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 <trace/events/power.h>
31 #include <asm/div64.h>
33 #include <asm/cpu_device_id.h>
35 #define BYT_RATIOS 0x66a
36 #define BYT_VIDS 0x66b
37 #define BYT_TURBO_RATIOS 0x66c
38 #define BYT_TURBO_VIDS 0x66d
41 #define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
42 #define fp_toint(X) ((X) >> FRAC_BITS)
45 static inline int32_t mul_fp(int32_t x, int32_t y)
47 return ((int64_t)x * (int64_t)y) >> FRAC_BITS;
50 static inline int32_t div_fp(int32_t x, int32_t y)
52 return div_s64((int64_t)x << FRAC_BITS, y);
55 static inline int ceiling_fp(int32_t x)
60 mask = (1 << FRAC_BITS) - 1;
67 int32_t core_pct_busy;
102 struct timer_list timer;
104 struct pstate_data pstate;
108 ktime_t last_sample_time;
111 struct sample sample;
114 static struct cpudata **all_cpu_data;
115 struct pstate_adjust_policy {
124 struct pstate_funcs {
125 int (*get_max)(void);
126 int (*get_min)(void);
127 int (*get_turbo)(void);
128 int (*get_scaling)(void);
129 void (*set)(struct cpudata*, int pstate);
130 void (*get_vid)(struct cpudata *);
133 struct cpu_defaults {
134 struct pstate_adjust_policy pid_policy;
135 struct pstate_funcs funcs;
138 static struct pstate_adjust_policy pid_params;
139 static struct pstate_funcs pstate_funcs;
140 static int hwp_active;
153 static struct perf_limits limits = {
157 .max_perf = int_tofp(1),
160 .max_policy_pct = 100,
161 .max_sysfs_pct = 100,
164 static inline void pid_reset(struct _pid *pid, int setpoint, int busy,
165 int deadband, int integral) {
166 pid->setpoint = setpoint;
167 pid->deadband = deadband;
168 pid->integral = int_tofp(integral);
169 pid->last_err = int_tofp(setpoint) - int_tofp(busy);
172 static inline void pid_p_gain_set(struct _pid *pid, int percent)
174 pid->p_gain = div_fp(int_tofp(percent), int_tofp(100));
177 static inline void pid_i_gain_set(struct _pid *pid, int percent)
179 pid->i_gain = div_fp(int_tofp(percent), int_tofp(100));
182 static inline void pid_d_gain_set(struct _pid *pid, int percent)
184 pid->d_gain = div_fp(int_tofp(percent), int_tofp(100));
187 static signed int pid_calc(struct _pid *pid, int32_t busy)
190 int32_t pterm, dterm, fp_error;
191 int32_t integral_limit;
193 fp_error = int_tofp(pid->setpoint) - busy;
195 if (abs(fp_error) <= int_tofp(pid->deadband))
198 pterm = mul_fp(pid->p_gain, fp_error);
200 pid->integral += fp_error;
203 * We limit the integral here so that it will never
204 * get higher than 30. This prevents it from becoming
205 * too large an input over long periods of time and allows
206 * it to get factored out sooner.
208 * The value of 30 was chosen through experimentation.
210 integral_limit = int_tofp(30);
211 if (pid->integral > integral_limit)
212 pid->integral = integral_limit;
213 if (pid->integral < -integral_limit)
214 pid->integral = -integral_limit;
216 dterm = mul_fp(pid->d_gain, fp_error - pid->last_err);
217 pid->last_err = fp_error;
219 result = pterm + mul_fp(pid->integral, pid->i_gain) + dterm;
220 result = result + (1 << (FRAC_BITS-1));
221 return (signed int)fp_toint(result);
224 static inline void intel_pstate_busy_pid_reset(struct cpudata *cpu)
226 pid_p_gain_set(&cpu->pid, pid_params.p_gain_pct);
227 pid_d_gain_set(&cpu->pid, pid_params.d_gain_pct);
228 pid_i_gain_set(&cpu->pid, pid_params.i_gain_pct);
230 pid_reset(&cpu->pid, pid_params.setpoint, 100, pid_params.deadband, 0);
233 static inline void intel_pstate_reset_all_pid(void)
237 for_each_online_cpu(cpu) {
238 if (all_cpu_data[cpu])
239 intel_pstate_busy_pid_reset(all_cpu_data[cpu]);
243 static inline void update_turbo_state(void)
248 cpu = all_cpu_data[0];
249 rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
250 limits.turbo_disabled =
251 (misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ||
252 cpu->pstate.max_pstate == cpu->pstate.turbo_pstate);
255 #define PCT_TO_HWP(x) (x * 255 / 100)
256 static void intel_pstate_hwp_set(void)
263 for_each_online_cpu(cpu) {
264 rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
265 min = PCT_TO_HWP(limits.min_perf_pct);
266 value &= ~HWP_MIN_PERF(~0L);
267 value |= HWP_MIN_PERF(min);
269 max = PCT_TO_HWP(limits.max_perf_pct);
270 if (limits.no_turbo) {
271 rdmsrl( MSR_HWP_CAPABILITIES, freq);
272 max = HWP_GUARANTEED_PERF(freq);
275 value &= ~HWP_MAX_PERF(~0L);
276 value |= HWP_MAX_PERF(max);
277 wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
283 /************************** debugfs begin ************************/
284 static int pid_param_set(void *data, u64 val)
287 intel_pstate_reset_all_pid();
291 static int pid_param_get(void *data, u64 *val)
296 DEFINE_SIMPLE_ATTRIBUTE(fops_pid_param, pid_param_get, pid_param_set, "%llu\n");
303 static struct pid_param pid_files[] = {
304 {"sample_rate_ms", &pid_params.sample_rate_ms},
305 {"d_gain_pct", &pid_params.d_gain_pct},
306 {"i_gain_pct", &pid_params.i_gain_pct},
307 {"deadband", &pid_params.deadband},
308 {"setpoint", &pid_params.setpoint},
309 {"p_gain_pct", &pid_params.p_gain_pct},
313 static void __init intel_pstate_debug_expose_params(void)
315 struct dentry *debugfs_parent;
320 debugfs_parent = debugfs_create_dir("pstate_snb", NULL);
321 if (IS_ERR_OR_NULL(debugfs_parent))
323 while (pid_files[i].name) {
324 debugfs_create_file(pid_files[i].name, 0660,
325 debugfs_parent, pid_files[i].value,
331 /************************** debugfs end ************************/
333 /************************** sysfs begin ************************/
334 #define show_one(file_name, object) \
335 static ssize_t show_##file_name \
336 (struct kobject *kobj, struct attribute *attr, char *buf) \
338 return sprintf(buf, "%u\n", limits.object); \
341 static ssize_t show_no_turbo(struct kobject *kobj,
342 struct attribute *attr, char *buf)
346 update_turbo_state();
347 if (limits.turbo_disabled)
348 ret = sprintf(buf, "%u\n", limits.turbo_disabled);
350 ret = sprintf(buf, "%u\n", limits.no_turbo);
355 static ssize_t store_no_turbo(struct kobject *a, struct attribute *b,
356 const char *buf, size_t count)
361 ret = sscanf(buf, "%u", &input);
365 update_turbo_state();
366 if (limits.turbo_disabled) {
367 pr_warn("Turbo disabled by BIOS or unavailable on processor\n");
371 limits.no_turbo = clamp_t(int, input, 0, 1);
374 intel_pstate_hwp_set();
379 static ssize_t store_max_perf_pct(struct kobject *a, struct attribute *b,
380 const char *buf, size_t count)
385 ret = sscanf(buf, "%u", &input);
389 limits.max_sysfs_pct = clamp_t(int, input, 0 , 100);
390 limits.max_perf_pct = min(limits.max_policy_pct, limits.max_sysfs_pct);
391 limits.max_perf = div_fp(int_tofp(limits.max_perf_pct), int_tofp(100));
394 intel_pstate_hwp_set();
398 static ssize_t store_min_perf_pct(struct kobject *a, struct attribute *b,
399 const char *buf, size_t count)
404 ret = sscanf(buf, "%u", &input);
407 limits.min_perf_pct = clamp_t(int, input, 0 , 100);
408 limits.min_perf = div_fp(int_tofp(limits.min_perf_pct), int_tofp(100));
411 intel_pstate_hwp_set();
415 show_one(max_perf_pct, max_perf_pct);
416 show_one(min_perf_pct, min_perf_pct);
418 define_one_global_rw(no_turbo);
419 define_one_global_rw(max_perf_pct);
420 define_one_global_rw(min_perf_pct);
422 static struct attribute *intel_pstate_attributes[] = {
429 static struct attribute_group intel_pstate_attr_group = {
430 .attrs = intel_pstate_attributes,
433 static void __init intel_pstate_sysfs_expose_params(void)
435 struct kobject *intel_pstate_kobject;
438 intel_pstate_kobject = kobject_create_and_add("intel_pstate",
439 &cpu_subsys.dev_root->kobj);
440 BUG_ON(!intel_pstate_kobject);
441 rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
444 /************************** sysfs end ************************/
446 static void intel_pstate_hwp_enable(void)
449 pr_info("intel_pstate HWP enabled\n");
451 wrmsrl( MSR_PM_ENABLE, 0x1);
454 static int byt_get_min_pstate(void)
458 rdmsrl(BYT_RATIOS, value);
459 return (value >> 8) & 0x7F;
462 static int byt_get_max_pstate(void)
466 rdmsrl(BYT_RATIOS, value);
467 return (value >> 16) & 0x7F;
470 static int byt_get_turbo_pstate(void)
474 rdmsrl(BYT_TURBO_RATIOS, value);
478 static void byt_set_pstate(struct cpudata *cpudata, int pstate)
485 if (limits.no_turbo && !limits.turbo_disabled)
488 vid_fp = cpudata->vid.min + mul_fp(
489 int_tofp(pstate - cpudata->pstate.min_pstate),
492 vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
493 vid = ceiling_fp(vid_fp);
495 if (pstate > cpudata->pstate.max_pstate)
496 vid = cpudata->vid.turbo;
500 wrmsrl(MSR_IA32_PERF_CTL, val);
503 #define BYT_BCLK_FREQS 5
504 static int byt_freq_table[BYT_BCLK_FREQS] = { 833, 1000, 1333, 1167, 800};
506 static int byt_get_scaling(void)
511 rdmsrl(MSR_FSB_FREQ, value);
514 BUG_ON(i > BYT_BCLK_FREQS);
516 return byt_freq_table[i] * 100;
519 static void byt_get_vid(struct cpudata *cpudata)
523 rdmsrl(BYT_VIDS, value);
524 cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
525 cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
526 cpudata->vid.ratio = div_fp(
527 cpudata->vid.max - cpudata->vid.min,
528 int_tofp(cpudata->pstate.max_pstate -
529 cpudata->pstate.min_pstate));
531 rdmsrl(BYT_TURBO_VIDS, value);
532 cpudata->vid.turbo = value & 0x7f;
535 static int core_get_min_pstate(void)
539 rdmsrl(MSR_PLATFORM_INFO, value);
540 return (value >> 40) & 0xFF;
543 static int core_get_max_pstate(void)
547 rdmsrl(MSR_PLATFORM_INFO, value);
548 return (value >> 8) & 0xFF;
551 static int core_get_turbo_pstate(void)
556 rdmsrl(MSR_NHM_TURBO_RATIO_LIMIT, value);
557 nont = core_get_max_pstate();
564 static inline int core_get_scaling(void)
569 static void core_set_pstate(struct cpudata *cpudata, int pstate)
574 if (limits.no_turbo && !limits.turbo_disabled)
577 wrmsrl_on_cpu(cpudata->cpu, MSR_IA32_PERF_CTL, val);
580 static struct cpu_defaults core_params = {
582 .sample_rate_ms = 10,
590 .get_max = core_get_max_pstate,
591 .get_min = core_get_min_pstate,
592 .get_turbo = core_get_turbo_pstate,
593 .get_scaling = core_get_scaling,
594 .set = core_set_pstate,
598 static struct cpu_defaults byt_params = {
600 .sample_rate_ms = 10,
608 .get_max = byt_get_max_pstate,
609 .get_min = byt_get_min_pstate,
610 .get_turbo = byt_get_turbo_pstate,
611 .set = byt_set_pstate,
612 .get_scaling = byt_get_scaling,
613 .get_vid = byt_get_vid,
617 static void intel_pstate_get_min_max(struct cpudata *cpu, int *min, int *max)
619 int max_perf = cpu->pstate.turbo_pstate;
623 if (limits.no_turbo || limits.turbo_disabled)
624 max_perf = cpu->pstate.max_pstate;
627 * performance can be limited by user through sysfs, by cpufreq
628 * policy, or by cpu specific default values determined through
631 max_perf_adj = fp_toint(mul_fp(int_tofp(max_perf), limits.max_perf));
632 *max = clamp_t(int, max_perf_adj,
633 cpu->pstate.min_pstate, cpu->pstate.turbo_pstate);
635 min_perf = fp_toint(mul_fp(int_tofp(max_perf), limits.min_perf));
636 *min = clamp_t(int, min_perf, cpu->pstate.min_pstate, max_perf);
639 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
641 int max_perf, min_perf;
643 update_turbo_state();
645 intel_pstate_get_min_max(cpu, &min_perf, &max_perf);
647 pstate = clamp_t(int, pstate, min_perf, max_perf);
649 if (pstate == cpu->pstate.current_pstate)
652 trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
654 cpu->pstate.current_pstate = pstate;
656 pstate_funcs.set(cpu, pstate);
659 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
661 cpu->pstate.min_pstate = pstate_funcs.get_min();
662 cpu->pstate.max_pstate = pstate_funcs.get_max();
663 cpu->pstate.turbo_pstate = pstate_funcs.get_turbo();
664 cpu->pstate.scaling = pstate_funcs.get_scaling();
666 if (pstate_funcs.get_vid)
667 pstate_funcs.get_vid(cpu);
668 intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
671 static inline void intel_pstate_calc_busy(struct cpudata *cpu)
673 struct sample *sample = &cpu->sample;
676 core_pct = int_tofp(sample->aperf) * int_tofp(100);
677 core_pct = div64_u64(core_pct, int_tofp(sample->mperf));
679 sample->freq = fp_toint(
681 cpu->pstate.max_pstate * cpu->pstate.scaling / 100),
684 sample->core_pct_busy = (int32_t)core_pct;
687 static inline void intel_pstate_sample(struct cpudata *cpu)
692 local_irq_save(flags);
693 rdmsrl(MSR_IA32_APERF, aperf);
694 rdmsrl(MSR_IA32_MPERF, mperf);
695 local_irq_restore(flags);
697 cpu->last_sample_time = cpu->sample.time;
698 cpu->sample.time = ktime_get();
699 cpu->sample.aperf = aperf;
700 cpu->sample.mperf = mperf;
701 cpu->sample.aperf -= cpu->prev_aperf;
702 cpu->sample.mperf -= cpu->prev_mperf;
704 intel_pstate_calc_busy(cpu);
706 cpu->prev_aperf = aperf;
707 cpu->prev_mperf = mperf;
710 static inline void intel_hwp_set_sample_time(struct cpudata *cpu)
714 delay = msecs_to_jiffies(50);
715 mod_timer_pinned(&cpu->timer, jiffies + delay);
718 static inline void intel_pstate_set_sample_time(struct cpudata *cpu)
722 delay = msecs_to_jiffies(pid_params.sample_rate_ms);
723 mod_timer_pinned(&cpu->timer, jiffies + delay);
726 static inline int32_t intel_pstate_get_scaled_busy(struct cpudata *cpu)
728 int32_t core_busy, max_pstate, current_pstate, sample_ratio;
733 * core_busy is the ratio of actual performance to max
734 * max_pstate is the max non turbo pstate available
735 * current_pstate was the pstate that was requested during
736 * the last sample period.
738 * We normalize core_busy, which was our actual percent
739 * performance to what we requested during the last sample
740 * period. The result will be a percentage of busy at a
743 core_busy = cpu->sample.core_pct_busy;
744 max_pstate = int_tofp(cpu->pstate.max_pstate);
745 current_pstate = int_tofp(cpu->pstate.current_pstate);
746 core_busy = mul_fp(core_busy, div_fp(max_pstate, current_pstate));
749 * Since we have a deferred timer, it will not fire unless
750 * we are in C0. So, determine if the actual elapsed time
751 * is significantly greater (3x) than our sample interval. If it
752 * is, then we were idle for a long enough period of time
753 * to adjust our busyness.
755 sample_time = pid_params.sample_rate_ms * USEC_PER_MSEC;
756 duration_us = (u32) ktime_us_delta(cpu->sample.time,
757 cpu->last_sample_time);
758 if (duration_us > sample_time * 3) {
759 sample_ratio = div_fp(int_tofp(sample_time),
760 int_tofp(duration_us));
761 core_busy = mul_fp(core_busy, sample_ratio);
767 static inline void intel_pstate_adjust_busy_pstate(struct cpudata *cpu)
774 busy_scaled = intel_pstate_get_scaled_busy(cpu);
776 ctl = pid_calc(pid, busy_scaled);
778 /* Negative values of ctl increase the pstate and vice versa */
779 intel_pstate_set_pstate(cpu, cpu->pstate.current_pstate - ctl);
782 static void intel_hwp_timer_func(unsigned long __data)
784 struct cpudata *cpu = (struct cpudata *) __data;
786 intel_pstate_sample(cpu);
787 intel_hwp_set_sample_time(cpu);
790 static void intel_pstate_timer_func(unsigned long __data)
792 struct cpudata *cpu = (struct cpudata *) __data;
793 struct sample *sample;
795 intel_pstate_sample(cpu);
797 sample = &cpu->sample;
799 intel_pstate_adjust_busy_pstate(cpu);
801 trace_pstate_sample(fp_toint(sample->core_pct_busy),
802 fp_toint(intel_pstate_get_scaled_busy(cpu)),
803 cpu->pstate.current_pstate,
808 intel_pstate_set_sample_time(cpu);
811 #define ICPU(model, policy) \
812 { X86_VENDOR_INTEL, 6, model, X86_FEATURE_APERFMPERF,\
813 (unsigned long)&policy }
815 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
816 ICPU(0x2a, core_params),
817 ICPU(0x2d, core_params),
818 ICPU(0x37, byt_params),
819 ICPU(0x3a, core_params),
820 ICPU(0x3c, core_params),
821 ICPU(0x3d, core_params),
822 ICPU(0x3e, core_params),
823 ICPU(0x3f, core_params),
824 ICPU(0x45, core_params),
825 ICPU(0x46, core_params),
826 ICPU(0x47, core_params),
827 ICPU(0x4c, byt_params),
828 ICPU(0x4e, core_params),
829 ICPU(0x4f, core_params),
830 ICPU(0x56, core_params),
833 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
835 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] = {
836 ICPU(0x56, core_params),
840 static int intel_pstate_init_cpu(unsigned int cpunum)
844 if (!all_cpu_data[cpunum])
845 all_cpu_data[cpunum] = kzalloc(sizeof(struct cpudata),
847 if (!all_cpu_data[cpunum])
850 cpu = all_cpu_data[cpunum];
853 intel_pstate_get_cpu_pstates(cpu);
855 init_timer_deferrable(&cpu->timer);
856 cpu->timer.data = (unsigned long)cpu;
857 cpu->timer.expires = jiffies + HZ/100;
860 cpu->timer.function = intel_pstate_timer_func;
862 cpu->timer.function = intel_hwp_timer_func;
864 intel_pstate_busy_pid_reset(cpu);
865 intel_pstate_sample(cpu);
867 add_timer_on(&cpu->timer, cpunum);
869 pr_debug("Intel pstate controlling: cpu %d\n", cpunum);
874 static unsigned int intel_pstate_get(unsigned int cpu_num)
876 struct sample *sample;
879 cpu = all_cpu_data[cpu_num];
882 sample = &cpu->sample;
886 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
888 if (!policy->cpuinfo.max_freq)
891 if (policy->policy == CPUFREQ_POLICY_PERFORMANCE) {
892 limits.min_perf_pct = 100;
893 limits.min_perf = int_tofp(1);
894 limits.max_policy_pct = 100;
895 limits.max_perf_pct = 100;
896 limits.max_perf = int_tofp(1);
901 limits.min_perf_pct = (policy->min * 100) / policy->cpuinfo.max_freq;
902 limits.min_perf_pct = clamp_t(int, limits.min_perf_pct, 0 , 100);
903 limits.min_perf = div_fp(int_tofp(limits.min_perf_pct), int_tofp(100));
905 limits.max_policy_pct = (policy->max * 100) / policy->cpuinfo.max_freq;
906 limits.max_policy_pct = clamp_t(int, limits.max_policy_pct, 0 , 100);
907 limits.max_perf_pct = min(limits.max_policy_pct, limits.max_sysfs_pct);
908 limits.max_perf = div_fp(int_tofp(limits.max_perf_pct), int_tofp(100));
911 intel_pstate_hwp_set();
916 static int intel_pstate_verify_policy(struct cpufreq_policy *policy)
918 cpufreq_verify_within_cpu_limits(policy);
920 if (policy->policy != CPUFREQ_POLICY_POWERSAVE &&
921 policy->policy != CPUFREQ_POLICY_PERFORMANCE)
927 static void intel_pstate_stop_cpu(struct cpufreq_policy *policy)
929 int cpu_num = policy->cpu;
930 struct cpudata *cpu = all_cpu_data[cpu_num];
932 pr_info("intel_pstate CPU %d exiting\n", cpu_num);
934 del_timer_sync(&all_cpu_data[cpu_num]->timer);
938 intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
941 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
946 rc = intel_pstate_init_cpu(policy->cpu);
950 cpu = all_cpu_data[policy->cpu];
952 if (limits.min_perf_pct == 100 && limits.max_perf_pct == 100)
953 policy->policy = CPUFREQ_POLICY_PERFORMANCE;
955 policy->policy = CPUFREQ_POLICY_POWERSAVE;
957 policy->min = cpu->pstate.min_pstate * cpu->pstate.scaling;
958 policy->max = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
960 /* cpuinfo and default policy values */
961 policy->cpuinfo.min_freq = cpu->pstate.min_pstate * cpu->pstate.scaling;
962 policy->cpuinfo.max_freq =
963 cpu->pstate.turbo_pstate * cpu->pstate.scaling;
964 policy->cpuinfo.transition_latency = CPUFREQ_ETERNAL;
965 cpumask_set_cpu(policy->cpu, policy->cpus);
970 static struct cpufreq_driver intel_pstate_driver = {
971 .flags = CPUFREQ_CONST_LOOPS,
972 .verify = intel_pstate_verify_policy,
973 .setpolicy = intel_pstate_set_policy,
974 .get = intel_pstate_get,
975 .init = intel_pstate_cpu_init,
976 .stop_cpu = intel_pstate_stop_cpu,
977 .name = "intel_pstate",
980 static int __initdata no_load;
981 static int __initdata no_hwp;
982 static unsigned int force_load;
984 static int intel_pstate_msrs_not_valid(void)
986 /* Check that all the msr's we are using are valid. */
987 u64 aperf, mperf, tmp;
989 rdmsrl(MSR_IA32_APERF, aperf);
990 rdmsrl(MSR_IA32_MPERF, mperf);
992 if (!pstate_funcs.get_max() ||
993 !pstate_funcs.get_min() ||
994 !pstate_funcs.get_turbo())
997 rdmsrl(MSR_IA32_APERF, tmp);
1001 rdmsrl(MSR_IA32_MPERF, tmp);
1008 static void copy_pid_params(struct pstate_adjust_policy *policy)
1010 pid_params.sample_rate_ms = policy->sample_rate_ms;
1011 pid_params.p_gain_pct = policy->p_gain_pct;
1012 pid_params.i_gain_pct = policy->i_gain_pct;
1013 pid_params.d_gain_pct = policy->d_gain_pct;
1014 pid_params.deadband = policy->deadband;
1015 pid_params.setpoint = policy->setpoint;
1018 static void copy_cpu_funcs(struct pstate_funcs *funcs)
1020 pstate_funcs.get_max = funcs->get_max;
1021 pstate_funcs.get_min = funcs->get_min;
1022 pstate_funcs.get_turbo = funcs->get_turbo;
1023 pstate_funcs.get_scaling = funcs->get_scaling;
1024 pstate_funcs.set = funcs->set;
1025 pstate_funcs.get_vid = funcs->get_vid;
1028 #if IS_ENABLED(CONFIG_ACPI)
1029 #include <acpi/processor.h>
1031 static bool intel_pstate_no_acpi_pss(void)
1035 for_each_possible_cpu(i) {
1037 union acpi_object *pss;
1038 struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
1039 struct acpi_processor *pr = per_cpu(processors, i);
1044 status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
1045 if (ACPI_FAILURE(status))
1048 pss = buffer.pointer;
1049 if (pss && pss->type == ACPI_TYPE_PACKAGE) {
1060 static bool intel_pstate_has_acpi_ppc(void)
1064 for_each_possible_cpu(i) {
1065 struct acpi_processor *pr = per_cpu(processors, i);
1069 if (acpi_has_method(pr->handle, "_PPC"))
1080 struct hw_vendor_info {
1082 char oem_id[ACPI_OEM_ID_SIZE];
1083 char oem_table_id[ACPI_OEM_TABLE_ID_SIZE];
1087 /* Hardware vendor-specific info that has its own power management modes */
1088 static struct hw_vendor_info vendor_info[] = {
1089 {1, "HP ", "ProLiant", PSS},
1090 {1, "ORACLE", "X4-2 ", PPC},
1091 {1, "ORACLE", "X4-2L ", PPC},
1092 {1, "ORACLE", "X4-2B ", PPC},
1093 {1, "ORACLE", "X3-2 ", PPC},
1094 {1, "ORACLE", "X3-2L ", PPC},
1095 {1, "ORACLE", "X3-2B ", PPC},
1096 {1, "ORACLE", "X4470M2 ", PPC},
1097 {1, "ORACLE", "X4270M3 ", PPC},
1098 {1, "ORACLE", "X4270M2 ", PPC},
1099 {1, "ORACLE", "X4170M2 ", PPC},
1103 static bool intel_pstate_platform_pwr_mgmt_exists(void)
1105 struct acpi_table_header hdr;
1106 struct hw_vendor_info *v_info;
1107 const struct x86_cpu_id *id;
1110 id = x86_match_cpu(intel_pstate_cpu_oob_ids);
1112 rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
1113 if ( misc_pwr & (1 << 8))
1117 if (acpi_disabled ||
1118 ACPI_FAILURE(acpi_get_table_header(ACPI_SIG_FADT, 0, &hdr)))
1121 for (v_info = vendor_info; v_info->valid; v_info++) {
1122 if (!strncmp(hdr.oem_id, v_info->oem_id, ACPI_OEM_ID_SIZE) &&
1123 !strncmp(hdr.oem_table_id, v_info->oem_table_id,
1124 ACPI_OEM_TABLE_ID_SIZE))
1125 switch (v_info->oem_pwr_table) {
1127 return intel_pstate_no_acpi_pss();
1129 return intel_pstate_has_acpi_ppc() &&
1136 #else /* CONFIG_ACPI not enabled */
1137 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
1138 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
1139 #endif /* CONFIG_ACPI */
1141 static int __init intel_pstate_init(void)
1144 const struct x86_cpu_id *id;
1145 struct cpu_defaults *cpu_info;
1146 struct cpuinfo_x86 *c = &boot_cpu_data;
1151 id = x86_match_cpu(intel_pstate_cpu_ids);
1156 * The Intel pstate driver will be ignored if the platform
1157 * firmware has its own power management modes.
1159 if (intel_pstate_platform_pwr_mgmt_exists())
1162 cpu_info = (struct cpu_defaults *)id->driver_data;
1164 copy_pid_params(&cpu_info->pid_policy);
1165 copy_cpu_funcs(&cpu_info->funcs);
1167 if (intel_pstate_msrs_not_valid())
1170 pr_info("Intel P-state driver initializing.\n");
1172 all_cpu_data = vzalloc(sizeof(void *) * num_possible_cpus());
1176 if (cpu_has(c,X86_FEATURE_HWP) && !no_hwp)
1177 intel_pstate_hwp_enable();
1179 rc = cpufreq_register_driver(&intel_pstate_driver);
1183 intel_pstate_debug_expose_params();
1184 intel_pstate_sysfs_expose_params();
1189 for_each_online_cpu(cpu) {
1190 if (all_cpu_data[cpu]) {
1191 del_timer_sync(&all_cpu_data[cpu]->timer);
1192 kfree(all_cpu_data[cpu]);
1197 vfree(all_cpu_data);
1200 device_initcall(intel_pstate_init);
1202 static int __init intel_pstate_setup(char *str)
1207 if (!strcmp(str, "disable"))
1209 if (!strcmp(str, "no_hwp"))
1211 if (!strcmp(str, "force"))
1215 early_param("intel_pstate", intel_pstate_setup);
1217 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
1218 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
1219 MODULE_LICENSE("GPL");