
908 limits->min_perf_pct);
909 limits->min_perf_pct = min(limits->max_perf_pct,
910 limits->min_perf_pct);
911 limits->min_perf = div_ext_fp(limits->min_perf_pct, 100);
913 mutex_unlock(&intel_pstate_limits_lock);
916 intel_pstate_hwp_set_online_cpus();
920 show_one(max_perf_pct, max_perf_pct);
921 show_one(min_perf_pct, min_perf_pct);
923 define_one_global_rw(no_turbo);
924 define_one_global_rw(max_perf_pct);
925 define_one_global_rw(min_perf_pct);
926 define_one_global_ro(turbo_pct);
927 define_one_global_ro(num_pstates);
929 static struct attribute *intel_pstate_attributes[] = {
936 static struct attribute_group intel_pstate_attr_group = {
937 .attrs = intel_pstate_attributes,
940 static void __init intel_pstate_sysfs_expose_params(void)
942 struct kobject *intel_pstate_kobject;
945 intel_pstate_kobject = kobject_create_and_add("intel_pstate",
946 &cpu_subsys.dev_root->kobj);
947 if (WARN_ON(!intel_pstate_kobject))
950 rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
955 * If per cpu limits are enforced there are no global limits, so
956 * return without creating max/min_perf_pct attributes
961 rc = sysfs_create_file(intel_pstate_kobject, &max_perf_pct.attr);
964 rc = sysfs_create_file(intel_pstate_kobject, &min_perf_pct.attr);
968 /************************** sysfs end ************************/
970 static void intel_pstate_hwp_enable(struct cpudata *cpudata)
972 /* First disable HWP notification interrupt as we don't process them */
973 if (static_cpu_has(X86_FEATURE_HWP_NOTIFY))
974 wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
976 wrmsrl_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1);
977 cpudata->epp_policy = 0;
978 cpudata->epp_saved = -EINVAL;
981 static int atom_get_min_pstate(void)
985 rdmsrl(ATOM_RATIOS, value);
986 return (value >> 8) & 0x7F;
989 static int atom_get_max_pstate(void)
993 rdmsrl(ATOM_RATIOS, value);
994 return (value >> 16) & 0x7F;
997 static int atom_get_turbo_pstate(void)
1001 rdmsrl(ATOM_TURBO_RATIOS, value);
1002 return value & 0x7F;
1005 static u64 atom_get_val(struct cpudata *cpudata, int pstate)
1011 val = (u64)pstate << 8;
1012 if (limits->no_turbo && !limits->turbo_disabled)
1013 val |= (u64)1 << 32;
1015 vid_fp = cpudata->vid.min + mul_fp(
1016 int_tofp(pstate - cpudata->pstate.min_pstate),
1017 cpudata->vid.ratio);
1019 vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
1020 vid = ceiling_fp(vid_fp);
1022 if (pstate > cpudata->pstate.max_pstate)
1023 vid = cpudata->vid.turbo;
1028 static int silvermont_get_scaling(void)
1032 /* Defined in Table 35-6 from SDM (Sept 2015) */
1033 static int silvermont_freq_table[] = {
1034 83300, 100000, 133300, 116700, 80000};
1036 rdmsrl(MSR_FSB_FREQ, value);
1040 return silvermont_freq_table[i];
1043 static int airmont_get_scaling(void)
1047 /* Defined in Table 35-10 from SDM (Sept 2015) */
1048 static int airmont_freq_table[] = {
1049 83300, 100000, 133300, 116700, 80000,
1050 93300, 90000, 88900, 87500};
1052 rdmsrl(MSR_FSB_FREQ, value);
1056 return airmont_freq_table[i];
1059 static void atom_get_vid(struct cpudata *cpudata)
1063 rdmsrl(ATOM_VIDS, value);
1064 cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
1065 cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
1066 cpudata->vid.ratio = div_fp(
1067 cpudata->vid.max - cpudata->vid.min,
1068 int_tofp(cpudata->pstate.max_pstate -
1069 cpudata->pstate.min_pstate));
1071 rdmsrl(ATOM_TURBO_VIDS, value);
1072 cpudata->vid.turbo = value & 0x7f;
1075 static int core_get_min_pstate(void)
1079 rdmsrl(MSR_PLATFORM_INFO, value);
1080 return (value >> 40) & 0xFF;
1083 static int core_get_max_pstate_physical(void)
1087 rdmsrl(MSR_PLATFORM_INFO, value);
1088 return (value >> 8) & 0xFF;
1091 static int core_get_max_pstate(void)
1098 rdmsrl(MSR_PLATFORM_INFO, plat_info);
1099 max_pstate = (plat_info >> 8) & 0xFF;
1101 err = rdmsrl_safe(MSR_TURBO_ACTIVATION_RATIO, &tar);
1103 /* Do some sanity checking for safety */
1104 if (plat_info & 0x600000000) {
1109 err = rdmsrl_safe(MSR_CONFIG_TDP_CONTROL, &tdp_ctrl);
1113 tdp_msr = MSR_CONFIG_TDP_NOMINAL + (tdp_ctrl & 0x3);
1114 err = rdmsrl_safe(tdp_msr, &tdp_ratio);
1118 /* For level 1 and 2, bits[23:16] contain the ratio */
1122 tdp_ratio &= 0xff; /* ratios are only 8 bits long */
1123 if (tdp_ratio - 1 == tar) {
1125 pr_debug("max_pstate=TAC %x\n", max_pstate);
1136 static int core_get_turbo_pstate(void)
1141 rdmsrl(MSR_TURBO_RATIO_LIMIT, value);
1142 nont = core_get_max_pstate();
1143 ret = (value) & 255;
1149 static inline int core_get_scaling(void)
1154 static u64 core_get_val(struct cpudata *cpudata, int pstate)
1158 val = (u64)pstate << 8;
1159 if (limits->no_turbo && !limits->turbo_disabled)
1160 val |= (u64)1 << 32;
1165 static int knl_get_turbo_pstate(void)
1170 rdmsrl(MSR_TURBO_RATIO_LIMIT, value);
1171 nont = core_get_max_pstate();
1172 ret = (((value) >> 8) & 0xFF);
1178 static struct cpu_defaults core_params = {
1180 .sample_rate_ms = 10,
1188 .get_max = core_get_max_pstate,
1189 .get_max_physical = core_get_max_pstate_physical,
1190 .get_min = core_get_min_pstate,
1191 .get_turbo = core_get_turbo_pstate,
1192 .get_scaling = core_get_scaling,
1193 .get_val = core_get_val,
1194 .get_target_pstate = get_target_pstate_use_performance,
1198 static const struct cpu_defaults silvermont_params = {
1200 .sample_rate_ms = 10,
1208 .get_max = atom_get_max_pstate,
1209 .get_max_physical = atom_get_max_pstate,
1210 .get_min = atom_get_min_pstate,
1211 .get_turbo = atom_get_turbo_pstate,
1212 .get_val = atom_get_val,
1213 .get_scaling = silvermont_get_scaling,
1214 .get_vid = atom_get_vid,
1215 .get_target_pstate = get_target_pstate_use_cpu_load,
1219 static const struct cpu_defaults airmont_params = {
1221 .sample_rate_ms = 10,
1229 .get_max = atom_get_max_pstate,
1230 .get_max_physical = atom_get_max_pstate,
1231 .get_min = atom_get_min_pstate,
1232 .get_turbo = atom_get_turbo_pstate,
1233 .get_val = atom_get_val,
1234 .get_scaling = airmont_get_scaling,
1235 .get_vid = atom_get_vid,
1236 .get_target_pstate = get_target_pstate_use_cpu_load,
1240 static const struct cpu_defaults knl_params = {
1242 .sample_rate_ms = 10,
1250 .get_max = core_get_max_pstate,
1251 .get_max_physical = core_get_max_pstate_physical,
1252 .get_min = core_get_min_pstate,
1253 .get_turbo = knl_get_turbo_pstate,
1254 .get_scaling = core_get_scaling,
1255 .get_val = core_get_val,
1256 .get_target_pstate = get_target_pstate_use_performance,
1260 static const struct cpu_defaults bxt_params = {
1262 .sample_rate_ms = 10,
1270 .get_max = core_get_max_pstate,
1271 .get_max_physical = core_get_max_pstate_physical,
1272 .get_min = core_get_min_pstate,
1273 .get_turbo = core_get_turbo_pstate,
1274 .get_scaling = core_get_scaling,
1275 .get_val = core_get_val,
1276 .get_target_pstate = get_target_pstate_use_cpu_load,
1280 static void intel_pstate_get_min_max(struct cpudata *cpu, int *min, int *max)
1282 int max_perf = cpu->pstate.turbo_pstate;
1285 struct perf_limits *perf_limits = limits;
1287 if (limits->no_turbo || limits->turbo_disabled)
1288 max_perf = cpu->pstate.max_pstate;
1291 perf_limits = cpu->perf_limits;
1294 * performance can be limited by user through sysfs, by cpufreq
1295 * policy, or by cpu specific default values determined through
1298 max_perf_adj = fp_ext_toint(max_perf * perf_limits->max_perf);
1299 *max = clamp_t(int, max_perf_adj,
1300 cpu->pstate.min_pstate, cpu->pstate.turbo_pstate);
1302 min_perf = fp_ext_toint(max_perf * perf_limits->min_perf);
1303 *min = clamp_t(int, min_perf, cpu->pstate.min_pstate, max_perf);
1306 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
1308 trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
1309 cpu->pstate.current_pstate = pstate;
1311 * Generally, there is no guarantee that this code will always run on
1312 * the CPU being updated, so force the register update to run on the
1315 wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
1316 pstate_funcs.get_val(cpu, pstate));
1319 static void intel_pstate_set_min_pstate(struct cpudata *cpu)
1321 intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
1324 static void intel_pstate_max_within_limits(struct cpudata *cpu)
1326 int min_pstate, max_pstate;
1328 update_turbo_state();
1329 intel_pstate_get_min_max(cpu, &min_pstate, &max_pstate);
1330 intel_pstate_set_pstate(cpu, max_pstate);
1333 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
1335 cpu->pstate.min_pstate = pstate_funcs.get_min();
1336 cpu->pstate.max_pstate = pstate_funcs.get_max();
1337 cpu->pstate.max_pstate_physical = pstate_funcs.get_max_physical();
1338 cpu->pstate.turbo_pstate = pstate_funcs.get_turbo();
1339 cpu->pstate.scaling = pstate_funcs.get_scaling();
1340 cpu->pstate.max_freq = cpu->pstate.max_pstate * cpu->pstate.scaling;
1341 cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
1343 if (pstate_funcs.get_vid)
1344 pstate_funcs.get_vid(cpu);
1346 intel_pstate_set_min_pstate(cpu);
1349 static inline void intel_pstate_calc_avg_perf(struct cpudata *cpu)
1351 struct sample *sample = &cpu->sample;
1353 sample->core_avg_perf = div_ext_fp(sample->aperf, sample->mperf);
1356 static inline bool intel_pstate_sample(struct cpudata *cpu, u64 time)
1359 unsigned long flags;
1362 local_irq_save(flags);
1363 rdmsrl(MSR_IA32_APERF, aperf);
1364 rdmsrl(MSR_IA32_MPERF, mperf);
1366 if (cpu->prev_mperf == mperf || cpu->prev_tsc == tsc) {
1367 local_irq_restore(flags);
1370 local_irq_restore(flags);
1372 cpu->last_sample_time = cpu->sample.time;
1373 cpu->sample.time = time;
1374 cpu->sample.aperf = aperf;
1375 cpu->sample.mperf = mperf;
1376 cpu->sample.tsc = tsc;
1377 cpu->sample.aperf -= cpu->prev_aperf;
1378 cpu->sample.mperf -= cpu->prev_mperf;
1379 cpu->sample.tsc -= cpu->prev_tsc;
1381 cpu->prev_aperf = aperf;
1382 cpu->prev_mperf = mperf;
1383 cpu->prev_tsc = tsc;
1385 * First time this function is invoked in a given cycle, all of the
1386 * previous sample data fields are equal to zero or stale and they must
1387 * be populated with meaningful numbers for things to work, so assume
1388 * that sample.time will always be reset before setting the utilization
1389 * update hook and make the caller skip the sample then.
1391 return !!cpu->last_sample_time;
1394 static inline int32_t get_avg_frequency(struct cpudata *cpu)
1396 return mul_ext_fp(cpu->sample.core_avg_perf,
1397 cpu->pstate.max_pstate_physical * cpu->pstate.scaling);
1400 static inline int32_t get_avg_pstate(struct cpudata *cpu)
1402 return mul_ext_fp(cpu->pstate.max_pstate_physical,
1403 cpu->sample.core_avg_perf);
1406 static inline int32_t get_target_pstate_use_cpu_load(struct cpudata *cpu)
1408 struct sample *sample = &cpu->sample;
1409 int32_t busy_frac, boost;
1410 int target, avg_pstate;
1412 busy_frac = div_fp(sample->mperf, sample->tsc);
1414 boost = cpu->iowait_boost;
1415 cpu->iowait_boost >>= 1;
1417 if (busy_frac < boost)
1420 sample->busy_scaled = busy_frac * 100;
1422 target = limits->no_turbo || limits->turbo_disabled ?
1423 cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
1424 target += target >> 2;
1425 target = mul_fp(target, busy_frac);
1426 if (target < cpu->pstate.min_pstate)
1427 target = cpu->pstate.min_pstate;
1430 * If the average P-state during the previous cycle was higher than the
1431 * current target, add 50% of the difference to the target to reduce
1432 * possible performance oscillations and offset possible performance
1433 * loss related to moving the workload from one CPU to another within
1436 avg_pstate = get_avg_pstate(cpu);
1437 if (avg_pstate > target)
1438 target += (avg_pstate - target) >> 1;
1443 static inline int32_t get_target_pstate_use_performance(struct cpudata *cpu)
1445 int32_t perf_scaled, max_pstate, current_pstate, sample_ratio;
1449 * perf_scaled is the ratio of the average P-state during the last
1450 * sampling period to the P-state requested last time (in percent).
1452 * That measures the system's response to the previous P-state
1455 max_pstate = cpu->pstate.max_pstate_physical;
1456 current_pstate = cpu->pstate.current_pstate;
1457 perf_scaled = mul_ext_fp(cpu->sample.core_avg_perf,
1458 div_fp(100 * max_pstate, current_pstate));
1461 * Since our utilization update callback will not run unless we are
1462 * in C0, check if the actual elapsed time is significantly greater (3x)
1463 * than our sample interval. If it is, then we were idle for a long
1464 * enough period of time to adjust our performance metric.
1466 duration_ns = cpu->sample.time - cpu->last_sample_time;
1467 if ((s64)duration_ns > pid_params.sample_rate_ns * 3) {
1468 sample_ratio = div_fp(pid_params.sample_rate_ns, duration_ns);
1469 perf_scaled = mul_fp(perf_scaled, sample_ratio);
1471 sample_ratio = div_fp(100 * cpu->sample.mperf, cpu->sample.tsc);
1472 if (sample_ratio < int_tofp(1))
1476 cpu->sample.busy_scaled = perf_scaled;
1477 return cpu->pstate.current_pstate - pid_calc(&cpu->pid, perf_scaled);
1480 static int intel_pstate_prepare_request(struct cpudata *cpu, int pstate)
1482 int max_perf, min_perf;
1484 intel_pstate_get_min_max(cpu, &min_perf, &max_perf);
1485 pstate = clamp_t(int, pstate, min_perf, max_perf);
1486 trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
1490 static void intel_pstate_update_pstate(struct cpudata *cpu, int pstate)
1492 pstate = intel_pstate_prepare_request(cpu, pstate);
1493 if (pstate == cpu->pstate.current_pstate)
1496 cpu->pstate.current_pstate = pstate;
1497 wrmsrl(MSR_IA32_PERF_CTL, pstate_funcs.get_val(cpu, pstate));
1500 static inline void intel_pstate_adjust_busy_pstate(struct cpudata *cpu)
1502 int from, target_pstate;
1503 struct sample *sample;
1505 from = cpu->pstate.current_pstate;
1507 target_pstate = cpu->policy == CPUFREQ_POLICY_PERFORMANCE ?
1508 cpu->pstate.turbo_pstate : pstate_funcs.get_target_pstate(cpu);
1510 update_turbo_state();
1512 intel_pstate_update_pstate(cpu, target_pstate);
1514 sample = &cpu->sample;
1515 trace_pstate_sample(mul_ext_fp(100, sample->core_avg_perf),
1516 fp_toint(sample->busy_scaled),
1518 cpu->pstate.current_pstate,
1522 get_avg_frequency(cpu),
1523 fp_toint(cpu->iowait_boost * 100));
1526 static void intel_pstate_update_util(struct update_util_data *data, u64 time,
1529 struct cpudata *cpu = container_of(data, struct cpudata, update_util);
1532 if (pstate_funcs.get_target_pstate == get_target_pstate_use_cpu_load) {
1533 if (flags & SCHED_CPUFREQ_IOWAIT) {
1534 cpu->iowait_boost = int_tofp(1);
1535 } else if (cpu->iowait_boost) {
1536 /* Clear iowait_boost if the CPU may have been idle. */
1537 delta_ns = time - cpu->last_update;
1538 if (delta_ns > TICK_NSEC)
1539 cpu->iowait_boost = 0;
1541 cpu->last_update = time;
1544 delta_ns = time - cpu->sample.time;
1545 if ((s64)delta_ns >= pid_params.sample_rate_ns) {
1546 bool sample_taken = intel_pstate_sample(cpu, time);
1549 intel_pstate_calc_avg_perf(cpu);
1551 intel_pstate_adjust_busy_pstate(cpu);
1556 #define ICPU(model, policy) \
1557 { X86_VENDOR_INTEL, 6, model, X86_FEATURE_APERFMPERF,\
1558 (unsigned long)&policy }
1560 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
1561 ICPU(INTEL_FAM6_SANDYBRIDGE, core_params),
1562 ICPU(INTEL_FAM6_SANDYBRIDGE_X, core_params),
1563 ICPU(INTEL_FAM6_ATOM_SILVERMONT1, silvermont_params),
1564 ICPU(INTEL_FAM6_IVYBRIDGE, core_params),
1565 ICPU(INTEL_FAM6_HASWELL_CORE, core_params),
1566 ICPU(INTEL_FAM6_BROADWELL_CORE, core_params),
1567 ICPU(INTEL_FAM6_IVYBRIDGE_X, core_params),
1568 ICPU(INTEL_FAM6_HASWELL_X, core_params),
1569 ICPU(INTEL_FAM6_HASWELL_ULT, core_params),
1570 ICPU(INTEL_FAM6_HASWELL_GT3E, core_params),
1571 ICPU(INTEL_FAM6_BROADWELL_GT3E, core_params),
1572 ICPU(INTEL_FAM6_ATOM_AIRMONT, airmont_params),
1573 ICPU(INTEL_FAM6_SKYLAKE_MOBILE, core_params),
1574 ICPU(INTEL_FAM6_BROADWELL_X, core_params),
1575 ICPU(INTEL_FAM6_SKYLAKE_DESKTOP, core_params),
1576 ICPU(INTEL_FAM6_BROADWELL_XEON_D, core_params),
1577 ICPU(INTEL_FAM6_XEON_PHI_KNL, knl_params),
1578 ICPU(INTEL_FAM6_XEON_PHI_KNM, knl_params),
1579 ICPU(INTEL_FAM6_ATOM_GOLDMONT, bxt_params),
1582 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
1584 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] __initconst = {
1585 ICPU(INTEL_FAM6_BROADWELL_XEON_D, core_params),
1586 ICPU(INTEL_FAM6_BROADWELL_X, core_params),
1587 ICPU(INTEL_FAM6_SKYLAKE_X, core_params),
1591 static int intel_pstate_init_cpu(unsigned int cpunum)
1593 struct cpudata *cpu;
1595 cpu = all_cpu_data[cpunum];
1598 unsigned int size = sizeof(struct cpudata);
1601 size += sizeof(struct perf_limits);
1603 cpu = kzalloc(size, GFP_KERNEL);
1607 all_cpu_data[cpunum] = cpu;
1609 cpu->perf_limits = (struct perf_limits *)(cpu + 1);
1613 cpu = all_cpu_data[cpunum];
1618 intel_pstate_hwp_enable(cpu);
1619 pid_params.sample_rate_ms = 50;
1620 pid_params.sample_rate_ns = 50 * NSEC_PER_MSEC;
1623 intel_pstate_get_cpu_pstates(cpu);
1625 intel_pstate_busy_pid_reset(cpu);
1627 pr_debug("controlling: cpu %d\n", cpunum);
1632 static unsigned int intel_pstate_get(unsigned int cpu_num)
1634 struct cpudata *cpu = all_cpu_data[cpu_num];
1636 return cpu ? get_avg_frequency(cpu) : 0;
1639 static void intel_pstate_set_update_util_hook(unsigned int cpu_num)
1641 struct cpudata *cpu = all_cpu_data[cpu_num];
1643 if (cpu->update_util_set)
1646 /* Prevent intel_pstate_update_util() from using stale data. */
1647 cpu->sample.time = 0;
1648 cpufreq_add_update_util_hook(cpu_num, &cpu->update_util,
1649 intel_pstate_update_util);
1650 cpu->update_util_set = true;
1653 static void intel_pstate_clear_update_util_hook(unsigned int cpu)
1655 struct cpudata *cpu_data = all_cpu_data[cpu];
1657 if (!cpu_data->update_util_set)
1660 cpufreq_remove_update_util_hook(cpu);
1661 cpu_data->update_util_set = false;
1662 synchronize_sched();
1665 static void intel_pstate_set_performance_limits(struct perf_limits *limits)
1667 mutex_lock(&intel_pstate_limits_lock);
1668 limits->no_turbo = 0;
1669 limits->turbo_disabled = 0;
1670 limits->max_perf_pct = 100;
1671 limits->max_perf = int_ext_tofp(1);
1672 limits->min_perf_pct = 100;
1673 limits->min_perf = int_ext_tofp(1);
1674 limits->max_policy_pct = 100;
1675 limits->max_sysfs_pct = 100;
1676 limits->min_policy_pct = 0;
1677 limits->min_sysfs_pct = 0;
1678 mutex_unlock(&intel_pstate_limits_lock);
1681 static void intel_pstate_update_perf_limits(struct cpufreq_policy *policy,
1682 struct perf_limits *limits)
1685 mutex_lock(&intel_pstate_limits_lock);
1687 limits->max_policy_pct = DIV_ROUND_UP(policy->max * 100,
1688 policy->cpuinfo.max_freq);
1689 limits->max_policy_pct = clamp_t(int, limits->max_policy_pct, 0, 100);
1690 if (policy->max == policy->min) {
1691 limits->min_policy_pct = limits->max_policy_pct;
1693 limits->min_policy_pct = DIV_ROUND_UP(policy->min * 100,
1694 policy->cpuinfo.max_freq);
1695 limits->min_policy_pct = clamp_t(int, limits->min_policy_pct,
1699 /* Normalize user input to [min_policy_pct, max_policy_pct] */
1700 limits->min_perf_pct = max(limits->min_policy_pct,
1701 limits->min_sysfs_pct);
1702 limits->min_perf_pct = min(limits->max_policy_pct,
1703 limits->min_perf_pct);
1704 limits->max_perf_pct = min(limits->max_policy_pct,
1705 limits->max_sysfs_pct);
1706 limits->max_perf_pct = max(limits->min_policy_pct,
1707 limits->max_perf_pct);
1709 /* Make sure min_perf_pct <= max_perf_pct */
1710 limits->min_perf_pct = min(limits->max_perf_pct, limits->min_perf_pct);
1712 limits->min_perf = div_ext_fp(limits->min_perf_pct, 100);
1713 limits->max_perf = div_ext_fp(limits->max_perf_pct, 100);
1714 limits->max_perf = round_up(limits->max_perf, EXT_FRAC_BITS);
1715 limits->min_perf = round_up(limits->min_perf, EXT_FRAC_BITS);
1717 mutex_unlock(&intel_pstate_limits_lock);
1719 pr_debug("cpu:%d max_perf_pct:%d min_perf_pct:%d\n", policy->cpu,
1720 limits->max_perf_pct, limits->min_perf_pct);
1723 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
1725 struct cpudata *cpu;
1726 struct perf_limits *perf_limits = NULL;
1728 if (!policy->cpuinfo.max_freq)
1731 pr_debug("set_policy cpuinfo.max %u policy->max %u\n",
1732 policy->cpuinfo.max_freq, policy->max);
1734 cpu = all_cpu_data[policy->cpu];
1735 cpu->policy = policy->policy;
1737 if (cpu->pstate.max_pstate_physical > cpu->pstate.max_pstate &&
1738 policy->max < policy->cpuinfo.max_freq &&
1739 policy->max > cpu->pstate.max_pstate * cpu->pstate.scaling) {
1740 pr_debug("policy->max > max non turbo frequency\n");
1741 policy->max = policy->cpuinfo.max_freq;
1745 perf_limits = cpu->perf_limits;
1747 if (policy->policy == CPUFREQ_POLICY_PERFORMANCE) {
1749 limits = &performance_limits;
1750 perf_limits = limits;
1752 if (policy->max >= policy->cpuinfo.max_freq) {
1753 pr_debug("set performance\n");
1754 intel_pstate_set_performance_limits(perf_limits);
1758 pr_debug("set powersave\n");
1760 limits = &powersave_limits;
1761 perf_limits = limits;
1766 intel_pstate_update_perf_limits(policy, perf_limits);
1768 if (cpu->policy == CPUFREQ_POLICY_PERFORMANCE) {
1770 * NOHZ_FULL CPUs need this as the governor callback may not
1771 * be invoked on them.
1773 intel_pstate_clear_update_util_hook(policy->cpu);
1774 intel_pstate_max_within_limits(cpu);
1777 intel_pstate_set_update_util_hook(policy->cpu);
1779 intel_pstate_hwp_set_policy(policy);
1784 static int intel_pstate_verify_policy(struct cpufreq_policy *policy)
1786 cpufreq_verify_within_cpu_limits(policy);
1788 if (policy->policy != CPUFREQ_POLICY_POWERSAVE &&
1789 policy->policy != CPUFREQ_POLICY_PERFORMANCE)
1795 static void intel_cpufreq_stop_cpu(struct cpufreq_policy *policy)
1797 intel_pstate_set_min_pstate(all_cpu_data[policy->cpu]);
1800 static void intel_pstate_stop_cpu(struct cpufreq_policy *policy)
1802 pr_debug("CPU %d exiting\n", policy->cpu);
1804 intel_pstate_clear_update_util_hook(policy->cpu);
1806 intel_cpufreq_stop_cpu(policy);
1809 static int intel_pstate_cpu_exit(struct cpufreq_policy *policy)
1811 intel_pstate_exit_perf_limits(policy);
1813 policy->fast_switch_possible = false;
1818 static int __intel_pstate_cpu_init(struct cpufreq_policy *policy)
1820 struct cpudata *cpu;
1823 rc = intel_pstate_init_cpu(policy->cpu);
1827 cpu = all_cpu_data[policy->cpu];
1830 * We need sane value in the cpu->perf_limits, so inherit from global
1831 * perf_limits limits, which are seeded with values based on the
1832 * CONFIG_CPU_FREQ_DEFAULT_GOV_*, during boot up.
1835 memcpy(cpu->perf_limits, limits, sizeof(struct perf_limits));
1837 policy->min = cpu->pstate.min_pstate * cpu->pstate.scaling;
1838 policy->max = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
1840 /* cpuinfo and default policy values */
1841 policy->cpuinfo.min_freq = cpu->pstate.min_pstate * cpu->pstate.scaling;
1842 update_turbo_state();
1843 policy->cpuinfo.max_freq = limits->turbo_disabled ?
1844 cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
1845 policy->cpuinfo.max_freq *= cpu->pstate.scaling;
1847 intel_pstate_init_acpi_perf_limits(policy);
1848 cpumask_set_cpu(policy->cpu, policy->cpus);
1850 policy->fast_switch_possible = true;
1855 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
1857 int ret = __intel_pstate_cpu_init(policy);
1862 policy->cpuinfo.transition_latency = CPUFREQ_ETERNAL;
1863 if (limits->min_perf_pct == 100 && limits->max_perf_pct == 100)
1864 policy->policy = CPUFREQ_POLICY_PERFORMANCE;
1866 policy->policy = CPUFREQ_POLICY_POWERSAVE;
1871 static struct cpufreq_driver intel_pstate = {
1872 .flags = CPUFREQ_CONST_LOOPS,
1873 .verify = intel_pstate_verify_policy,
1874 .setpolicy = intel_pstate_set_policy,
1875 .resume = intel_pstate_resume,
1876 .get = intel_pstate_get,
1877 .init = intel_pstate_cpu_init,
1878 .exit = intel_pstate_cpu_exit,
1879 .stop_cpu = intel_pstate_stop_cpu,
1880 .name = "intel_pstate",
1883 static int intel_cpufreq_verify_policy(struct cpufreq_policy *policy)
1885 struct cpudata *cpu = all_cpu_data[policy->cpu];
1886 struct perf_limits *perf_limits = limits;
1888 update_turbo_state();
1889 policy->cpuinfo.max_freq = limits->turbo_disabled ?
1890 cpu->pstate.max_freq : cpu->pstate.turbo_freq;
1892 cpufreq_verify_within_cpu_limits(policy);
1895 perf_limits = cpu->perf_limits;
1897 intel_pstate_update_perf_limits(policy, perf_limits);
1902 static unsigned int intel_cpufreq_turbo_update(struct cpudata *cpu,
1903 struct cpufreq_policy *policy,
1904 unsigned int target_freq)
1906 unsigned int max_freq;
1908 update_turbo_state();
1910 max_freq = limits->no_turbo || limits->turbo_disabled ?
1911 cpu->pstate.max_freq : cpu->pstate.turbo_freq;
1912 policy->cpuinfo.max_freq = max_freq;
1913 if (policy->max > max_freq)
1914 policy->max = max_freq;
1916 if (target_freq > max_freq)
1917 target_freq = max_freq;
1922 static int intel_cpufreq_target(struct cpufreq_policy *policy,
1923 unsigned int target_freq,
1924 unsigned int relation)
1926 struct cpudata *cpu = all_cpu_data[policy->cpu];
1927 struct cpufreq_freqs freqs;
1930 freqs.old = policy->cur;
1931 freqs.new = intel_cpufreq_turbo_update(cpu, policy, target_freq);
1933 cpufreq_freq_transition_begin(policy, &freqs);
1935 case CPUFREQ_RELATION_L:
1936 target_pstate = DIV_ROUND_UP(freqs.new, cpu->pstate.scaling);
1938 case CPUFREQ_RELATION_H:
1939 target_pstate = freqs.new / cpu->pstate.scaling;
1942 target_pstate = DIV_ROUND_CLOSEST(freqs.new, cpu->pstate.scaling);
1945 target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
1946 if (target_pstate != cpu->pstate.current_pstate) {
1947 cpu->pstate.current_pstate = target_pstate;
1948 wrmsrl_on_cpu(policy->cpu, MSR_IA32_PERF_CTL,
1949 pstate_funcs.get_val(cpu, target_pstate));
1951 cpufreq_freq_transition_end(policy, &freqs, false);
1956 static unsigned int intel_cpufreq_fast_switch(struct cpufreq_policy *policy,
1957 unsigned int target_freq)
1959 struct cpudata *cpu = all_cpu_data[policy->cpu];
1962 target_freq = intel_cpufreq_turbo_update(cpu, policy, target_freq);
1963 target_pstate = DIV_ROUND_UP(target_freq, cpu->pstate.scaling);
1964 intel_pstate_update_pstate(cpu, target_pstate);
1968 static int intel_cpufreq_cpu_init(struct cpufreq_policy *policy)
1970 int ret = __intel_pstate_cpu_init(policy);
1975 policy->cpuinfo.transition_latency = INTEL_CPUFREQ_TRANSITION_LATENCY;
1976 /* This reflects the intel_pstate_get_cpu_pstates() setting. */
1977 policy->cur = policy->cpuinfo.min_freq;
1982 static struct cpufreq_driver intel_cpufreq = {
1983 .flags = CPUFREQ_CONST_LOOPS,
1984 .verify = intel_cpufreq_verify_policy,
1985 .target = intel_cpufreq_target,
1986 .fast_switch = intel_cpufreq_fast_switch,
1987 .init = intel_cpufreq_cpu_init,
1988 .exit = intel_pstate_cpu_exit,
1989 .stop_cpu = intel_cpufreq_stop_cpu,
1990 .name = "intel_cpufreq",
1993 static struct cpufreq_driver *intel_pstate_driver = &intel_pstate;
1995 static int no_load __initdata;
1996 static int no_hwp __initdata;
1997 static int hwp_only __initdata;
1998 static unsigned int force_load __initdata;
2000 static int __init intel_pstate_msrs_not_valid(void)
2002 if (!pstate_funcs.get_max() ||
2003 !pstate_funcs.get_min() ||
2004 !pstate_funcs.get_turbo())
2010 static void __init copy_pid_params(struct pstate_adjust_policy *policy)
2012 pid_params.sample_rate_ms = policy->sample_rate_ms;
2013 pid_params.sample_rate_ns = pid_params.sample_rate_ms * NSEC_PER_MSEC;
2014 pid_params.p_gain_pct = policy->p_gain_pct;
2015 pid_params.i_gain_pct = policy->i_gain_pct;
2016 pid_params.d_gain_pct = policy->d_gain_pct;
2017 pid_params.deadband = policy->deadband;
2018 pid_params.setpoint = policy->setpoint;
2022 static void intel_pstate_use_acpi_profile(void)
2024 if (acpi_gbl_FADT.preferred_profile == PM_MOBILE)
2025 pstate_funcs.get_target_pstate =
2026 get_target_pstate_use_cpu_load;
2029 static void intel_pstate_use_acpi_profile(void)
2034 static void __init copy_cpu_funcs(struct pstate_funcs *funcs)
2036 pstate_funcs.get_max = funcs->get_max;
2037 pstate_funcs.get_max_physical = funcs->get_max_physical;
2038 pstate_funcs.get_min = funcs->get_min;
2039 pstate_funcs.get_turbo = funcs->get_turbo;
2040 pstate_funcs.get_scaling = funcs->get_scaling;
2041 pstate_funcs.get_val = funcs->get_val;
2042 pstate_funcs.get_vid = funcs->get_vid;
2043 pstate_funcs.get_target_pstate = funcs->get_target_pstate;
2045 intel_pstate_use_acpi_profile();
2050 static bool __init intel_pstate_no_acpi_pss(void)
2054 for_each_possible_cpu(i) {
2056 union acpi_object *pss;
2057 struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
2058 struct acpi_processor *pr = per_cpu(processors, i);
2063 status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
2064 if (ACPI_FAILURE(status))
2067 pss = buffer.pointer;
2068 if (pss && pss->type == ACPI_TYPE_PACKAGE) {
2079 static bool __init intel_pstate_has_acpi_ppc(void)
2083 for_each_possible_cpu(i) {
2084 struct acpi_processor *pr = per_cpu(processors, i);
2088 if (acpi_has_method(pr->handle, "_PPC"))
2099 struct hw_vendor_info {
2101 char oem_id[ACPI_OEM_ID_SIZE];
2102 char oem_table_id[ACPI_OEM_TABLE_ID_SIZE];
2106 /* Hardware vendor-specific info that has its own power management modes */
2107 static struct hw_vendor_info vendor_info[] __initdata = {
2108 {1, "HP ", "ProLiant", PSS},
2109 {1, "ORACLE", "X4-2 ", PPC},
2110 {1, "ORACLE", "X4-2L ", PPC},
2111 {1, "ORACLE", "X4-2B ", PPC},
2112 {1, "ORACLE", "X3-2 ", PPC},
2113 {1, "ORACLE", "X3-2L ", PPC},
2114 {1, "ORACLE", "X3-2B ", PPC},
2115 {1, "ORACLE", "X4470M2 ", PPC},
2116 {1, "ORACLE", "X4270M3 ", PPC},
2117 {1, "ORACLE", "X4270M2 ", PPC},
2118 {1, "ORACLE", "X4170M2 ", PPC},
2119 {1, "ORACLE", "X4170 M3", PPC},
2120 {1, "ORACLE", "X4275 M3", PPC},
2121 {1, "ORACLE", "X6-2 ", PPC},
2122 {1, "ORACLE", "Sudbury ", PPC},
2126 static bool __init intel_pstate_platform_pwr_mgmt_exists(void)
2128 struct acpi_table_header hdr;
2129 struct hw_vendor_info *v_info;
2130 const struct x86_cpu_id *id;
2133 id = x86_match_cpu(intel_pstate_cpu_oob_ids);
2135 rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
2136 if ( misc_pwr & (1 << 8))
2140 if (acpi_disabled ||
2141 ACPI_FAILURE(acpi_get_table_header(ACPI_SIG_FADT, 0, &hdr)))
2144 for (v_info = vendor_info; v_info->valid; v_info++) {
2145 if (!strncmp(hdr.oem_id, v_info->oem_id, ACPI_OEM_ID_SIZE) &&
2146 !strncmp(hdr.oem_table_id, v_info->oem_table_id,
2147 ACPI_OEM_TABLE_ID_SIZE))
2148 switch (v_info->oem_pwr_table) {
2150 return intel_pstate_no_acpi_pss();
2152 return intel_pstate_has_acpi_ppc() &&
2160 static void intel_pstate_request_control_from_smm(void)
2163 * It may be unsafe to request P-states control from SMM if _PPC support
2164 * has not been enabled.
2167 acpi_processor_pstate_control();
2169 #else /* CONFIG_ACPI not enabled */
2170 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
2171 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
2172 static inline void intel_pstate_request_control_from_smm(void) {}
2173 #endif /* CONFIG_ACPI */
2175 static const struct x86_cpu_id hwp_support_ids[] __initconst = {
2176 { X86_VENDOR_INTEL, 6, X86_MODEL_ANY, X86_FEATURE_HWP },
2180 static int __init intel_pstate_init(void)
2183 const struct x86_cpu_id *id;
2184 struct cpu_defaults *cpu_def;
2189 if (x86_match_cpu(hwp_support_ids) && !no_hwp) {
2190 copy_cpu_funcs(&core_params.funcs);
2192 goto hwp_cpu_matched;
2195 id = x86_match_cpu(intel_pstate_cpu_ids);
2199 cpu_def = (struct cpu_defaults *)id->driver_data;
2201 copy_pid_params(&cpu_def->pid_policy);
2202 copy_cpu_funcs(&cpu_def->funcs);
2204 if (intel_pstate_msrs_not_valid())
2209 * The Intel pstate driver will be ignored if the platform
2210 * firmware has its own power management modes.
2212 if (intel_pstate_platform_pwr_mgmt_exists())
2215 pr_info("Intel P-state driver initializing\n");
2217 all_cpu_data = vzalloc(sizeof(void *) * num_possible_cpus());
2221 if (!hwp_active && hwp_only)
2224 intel_pstate_request_control_from_smm();
2226 rc = cpufreq_register_driver(intel_pstate_driver);
2230 intel_pstate_debug_expose_params();
2231 intel_pstate_sysfs_expose_params();
2234 pr_info("HWP enabled\n");
2239 for_each_online_cpu(cpu) {
2240 if (all_cpu_data[cpu]) {
2241 if (intel_pstate_driver == &intel_pstate)
2242 intel_pstate_clear_update_util_hook(cpu);
2244 kfree(all_cpu_data[cpu]);
2249 vfree(all_cpu_data);
2252 device_initcall(intel_pstate_init);
2254 static int __init intel_pstate_setup(char *str)
2259 if (!strcmp(str, "disable")) {
2261 } else if (!strcmp(str, "passive")) {
2262 pr_info("Passive mode enabled\n");
2263 intel_pstate_driver = &intel_cpufreq;
2266 if (!strcmp(str, "no_hwp")) {
2267 pr_info("HWP disabled\n");
2270 if (!strcmp(str, "force"))
2272 if (!strcmp(str, "hwp_only"))
2274 if (!strcmp(str, "per_cpu_perf_limits"))
2275 per_cpu_limits = true;
2278 if (!strcmp(str, "support_acpi_ppc"))
2284 early_param("intel_pstate", intel_pstate_setup);
2286 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
2287 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
2288 MODULE_LICENSE("GPL");