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>
39 #define ATOM_RATIOS 0x66a
40 #define ATOM_VIDS 0x66b
41 #define ATOM_TURBO_RATIOS 0x66c
42 #define ATOM_TURBO_VIDS 0x66d
45 #define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
46 #define fp_toint(X) ((X) >> FRAC_BITS)
48 static inline int32_t mul_fp(int32_t x, int32_t y)
50 return ((int64_t)x * (int64_t)y) >> FRAC_BITS;
53 static inline int32_t div_fp(s64 x, s64 y)
55 return div64_s64((int64_t)x << FRAC_BITS, y);
58 static inline int ceiling_fp(int32_t x)
63 mask = (1 << FRAC_BITS) - 1;
70 * struct sample - Store performance sample
71 * @core_pct_busy: Ratio of APERF/MPERF in percent, which is actual
72 * performance during last sample period
73 * @busy_scaled: Scaled busy value which is used to calculate next
74 * P state. This can be different than core_pct_busy
75 * to account for cpu idle period
76 * @aperf: Difference of actual performance frequency clock count
77 * read from APERF MSR between last and current sample
78 * @mperf: Difference of maximum performance frequency clock count
79 * read from MPERF MSR between last and current sample
80 * @tsc: Difference of time stamp counter between last and
82 * @freq: Effective frequency calculated from APERF/MPERF
83 * @time: Current time from scheduler
85 * This structure is used in the cpudata structure to store performance sample
86 * data for choosing next P State.
89 int32_t core_pct_busy;
99 * struct pstate_data - Store P state data
100 * @current_pstate: Current requested P state
101 * @min_pstate: Min P state possible for this platform
102 * @max_pstate: Max P state possible for this platform
103 * @max_pstate_physical:This is physical Max P state for a processor
104 * This can be higher than the max_pstate which can
105 * be limited by platform thermal design power limits
106 * @scaling: Scaling factor to convert frequency to cpufreq
108 * @turbo_pstate: Max Turbo P state possible for this platform
110 * Stores the per cpu model P state limits and current P state.
116 int max_pstate_physical;
122 * struct vid_data - Stores voltage information data
123 * @min: VID data for this platform corresponding to
125 * @max: VID data corresponding to the highest P State.
126 * @turbo: VID data for turbo P state
127 * @ratio: Ratio of (vid max - vid min) /
128 * (max P state - Min P State)
130 * Stores the voltage data for DVFS (Dynamic Voltage and Frequency Scaling)
131 * This data is used in Atom platforms, where in addition to target P state,
132 * the voltage data needs to be specified to select next P State.
142 * struct _pid - Stores PID data
143 * @setpoint: Target set point for busyness or performance
144 * @integral: Storage for accumulated error values
145 * @p_gain: PID proportional gain
146 * @i_gain: PID integral gain
147 * @d_gain: PID derivative gain
148 * @deadband: PID deadband
149 * @last_err: Last error storage for integral part of PID calculation
151 * Stores PID coefficients and last error for PID controller.
164 * struct cpudata - Per CPU instance data storage
165 * @cpu: CPU number for this instance data
166 * @update_util: CPUFreq utility callback information
167 * @pstate: Stores P state limits for this CPU
168 * @vid: Stores VID limits for this CPU
169 * @pid: Stores PID parameters for this CPU
170 * @last_sample_time: Last Sample time
171 * @prev_aperf: Last APERF value read from APERF MSR
172 * @prev_mperf: Last MPERF value read from MPERF MSR
173 * @prev_tsc: Last timestamp counter (TSC) value
174 * @prev_cummulative_iowait: IO Wait time difference from last and
176 * @sample: Storage for storing last Sample data
178 * This structure stores per CPU instance data for all CPUs.
183 struct update_util_data update_util;
185 struct pstate_data pstate;
189 u64 last_sample_time;
193 u64 prev_cummulative_iowait;
194 struct sample sample;
197 static struct cpudata **all_cpu_data;
200 * struct pid_adjust_policy - Stores static PID configuration data
201 * @sample_rate_ms: PID calculation sample rate in ms
202 * @sample_rate_ns: Sample rate calculation in ns
203 * @deadband: PID deadband
204 * @setpoint: PID Setpoint
205 * @p_gain_pct: PID proportional gain
206 * @i_gain_pct: PID integral gain
207 * @d_gain_pct: PID derivative gain
209 * Stores per CPU model static PID configuration data.
211 struct pstate_adjust_policy {
222 * struct pstate_funcs - Per CPU model specific callbacks
223 * @get_max: Callback to get maximum non turbo effective P state
224 * @get_max_physical: Callback to get maximum non turbo physical P state
225 * @get_min: Callback to get minimum P state
226 * @get_turbo: Callback to get turbo P state
227 * @get_scaling: Callback to get frequency scaling factor
228 * @get_val: Callback to convert P state to actual MSR write value
229 * @get_vid: Callback to get VID data for Atom platforms
230 * @get_target_pstate: Callback to a function to calculate next P state to use
232 * Core and Atom CPU models have different way to get P State limits. This
233 * structure is used to store those callbacks.
235 struct pstate_funcs {
236 int (*get_max)(void);
237 int (*get_max_physical)(void);
238 int (*get_min)(void);
239 int (*get_turbo)(void);
240 int (*get_scaling)(void);
241 u64 (*get_val)(struct cpudata*, int pstate);
242 void (*get_vid)(struct cpudata *);
243 int32_t (*get_target_pstate)(struct cpudata *);
247 * struct cpu_defaults- Per CPU model default config data
248 * @pid_policy: PID config data
249 * @funcs: Callback function data
251 struct cpu_defaults {
252 struct pstate_adjust_policy pid_policy;
253 struct pstate_funcs funcs;
256 static inline int32_t get_target_pstate_use_performance(struct cpudata *cpu);
257 static inline int32_t get_target_pstate_use_cpu_load(struct cpudata *cpu);
259 static struct pstate_adjust_policy pid_params;
260 static struct pstate_funcs pstate_funcs;
261 static int hwp_active;
265 * struct perf_limits - Store user and policy limits
266 * @no_turbo: User requested turbo state from intel_pstate sysfs
267 * @turbo_disabled: Platform turbo status either from msr
268 * MSR_IA32_MISC_ENABLE or when maximum available pstate
269 * matches the maximum turbo pstate
270 * @max_perf_pct: Effective maximum performance limit in percentage, this
271 * is minimum of either limits enforced by cpufreq policy
272 * or limits from user set limits via intel_pstate sysfs
273 * @min_perf_pct: Effective minimum performance limit in percentage, this
274 * is maximum of either limits enforced by cpufreq policy
275 * or limits from user set limits via intel_pstate sysfs
276 * @max_perf: This is a scaled value between 0 to 255 for max_perf_pct
277 * This value is used to limit max pstate
278 * @min_perf: This is a scaled value between 0 to 255 for min_perf_pct
279 * This value is used to limit min pstate
280 * @max_policy_pct: The maximum performance in percentage enforced by
281 * cpufreq setpolicy interface
282 * @max_sysfs_pct: The maximum performance in percentage enforced by
283 * intel pstate sysfs interface
284 * @min_policy_pct: The minimum performance in percentage enforced by
285 * cpufreq setpolicy interface
286 * @min_sysfs_pct: The minimum performance in percentage enforced by
287 * intel pstate sysfs interface
289 * Storage for user and policy defined limits.
304 static struct perf_limits performance_limits = {
308 .max_perf = int_tofp(1),
310 .min_perf = int_tofp(1),
311 .max_policy_pct = 100,
312 .max_sysfs_pct = 100,
317 static struct perf_limits powersave_limits = {
321 .max_perf = int_tofp(1),
324 .max_policy_pct = 100,
325 .max_sysfs_pct = 100,
330 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_PERFORMANCE
331 static struct perf_limits *limits = &performance_limits;
333 static struct perf_limits *limits = &powersave_limits;
336 static inline void pid_reset(struct _pid *pid, int setpoint, int busy,
337 int deadband, int integral) {
338 pid->setpoint = int_tofp(setpoint);
339 pid->deadband = int_tofp(deadband);
340 pid->integral = int_tofp(integral);
341 pid->last_err = int_tofp(setpoint) - int_tofp(busy);
344 static inline void pid_p_gain_set(struct _pid *pid, int percent)
346 pid->p_gain = div_fp(percent, 100);
349 static inline void pid_i_gain_set(struct _pid *pid, int percent)
351 pid->i_gain = div_fp(percent, 100);
354 static inline void pid_d_gain_set(struct _pid *pid, int percent)
356 pid->d_gain = div_fp(percent, 100);
359 static signed int pid_calc(struct _pid *pid, int32_t busy)
362 int32_t pterm, dterm, fp_error;
363 int32_t integral_limit;
365 fp_error = pid->setpoint - busy;
367 if (abs(fp_error) <= pid->deadband)
370 pterm = mul_fp(pid->p_gain, fp_error);
372 pid->integral += fp_error;
375 * We limit the integral here so that it will never
376 * get higher than 30. This prevents it from becoming
377 * too large an input over long periods of time and allows
378 * it to get factored out sooner.
380 * The value of 30 was chosen through experimentation.
382 integral_limit = int_tofp(30);
383 if (pid->integral > integral_limit)
384 pid->integral = integral_limit;
385 if (pid->integral < -integral_limit)
386 pid->integral = -integral_limit;
388 dterm = mul_fp(pid->d_gain, fp_error - pid->last_err);
389 pid->last_err = fp_error;
391 result = pterm + mul_fp(pid->integral, pid->i_gain) + dterm;
392 result = result + (1 << (FRAC_BITS-1));
393 return (signed int)fp_toint(result);
396 static inline void intel_pstate_busy_pid_reset(struct cpudata *cpu)
398 pid_p_gain_set(&cpu->pid, pid_params.p_gain_pct);
399 pid_d_gain_set(&cpu->pid, pid_params.d_gain_pct);
400 pid_i_gain_set(&cpu->pid, pid_params.i_gain_pct);
402 pid_reset(&cpu->pid, pid_params.setpoint, 100, pid_params.deadband, 0);
405 static inline void intel_pstate_reset_all_pid(void)
409 for_each_online_cpu(cpu) {
410 if (all_cpu_data[cpu])
411 intel_pstate_busy_pid_reset(all_cpu_data[cpu]);
415 static inline void update_turbo_state(void)
420 cpu = all_cpu_data[0];
421 rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
422 limits->turbo_disabled =
423 (misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ||
424 cpu->pstate.max_pstate == cpu->pstate.turbo_pstate);
427 static void intel_pstate_hwp_set(const struct cpumask *cpumask)
429 int min, hw_min, max, hw_max, cpu, range, adj_range;
432 rdmsrl(MSR_HWP_CAPABILITIES, cap);
433 hw_min = HWP_LOWEST_PERF(cap);
434 hw_max = HWP_HIGHEST_PERF(cap);
435 range = hw_max - hw_min;
437 for_each_cpu(cpu, cpumask) {
438 rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
439 adj_range = limits->min_perf_pct * range / 100;
440 min = hw_min + adj_range;
441 value &= ~HWP_MIN_PERF(~0L);
442 value |= HWP_MIN_PERF(min);
444 adj_range = limits->max_perf_pct * range / 100;
445 max = hw_min + adj_range;
446 if (limits->no_turbo) {
447 hw_max = HWP_GUARANTEED_PERF(cap);
452 value &= ~HWP_MAX_PERF(~0L);
453 value |= HWP_MAX_PERF(max);
454 wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
458 static void intel_pstate_hwp_set_online_cpus(void)
461 intel_pstate_hwp_set(cpu_online_mask);
465 /************************** debugfs begin ************************/
466 static int pid_param_set(void *data, u64 val)
469 intel_pstate_reset_all_pid();
473 static int pid_param_get(void *data, u64 *val)
478 DEFINE_SIMPLE_ATTRIBUTE(fops_pid_param, pid_param_get, pid_param_set, "%llu\n");
485 static struct pid_param pid_files[] = {
486 {"sample_rate_ms", &pid_params.sample_rate_ms},
487 {"d_gain_pct", &pid_params.d_gain_pct},
488 {"i_gain_pct", &pid_params.i_gain_pct},
489 {"deadband", &pid_params.deadband},
490 {"setpoint", &pid_params.setpoint},
491 {"p_gain_pct", &pid_params.p_gain_pct},
495 static void __init intel_pstate_debug_expose_params(void)
497 struct dentry *debugfs_parent;
502 debugfs_parent = debugfs_create_dir("pstate_snb", NULL);
503 if (IS_ERR_OR_NULL(debugfs_parent))
505 while (pid_files[i].name) {
506 debugfs_create_file(pid_files[i].name, 0660,
507 debugfs_parent, pid_files[i].value,
513 /************************** debugfs end ************************/
515 /************************** sysfs begin ************************/
516 #define show_one(file_name, object) \
517 static ssize_t show_##file_name \
518 (struct kobject *kobj, struct attribute *attr, char *buf) \
520 return sprintf(buf, "%u\n", limits->object); \
523 static ssize_t show_turbo_pct(struct kobject *kobj,
524 struct attribute *attr, char *buf)
527 int total, no_turbo, turbo_pct;
530 cpu = all_cpu_data[0];
532 total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
533 no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
534 turbo_fp = div_fp(no_turbo, total);
535 turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
536 return sprintf(buf, "%u\n", turbo_pct);
539 static ssize_t show_num_pstates(struct kobject *kobj,
540 struct attribute *attr, char *buf)
545 cpu = all_cpu_data[0];
546 total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
547 return sprintf(buf, "%u\n", total);
550 static ssize_t show_no_turbo(struct kobject *kobj,
551 struct attribute *attr, char *buf)
555 update_turbo_state();
556 if (limits->turbo_disabled)
557 ret = sprintf(buf, "%u\n", limits->turbo_disabled);
559 ret = sprintf(buf, "%u\n", limits->no_turbo);
564 static ssize_t store_no_turbo(struct kobject *a, struct attribute *b,
565 const char *buf, size_t count)
570 ret = sscanf(buf, "%u", &input);
574 update_turbo_state();
575 if (limits->turbo_disabled) {
576 pr_warn("Turbo disabled by BIOS or unavailable on processor\n");
580 limits->no_turbo = clamp_t(int, input, 0, 1);
583 intel_pstate_hwp_set_online_cpus();
588 static ssize_t store_max_perf_pct(struct kobject *a, struct attribute *b,
589 const char *buf, size_t count)
594 ret = sscanf(buf, "%u", &input);
598 limits->max_sysfs_pct = clamp_t(int, input, 0 , 100);
599 limits->max_perf_pct = min(limits->max_policy_pct,
600 limits->max_sysfs_pct);
601 limits->max_perf_pct = max(limits->min_policy_pct,
602 limits->max_perf_pct);
603 limits->max_perf_pct = max(limits->min_perf_pct,
604 limits->max_perf_pct);
605 limits->max_perf = div_fp(limits->max_perf_pct, 100);
608 intel_pstate_hwp_set_online_cpus();
612 static ssize_t store_min_perf_pct(struct kobject *a, struct attribute *b,
613 const char *buf, size_t count)
618 ret = sscanf(buf, "%u", &input);
622 limits->min_sysfs_pct = clamp_t(int, input, 0 , 100);
623 limits->min_perf_pct = max(limits->min_policy_pct,
624 limits->min_sysfs_pct);
625 limits->min_perf_pct = min(limits->max_policy_pct,
626 limits->min_perf_pct);
627 limits->min_perf_pct = min(limits->max_perf_pct,
628 limits->min_perf_pct);
629 limits->min_perf = div_fp(limits->min_perf_pct, 100);
632 intel_pstate_hwp_set_online_cpus();
636 show_one(max_perf_pct, max_perf_pct);
637 show_one(min_perf_pct, min_perf_pct);
639 define_one_global_rw(no_turbo);
640 define_one_global_rw(max_perf_pct);
641 define_one_global_rw(min_perf_pct);
642 define_one_global_ro(turbo_pct);
643 define_one_global_ro(num_pstates);
645 static struct attribute *intel_pstate_attributes[] = {
654 static struct attribute_group intel_pstate_attr_group = {
655 .attrs = intel_pstate_attributes,
658 static void __init intel_pstate_sysfs_expose_params(void)
660 struct kobject *intel_pstate_kobject;
663 intel_pstate_kobject = kobject_create_and_add("intel_pstate",
664 &cpu_subsys.dev_root->kobj);
665 BUG_ON(!intel_pstate_kobject);
666 rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
669 /************************** sysfs end ************************/
671 static void intel_pstate_hwp_enable(struct cpudata *cpudata)
673 /* First disable HWP notification interrupt as we don't process them */
674 wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
676 wrmsrl_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1);
679 static int atom_get_min_pstate(void)
683 rdmsrl(ATOM_RATIOS, value);
684 return (value >> 8) & 0x7F;
687 static int atom_get_max_pstate(void)
691 rdmsrl(ATOM_RATIOS, value);
692 return (value >> 16) & 0x7F;
695 static int atom_get_turbo_pstate(void)
699 rdmsrl(ATOM_TURBO_RATIOS, value);
703 static u64 atom_get_val(struct cpudata *cpudata, int pstate)
709 val = (u64)pstate << 8;
710 if (limits->no_turbo && !limits->turbo_disabled)
713 vid_fp = cpudata->vid.min + mul_fp(
714 int_tofp(pstate - cpudata->pstate.min_pstate),
717 vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
718 vid = ceiling_fp(vid_fp);
720 if (pstate > cpudata->pstate.max_pstate)
721 vid = cpudata->vid.turbo;
726 static int silvermont_get_scaling(void)
730 /* Defined in Table 35-6 from SDM (Sept 2015) */
731 static int silvermont_freq_table[] = {
732 83300, 100000, 133300, 116700, 80000};
734 rdmsrl(MSR_FSB_FREQ, value);
738 return silvermont_freq_table[i];
741 static int airmont_get_scaling(void)
745 /* Defined in Table 35-10 from SDM (Sept 2015) */
746 static int airmont_freq_table[] = {
747 83300, 100000, 133300, 116700, 80000,
748 93300, 90000, 88900, 87500};
750 rdmsrl(MSR_FSB_FREQ, value);
754 return airmont_freq_table[i];
757 static void atom_get_vid(struct cpudata *cpudata)
761 rdmsrl(ATOM_VIDS, value);
762 cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
763 cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
764 cpudata->vid.ratio = div_fp(
765 cpudata->vid.max - cpudata->vid.min,
766 int_tofp(cpudata->pstate.max_pstate -
767 cpudata->pstate.min_pstate));
769 rdmsrl(ATOM_TURBO_VIDS, value);
770 cpudata->vid.turbo = value & 0x7f;
773 static int core_get_min_pstate(void)
777 rdmsrl(MSR_PLATFORM_INFO, value);
778 return (value >> 40) & 0xFF;
781 static int core_get_max_pstate_physical(void)
785 rdmsrl(MSR_PLATFORM_INFO, value);
786 return (value >> 8) & 0xFF;
789 static int core_get_max_pstate(void)
796 rdmsrl(MSR_PLATFORM_INFO, plat_info);
797 max_pstate = (plat_info >> 8) & 0xFF;
799 err = rdmsrl_safe(MSR_TURBO_ACTIVATION_RATIO, &tar);
801 /* Do some sanity checking for safety */
802 if (plat_info & 0x600000000) {
807 err = rdmsrl_safe(MSR_CONFIG_TDP_CONTROL, &tdp_ctrl);
811 tdp_msr = MSR_CONFIG_TDP_NOMINAL + tdp_ctrl;
812 err = rdmsrl_safe(tdp_msr, &tdp_ratio);
816 if (tdp_ratio - 1 == tar) {
818 pr_debug("max_pstate=TAC %x\n", max_pstate);
829 static int core_get_turbo_pstate(void)
834 rdmsrl(MSR_NHM_TURBO_RATIO_LIMIT, value);
835 nont = core_get_max_pstate();
842 static inline int core_get_scaling(void)
847 static u64 core_get_val(struct cpudata *cpudata, int pstate)
851 val = (u64)pstate << 8;
852 if (limits->no_turbo && !limits->turbo_disabled)
858 static int knl_get_turbo_pstate(void)
863 rdmsrl(MSR_NHM_TURBO_RATIO_LIMIT, value);
864 nont = core_get_max_pstate();
865 ret = (((value) >> 8) & 0xFF);
871 static struct cpu_defaults core_params = {
873 .sample_rate_ms = 10,
881 .get_max = core_get_max_pstate,
882 .get_max_physical = core_get_max_pstate_physical,
883 .get_min = core_get_min_pstate,
884 .get_turbo = core_get_turbo_pstate,
885 .get_scaling = core_get_scaling,
886 .get_val = core_get_val,
887 .get_target_pstate = get_target_pstate_use_performance,
891 static struct cpu_defaults silvermont_params = {
893 .sample_rate_ms = 10,
901 .get_max = atom_get_max_pstate,
902 .get_max_physical = atom_get_max_pstate,
903 .get_min = atom_get_min_pstate,
904 .get_turbo = atom_get_turbo_pstate,
905 .get_val = atom_get_val,
906 .get_scaling = silvermont_get_scaling,
907 .get_vid = atom_get_vid,
908 .get_target_pstate = get_target_pstate_use_cpu_load,
912 static struct cpu_defaults airmont_params = {
914 .sample_rate_ms = 10,
922 .get_max = atom_get_max_pstate,
923 .get_max_physical = atom_get_max_pstate,
924 .get_min = atom_get_min_pstate,
925 .get_turbo = atom_get_turbo_pstate,
926 .get_val = atom_get_val,
927 .get_scaling = airmont_get_scaling,
928 .get_vid = atom_get_vid,
929 .get_target_pstate = get_target_pstate_use_cpu_load,
933 static struct cpu_defaults knl_params = {
935 .sample_rate_ms = 10,
943 .get_max = core_get_max_pstate,
944 .get_max_physical = core_get_max_pstate_physical,
945 .get_min = core_get_min_pstate,
946 .get_turbo = knl_get_turbo_pstate,
947 .get_scaling = core_get_scaling,
948 .get_val = core_get_val,
949 .get_target_pstate = get_target_pstate_use_performance,
953 static void intel_pstate_get_min_max(struct cpudata *cpu, int *min, int *max)
955 int max_perf = cpu->pstate.turbo_pstate;
959 if (limits->no_turbo || limits->turbo_disabled)
960 max_perf = cpu->pstate.max_pstate;
963 * performance can be limited by user through sysfs, by cpufreq
964 * policy, or by cpu specific default values determined through
967 max_perf_adj = fp_toint(max_perf * limits->max_perf);
968 *max = clamp_t(int, max_perf_adj,
969 cpu->pstate.min_pstate, cpu->pstate.turbo_pstate);
971 min_perf = fp_toint(max_perf * limits->min_perf);
972 *min = clamp_t(int, min_perf, cpu->pstate.min_pstate, max_perf);
975 static inline void intel_pstate_record_pstate(struct cpudata *cpu, int pstate)
977 trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
978 cpu->pstate.current_pstate = pstate;
981 static void intel_pstate_set_min_pstate(struct cpudata *cpu)
983 int pstate = cpu->pstate.min_pstate;
985 intel_pstate_record_pstate(cpu, pstate);
987 * Generally, there is no guarantee that this code will always run on
988 * the CPU being updated, so force the register update to run on the
991 wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
992 pstate_funcs.get_val(cpu, pstate));
995 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
997 cpu->pstate.min_pstate = pstate_funcs.get_min();
998 cpu->pstate.max_pstate = pstate_funcs.get_max();
999 cpu->pstate.max_pstate_physical = pstate_funcs.get_max_physical();
1000 cpu->pstate.turbo_pstate = pstate_funcs.get_turbo();
1001 cpu->pstate.scaling = pstate_funcs.get_scaling();
1003 if (pstate_funcs.get_vid)
1004 pstate_funcs.get_vid(cpu);
1006 intel_pstate_set_min_pstate(cpu);
1009 static inline void intel_pstate_calc_busy(struct cpudata *cpu)
1011 struct sample *sample = &cpu->sample;
1014 core_pct = sample->aperf * int_tofp(100);
1015 core_pct = div64_u64(core_pct, sample->mperf);
1017 sample->core_pct_busy = (int32_t)core_pct;
1020 static inline bool intel_pstate_sample(struct cpudata *cpu, u64 time)
1023 unsigned long flags;
1026 local_irq_save(flags);
1027 rdmsrl(MSR_IA32_APERF, aperf);
1028 rdmsrl(MSR_IA32_MPERF, mperf);
1030 if (cpu->prev_mperf == mperf || cpu->prev_tsc == tsc) {
1031 local_irq_restore(flags);
1034 local_irq_restore(flags);
1036 cpu->last_sample_time = cpu->sample.time;
1037 cpu->sample.time = time;
1038 cpu->sample.aperf = aperf;
1039 cpu->sample.mperf = mperf;
1040 cpu->sample.tsc = tsc;
1041 cpu->sample.aperf -= cpu->prev_aperf;
1042 cpu->sample.mperf -= cpu->prev_mperf;
1043 cpu->sample.tsc -= cpu->prev_tsc;
1045 cpu->prev_aperf = aperf;
1046 cpu->prev_mperf = mperf;
1047 cpu->prev_tsc = tsc;
1049 * First time this function is invoked in a given cycle, all of the
1050 * previous sample data fields are equal to zero or stale and they must
1051 * be populated with meaningful numbers for things to work, so assume
1052 * that sample.time will always be reset before setting the utilization
1053 * update hook and make the caller skip the sample then.
1055 return !!cpu->last_sample_time;
1058 static inline int32_t get_avg_frequency(struct cpudata *cpu)
1060 return div64_u64(cpu->pstate.max_pstate_physical * cpu->sample.aperf *
1061 cpu->pstate.scaling, cpu->sample.mperf);
1064 static inline int32_t get_target_pstate_use_cpu_load(struct cpudata *cpu)
1066 struct sample *sample = &cpu->sample;
1067 u64 cummulative_iowait, delta_iowait_us;
1068 u64 delta_iowait_mperf;
1072 cummulative_iowait = get_cpu_iowait_time_us(cpu->cpu, &now);
1075 * Convert iowait time into number of IO cycles spent at max_freq.
1076 * IO is considered as busy only for the cpu_load algorithm. For
1077 * performance this is not needed since we always try to reach the
1078 * maximum P-State, so we are already boosting the IOs.
1080 delta_iowait_us = cummulative_iowait - cpu->prev_cummulative_iowait;
1081 delta_iowait_mperf = div64_u64(delta_iowait_us * cpu->pstate.scaling *
1082 cpu->pstate.max_pstate, MSEC_PER_SEC);
1084 mperf = cpu->sample.mperf + delta_iowait_mperf;
1085 cpu->prev_cummulative_iowait = cummulative_iowait;
1088 * The load can be estimated as the ratio of the mperf counter
1089 * running at a constant frequency during active periods
1090 * (C0) and the time stamp counter running at the same frequency
1091 * also during C-states.
1093 cpu_load = div64_u64(int_tofp(100) * mperf, sample->tsc);
1094 cpu->sample.busy_scaled = cpu_load;
1096 return cpu->pstate.current_pstate - pid_calc(&cpu->pid, cpu_load);
1099 static inline int32_t get_target_pstate_use_performance(struct cpudata *cpu)
1101 int32_t core_busy, max_pstate, current_pstate, sample_ratio;
1104 intel_pstate_calc_busy(cpu);
1107 * core_busy is the ratio of actual performance to max
1108 * max_pstate is the max non turbo pstate available
1109 * current_pstate was the pstate that was requested during
1110 * the last sample period.
1112 * We normalize core_busy, which was our actual percent
1113 * performance to what we requested during the last sample
1114 * period. The result will be a percentage of busy at a
1117 core_busy = cpu->sample.core_pct_busy;
1118 max_pstate = cpu->pstate.max_pstate_physical;
1119 current_pstate = cpu->pstate.current_pstate;
1120 core_busy = mul_fp(core_busy, div_fp(max_pstate, current_pstate));
1123 * Since our utilization update callback will not run unless we are
1124 * in C0, check if the actual elapsed time is significantly greater (3x)
1125 * than our sample interval. If it is, then we were idle for a long
1126 * enough period of time to adjust our busyness.
1128 duration_ns = cpu->sample.time - cpu->last_sample_time;
1129 if ((s64)duration_ns > pid_params.sample_rate_ns * 3) {
1130 sample_ratio = div_fp(pid_params.sample_rate_ns, duration_ns);
1131 core_busy = mul_fp(core_busy, sample_ratio);
1134 cpu->sample.busy_scaled = core_busy;
1135 return cpu->pstate.current_pstate - pid_calc(&cpu->pid, core_busy);
1138 static inline void intel_pstate_update_pstate(struct cpudata *cpu, int pstate)
1140 int max_perf, min_perf;
1142 update_turbo_state();
1144 intel_pstate_get_min_max(cpu, &min_perf, &max_perf);
1145 pstate = clamp_t(int, pstate, min_perf, max_perf);
1146 if (pstate == cpu->pstate.current_pstate)
1149 intel_pstate_record_pstate(cpu, pstate);
1150 wrmsrl(MSR_IA32_PERF_CTL, pstate_funcs.get_val(cpu, pstate));
1153 static inline void intel_pstate_adjust_busy_pstate(struct cpudata *cpu)
1155 int from, target_pstate;
1156 struct sample *sample;
1158 from = cpu->pstate.current_pstate;
1160 target_pstate = pstate_funcs.get_target_pstate(cpu);
1162 intel_pstate_update_pstate(cpu, target_pstate);
1164 sample = &cpu->sample;
1165 trace_pstate_sample(fp_toint(sample->core_pct_busy),
1166 fp_toint(sample->busy_scaled),
1168 cpu->pstate.current_pstate,
1172 get_avg_frequency(cpu));
1175 static void intel_pstate_update_util(struct update_util_data *data, u64 time,
1176 unsigned long util, unsigned long max)
1178 struct cpudata *cpu = container_of(data, struct cpudata, update_util);
1179 u64 delta_ns = time - cpu->sample.time;
1181 if ((s64)delta_ns >= pid_params.sample_rate_ns) {
1182 bool sample_taken = intel_pstate_sample(cpu, time);
1184 if (sample_taken && !hwp_active)
1185 intel_pstate_adjust_busy_pstate(cpu);
1189 #define ICPU(model, policy) \
1190 { X86_VENDOR_INTEL, 6, model, X86_FEATURE_APERFMPERF,\
1191 (unsigned long)&policy }
1193 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
1194 ICPU(0x2a, core_params),
1195 ICPU(0x2d, core_params),
1196 ICPU(0x37, silvermont_params),
1197 ICPU(0x3a, core_params),
1198 ICPU(0x3c, core_params),
1199 ICPU(0x3d, core_params),
1200 ICPU(0x3e, core_params),
1201 ICPU(0x3f, core_params),
1202 ICPU(0x45, core_params),
1203 ICPU(0x46, core_params),
1204 ICPU(0x47, core_params),
1205 ICPU(0x4c, airmont_params),
1206 ICPU(0x4e, core_params),
1207 ICPU(0x4f, core_params),
1208 ICPU(0x5e, core_params),
1209 ICPU(0x56, core_params),
1210 ICPU(0x57, knl_params),
1213 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
1215 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] = {
1216 ICPU(0x56, core_params),
1220 static int intel_pstate_init_cpu(unsigned int cpunum)
1222 struct cpudata *cpu;
1224 if (!all_cpu_data[cpunum])
1225 all_cpu_data[cpunum] = kzalloc(sizeof(struct cpudata),
1227 if (!all_cpu_data[cpunum])
1230 cpu = all_cpu_data[cpunum];
1235 intel_pstate_hwp_enable(cpu);
1236 pid_params.sample_rate_ms = 50;
1237 pid_params.sample_rate_ns = 50 * NSEC_PER_MSEC;
1240 intel_pstate_get_cpu_pstates(cpu);
1242 intel_pstate_busy_pid_reset(cpu);
1244 pr_debug("controlling: cpu %d\n", cpunum);
1249 static unsigned int intel_pstate_get(unsigned int cpu_num)
1251 struct sample *sample;
1252 struct cpudata *cpu;
1254 cpu = all_cpu_data[cpu_num];
1257 sample = &cpu->sample;
1258 return get_avg_frequency(cpu);
1261 static void intel_pstate_set_update_util_hook(unsigned int cpu_num)
1263 struct cpudata *cpu = all_cpu_data[cpu_num];
1265 /* Prevent intel_pstate_update_util() from using stale data. */
1266 cpu->sample.time = 0;
1267 cpufreq_add_update_util_hook(cpu_num, &cpu->update_util,
1268 intel_pstate_update_util);
1271 static void intel_pstate_clear_update_util_hook(unsigned int cpu)
1273 cpufreq_remove_update_util_hook(cpu);
1274 synchronize_sched();
1277 static void intel_pstate_set_performance_limits(struct perf_limits *limits)
1279 limits->no_turbo = 0;
1280 limits->turbo_disabled = 0;
1281 limits->max_perf_pct = 100;
1282 limits->max_perf = int_tofp(1);
1283 limits->min_perf_pct = 100;
1284 limits->min_perf = int_tofp(1);
1285 limits->max_policy_pct = 100;
1286 limits->max_sysfs_pct = 100;
1287 limits->min_policy_pct = 0;
1288 limits->min_sysfs_pct = 0;
1291 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
1293 if (!policy->cpuinfo.max_freq)
1296 intel_pstate_clear_update_util_hook(policy->cpu);
1298 if (policy->policy == CPUFREQ_POLICY_PERFORMANCE) {
1299 limits = &performance_limits;
1300 if (policy->max >= policy->cpuinfo.max_freq) {
1301 pr_debug("set performance\n");
1302 intel_pstate_set_performance_limits(limits);
1306 pr_debug("set powersave\n");
1307 limits = &powersave_limits;
1310 limits->min_policy_pct = (policy->min * 100) / policy->cpuinfo.max_freq;
1311 limits->min_policy_pct = clamp_t(int, limits->min_policy_pct, 0 , 100);
1312 limits->max_policy_pct = DIV_ROUND_UP(policy->max * 100,
1313 policy->cpuinfo.max_freq);
1314 limits->max_policy_pct = clamp_t(int, limits->max_policy_pct, 0 , 100);
1316 /* Normalize user input to [min_policy_pct, max_policy_pct] */
1317 limits->min_perf_pct = max(limits->min_policy_pct,
1318 limits->min_sysfs_pct);
1319 limits->min_perf_pct = min(limits->max_policy_pct,
1320 limits->min_perf_pct);
1321 limits->max_perf_pct = min(limits->max_policy_pct,
1322 limits->max_sysfs_pct);
1323 limits->max_perf_pct = max(limits->min_policy_pct,
1324 limits->max_perf_pct);
1325 limits->max_perf = round_up(limits->max_perf, FRAC_BITS);
1327 /* Make sure min_perf_pct <= max_perf_pct */
1328 limits->min_perf_pct = min(limits->max_perf_pct, limits->min_perf_pct);
1330 limits->min_perf = div_fp(limits->min_perf_pct, 100);
1331 limits->max_perf = div_fp(limits->max_perf_pct, 100);
1334 intel_pstate_set_update_util_hook(policy->cpu);
1337 intel_pstate_hwp_set(policy->cpus);
1342 static int intel_pstate_verify_policy(struct cpufreq_policy *policy)
1344 cpufreq_verify_within_cpu_limits(policy);
1346 if (policy->policy != CPUFREQ_POLICY_POWERSAVE &&
1347 policy->policy != CPUFREQ_POLICY_PERFORMANCE)
1353 static void intel_pstate_stop_cpu(struct cpufreq_policy *policy)
1355 int cpu_num = policy->cpu;
1356 struct cpudata *cpu = all_cpu_data[cpu_num];
1358 pr_debug("CPU %d exiting\n", cpu_num);
1360 intel_pstate_clear_update_util_hook(cpu_num);
1365 intel_pstate_set_min_pstate(cpu);
1368 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
1370 struct cpudata *cpu;
1373 rc = intel_pstate_init_cpu(policy->cpu);
1377 cpu = all_cpu_data[policy->cpu];
1379 if (limits->min_perf_pct == 100 && limits->max_perf_pct == 100)
1380 policy->policy = CPUFREQ_POLICY_PERFORMANCE;
1382 policy->policy = CPUFREQ_POLICY_POWERSAVE;
1384 policy->min = cpu->pstate.min_pstate * cpu->pstate.scaling;
1385 policy->max = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
1387 /* cpuinfo and default policy values */
1388 policy->cpuinfo.min_freq = cpu->pstate.min_pstate * cpu->pstate.scaling;
1389 policy->cpuinfo.max_freq =
1390 cpu->pstate.turbo_pstate * cpu->pstate.scaling;
1391 policy->cpuinfo.transition_latency = CPUFREQ_ETERNAL;
1392 cpumask_set_cpu(policy->cpu, policy->cpus);
1397 static struct cpufreq_driver intel_pstate_driver = {
1398 .flags = CPUFREQ_CONST_LOOPS,
1399 .verify = intel_pstate_verify_policy,
1400 .setpolicy = intel_pstate_set_policy,
1401 .get = intel_pstate_get,
1402 .init = intel_pstate_cpu_init,
1403 .stop_cpu = intel_pstate_stop_cpu,
1404 .name = "intel_pstate",
1407 static int __initdata no_load;
1408 static int __initdata no_hwp;
1409 static int __initdata hwp_only;
1410 static unsigned int force_load;
1412 static int intel_pstate_msrs_not_valid(void)
1414 if (!pstate_funcs.get_max() ||
1415 !pstate_funcs.get_min() ||
1416 !pstate_funcs.get_turbo())
1422 static void copy_pid_params(struct pstate_adjust_policy *policy)
1424 pid_params.sample_rate_ms = policy->sample_rate_ms;
1425 pid_params.sample_rate_ns = pid_params.sample_rate_ms * NSEC_PER_MSEC;
1426 pid_params.p_gain_pct = policy->p_gain_pct;
1427 pid_params.i_gain_pct = policy->i_gain_pct;
1428 pid_params.d_gain_pct = policy->d_gain_pct;
1429 pid_params.deadband = policy->deadband;
1430 pid_params.setpoint = policy->setpoint;
1433 static void copy_cpu_funcs(struct pstate_funcs *funcs)
1435 pstate_funcs.get_max = funcs->get_max;
1436 pstate_funcs.get_max_physical = funcs->get_max_physical;
1437 pstate_funcs.get_min = funcs->get_min;
1438 pstate_funcs.get_turbo = funcs->get_turbo;
1439 pstate_funcs.get_scaling = funcs->get_scaling;
1440 pstate_funcs.get_val = funcs->get_val;
1441 pstate_funcs.get_vid = funcs->get_vid;
1442 pstate_funcs.get_target_pstate = funcs->get_target_pstate;
1446 #if IS_ENABLED(CONFIG_ACPI)
1447 #include <acpi/processor.h>
1449 static bool intel_pstate_no_acpi_pss(void)
1453 for_each_possible_cpu(i) {
1455 union acpi_object *pss;
1456 struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
1457 struct acpi_processor *pr = per_cpu(processors, i);
1462 status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
1463 if (ACPI_FAILURE(status))
1466 pss = buffer.pointer;
1467 if (pss && pss->type == ACPI_TYPE_PACKAGE) {
1478 static bool intel_pstate_has_acpi_ppc(void)
1482 for_each_possible_cpu(i) {
1483 struct acpi_processor *pr = per_cpu(processors, i);
1487 if (acpi_has_method(pr->handle, "_PPC"))
1498 struct hw_vendor_info {
1500 char oem_id[ACPI_OEM_ID_SIZE];
1501 char oem_table_id[ACPI_OEM_TABLE_ID_SIZE];
1505 /* Hardware vendor-specific info that has its own power management modes */
1506 static struct hw_vendor_info vendor_info[] = {
1507 {1, "HP ", "ProLiant", PSS},
1508 {1, "ORACLE", "X4-2 ", PPC},
1509 {1, "ORACLE", "X4-2L ", PPC},
1510 {1, "ORACLE", "X4-2B ", PPC},
1511 {1, "ORACLE", "X3-2 ", PPC},
1512 {1, "ORACLE", "X3-2L ", PPC},
1513 {1, "ORACLE", "X3-2B ", PPC},
1514 {1, "ORACLE", "X4470M2 ", PPC},
1515 {1, "ORACLE", "X4270M3 ", PPC},
1516 {1, "ORACLE", "X4270M2 ", PPC},
1517 {1, "ORACLE", "X4170M2 ", PPC},
1518 {1, "ORACLE", "X4170 M3", PPC},
1519 {1, "ORACLE", "X4275 M3", PPC},
1520 {1, "ORACLE", "X6-2 ", PPC},
1521 {1, "ORACLE", "Sudbury ", PPC},
1525 static bool intel_pstate_platform_pwr_mgmt_exists(void)
1527 struct acpi_table_header hdr;
1528 struct hw_vendor_info *v_info;
1529 const struct x86_cpu_id *id;
1532 id = x86_match_cpu(intel_pstate_cpu_oob_ids);
1534 rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
1535 if ( misc_pwr & (1 << 8))
1539 if (acpi_disabled ||
1540 ACPI_FAILURE(acpi_get_table_header(ACPI_SIG_FADT, 0, &hdr)))
1543 for (v_info = vendor_info; v_info->valid; v_info++) {
1544 if (!strncmp(hdr.oem_id, v_info->oem_id, ACPI_OEM_ID_SIZE) &&
1545 !strncmp(hdr.oem_table_id, v_info->oem_table_id,
1546 ACPI_OEM_TABLE_ID_SIZE))
1547 switch (v_info->oem_pwr_table) {
1549 return intel_pstate_no_acpi_pss();
1551 return intel_pstate_has_acpi_ppc() &&
1558 #else /* CONFIG_ACPI not enabled */
1559 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
1560 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
1561 #endif /* CONFIG_ACPI */
1563 static const struct x86_cpu_id hwp_support_ids[] __initconst = {
1564 { X86_VENDOR_INTEL, 6, X86_MODEL_ANY, X86_FEATURE_HWP },
1568 static int __init intel_pstate_init(void)
1571 const struct x86_cpu_id *id;
1572 struct cpu_defaults *cpu_def;
1577 if (x86_match_cpu(hwp_support_ids) && !no_hwp) {
1578 copy_cpu_funcs(&core_params.funcs);
1580 goto hwp_cpu_matched;
1583 id = x86_match_cpu(intel_pstate_cpu_ids);
1587 cpu_def = (struct cpu_defaults *)id->driver_data;
1589 copy_pid_params(&cpu_def->pid_policy);
1590 copy_cpu_funcs(&cpu_def->funcs);
1592 if (intel_pstate_msrs_not_valid())
1597 * The Intel pstate driver will be ignored if the platform
1598 * firmware has its own power management modes.
1600 if (intel_pstate_platform_pwr_mgmt_exists())
1603 pr_info("Intel P-state driver initializing\n");
1605 all_cpu_data = vzalloc(sizeof(void *) * num_possible_cpus());
1609 if (!hwp_active && hwp_only)
1612 rc = cpufreq_register_driver(&intel_pstate_driver);
1616 intel_pstate_debug_expose_params();
1617 intel_pstate_sysfs_expose_params();
1620 pr_info("HWP enabled\n");
1625 for_each_online_cpu(cpu) {
1626 if (all_cpu_data[cpu]) {
1627 intel_pstate_clear_update_util_hook(cpu);
1628 kfree(all_cpu_data[cpu]);
1633 vfree(all_cpu_data);
1636 device_initcall(intel_pstate_init);
1638 static int __init intel_pstate_setup(char *str)
1643 if (!strcmp(str, "disable"))
1645 if (!strcmp(str, "no_hwp")) {
1646 pr_info("HWP disabled\n");
1649 if (!strcmp(str, "force"))
1651 if (!strcmp(str, "hwp_only"))
1655 early_param("intel_pstate", intel_pstate_setup);
1657 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
1658 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
1659 MODULE_LICENSE("GPL");