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 #include <linux/kernel.h>
14 #include <linux/kernel_stat.h>
15 #include <linux/module.h>
16 #include <linux/ktime.h>
17 #include <linux/hrtimer.h>
18 #include <linux/tick.h>
19 #include <linux/slab.h>
20 #include <linux/sched.h>
21 #include <linux/list.h>
22 #include <linux/cpu.h>
23 #include <linux/cpufreq.h>
24 #include <linux/sysfs.h>
25 #include <linux/types.h>
27 #include <linux/debugfs.h>
28 #include <linux/acpi.h>
29 #include <linux/vmalloc.h>
30 #include <trace/events/power.h>
32 #include <asm/div64.h>
34 #include <asm/cpu_device_id.h>
35 #include <asm/cpufeature.h>
37 #define ATOM_RATIOS 0x66a
38 #define ATOM_VIDS 0x66b
39 #define ATOM_TURBO_RATIOS 0x66c
40 #define ATOM_TURBO_VIDS 0x66d
43 #define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
44 #define fp_toint(X) ((X) >> FRAC_BITS)
46 static inline int32_t mul_fp(int32_t x, int32_t y)
48 return ((int64_t)x * (int64_t)y) >> FRAC_BITS;
51 static inline int32_t div_fp(s64 x, s64 y)
53 return div64_s64((int64_t)x << FRAC_BITS, y);
56 static inline int ceiling_fp(int32_t x)
61 mask = (1 << FRAC_BITS) - 1;
68 * struct sample - Store performance sample
69 * @core_pct_busy: Ratio of APERF/MPERF in percent, which is actual
70 * performance during last sample period
71 * @busy_scaled: Scaled busy value which is used to calculate next
72 * P state. This can be different than core_pct_busy
73 * to account for cpu idle period
74 * @aperf: Difference of actual performance frequency clock count
75 * read from APERF MSR between last and current sample
76 * @mperf: Difference of maximum performance frequency clock count
77 * read from MPERF MSR between last and current sample
78 * @tsc: Difference of time stamp counter between last and
80 * @freq: Effective frequency calculated from APERF/MPERF
81 * @time: Current time from scheduler
83 * This structure is used in the cpudata structure to store performance sample
84 * data for choosing next P State.
87 int32_t core_pct_busy;
97 * struct pstate_data - Store P state data
98 * @current_pstate: Current requested P state
99 * @min_pstate: Min P state possible for this platform
100 * @max_pstate: Max P state possible for this platform
101 * @max_pstate_physical:This is physical Max P state for a processor
102 * This can be higher than the max_pstate which can
103 * be limited by platform thermal design power limits
104 * @scaling: Scaling factor to convert frequency to cpufreq
106 * @turbo_pstate: Max Turbo P state possible for this platform
108 * Stores the per cpu model P state limits and current P state.
114 int max_pstate_physical;
120 * struct vid_data - Stores voltage information data
121 * @min: VID data for this platform corresponding to
123 * @max: VID data corresponding to the highest P State.
124 * @turbo: VID data for turbo P state
125 * @ratio: Ratio of (vid max - vid min) /
126 * (max P state - Min P State)
128 * Stores the voltage data for DVFS (Dynamic Voltage and Frequency Scaling)
129 * This data is used in Atom platforms, where in addition to target P state,
130 * the voltage data needs to be specified to select next P State.
140 * struct _pid - Stores PID data
141 * @setpoint: Target set point for busyness or performance
142 * @integral: Storage for accumulated error values
143 * @p_gain: PID proportional gain
144 * @i_gain: PID integral gain
145 * @d_gain: PID derivative gain
146 * @deadband: PID deadband
147 * @last_err: Last error storage for integral part of PID calculation
149 * Stores PID coefficients and last error for PID controller.
162 * struct cpudata - Per CPU instance data storage
163 * @cpu: CPU number for this instance data
164 * @update_util: CPUFreq utility callback information
165 * @pstate: Stores P state limits for this CPU
166 * @vid: Stores VID limits for this CPU
167 * @pid: Stores PID parameters for this CPU
168 * @last_sample_time: Last Sample time
169 * @prev_aperf: Last APERF value read from APERF MSR
170 * @prev_mperf: Last MPERF value read from MPERF MSR
171 * @prev_tsc: Last timestamp counter (TSC) value
172 * @prev_cummulative_iowait: IO Wait time difference from last and
174 * @sample: Storage for storing last Sample data
176 * This structure stores per CPU instance data for all CPUs.
181 struct update_util_data update_util;
183 struct pstate_data pstate;
187 u64 last_sample_time;
191 u64 prev_cummulative_iowait;
192 struct sample sample;
195 static struct cpudata **all_cpu_data;
198 * struct pid_adjust_policy - Stores static PID configuration data
199 * @sample_rate_ms: PID calculation sample rate in ms
200 * @sample_rate_ns: Sample rate calculation in ns
201 * @deadband: PID deadband
202 * @setpoint: PID Setpoint
203 * @p_gain_pct: PID proportional gain
204 * @i_gain_pct: PID integral gain
205 * @d_gain_pct: PID derivative gain
207 * Stores per CPU model static PID configuration data.
209 struct pstate_adjust_policy {
220 * struct pstate_funcs - Per CPU model specific callbacks
221 * @get_max: Callback to get maximum non turbo effective P state
222 * @get_max_physical: Callback to get maximum non turbo physical P state
223 * @get_min: Callback to get minimum P state
224 * @get_turbo: Callback to get turbo P state
225 * @get_scaling: Callback to get frequency scaling factor
226 * @get_val: Callback to convert P state to actual MSR write value
227 * @get_vid: Callback to get VID data for Atom platforms
228 * @get_target_pstate: Callback to a function to calculate next P state to use
230 * Core and Atom CPU models have different way to get P State limits. This
231 * structure is used to store those callbacks.
233 struct pstate_funcs {
234 int (*get_max)(void);
235 int (*get_max_physical)(void);
236 int (*get_min)(void);
237 int (*get_turbo)(void);
238 int (*get_scaling)(void);
239 u64 (*get_val)(struct cpudata*, int pstate);
240 void (*get_vid)(struct cpudata *);
241 int32_t (*get_target_pstate)(struct cpudata *);
245 * struct cpu_defaults- Per CPU model default config data
246 * @pid_policy: PID config data
247 * @funcs: Callback function data
249 struct cpu_defaults {
250 struct pstate_adjust_policy pid_policy;
251 struct pstate_funcs funcs;
254 static inline int32_t get_target_pstate_use_performance(struct cpudata *cpu);
255 static inline int32_t get_target_pstate_use_cpu_load(struct cpudata *cpu);
257 static struct pstate_adjust_policy pid_params;
258 static struct pstate_funcs pstate_funcs;
259 static int hwp_active;
263 * struct perf_limits - Store user and policy limits
264 * @no_turbo: User requested turbo state from intel_pstate sysfs
265 * @turbo_disabled: Platform turbo status either from msr
266 * MSR_IA32_MISC_ENABLE or when maximum available pstate
267 * matches the maximum turbo pstate
268 * @max_perf_pct: Effective maximum performance limit in percentage, this
269 * is minimum of either limits enforced by cpufreq policy
270 * or limits from user set limits via intel_pstate sysfs
271 * @min_perf_pct: Effective minimum performance limit in percentage, this
272 * is maximum of either limits enforced by cpufreq policy
273 * or limits from user set limits via intel_pstate sysfs
274 * @max_perf: This is a scaled value between 0 to 255 for max_perf_pct
275 * This value is used to limit max pstate
276 * @min_perf: This is a scaled value between 0 to 255 for min_perf_pct
277 * This value is used to limit min pstate
278 * @max_policy_pct: The maximum performance in percentage enforced by
279 * cpufreq setpolicy interface
280 * @max_sysfs_pct: The maximum performance in percentage enforced by
281 * intel pstate sysfs interface
282 * @min_policy_pct: The minimum performance in percentage enforced by
283 * cpufreq setpolicy interface
284 * @min_sysfs_pct: The minimum performance in percentage enforced by
285 * intel pstate sysfs interface
287 * Storage for user and policy defined limits.
302 static struct perf_limits performance_limits = {
306 .max_perf = int_tofp(1),
308 .min_perf = int_tofp(1),
309 .max_policy_pct = 100,
310 .max_sysfs_pct = 100,
315 static struct perf_limits powersave_limits = {
319 .max_perf = int_tofp(1),
322 .max_policy_pct = 100,
323 .max_sysfs_pct = 100,
328 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_PERFORMANCE
329 static struct perf_limits *limits = &performance_limits;
331 static struct perf_limits *limits = &powersave_limits;
334 static inline void pid_reset(struct _pid *pid, int setpoint, int busy,
335 int deadband, int integral) {
336 pid->setpoint = int_tofp(setpoint);
337 pid->deadband = int_tofp(deadband);
338 pid->integral = int_tofp(integral);
339 pid->last_err = int_tofp(setpoint) - int_tofp(busy);
342 static inline void pid_p_gain_set(struct _pid *pid, int percent)
344 pid->p_gain = div_fp(percent, 100);
347 static inline void pid_i_gain_set(struct _pid *pid, int percent)
349 pid->i_gain = div_fp(percent, 100);
352 static inline void pid_d_gain_set(struct _pid *pid, int percent)
354 pid->d_gain = div_fp(percent, 100);
357 static signed int pid_calc(struct _pid *pid, int32_t busy)
360 int32_t pterm, dterm, fp_error;
361 int32_t integral_limit;
363 fp_error = pid->setpoint - busy;
365 if (abs(fp_error) <= pid->deadband)
368 pterm = mul_fp(pid->p_gain, fp_error);
370 pid->integral += fp_error;
373 * We limit the integral here so that it will never
374 * get higher than 30. This prevents it from becoming
375 * too large an input over long periods of time and allows
376 * it to get factored out sooner.
378 * The value of 30 was chosen through experimentation.
380 integral_limit = int_tofp(30);
381 if (pid->integral > integral_limit)
382 pid->integral = integral_limit;
383 if (pid->integral < -integral_limit)
384 pid->integral = -integral_limit;
386 dterm = mul_fp(pid->d_gain, fp_error - pid->last_err);
387 pid->last_err = fp_error;
389 result = pterm + mul_fp(pid->integral, pid->i_gain) + dterm;
390 result = result + (1 << (FRAC_BITS-1));
391 return (signed int)fp_toint(result);
394 static inline void intel_pstate_busy_pid_reset(struct cpudata *cpu)
396 pid_p_gain_set(&cpu->pid, pid_params.p_gain_pct);
397 pid_d_gain_set(&cpu->pid, pid_params.d_gain_pct);
398 pid_i_gain_set(&cpu->pid, pid_params.i_gain_pct);
400 pid_reset(&cpu->pid, pid_params.setpoint, 100, pid_params.deadband, 0);
403 static inline void intel_pstate_reset_all_pid(void)
407 for_each_online_cpu(cpu) {
408 if (all_cpu_data[cpu])
409 intel_pstate_busy_pid_reset(all_cpu_data[cpu]);
413 static inline void update_turbo_state(void)
418 cpu = all_cpu_data[0];
419 rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
420 limits->turbo_disabled =
421 (misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ||
422 cpu->pstate.max_pstate == cpu->pstate.turbo_pstate);
425 static void intel_pstate_hwp_set(const struct cpumask *cpumask)
427 int min, hw_min, max, hw_max, cpu, range, adj_range;
430 rdmsrl(MSR_HWP_CAPABILITIES, cap);
431 hw_min = HWP_LOWEST_PERF(cap);
432 hw_max = HWP_HIGHEST_PERF(cap);
433 range = hw_max - hw_min;
435 for_each_cpu(cpu, cpumask) {
436 rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
437 adj_range = limits->min_perf_pct * range / 100;
438 min = hw_min + adj_range;
439 value &= ~HWP_MIN_PERF(~0L);
440 value |= HWP_MIN_PERF(min);
442 adj_range = limits->max_perf_pct * range / 100;
443 max = hw_min + adj_range;
444 if (limits->no_turbo) {
445 hw_max = HWP_GUARANTEED_PERF(cap);
450 value &= ~HWP_MAX_PERF(~0L);
451 value |= HWP_MAX_PERF(max);
452 wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
456 static void intel_pstate_hwp_set_online_cpus(void)
459 intel_pstate_hwp_set(cpu_online_mask);
463 /************************** debugfs begin ************************/
464 static int pid_param_set(void *data, u64 val)
467 intel_pstate_reset_all_pid();
471 static int pid_param_get(void *data, u64 *val)
476 DEFINE_SIMPLE_ATTRIBUTE(fops_pid_param, pid_param_get, pid_param_set, "%llu\n");
483 static struct pid_param pid_files[] = {
484 {"sample_rate_ms", &pid_params.sample_rate_ms},
485 {"d_gain_pct", &pid_params.d_gain_pct},
486 {"i_gain_pct", &pid_params.i_gain_pct},
487 {"deadband", &pid_params.deadband},
488 {"setpoint", &pid_params.setpoint},
489 {"p_gain_pct", &pid_params.p_gain_pct},
493 static void __init intel_pstate_debug_expose_params(void)
495 struct dentry *debugfs_parent;
500 debugfs_parent = debugfs_create_dir("pstate_snb", NULL);
501 if (IS_ERR_OR_NULL(debugfs_parent))
503 while (pid_files[i].name) {
504 debugfs_create_file(pid_files[i].name, 0660,
505 debugfs_parent, pid_files[i].value,
511 /************************** debugfs end ************************/
513 /************************** sysfs begin ************************/
514 #define show_one(file_name, object) \
515 static ssize_t show_##file_name \
516 (struct kobject *kobj, struct attribute *attr, char *buf) \
518 return sprintf(buf, "%u\n", limits->object); \
521 static ssize_t show_turbo_pct(struct kobject *kobj,
522 struct attribute *attr, char *buf)
525 int total, no_turbo, turbo_pct;
528 cpu = all_cpu_data[0];
530 total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
531 no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
532 turbo_fp = div_fp(no_turbo, total);
533 turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
534 return sprintf(buf, "%u\n", turbo_pct);
537 static ssize_t show_num_pstates(struct kobject *kobj,
538 struct attribute *attr, char *buf)
543 cpu = all_cpu_data[0];
544 total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
545 return sprintf(buf, "%u\n", total);
548 static ssize_t show_no_turbo(struct kobject *kobj,
549 struct attribute *attr, char *buf)
553 update_turbo_state();
554 if (limits->turbo_disabled)
555 ret = sprintf(buf, "%u\n", limits->turbo_disabled);
557 ret = sprintf(buf, "%u\n", limits->no_turbo);
562 static ssize_t store_no_turbo(struct kobject *a, struct attribute *b,
563 const char *buf, size_t count)
568 ret = sscanf(buf, "%u", &input);
572 update_turbo_state();
573 if (limits->turbo_disabled) {
574 pr_warn("intel_pstate: Turbo disabled by BIOS or unavailable on processor\n");
578 limits->no_turbo = clamp_t(int, input, 0, 1);
581 intel_pstate_hwp_set_online_cpus();
586 static ssize_t store_max_perf_pct(struct kobject *a, struct attribute *b,
587 const char *buf, size_t count)
592 ret = sscanf(buf, "%u", &input);
596 limits->max_sysfs_pct = clamp_t(int, input, 0 , 100);
597 limits->max_perf_pct = min(limits->max_policy_pct,
598 limits->max_sysfs_pct);
599 limits->max_perf_pct = max(limits->min_policy_pct,
600 limits->max_perf_pct);
601 limits->max_perf_pct = max(limits->min_perf_pct,
602 limits->max_perf_pct);
603 limits->max_perf = div_fp(limits->max_perf_pct, 100);
606 intel_pstate_hwp_set_online_cpus();
610 static ssize_t store_min_perf_pct(struct kobject *a, struct attribute *b,
611 const char *buf, size_t count)
616 ret = sscanf(buf, "%u", &input);
620 limits->min_sysfs_pct = clamp_t(int, input, 0 , 100);
621 limits->min_perf_pct = max(limits->min_policy_pct,
622 limits->min_sysfs_pct);
623 limits->min_perf_pct = min(limits->max_policy_pct,
624 limits->min_perf_pct);
625 limits->min_perf_pct = min(limits->max_perf_pct,
626 limits->min_perf_pct);
627 limits->min_perf = div_fp(limits->min_perf_pct, 100);
630 intel_pstate_hwp_set_online_cpus();
634 show_one(max_perf_pct, max_perf_pct);
635 show_one(min_perf_pct, min_perf_pct);
637 define_one_global_rw(no_turbo);
638 define_one_global_rw(max_perf_pct);
639 define_one_global_rw(min_perf_pct);
640 define_one_global_ro(turbo_pct);
641 define_one_global_ro(num_pstates);
643 static struct attribute *intel_pstate_attributes[] = {
652 static struct attribute_group intel_pstate_attr_group = {
653 .attrs = intel_pstate_attributes,
656 static void __init intel_pstate_sysfs_expose_params(void)
658 struct kobject *intel_pstate_kobject;
661 intel_pstate_kobject = kobject_create_and_add("intel_pstate",
662 &cpu_subsys.dev_root->kobj);
663 BUG_ON(!intel_pstate_kobject);
664 rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
667 /************************** sysfs end ************************/
669 static void intel_pstate_hwp_enable(struct cpudata *cpudata)
671 /* First disable HWP notification interrupt as we don't process them */
672 wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
674 wrmsrl_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1);
677 static int atom_get_min_pstate(void)
681 rdmsrl(ATOM_RATIOS, value);
682 return (value >> 8) & 0x7F;
685 static int atom_get_max_pstate(void)
689 rdmsrl(ATOM_RATIOS, value);
690 return (value >> 16) & 0x7F;
693 static int atom_get_turbo_pstate(void)
697 rdmsrl(ATOM_TURBO_RATIOS, value);
701 static u64 atom_get_val(struct cpudata *cpudata, int pstate)
707 val = (u64)pstate << 8;
708 if (limits->no_turbo && !limits->turbo_disabled)
711 vid_fp = cpudata->vid.min + mul_fp(
712 int_tofp(pstate - cpudata->pstate.min_pstate),
715 vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
716 vid = ceiling_fp(vid_fp);
718 if (pstate > cpudata->pstate.max_pstate)
719 vid = cpudata->vid.turbo;
724 static int silvermont_get_scaling(void)
728 /* Defined in Table 35-6 from SDM (Sept 2015) */
729 static int silvermont_freq_table[] = {
730 83300, 100000, 133300, 116700, 80000};
732 rdmsrl(MSR_FSB_FREQ, value);
736 return silvermont_freq_table[i];
739 static int airmont_get_scaling(void)
743 /* Defined in Table 35-10 from SDM (Sept 2015) */
744 static int airmont_freq_table[] = {
745 83300, 100000, 133300, 116700, 80000,
746 93300, 90000, 88900, 87500};
748 rdmsrl(MSR_FSB_FREQ, value);
752 return airmont_freq_table[i];
755 static void atom_get_vid(struct cpudata *cpudata)
759 rdmsrl(ATOM_VIDS, value);
760 cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
761 cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
762 cpudata->vid.ratio = div_fp(
763 cpudata->vid.max - cpudata->vid.min,
764 int_tofp(cpudata->pstate.max_pstate -
765 cpudata->pstate.min_pstate));
767 rdmsrl(ATOM_TURBO_VIDS, value);
768 cpudata->vid.turbo = value & 0x7f;
771 static int core_get_min_pstate(void)
775 rdmsrl(MSR_PLATFORM_INFO, value);
776 return (value >> 40) & 0xFF;
779 static int core_get_max_pstate_physical(void)
783 rdmsrl(MSR_PLATFORM_INFO, value);
784 return (value >> 8) & 0xFF;
787 static int core_get_max_pstate(void)
794 rdmsrl(MSR_PLATFORM_INFO, plat_info);
795 max_pstate = (plat_info >> 8) & 0xFF;
797 err = rdmsrl_safe(MSR_TURBO_ACTIVATION_RATIO, &tar);
799 /* Do some sanity checking for safety */
800 if (plat_info & 0x600000000) {
805 err = rdmsrl_safe(MSR_CONFIG_TDP_CONTROL, &tdp_ctrl);
809 tdp_msr = MSR_CONFIG_TDP_NOMINAL + tdp_ctrl;
810 err = rdmsrl_safe(tdp_msr, &tdp_ratio);
814 if (tdp_ratio - 1 == tar) {
816 pr_debug("max_pstate=TAC %x\n", max_pstate);
827 static int core_get_turbo_pstate(void)
832 rdmsrl(MSR_NHM_TURBO_RATIO_LIMIT, value);
833 nont = core_get_max_pstate();
840 static inline int core_get_scaling(void)
845 static u64 core_get_val(struct cpudata *cpudata, int pstate)
849 val = (u64)pstate << 8;
850 if (limits->no_turbo && !limits->turbo_disabled)
856 static int knl_get_turbo_pstate(void)
861 rdmsrl(MSR_NHM_TURBO_RATIO_LIMIT, value);
862 nont = core_get_max_pstate();
863 ret = (((value) >> 8) & 0xFF);
869 static struct cpu_defaults core_params = {
871 .sample_rate_ms = 10,
879 .get_max = core_get_max_pstate,
880 .get_max_physical = core_get_max_pstate_physical,
881 .get_min = core_get_min_pstate,
882 .get_turbo = core_get_turbo_pstate,
883 .get_scaling = core_get_scaling,
884 .get_val = core_get_val,
885 .get_target_pstate = get_target_pstate_use_performance,
889 static struct cpu_defaults silvermont_params = {
891 .sample_rate_ms = 10,
899 .get_max = atom_get_max_pstate,
900 .get_max_physical = atom_get_max_pstate,
901 .get_min = atom_get_min_pstate,
902 .get_turbo = atom_get_turbo_pstate,
903 .get_val = atom_get_val,
904 .get_scaling = silvermont_get_scaling,
905 .get_vid = atom_get_vid,
906 .get_target_pstate = get_target_pstate_use_cpu_load,
910 static struct cpu_defaults airmont_params = {
912 .sample_rate_ms = 10,
920 .get_max = atom_get_max_pstate,
921 .get_max_physical = atom_get_max_pstate,
922 .get_min = atom_get_min_pstate,
923 .get_turbo = atom_get_turbo_pstate,
924 .get_val = atom_get_val,
925 .get_scaling = airmont_get_scaling,
926 .get_vid = atom_get_vid,
927 .get_target_pstate = get_target_pstate_use_cpu_load,
931 static struct cpu_defaults knl_params = {
933 .sample_rate_ms = 10,
941 .get_max = core_get_max_pstate,
942 .get_max_physical = core_get_max_pstate_physical,
943 .get_min = core_get_min_pstate,
944 .get_turbo = knl_get_turbo_pstate,
945 .get_scaling = core_get_scaling,
946 .get_val = core_get_val,
947 .get_target_pstate = get_target_pstate_use_performance,
951 static void intel_pstate_get_min_max(struct cpudata *cpu, int *min, int *max)
953 int max_perf = cpu->pstate.turbo_pstate;
957 if (limits->no_turbo || limits->turbo_disabled)
958 max_perf = cpu->pstate.max_pstate;
961 * performance can be limited by user through sysfs, by cpufreq
962 * policy, or by cpu specific default values determined through
965 max_perf_adj = fp_toint(max_perf * limits->max_perf);
966 *max = clamp_t(int, max_perf_adj,
967 cpu->pstate.min_pstate, cpu->pstate.turbo_pstate);
969 min_perf = fp_toint(max_perf * limits->min_perf);
970 *min = clamp_t(int, min_perf, cpu->pstate.min_pstate, max_perf);
973 static inline void intel_pstate_record_pstate(struct cpudata *cpu, int pstate)
975 trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
976 cpu->pstate.current_pstate = pstate;
979 static void intel_pstate_set_min_pstate(struct cpudata *cpu)
981 int pstate = cpu->pstate.min_pstate;
983 intel_pstate_record_pstate(cpu, pstate);
985 * Generally, there is no guarantee that this code will always run on
986 * the CPU being updated, so force the register update to run on the
989 wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
990 pstate_funcs.get_val(cpu, pstate));
993 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
995 cpu->pstate.min_pstate = pstate_funcs.get_min();
996 cpu->pstate.max_pstate = pstate_funcs.get_max();
997 cpu->pstate.max_pstate_physical = pstate_funcs.get_max_physical();
998 cpu->pstate.turbo_pstate = pstate_funcs.get_turbo();
999 cpu->pstate.scaling = pstate_funcs.get_scaling();
1001 if (pstate_funcs.get_vid)
1002 pstate_funcs.get_vid(cpu);
1004 intel_pstate_set_min_pstate(cpu);
1007 static inline void intel_pstate_calc_busy(struct cpudata *cpu)
1009 struct sample *sample = &cpu->sample;
1012 core_pct = sample->aperf * int_tofp(100);
1013 core_pct = div64_u64(core_pct, sample->mperf);
1015 sample->core_pct_busy = (int32_t)core_pct;
1018 static inline bool intel_pstate_sample(struct cpudata *cpu, u64 time)
1021 unsigned long flags;
1024 local_irq_save(flags);
1025 rdmsrl(MSR_IA32_APERF, aperf);
1026 rdmsrl(MSR_IA32_MPERF, mperf);
1028 if (cpu->prev_mperf == mperf || cpu->prev_tsc == tsc) {
1029 local_irq_restore(flags);
1032 local_irq_restore(flags);
1034 cpu->last_sample_time = cpu->sample.time;
1035 cpu->sample.time = time;
1036 cpu->sample.aperf = aperf;
1037 cpu->sample.mperf = mperf;
1038 cpu->sample.tsc = tsc;
1039 cpu->sample.aperf -= cpu->prev_aperf;
1040 cpu->sample.mperf -= cpu->prev_mperf;
1041 cpu->sample.tsc -= cpu->prev_tsc;
1043 cpu->prev_aperf = aperf;
1044 cpu->prev_mperf = mperf;
1045 cpu->prev_tsc = tsc;
1047 * First time this function is invoked in a given cycle, all of the
1048 * previous sample data fields are equal to zero or stale and they must
1049 * be populated with meaningful numbers for things to work, so assume
1050 * that sample.time will always be reset before setting the utilization
1051 * update hook and make the caller skip the sample then.
1053 return !!cpu->last_sample_time;
1056 static inline int32_t get_avg_frequency(struct cpudata *cpu)
1058 return div64_u64(cpu->pstate.max_pstate_physical * cpu->sample.aperf *
1059 cpu->pstate.scaling, cpu->sample.mperf);
1062 static inline int32_t get_target_pstate_use_cpu_load(struct cpudata *cpu)
1064 struct sample *sample = &cpu->sample;
1065 u64 cummulative_iowait, delta_iowait_us;
1066 u64 delta_iowait_mperf;
1070 cummulative_iowait = get_cpu_iowait_time_us(cpu->cpu, &now);
1073 * Convert iowait time into number of IO cycles spent at max_freq.
1074 * IO is considered as busy only for the cpu_load algorithm. For
1075 * performance this is not needed since we always try to reach the
1076 * maximum P-State, so we are already boosting the IOs.
1078 delta_iowait_us = cummulative_iowait - cpu->prev_cummulative_iowait;
1079 delta_iowait_mperf = div64_u64(delta_iowait_us * cpu->pstate.scaling *
1080 cpu->pstate.max_pstate, MSEC_PER_SEC);
1082 mperf = cpu->sample.mperf + delta_iowait_mperf;
1083 cpu->prev_cummulative_iowait = cummulative_iowait;
1086 * The load can be estimated as the ratio of the mperf counter
1087 * running at a constant frequency during active periods
1088 * (C0) and the time stamp counter running at the same frequency
1089 * also during C-states.
1091 cpu_load = div64_u64(int_tofp(100) * mperf, sample->tsc);
1092 cpu->sample.busy_scaled = cpu_load;
1094 return cpu->pstate.current_pstate - pid_calc(&cpu->pid, cpu_load);
1097 static inline int32_t get_target_pstate_use_performance(struct cpudata *cpu)
1099 int32_t core_busy, max_pstate, current_pstate, sample_ratio;
1102 intel_pstate_calc_busy(cpu);
1105 * core_busy is the ratio of actual performance to max
1106 * max_pstate is the max non turbo pstate available
1107 * current_pstate was the pstate that was requested during
1108 * the last sample period.
1110 * We normalize core_busy, which was our actual percent
1111 * performance to what we requested during the last sample
1112 * period. The result will be a percentage of busy at a
1115 core_busy = cpu->sample.core_pct_busy;
1116 max_pstate = cpu->pstate.max_pstate_physical;
1117 current_pstate = cpu->pstate.current_pstate;
1118 core_busy = mul_fp(core_busy, div_fp(max_pstate, current_pstate));
1121 * Since our utilization update callback will not run unless we are
1122 * in C0, check if the actual elapsed time is significantly greater (3x)
1123 * than our sample interval. If it is, then we were idle for a long
1124 * enough period of time to adjust our busyness.
1126 duration_ns = cpu->sample.time - cpu->last_sample_time;
1127 if ((s64)duration_ns > pid_params.sample_rate_ns * 3) {
1128 sample_ratio = div_fp(pid_params.sample_rate_ns, duration_ns);
1129 core_busy = mul_fp(core_busy, sample_ratio);
1132 cpu->sample.busy_scaled = core_busy;
1133 return cpu->pstate.current_pstate - pid_calc(&cpu->pid, core_busy);
1136 static inline void intel_pstate_update_pstate(struct cpudata *cpu, int pstate)
1138 int max_perf, min_perf;
1140 update_turbo_state();
1142 intel_pstate_get_min_max(cpu, &min_perf, &max_perf);
1143 pstate = clamp_t(int, pstate, min_perf, max_perf);
1144 if (pstate == cpu->pstate.current_pstate)
1147 intel_pstate_record_pstate(cpu, pstate);
1148 wrmsrl(MSR_IA32_PERF_CTL, pstate_funcs.get_val(cpu, pstate));
1151 static inline void intel_pstate_adjust_busy_pstate(struct cpudata *cpu)
1153 int from, target_pstate;
1154 struct sample *sample;
1156 from = cpu->pstate.current_pstate;
1158 target_pstate = pstate_funcs.get_target_pstate(cpu);
1160 intel_pstate_update_pstate(cpu, target_pstate);
1162 sample = &cpu->sample;
1163 trace_pstate_sample(fp_toint(sample->core_pct_busy),
1164 fp_toint(sample->busy_scaled),
1166 cpu->pstate.current_pstate,
1170 get_avg_frequency(cpu));
1173 static void intel_pstate_update_util(struct update_util_data *data, u64 time,
1174 unsigned long util, unsigned long max)
1176 struct cpudata *cpu = container_of(data, struct cpudata, update_util);
1177 u64 delta_ns = time - cpu->sample.time;
1179 if ((s64)delta_ns >= pid_params.sample_rate_ns) {
1180 bool sample_taken = intel_pstate_sample(cpu, time);
1182 if (sample_taken && !hwp_active)
1183 intel_pstate_adjust_busy_pstate(cpu);
1187 #define ICPU(model, policy) \
1188 { X86_VENDOR_INTEL, 6, model, X86_FEATURE_APERFMPERF,\
1189 (unsigned long)&policy }
1191 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
1192 ICPU(0x2a, core_params),
1193 ICPU(0x2d, core_params),
1194 ICPU(0x37, silvermont_params),
1195 ICPU(0x3a, core_params),
1196 ICPU(0x3c, core_params),
1197 ICPU(0x3d, core_params),
1198 ICPU(0x3e, core_params),
1199 ICPU(0x3f, core_params),
1200 ICPU(0x45, core_params),
1201 ICPU(0x46, core_params),
1202 ICPU(0x47, core_params),
1203 ICPU(0x4c, airmont_params),
1204 ICPU(0x4e, core_params),
1205 ICPU(0x4f, core_params),
1206 ICPU(0x5e, core_params),
1207 ICPU(0x56, core_params),
1208 ICPU(0x57, knl_params),
1211 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
1213 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] = {
1214 ICPU(0x56, core_params),
1218 static int intel_pstate_init_cpu(unsigned int cpunum)
1220 struct cpudata *cpu;
1222 if (!all_cpu_data[cpunum])
1223 all_cpu_data[cpunum] = kzalloc(sizeof(struct cpudata),
1225 if (!all_cpu_data[cpunum])
1228 cpu = all_cpu_data[cpunum];
1233 intel_pstate_hwp_enable(cpu);
1234 pid_params.sample_rate_ms = 50;
1235 pid_params.sample_rate_ns = 50 * NSEC_PER_MSEC;
1238 intel_pstate_get_cpu_pstates(cpu);
1240 intel_pstate_busy_pid_reset(cpu);
1242 pr_debug("intel_pstate: controlling: cpu %d\n", cpunum);
1247 static unsigned int intel_pstate_get(unsigned int cpu_num)
1249 struct sample *sample;
1250 struct cpudata *cpu;
1252 cpu = all_cpu_data[cpu_num];
1255 sample = &cpu->sample;
1256 return get_avg_frequency(cpu);
1259 static void intel_pstate_set_update_util_hook(unsigned int cpu_num)
1261 struct cpudata *cpu = all_cpu_data[cpu_num];
1263 /* Prevent intel_pstate_update_util() from using stale data. */
1264 cpu->sample.time = 0;
1265 cpufreq_add_update_util_hook(cpu_num, &cpu->update_util,
1266 intel_pstate_update_util);
1269 static void intel_pstate_clear_update_util_hook(unsigned int cpu)
1271 cpufreq_remove_update_util_hook(cpu);
1272 synchronize_sched();
1275 static void intel_pstate_set_performance_limits(struct perf_limits *limits)
1277 limits->no_turbo = 0;
1278 limits->turbo_disabled = 0;
1279 limits->max_perf_pct = 100;
1280 limits->max_perf = int_tofp(1);
1281 limits->min_perf_pct = 100;
1282 limits->min_perf = int_tofp(1);
1283 limits->max_policy_pct = 100;
1284 limits->max_sysfs_pct = 100;
1285 limits->min_policy_pct = 0;
1286 limits->min_sysfs_pct = 0;
1289 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
1291 if (!policy->cpuinfo.max_freq)
1294 intel_pstate_clear_update_util_hook(policy->cpu);
1296 if (policy->policy == CPUFREQ_POLICY_PERFORMANCE) {
1297 limits = &performance_limits;
1298 if (policy->max >= policy->cpuinfo.max_freq) {
1299 pr_debug("intel_pstate: set performance\n");
1300 intel_pstate_set_performance_limits(limits);
1304 pr_debug("intel_pstate: set powersave\n");
1305 limits = &powersave_limits;
1308 limits->min_policy_pct = (policy->min * 100) / policy->cpuinfo.max_freq;
1309 limits->min_policy_pct = clamp_t(int, limits->min_policy_pct, 0 , 100);
1310 limits->max_policy_pct = DIV_ROUND_UP(policy->max * 100,
1311 policy->cpuinfo.max_freq);
1312 limits->max_policy_pct = clamp_t(int, limits->max_policy_pct, 0 , 100);
1314 /* Normalize user input to [min_policy_pct, max_policy_pct] */
1315 limits->min_perf_pct = max(limits->min_policy_pct,
1316 limits->min_sysfs_pct);
1317 limits->min_perf_pct = min(limits->max_policy_pct,
1318 limits->min_perf_pct);
1319 limits->max_perf_pct = min(limits->max_policy_pct,
1320 limits->max_sysfs_pct);
1321 limits->max_perf_pct = max(limits->min_policy_pct,
1322 limits->max_perf_pct);
1323 limits->max_perf = round_up(limits->max_perf, FRAC_BITS);
1325 /* Make sure min_perf_pct <= max_perf_pct */
1326 limits->min_perf_pct = min(limits->max_perf_pct, limits->min_perf_pct);
1328 limits->min_perf = div_fp(limits->min_perf_pct, 100);
1329 limits->max_perf = div_fp(limits->max_perf_pct, 100);
1332 intel_pstate_set_update_util_hook(policy->cpu);
1335 intel_pstate_hwp_set(policy->cpus);
1340 static int intel_pstate_verify_policy(struct cpufreq_policy *policy)
1342 cpufreq_verify_within_cpu_limits(policy);
1344 if (policy->policy != CPUFREQ_POLICY_POWERSAVE &&
1345 policy->policy != CPUFREQ_POLICY_PERFORMANCE)
1351 static void intel_pstate_stop_cpu(struct cpufreq_policy *policy)
1353 int cpu_num = policy->cpu;
1354 struct cpudata *cpu = all_cpu_data[cpu_num];
1356 pr_debug("intel_pstate: CPU %d exiting\n", cpu_num);
1358 intel_pstate_clear_update_util_hook(cpu_num);
1363 intel_pstate_set_min_pstate(cpu);
1366 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
1368 struct cpudata *cpu;
1371 rc = intel_pstate_init_cpu(policy->cpu);
1375 cpu = all_cpu_data[policy->cpu];
1377 if (limits->min_perf_pct == 100 && limits->max_perf_pct == 100)
1378 policy->policy = CPUFREQ_POLICY_PERFORMANCE;
1380 policy->policy = CPUFREQ_POLICY_POWERSAVE;
1382 policy->min = cpu->pstate.min_pstate * cpu->pstate.scaling;
1383 policy->max = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
1385 /* cpuinfo and default policy values */
1386 policy->cpuinfo.min_freq = cpu->pstate.min_pstate * cpu->pstate.scaling;
1387 policy->cpuinfo.max_freq =
1388 cpu->pstate.turbo_pstate * cpu->pstate.scaling;
1389 policy->cpuinfo.transition_latency = CPUFREQ_ETERNAL;
1390 cpumask_set_cpu(policy->cpu, policy->cpus);
1395 static struct cpufreq_driver intel_pstate_driver = {
1396 .flags = CPUFREQ_CONST_LOOPS,
1397 .verify = intel_pstate_verify_policy,
1398 .setpolicy = intel_pstate_set_policy,
1399 .get = intel_pstate_get,
1400 .init = intel_pstate_cpu_init,
1401 .stop_cpu = intel_pstate_stop_cpu,
1402 .name = "intel_pstate",
1405 static int __initdata no_load;
1406 static int __initdata no_hwp;
1407 static int __initdata hwp_only;
1408 static unsigned int force_load;
1410 static int intel_pstate_msrs_not_valid(void)
1412 if (!pstate_funcs.get_max() ||
1413 !pstate_funcs.get_min() ||
1414 !pstate_funcs.get_turbo())
1420 static void copy_pid_params(struct pstate_adjust_policy *policy)
1422 pid_params.sample_rate_ms = policy->sample_rate_ms;
1423 pid_params.sample_rate_ns = pid_params.sample_rate_ms * NSEC_PER_MSEC;
1424 pid_params.p_gain_pct = policy->p_gain_pct;
1425 pid_params.i_gain_pct = policy->i_gain_pct;
1426 pid_params.d_gain_pct = policy->d_gain_pct;
1427 pid_params.deadband = policy->deadband;
1428 pid_params.setpoint = policy->setpoint;
1431 static void copy_cpu_funcs(struct pstate_funcs *funcs)
1433 pstate_funcs.get_max = funcs->get_max;
1434 pstate_funcs.get_max_physical = funcs->get_max_physical;
1435 pstate_funcs.get_min = funcs->get_min;
1436 pstate_funcs.get_turbo = funcs->get_turbo;
1437 pstate_funcs.get_scaling = funcs->get_scaling;
1438 pstate_funcs.get_val = funcs->get_val;
1439 pstate_funcs.get_vid = funcs->get_vid;
1440 pstate_funcs.get_target_pstate = funcs->get_target_pstate;
1444 #if IS_ENABLED(CONFIG_ACPI)
1445 #include <acpi/processor.h>
1447 static bool intel_pstate_no_acpi_pss(void)
1451 for_each_possible_cpu(i) {
1453 union acpi_object *pss;
1454 struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
1455 struct acpi_processor *pr = per_cpu(processors, i);
1460 status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
1461 if (ACPI_FAILURE(status))
1464 pss = buffer.pointer;
1465 if (pss && pss->type == ACPI_TYPE_PACKAGE) {
1476 static bool intel_pstate_has_acpi_ppc(void)
1480 for_each_possible_cpu(i) {
1481 struct acpi_processor *pr = per_cpu(processors, i);
1485 if (acpi_has_method(pr->handle, "_PPC"))
1496 struct hw_vendor_info {
1498 char oem_id[ACPI_OEM_ID_SIZE];
1499 char oem_table_id[ACPI_OEM_TABLE_ID_SIZE];
1503 /* Hardware vendor-specific info that has its own power management modes */
1504 static struct hw_vendor_info vendor_info[] = {
1505 {1, "HP ", "ProLiant", PSS},
1506 {1, "ORACLE", "X4-2 ", PPC},
1507 {1, "ORACLE", "X4-2L ", PPC},
1508 {1, "ORACLE", "X4-2B ", PPC},
1509 {1, "ORACLE", "X3-2 ", PPC},
1510 {1, "ORACLE", "X3-2L ", PPC},
1511 {1, "ORACLE", "X3-2B ", PPC},
1512 {1, "ORACLE", "X4470M2 ", PPC},
1513 {1, "ORACLE", "X4270M3 ", PPC},
1514 {1, "ORACLE", "X4270M2 ", PPC},
1515 {1, "ORACLE", "X4170M2 ", PPC},
1516 {1, "ORACLE", "X4170 M3", PPC},
1517 {1, "ORACLE", "X4275 M3", PPC},
1518 {1, "ORACLE", "X6-2 ", PPC},
1519 {1, "ORACLE", "Sudbury ", PPC},
1523 static bool intel_pstate_platform_pwr_mgmt_exists(void)
1525 struct acpi_table_header hdr;
1526 struct hw_vendor_info *v_info;
1527 const struct x86_cpu_id *id;
1530 id = x86_match_cpu(intel_pstate_cpu_oob_ids);
1532 rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
1533 if ( misc_pwr & (1 << 8))
1537 if (acpi_disabled ||
1538 ACPI_FAILURE(acpi_get_table_header(ACPI_SIG_FADT, 0, &hdr)))
1541 for (v_info = vendor_info; v_info->valid; v_info++) {
1542 if (!strncmp(hdr.oem_id, v_info->oem_id, ACPI_OEM_ID_SIZE) &&
1543 !strncmp(hdr.oem_table_id, v_info->oem_table_id,
1544 ACPI_OEM_TABLE_ID_SIZE))
1545 switch (v_info->oem_pwr_table) {
1547 return intel_pstate_no_acpi_pss();
1549 return intel_pstate_has_acpi_ppc() &&
1556 #else /* CONFIG_ACPI not enabled */
1557 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
1558 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
1559 #endif /* CONFIG_ACPI */
1561 static const struct x86_cpu_id hwp_support_ids[] __initconst = {
1562 { X86_VENDOR_INTEL, 6, X86_MODEL_ANY, X86_FEATURE_HWP },
1566 static int __init intel_pstate_init(void)
1569 const struct x86_cpu_id *id;
1570 struct cpu_defaults *cpu_def;
1575 if (x86_match_cpu(hwp_support_ids) && !no_hwp) {
1576 copy_cpu_funcs(&core_params.funcs);
1578 goto hwp_cpu_matched;
1581 id = x86_match_cpu(intel_pstate_cpu_ids);
1585 cpu_def = (struct cpu_defaults *)id->driver_data;
1587 copy_pid_params(&cpu_def->pid_policy);
1588 copy_cpu_funcs(&cpu_def->funcs);
1590 if (intel_pstate_msrs_not_valid())
1595 * The Intel pstate driver will be ignored if the platform
1596 * firmware has its own power management modes.
1598 if (intel_pstate_platform_pwr_mgmt_exists())
1601 pr_info("Intel P-state driver initializing.\n");
1603 all_cpu_data = vzalloc(sizeof(void *) * num_possible_cpus());
1607 if (!hwp_active && hwp_only)
1610 rc = cpufreq_register_driver(&intel_pstate_driver);
1614 intel_pstate_debug_expose_params();
1615 intel_pstate_sysfs_expose_params();
1618 pr_info("intel_pstate: HWP enabled\n");
1623 for_each_online_cpu(cpu) {
1624 if (all_cpu_data[cpu]) {
1625 intel_pstate_clear_update_util_hook(cpu);
1626 kfree(all_cpu_data[cpu]);
1631 vfree(all_cpu_data);
1634 device_initcall(intel_pstate_init);
1636 static int __init intel_pstate_setup(char *str)
1641 if (!strcmp(str, "disable"))
1643 if (!strcmp(str, "no_hwp")) {
1644 pr_info("intel_pstate: HWP disabled\n");
1647 if (!strcmp(str, "force"))
1649 if (!strcmp(str, "hwp_only"))
1653 early_param("intel_pstate", intel_pstate_setup);
1655 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
1656 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
1657 MODULE_LICENSE("GPL");