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 int32_t core_pct_busy;
80 int max_pstate_physical;
105 struct timer_list timer;
107 struct pstate_data pstate;
111 ktime_t last_sample_time;
115 struct sample sample;
118 static struct cpudata **all_cpu_data;
119 struct pstate_adjust_policy {
128 struct pstate_funcs {
129 int (*get_max)(void);
130 int (*get_max_physical)(void);
131 int (*get_min)(void);
132 int (*get_turbo)(void);
133 int (*get_scaling)(void);
134 void (*set)(struct cpudata*, int pstate);
135 void (*get_vid)(struct cpudata *);
138 struct cpu_defaults {
139 struct pstate_adjust_policy pid_policy;
140 struct pstate_funcs funcs;
143 static struct pstate_adjust_policy pid_params;
144 static struct pstate_funcs pstate_funcs;
145 static int hwp_active;
160 static struct perf_limits performance_limits = {
164 .max_perf = int_tofp(1),
166 .min_perf = int_tofp(1),
167 .max_policy_pct = 100,
168 .max_sysfs_pct = 100,
173 static struct perf_limits powersave_limits = {
177 .max_perf = int_tofp(1),
180 .max_policy_pct = 100,
181 .max_sysfs_pct = 100,
186 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_PERFORMANCE
187 static struct perf_limits *limits = &performance_limits;
189 static struct perf_limits *limits = &powersave_limits;
192 static inline void pid_reset(struct _pid *pid, int setpoint, int busy,
193 int deadband, int integral) {
194 pid->setpoint = setpoint;
195 pid->deadband = deadband;
196 pid->integral = int_tofp(integral);
197 pid->last_err = int_tofp(setpoint) - int_tofp(busy);
200 static inline void pid_p_gain_set(struct _pid *pid, int percent)
202 pid->p_gain = div_fp(int_tofp(percent), int_tofp(100));
205 static inline void pid_i_gain_set(struct _pid *pid, int percent)
207 pid->i_gain = div_fp(int_tofp(percent), int_tofp(100));
210 static inline void pid_d_gain_set(struct _pid *pid, int percent)
212 pid->d_gain = div_fp(int_tofp(percent), int_tofp(100));
215 static signed int pid_calc(struct _pid *pid, int32_t busy)
218 int32_t pterm, dterm, fp_error;
219 int32_t integral_limit;
221 fp_error = int_tofp(pid->setpoint) - busy;
223 if (abs(fp_error) <= int_tofp(pid->deadband))
226 pterm = mul_fp(pid->p_gain, fp_error);
228 pid->integral += fp_error;
231 * We limit the integral here so that it will never
232 * get higher than 30. This prevents it from becoming
233 * too large an input over long periods of time and allows
234 * it to get factored out sooner.
236 * The value of 30 was chosen through experimentation.
238 integral_limit = int_tofp(30);
239 if (pid->integral > integral_limit)
240 pid->integral = integral_limit;
241 if (pid->integral < -integral_limit)
242 pid->integral = -integral_limit;
244 dterm = mul_fp(pid->d_gain, fp_error - pid->last_err);
245 pid->last_err = fp_error;
247 result = pterm + mul_fp(pid->integral, pid->i_gain) + dterm;
248 result = result + (1 << (FRAC_BITS-1));
249 return (signed int)fp_toint(result);
252 static inline void intel_pstate_busy_pid_reset(struct cpudata *cpu)
254 pid_p_gain_set(&cpu->pid, pid_params.p_gain_pct);
255 pid_d_gain_set(&cpu->pid, pid_params.d_gain_pct);
256 pid_i_gain_set(&cpu->pid, pid_params.i_gain_pct);
258 pid_reset(&cpu->pid, pid_params.setpoint, 100, pid_params.deadband, 0);
261 static inline void intel_pstate_reset_all_pid(void)
265 for_each_online_cpu(cpu) {
266 if (all_cpu_data[cpu])
267 intel_pstate_busy_pid_reset(all_cpu_data[cpu]);
271 static inline void update_turbo_state(void)
276 cpu = all_cpu_data[0];
277 rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
278 limits->turbo_disabled =
279 (misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ||
280 cpu->pstate.max_pstate == cpu->pstate.turbo_pstate);
283 static void intel_pstate_hwp_set(void)
285 int min, hw_min, max, hw_max, cpu, range, adj_range;
288 rdmsrl(MSR_HWP_CAPABILITIES, cap);
289 hw_min = HWP_LOWEST_PERF(cap);
290 hw_max = HWP_HIGHEST_PERF(cap);
291 range = hw_max - hw_min;
295 for_each_online_cpu(cpu) {
296 rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
297 adj_range = limits->min_perf_pct * range / 100;
298 min = hw_min + adj_range;
299 value &= ~HWP_MIN_PERF(~0L);
300 value |= HWP_MIN_PERF(min);
302 adj_range = limits->max_perf_pct * range / 100;
303 max = hw_min + adj_range;
304 if (limits->no_turbo) {
305 hw_max = HWP_GUARANTEED_PERF(cap);
310 value &= ~HWP_MAX_PERF(~0L);
311 value |= HWP_MAX_PERF(max);
312 wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
318 /************************** debugfs begin ************************/
319 static int pid_param_set(void *data, u64 val)
322 intel_pstate_reset_all_pid();
326 static int pid_param_get(void *data, u64 *val)
331 DEFINE_SIMPLE_ATTRIBUTE(fops_pid_param, pid_param_get, pid_param_set, "%llu\n");
338 static struct pid_param pid_files[] = {
339 {"sample_rate_ms", &pid_params.sample_rate_ms},
340 {"d_gain_pct", &pid_params.d_gain_pct},
341 {"i_gain_pct", &pid_params.i_gain_pct},
342 {"deadband", &pid_params.deadband},
343 {"setpoint", &pid_params.setpoint},
344 {"p_gain_pct", &pid_params.p_gain_pct},
348 static void __init intel_pstate_debug_expose_params(void)
350 struct dentry *debugfs_parent;
355 debugfs_parent = debugfs_create_dir("pstate_snb", NULL);
356 if (IS_ERR_OR_NULL(debugfs_parent))
358 while (pid_files[i].name) {
359 debugfs_create_file(pid_files[i].name, 0660,
360 debugfs_parent, pid_files[i].value,
366 /************************** debugfs end ************************/
368 /************************** sysfs begin ************************/
369 #define show_one(file_name, object) \
370 static ssize_t show_##file_name \
371 (struct kobject *kobj, struct attribute *attr, char *buf) \
373 return sprintf(buf, "%u\n", limits->object); \
376 static ssize_t show_turbo_pct(struct kobject *kobj,
377 struct attribute *attr, char *buf)
380 int total, no_turbo, turbo_pct;
383 cpu = all_cpu_data[0];
385 total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
386 no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
387 turbo_fp = div_fp(int_tofp(no_turbo), int_tofp(total));
388 turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
389 return sprintf(buf, "%u\n", turbo_pct);
392 static ssize_t show_num_pstates(struct kobject *kobj,
393 struct attribute *attr, char *buf)
398 cpu = all_cpu_data[0];
399 total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
400 return sprintf(buf, "%u\n", total);
403 static ssize_t show_no_turbo(struct kobject *kobj,
404 struct attribute *attr, char *buf)
408 update_turbo_state();
409 if (limits->turbo_disabled)
410 ret = sprintf(buf, "%u\n", limits->turbo_disabled);
412 ret = sprintf(buf, "%u\n", limits->no_turbo);
417 static ssize_t store_no_turbo(struct kobject *a, struct attribute *b,
418 const char *buf, size_t count)
423 ret = sscanf(buf, "%u", &input);
427 update_turbo_state();
428 if (limits->turbo_disabled) {
429 pr_warn("intel_pstate: Turbo disabled by BIOS or unavailable on processor\n");
433 limits->no_turbo = clamp_t(int, input, 0, 1);
436 intel_pstate_hwp_set();
441 static ssize_t store_max_perf_pct(struct kobject *a, struct attribute *b,
442 const char *buf, size_t count)
447 ret = sscanf(buf, "%u", &input);
451 limits->max_sysfs_pct = clamp_t(int, input, 0 , 100);
452 limits->max_perf_pct = min(limits->max_policy_pct,
453 limits->max_sysfs_pct);
454 limits->max_perf_pct = max(limits->min_policy_pct,
455 limits->max_perf_pct);
456 limits->max_perf_pct = max(limits->min_perf_pct,
457 limits->max_perf_pct);
458 limits->max_perf = div_fp(int_tofp(limits->max_perf_pct),
462 intel_pstate_hwp_set();
466 static ssize_t store_min_perf_pct(struct kobject *a, struct attribute *b,
467 const char *buf, size_t count)
472 ret = sscanf(buf, "%u", &input);
476 limits->min_sysfs_pct = clamp_t(int, input, 0 , 100);
477 limits->min_perf_pct = max(limits->min_policy_pct,
478 limits->min_sysfs_pct);
479 limits->min_perf_pct = min(limits->max_policy_pct,
480 limits->min_perf_pct);
481 limits->min_perf_pct = min(limits->max_perf_pct,
482 limits->min_perf_pct);
483 limits->min_perf = div_fp(int_tofp(limits->min_perf_pct),
487 intel_pstate_hwp_set();
491 show_one(max_perf_pct, max_perf_pct);
492 show_one(min_perf_pct, min_perf_pct);
494 define_one_global_rw(no_turbo);
495 define_one_global_rw(max_perf_pct);
496 define_one_global_rw(min_perf_pct);
497 define_one_global_ro(turbo_pct);
498 define_one_global_ro(num_pstates);
500 static struct attribute *intel_pstate_attributes[] = {
509 static struct attribute_group intel_pstate_attr_group = {
510 .attrs = intel_pstate_attributes,
513 static void __init intel_pstate_sysfs_expose_params(void)
515 struct kobject *intel_pstate_kobject;
518 intel_pstate_kobject = kobject_create_and_add("intel_pstate",
519 &cpu_subsys.dev_root->kobj);
520 BUG_ON(!intel_pstate_kobject);
521 rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
524 /************************** sysfs end ************************/
526 static void intel_pstate_hwp_enable(struct cpudata *cpudata)
528 wrmsrl_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1);
531 static int atom_get_min_pstate(void)
535 rdmsrl(ATOM_RATIOS, value);
536 return (value >> 8) & 0x7F;
539 static int atom_get_max_pstate(void)
543 rdmsrl(ATOM_RATIOS, value);
544 return (value >> 16) & 0x7F;
547 static int atom_get_turbo_pstate(void)
551 rdmsrl(ATOM_TURBO_RATIOS, value);
555 static void atom_set_pstate(struct cpudata *cpudata, int pstate)
561 val = (u64)pstate << 8;
562 if (limits->no_turbo && !limits->turbo_disabled)
565 vid_fp = cpudata->vid.min + mul_fp(
566 int_tofp(pstate - cpudata->pstate.min_pstate),
569 vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
570 vid = ceiling_fp(vid_fp);
572 if (pstate > cpudata->pstate.max_pstate)
573 vid = cpudata->vid.turbo;
577 wrmsrl_on_cpu(cpudata->cpu, MSR_IA32_PERF_CTL, val);
580 #define ATOM_BCLK_FREQS 5
581 static int atom_freq_table[ATOM_BCLK_FREQS] = { 833, 1000, 1333, 1167, 800};
583 static int atom_get_scaling(void)
588 rdmsrl(MSR_FSB_FREQ, value);
591 WARN_ON(i > ATOM_BCLK_FREQS);
593 return atom_freq_table[i] * 100;
596 static void atom_get_vid(struct cpudata *cpudata)
600 rdmsrl(ATOM_VIDS, value);
601 cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
602 cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
603 cpudata->vid.ratio = div_fp(
604 cpudata->vid.max - cpudata->vid.min,
605 int_tofp(cpudata->pstate.max_pstate -
606 cpudata->pstate.min_pstate));
608 rdmsrl(ATOM_TURBO_VIDS, value);
609 cpudata->vid.turbo = value & 0x7f;
612 static int core_get_min_pstate(void)
616 rdmsrl(MSR_PLATFORM_INFO, value);
617 return (value >> 40) & 0xFF;
620 static int core_get_max_pstate_physical(void)
624 rdmsrl(MSR_PLATFORM_INFO, value);
625 return (value >> 8) & 0xFF;
628 static int core_get_max_pstate(void)
635 rdmsrl(MSR_PLATFORM_INFO, plat_info);
636 max_pstate = (plat_info >> 8) & 0xFF;
638 err = rdmsrl_safe(MSR_TURBO_ACTIVATION_RATIO, &tar);
640 /* Do some sanity checking for safety */
641 if (plat_info & 0x600000000) {
646 err = rdmsrl_safe(MSR_CONFIG_TDP_CONTROL, &tdp_ctrl);
650 tdp_msr = MSR_CONFIG_TDP_NOMINAL + tdp_ctrl;
651 err = rdmsrl_safe(tdp_msr, &tdp_ratio);
655 if (tdp_ratio - 1 == tar) {
657 pr_debug("max_pstate=TAC %x\n", max_pstate);
668 static int core_get_turbo_pstate(void)
673 rdmsrl(MSR_NHM_TURBO_RATIO_LIMIT, value);
674 nont = core_get_max_pstate();
681 static inline int core_get_scaling(void)
686 static void core_set_pstate(struct cpudata *cpudata, int pstate)
690 val = (u64)pstate << 8;
691 if (limits->no_turbo && !limits->turbo_disabled)
694 wrmsrl_on_cpu(cpudata->cpu, MSR_IA32_PERF_CTL, val);
697 static int knl_get_turbo_pstate(void)
702 rdmsrl(MSR_NHM_TURBO_RATIO_LIMIT, value);
703 nont = core_get_max_pstate();
704 ret = (((value) >> 8) & 0xFF);
710 static struct cpu_defaults core_params = {
712 .sample_rate_ms = 10,
720 .get_max = core_get_max_pstate,
721 .get_max_physical = core_get_max_pstate_physical,
722 .get_min = core_get_min_pstate,
723 .get_turbo = core_get_turbo_pstate,
724 .get_scaling = core_get_scaling,
725 .set = core_set_pstate,
729 static struct cpu_defaults atom_params = {
731 .sample_rate_ms = 10,
739 .get_max = atom_get_max_pstate,
740 .get_max_physical = atom_get_max_pstate,
741 .get_min = atom_get_min_pstate,
742 .get_turbo = atom_get_turbo_pstate,
743 .set = atom_set_pstate,
744 .get_scaling = atom_get_scaling,
745 .get_vid = atom_get_vid,
749 static struct cpu_defaults knl_params = {
751 .sample_rate_ms = 10,
759 .get_max = core_get_max_pstate,
760 .get_max_physical = core_get_max_pstate_physical,
761 .get_min = core_get_min_pstate,
762 .get_turbo = knl_get_turbo_pstate,
763 .get_scaling = core_get_scaling,
764 .set = core_set_pstate,
768 static void intel_pstate_get_min_max(struct cpudata *cpu, int *min, int *max)
770 int max_perf = cpu->pstate.turbo_pstate;
774 if (limits->no_turbo || limits->turbo_disabled)
775 max_perf = cpu->pstate.max_pstate;
778 * performance can be limited by user through sysfs, by cpufreq
779 * policy, or by cpu specific default values determined through
782 max_perf_adj = fp_toint(mul_fp(int_tofp(max_perf), limits->max_perf));
783 *max = clamp_t(int, max_perf_adj,
784 cpu->pstate.min_pstate, cpu->pstate.turbo_pstate);
786 min_perf = fp_toint(mul_fp(int_tofp(max_perf), limits->min_perf));
787 *min = clamp_t(int, min_perf, cpu->pstate.min_pstate, max_perf);
790 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate, bool force)
792 int max_perf, min_perf;
795 update_turbo_state();
797 intel_pstate_get_min_max(cpu, &min_perf, &max_perf);
799 pstate = clamp_t(int, pstate, min_perf, max_perf);
801 if (pstate == cpu->pstate.current_pstate)
804 trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
806 cpu->pstate.current_pstate = pstate;
808 pstate_funcs.set(cpu, pstate);
811 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
813 cpu->pstate.min_pstate = pstate_funcs.get_min();
814 cpu->pstate.max_pstate = pstate_funcs.get_max();
815 cpu->pstate.max_pstate_physical = pstate_funcs.get_max_physical();
816 cpu->pstate.turbo_pstate = pstate_funcs.get_turbo();
817 cpu->pstate.scaling = pstate_funcs.get_scaling();
819 if (pstate_funcs.get_vid)
820 pstate_funcs.get_vid(cpu);
821 intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate, false);
824 static inline void intel_pstate_calc_busy(struct cpudata *cpu)
826 struct sample *sample = &cpu->sample;
829 core_pct = int_tofp(sample->aperf) * int_tofp(100);
830 core_pct = div64_u64(core_pct, int_tofp(sample->mperf));
832 sample->freq = fp_toint(
834 cpu->pstate.max_pstate_physical *
835 cpu->pstate.scaling / 100),
838 sample->core_pct_busy = (int32_t)core_pct;
841 static inline void intel_pstate_sample(struct cpudata *cpu)
847 local_irq_save(flags);
848 rdmsrl(MSR_IA32_APERF, aperf);
849 rdmsrl(MSR_IA32_MPERF, mperf);
850 if (cpu->prev_mperf == mperf) {
851 local_irq_restore(flags);
856 local_irq_restore(flags);
858 cpu->last_sample_time = cpu->sample.time;
859 cpu->sample.time = ktime_get();
860 cpu->sample.aperf = aperf;
861 cpu->sample.mperf = mperf;
862 cpu->sample.tsc = tsc;
863 cpu->sample.aperf -= cpu->prev_aperf;
864 cpu->sample.mperf -= cpu->prev_mperf;
865 cpu->sample.tsc -= cpu->prev_tsc;
867 intel_pstate_calc_busy(cpu);
869 cpu->prev_aperf = aperf;
870 cpu->prev_mperf = mperf;
874 static inline void intel_hwp_set_sample_time(struct cpudata *cpu)
878 delay = msecs_to_jiffies(50);
879 mod_timer_pinned(&cpu->timer, jiffies + delay);
882 static inline void intel_pstate_set_sample_time(struct cpudata *cpu)
886 delay = msecs_to_jiffies(pid_params.sample_rate_ms);
887 mod_timer_pinned(&cpu->timer, jiffies + delay);
890 static inline int32_t intel_pstate_get_scaled_busy(struct cpudata *cpu)
892 int32_t core_busy, max_pstate, current_pstate, sample_ratio;
897 * core_busy is the ratio of actual performance to max
898 * max_pstate is the max non turbo pstate available
899 * current_pstate was the pstate that was requested during
900 * the last sample period.
902 * We normalize core_busy, which was our actual percent
903 * performance to what we requested during the last sample
904 * period. The result will be a percentage of busy at a
907 core_busy = cpu->sample.core_pct_busy;
908 max_pstate = int_tofp(cpu->pstate.max_pstate_physical);
909 current_pstate = int_tofp(cpu->pstate.current_pstate);
910 core_busy = mul_fp(core_busy, div_fp(max_pstate, current_pstate));
913 * Since we have a deferred timer, it will not fire unless
914 * we are in C0. So, determine if the actual elapsed time
915 * is significantly greater (3x) than our sample interval. If it
916 * is, then we were idle for a long enough period of time
917 * to adjust our busyness.
919 sample_time = pid_params.sample_rate_ms * USEC_PER_MSEC;
920 duration_us = ktime_us_delta(cpu->sample.time,
921 cpu->last_sample_time);
922 if (duration_us > sample_time * 3) {
923 sample_ratio = div_fp(int_tofp(sample_time),
924 int_tofp(duration_us));
925 core_busy = mul_fp(core_busy, sample_ratio);
931 static inline void intel_pstate_adjust_busy_pstate(struct cpudata *cpu)
937 struct sample *sample;
939 from = cpu->pstate.current_pstate;
942 busy_scaled = intel_pstate_get_scaled_busy(cpu);
944 ctl = pid_calc(pid, busy_scaled);
946 /* Negative values of ctl increase the pstate and vice versa */
947 intel_pstate_set_pstate(cpu, cpu->pstate.current_pstate - ctl, true);
949 sample = &cpu->sample;
950 trace_pstate_sample(fp_toint(sample->core_pct_busy),
951 fp_toint(busy_scaled),
953 cpu->pstate.current_pstate,
960 static void intel_hwp_timer_func(unsigned long __data)
962 struct cpudata *cpu = (struct cpudata *) __data;
964 intel_pstate_sample(cpu);
965 intel_hwp_set_sample_time(cpu);
968 static void intel_pstate_timer_func(unsigned long __data)
970 struct cpudata *cpu = (struct cpudata *) __data;
972 intel_pstate_sample(cpu);
974 intel_pstate_adjust_busy_pstate(cpu);
976 intel_pstate_set_sample_time(cpu);
979 #define ICPU(model, policy) \
980 { X86_VENDOR_INTEL, 6, model, X86_FEATURE_APERFMPERF,\
981 (unsigned long)&policy }
983 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
984 ICPU(0x2a, core_params),
985 ICPU(0x2d, core_params),
986 ICPU(0x37, atom_params),
987 ICPU(0x3a, core_params),
988 ICPU(0x3c, core_params),
989 ICPU(0x3d, core_params),
990 ICPU(0x3e, core_params),
991 ICPU(0x3f, core_params),
992 ICPU(0x45, core_params),
993 ICPU(0x46, core_params),
994 ICPU(0x47, core_params),
995 ICPU(0x4c, atom_params),
996 ICPU(0x4e, core_params),
997 ICPU(0x4f, core_params),
998 ICPU(0x5e, core_params),
999 ICPU(0x56, core_params),
1000 ICPU(0x57, knl_params),
1003 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
1005 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] = {
1006 ICPU(0x56, core_params),
1010 static int intel_pstate_init_cpu(unsigned int cpunum)
1012 struct cpudata *cpu;
1014 if (!all_cpu_data[cpunum])
1015 all_cpu_data[cpunum] = kzalloc(sizeof(struct cpudata),
1017 if (!all_cpu_data[cpunum])
1020 cpu = all_cpu_data[cpunum];
1025 intel_pstate_hwp_enable(cpu);
1027 intel_pstate_get_cpu_pstates(cpu);
1029 init_timer_deferrable(&cpu->timer);
1030 cpu->timer.data = (unsigned long)cpu;
1031 cpu->timer.expires = jiffies + HZ/100;
1034 cpu->timer.function = intel_pstate_timer_func;
1036 cpu->timer.function = intel_hwp_timer_func;
1038 intel_pstate_busy_pid_reset(cpu);
1039 intel_pstate_sample(cpu);
1041 add_timer_on(&cpu->timer, cpunum);
1043 pr_debug("intel_pstate: controlling: cpu %d\n", cpunum);
1048 static unsigned int intel_pstate_get(unsigned int cpu_num)
1050 struct sample *sample;
1051 struct cpudata *cpu;
1053 cpu = all_cpu_data[cpu_num];
1056 sample = &cpu->sample;
1057 return sample->freq;
1060 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
1062 if (!policy->cpuinfo.max_freq)
1065 if (policy->policy == CPUFREQ_POLICY_PERFORMANCE &&
1066 policy->max >= policy->cpuinfo.max_freq) {
1067 pr_debug("intel_pstate: set performance\n");
1068 limits = &performance_limits;
1072 pr_debug("intel_pstate: set powersave\n");
1073 limits = &powersave_limits;
1074 limits->min_policy_pct = (policy->min * 100) / policy->cpuinfo.max_freq;
1075 limits->min_policy_pct = clamp_t(int, limits->min_policy_pct, 0 , 100);
1076 limits->max_policy_pct = (policy->max * 100) / policy->cpuinfo.max_freq;
1077 limits->max_policy_pct = clamp_t(int, limits->max_policy_pct, 0 , 100);
1079 /* Normalize user input to [min_policy_pct, max_policy_pct] */
1080 limits->min_perf_pct = max(limits->min_policy_pct,
1081 limits->min_sysfs_pct);
1082 limits->min_perf_pct = min(limits->max_policy_pct,
1083 limits->min_perf_pct);
1084 limits->max_perf_pct = min(limits->max_policy_pct,
1085 limits->max_sysfs_pct);
1086 limits->max_perf_pct = max(limits->min_policy_pct,
1087 limits->max_perf_pct);
1089 /* Make sure min_perf_pct <= max_perf_pct */
1090 limits->min_perf_pct = min(limits->max_perf_pct, limits->min_perf_pct);
1092 limits->min_perf = div_fp(int_tofp(limits->min_perf_pct),
1094 limits->max_perf = div_fp(int_tofp(limits->max_perf_pct),
1098 intel_pstate_hwp_set();
1103 static int intel_pstate_verify_policy(struct cpufreq_policy *policy)
1105 cpufreq_verify_within_cpu_limits(policy);
1107 if (policy->policy != CPUFREQ_POLICY_POWERSAVE &&
1108 policy->policy != CPUFREQ_POLICY_PERFORMANCE)
1114 static void intel_pstate_stop_cpu(struct cpufreq_policy *policy)
1116 int cpu_num = policy->cpu;
1117 struct cpudata *cpu = all_cpu_data[cpu_num];
1119 pr_debug("intel_pstate: CPU %d exiting\n", cpu_num);
1121 del_timer_sync(&all_cpu_data[cpu_num]->timer);
1125 intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate, false);
1128 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
1130 struct cpudata *cpu;
1133 rc = intel_pstate_init_cpu(policy->cpu);
1137 cpu = all_cpu_data[policy->cpu];
1139 if (limits->min_perf_pct == 100 && limits->max_perf_pct == 100)
1140 policy->policy = CPUFREQ_POLICY_PERFORMANCE;
1142 policy->policy = CPUFREQ_POLICY_POWERSAVE;
1144 policy->min = cpu->pstate.min_pstate * cpu->pstate.scaling;
1145 policy->max = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
1147 /* cpuinfo and default policy values */
1148 policy->cpuinfo.min_freq = cpu->pstate.min_pstate * cpu->pstate.scaling;
1149 policy->cpuinfo.max_freq =
1150 cpu->pstate.turbo_pstate * cpu->pstate.scaling;
1151 policy->cpuinfo.transition_latency = CPUFREQ_ETERNAL;
1152 cpumask_set_cpu(policy->cpu, policy->cpus);
1157 static struct cpufreq_driver intel_pstate_driver = {
1158 .flags = CPUFREQ_CONST_LOOPS,
1159 .verify = intel_pstate_verify_policy,
1160 .setpolicy = intel_pstate_set_policy,
1161 .get = intel_pstate_get,
1162 .init = intel_pstate_cpu_init,
1163 .stop_cpu = intel_pstate_stop_cpu,
1164 .name = "intel_pstate",
1167 static int __initdata no_load;
1168 static int __initdata no_hwp;
1169 static int __initdata hwp_only;
1170 static unsigned int force_load;
1172 static int intel_pstate_msrs_not_valid(void)
1174 if (!pstate_funcs.get_max() ||
1175 !pstate_funcs.get_min() ||
1176 !pstate_funcs.get_turbo())
1182 static void copy_pid_params(struct pstate_adjust_policy *policy)
1184 pid_params.sample_rate_ms = policy->sample_rate_ms;
1185 pid_params.p_gain_pct = policy->p_gain_pct;
1186 pid_params.i_gain_pct = policy->i_gain_pct;
1187 pid_params.d_gain_pct = policy->d_gain_pct;
1188 pid_params.deadband = policy->deadband;
1189 pid_params.setpoint = policy->setpoint;
1192 static void copy_cpu_funcs(struct pstate_funcs *funcs)
1194 pstate_funcs.get_max = funcs->get_max;
1195 pstate_funcs.get_max_physical = funcs->get_max_physical;
1196 pstate_funcs.get_min = funcs->get_min;
1197 pstate_funcs.get_turbo = funcs->get_turbo;
1198 pstate_funcs.get_scaling = funcs->get_scaling;
1199 pstate_funcs.set = funcs->set;
1200 pstate_funcs.get_vid = funcs->get_vid;
1203 #if IS_ENABLED(CONFIG_ACPI)
1204 #include <acpi/processor.h>
1206 static bool intel_pstate_no_acpi_pss(void)
1210 for_each_possible_cpu(i) {
1212 union acpi_object *pss;
1213 struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
1214 struct acpi_processor *pr = per_cpu(processors, i);
1219 status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
1220 if (ACPI_FAILURE(status))
1223 pss = buffer.pointer;
1224 if (pss && pss->type == ACPI_TYPE_PACKAGE) {
1235 static bool intel_pstate_has_acpi_ppc(void)
1239 for_each_possible_cpu(i) {
1240 struct acpi_processor *pr = per_cpu(processors, i);
1244 if (acpi_has_method(pr->handle, "_PPC"))
1255 struct hw_vendor_info {
1257 char oem_id[ACPI_OEM_ID_SIZE];
1258 char oem_table_id[ACPI_OEM_TABLE_ID_SIZE];
1262 /* Hardware vendor-specific info that has its own power management modes */
1263 static struct hw_vendor_info vendor_info[] = {
1264 {1, "HP ", "ProLiant", PSS},
1265 {1, "ORACLE", "X4-2 ", PPC},
1266 {1, "ORACLE", "X4-2L ", PPC},
1267 {1, "ORACLE", "X4-2B ", PPC},
1268 {1, "ORACLE", "X3-2 ", PPC},
1269 {1, "ORACLE", "X3-2L ", PPC},
1270 {1, "ORACLE", "X3-2B ", PPC},
1271 {1, "ORACLE", "X4470M2 ", PPC},
1272 {1, "ORACLE", "X4270M3 ", PPC},
1273 {1, "ORACLE", "X4270M2 ", PPC},
1274 {1, "ORACLE", "X4170M2 ", PPC},
1275 {1, "ORACLE", "X4170 M3", PPC},
1276 {1, "ORACLE", "X4275 M3", PPC},
1277 {1, "ORACLE", "X6-2 ", PPC},
1278 {1, "ORACLE", "Sudbury ", PPC},
1282 static bool intel_pstate_platform_pwr_mgmt_exists(void)
1284 struct acpi_table_header hdr;
1285 struct hw_vendor_info *v_info;
1286 const struct x86_cpu_id *id;
1289 id = x86_match_cpu(intel_pstate_cpu_oob_ids);
1291 rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
1292 if ( misc_pwr & (1 << 8))
1296 if (acpi_disabled ||
1297 ACPI_FAILURE(acpi_get_table_header(ACPI_SIG_FADT, 0, &hdr)))
1300 for (v_info = vendor_info; v_info->valid; v_info++) {
1301 if (!strncmp(hdr.oem_id, v_info->oem_id, ACPI_OEM_ID_SIZE) &&
1302 !strncmp(hdr.oem_table_id, v_info->oem_table_id,
1303 ACPI_OEM_TABLE_ID_SIZE))
1304 switch (v_info->oem_pwr_table) {
1306 return intel_pstate_no_acpi_pss();
1308 return intel_pstate_has_acpi_ppc() &&
1315 #else /* CONFIG_ACPI not enabled */
1316 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
1317 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
1318 #endif /* CONFIG_ACPI */
1320 static int __init intel_pstate_init(void)
1323 const struct x86_cpu_id *id;
1324 struct cpu_defaults *cpu_def;
1329 id = x86_match_cpu(intel_pstate_cpu_ids);
1334 * The Intel pstate driver will be ignored if the platform
1335 * firmware has its own power management modes.
1337 if (intel_pstate_platform_pwr_mgmt_exists())
1340 cpu_def = (struct cpu_defaults *)id->driver_data;
1342 copy_pid_params(&cpu_def->pid_policy);
1343 copy_cpu_funcs(&cpu_def->funcs);
1345 if (intel_pstate_msrs_not_valid())
1348 pr_info("Intel P-state driver initializing.\n");
1350 all_cpu_data = vzalloc(sizeof(void *) * num_possible_cpus());
1354 if (static_cpu_has_safe(X86_FEATURE_HWP) && !no_hwp) {
1355 pr_info("intel_pstate: HWP enabled\n");
1359 if (!hwp_active && hwp_only)
1362 rc = cpufreq_register_driver(&intel_pstate_driver);
1366 intel_pstate_debug_expose_params();
1367 intel_pstate_sysfs_expose_params();
1372 for_each_online_cpu(cpu) {
1373 if (all_cpu_data[cpu]) {
1374 del_timer_sync(&all_cpu_data[cpu]->timer);
1375 kfree(all_cpu_data[cpu]);
1380 vfree(all_cpu_data);
1383 device_initcall(intel_pstate_init);
1385 static int __init intel_pstate_setup(char *str)
1390 if (!strcmp(str, "disable"))
1392 if (!strcmp(str, "no_hwp")) {
1393 pr_info("intel_pstate: HWP disabled\n");
1396 if (!strcmp(str, "force"))
1398 if (!strcmp(str, "hwp_only"))
1402 early_param("intel_pstate", intel_pstate_setup);
1404 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
1405 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
1406 MODULE_LICENSE("GPL");