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 <trace/events/power.h>
30 #include <asm/div64.h>
32 #include <asm/cpu_device_id.h>
34 #define SAMPLE_COUNT 3
36 #define BYT_RATIOS 0x66a
39 #define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
40 #define fp_toint(X) ((X) >> FRAC_BITS)
42 static inline int32_t mul_fp(int32_t x, int32_t y)
44 return ((int64_t)x * (int64_t)y) >> FRAC_BITS;
47 static inline int32_t div_fp(int32_t x, int32_t y)
49 return div_s64((int64_t)x << FRAC_BITS, (int64_t)y);
53 int32_t core_pct_busy;
81 struct timer_list timer;
83 struct pstate_data pstate;
91 struct sample samples[SAMPLE_COUNT];
94 static struct cpudata **all_cpu_data;
95 struct pstate_adjust_policy {
104 struct pstate_funcs {
105 int (*get_max)(void);
106 int (*get_min)(void);
107 int (*get_turbo)(void);
108 void (*set)(int pstate);
111 struct cpu_defaults {
112 struct pstate_adjust_policy pid_policy;
113 struct pstate_funcs funcs;
116 static struct pstate_adjust_policy pid_params;
117 static struct pstate_funcs pstate_funcs;
129 static struct perf_limits limits = {
132 .max_perf = int_tofp(1),
135 .max_policy_pct = 100,
136 .max_sysfs_pct = 100,
139 static inline void pid_reset(struct _pid *pid, int setpoint, int busy,
140 int deadband, int integral) {
141 pid->setpoint = setpoint;
142 pid->deadband = deadband;
143 pid->integral = int_tofp(integral);
144 pid->last_err = setpoint - busy;
147 static inline void pid_p_gain_set(struct _pid *pid, int percent)
149 pid->p_gain = div_fp(int_tofp(percent), int_tofp(100));
152 static inline void pid_i_gain_set(struct _pid *pid, int percent)
154 pid->i_gain = div_fp(int_tofp(percent), int_tofp(100));
157 static inline void pid_d_gain_set(struct _pid *pid, int percent)
160 pid->d_gain = div_fp(int_tofp(percent), int_tofp(100));
163 static signed int pid_calc(struct _pid *pid, int32_t busy)
166 int32_t pterm, dterm, fp_error;
167 int32_t integral_limit;
169 fp_error = int_tofp(pid->setpoint) - busy;
171 if (abs(fp_error) <= int_tofp(pid->deadband))
174 pterm = mul_fp(pid->p_gain, fp_error);
176 pid->integral += fp_error;
178 /* limit the integral term */
179 integral_limit = int_tofp(30);
180 if (pid->integral > integral_limit)
181 pid->integral = integral_limit;
182 if (pid->integral < -integral_limit)
183 pid->integral = -integral_limit;
185 dterm = mul_fp(pid->d_gain, fp_error - pid->last_err);
186 pid->last_err = fp_error;
188 result = pterm + mul_fp(pid->integral, pid->i_gain) + dterm;
190 return (signed int)fp_toint(result);
193 static inline void intel_pstate_busy_pid_reset(struct cpudata *cpu)
195 pid_p_gain_set(&cpu->pid, pid_params.p_gain_pct);
196 pid_d_gain_set(&cpu->pid, pid_params.d_gain_pct);
197 pid_i_gain_set(&cpu->pid, pid_params.i_gain_pct);
206 static inline void intel_pstate_reset_all_pid(void)
209 for_each_online_cpu(cpu) {
210 if (all_cpu_data[cpu])
211 intel_pstate_busy_pid_reset(all_cpu_data[cpu]);
215 /************************** debugfs begin ************************/
216 static int pid_param_set(void *data, u64 val)
219 intel_pstate_reset_all_pid();
222 static int pid_param_get(void *data, u64 *val)
227 DEFINE_SIMPLE_ATTRIBUTE(fops_pid_param, pid_param_get,
228 pid_param_set, "%llu\n");
235 static struct pid_param pid_files[] = {
236 {"sample_rate_ms", &pid_params.sample_rate_ms},
237 {"d_gain_pct", &pid_params.d_gain_pct},
238 {"i_gain_pct", &pid_params.i_gain_pct},
239 {"deadband", &pid_params.deadband},
240 {"setpoint", &pid_params.setpoint},
241 {"p_gain_pct", &pid_params.p_gain_pct},
245 static struct dentry *debugfs_parent;
246 static void intel_pstate_debug_expose_params(void)
250 debugfs_parent = debugfs_create_dir("pstate_snb", NULL);
251 if (IS_ERR_OR_NULL(debugfs_parent))
253 while (pid_files[i].name) {
254 debugfs_create_file(pid_files[i].name, 0660,
255 debugfs_parent, pid_files[i].value,
261 /************************** debugfs end ************************/
263 /************************** sysfs begin ************************/
264 #define show_one(file_name, object) \
265 static ssize_t show_##file_name \
266 (struct kobject *kobj, struct attribute *attr, char *buf) \
268 return sprintf(buf, "%u\n", limits.object); \
271 static ssize_t store_no_turbo(struct kobject *a, struct attribute *b,
272 const char *buf, size_t count)
276 ret = sscanf(buf, "%u", &input);
279 limits.no_turbo = clamp_t(int, input, 0 , 1);
284 static ssize_t store_max_perf_pct(struct kobject *a, struct attribute *b,
285 const char *buf, size_t count)
289 ret = sscanf(buf, "%u", &input);
293 limits.max_sysfs_pct = clamp_t(int, input, 0 , 100);
294 limits.max_perf_pct = min(limits.max_policy_pct, limits.max_sysfs_pct);
295 limits.max_perf = div_fp(int_tofp(limits.max_perf_pct), int_tofp(100));
299 static ssize_t store_min_perf_pct(struct kobject *a, struct attribute *b,
300 const char *buf, size_t count)
304 ret = sscanf(buf, "%u", &input);
307 limits.min_perf_pct = clamp_t(int, input, 0 , 100);
308 limits.min_perf = div_fp(int_tofp(limits.min_perf_pct), int_tofp(100));
313 show_one(no_turbo, no_turbo);
314 show_one(max_perf_pct, max_perf_pct);
315 show_one(min_perf_pct, min_perf_pct);
317 define_one_global_rw(no_turbo);
318 define_one_global_rw(max_perf_pct);
319 define_one_global_rw(min_perf_pct);
321 static struct attribute *intel_pstate_attributes[] = {
328 static struct attribute_group intel_pstate_attr_group = {
329 .attrs = intel_pstate_attributes,
331 static struct kobject *intel_pstate_kobject;
333 static void intel_pstate_sysfs_expose_params(void)
337 intel_pstate_kobject = kobject_create_and_add("intel_pstate",
338 &cpu_subsys.dev_root->kobj);
339 BUG_ON(!intel_pstate_kobject);
340 rc = sysfs_create_group(intel_pstate_kobject,
341 &intel_pstate_attr_group);
345 /************************** sysfs end ************************/
346 static int byt_get_min_pstate(void)
349 rdmsrl(BYT_RATIOS, value);
353 static int byt_get_max_pstate(void)
356 rdmsrl(BYT_RATIOS, value);
357 return (value >> 16) & 0xFF;
360 static int core_get_min_pstate(void)
363 rdmsrl(MSR_PLATFORM_INFO, value);
364 return (value >> 40) & 0xFF;
367 static int core_get_max_pstate(void)
370 rdmsrl(MSR_PLATFORM_INFO, value);
371 return (value >> 8) & 0xFF;
374 static int core_get_turbo_pstate(void)
378 rdmsrl(MSR_NHM_TURBO_RATIO_LIMIT, value);
379 nont = core_get_max_pstate();
380 ret = ((value) & 255);
386 static void core_set_pstate(int pstate)
394 wrmsrl(MSR_IA32_PERF_CTL, val);
397 static struct cpu_defaults core_params = {
399 .sample_rate_ms = 10,
407 .get_max = core_get_max_pstate,
408 .get_min = core_get_min_pstate,
409 .get_turbo = core_get_turbo_pstate,
410 .set = core_set_pstate,
414 static struct cpu_defaults byt_params = {
416 .sample_rate_ms = 10,
424 .get_max = byt_get_max_pstate,
425 .get_min = byt_get_min_pstate,
426 .get_turbo = byt_get_max_pstate,
427 .set = core_set_pstate,
432 static void intel_pstate_get_min_max(struct cpudata *cpu, int *min, int *max)
434 int max_perf = cpu->pstate.turbo_pstate;
438 max_perf = cpu->pstate.max_pstate;
440 max_perf_adj = fp_toint(mul_fp(int_tofp(max_perf), limits.max_perf));
441 *max = clamp_t(int, max_perf_adj,
442 cpu->pstate.min_pstate, cpu->pstate.turbo_pstate);
444 min_perf = fp_toint(mul_fp(int_tofp(max_perf), limits.min_perf));
445 *min = clamp_t(int, min_perf,
446 cpu->pstate.min_pstate, max_perf);
449 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
451 int max_perf, min_perf;
453 intel_pstate_get_min_max(cpu, &min_perf, &max_perf);
455 pstate = clamp_t(int, pstate, min_perf, max_perf);
457 if (pstate == cpu->pstate.current_pstate)
460 trace_cpu_frequency(pstate * 100000, cpu->cpu);
462 cpu->pstate.current_pstate = pstate;
464 pstate_funcs.set(pstate);
467 static inline void intel_pstate_pstate_increase(struct cpudata *cpu, int steps)
470 target = cpu->pstate.current_pstate + steps;
472 intel_pstate_set_pstate(cpu, target);
475 static inline void intel_pstate_pstate_decrease(struct cpudata *cpu, int steps)
478 target = cpu->pstate.current_pstate - steps;
479 intel_pstate_set_pstate(cpu, target);
482 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
484 sprintf(cpu->name, "Intel 2nd generation core");
486 cpu->pstate.min_pstate = pstate_funcs.get_min();
487 cpu->pstate.max_pstate = pstate_funcs.get_max();
488 cpu->pstate.turbo_pstate = pstate_funcs.get_turbo();
491 * goto max pstate so we don't slow up boot if we are built-in if we are
492 * a module we will take care of it during normal operation
494 intel_pstate_set_pstate(cpu, cpu->pstate.max_pstate);
497 static inline void intel_pstate_calc_busy(struct cpudata *cpu,
498 struct sample *sample)
501 core_pct = div64_u64(int_tofp(sample->aperf * 100),
503 sample->freq = fp_toint(cpu->pstate.max_pstate * core_pct * 1000);
505 sample->core_pct_busy = core_pct;
508 static inline void intel_pstate_sample(struct cpudata *cpu)
512 rdmsrl(MSR_IA32_APERF, aperf);
513 rdmsrl(MSR_IA32_MPERF, mperf);
514 cpu->sample_ptr = (cpu->sample_ptr + 1) % SAMPLE_COUNT;
515 cpu->samples[cpu->sample_ptr].aperf = aperf;
516 cpu->samples[cpu->sample_ptr].mperf = mperf;
517 cpu->samples[cpu->sample_ptr].aperf -= cpu->prev_aperf;
518 cpu->samples[cpu->sample_ptr].mperf -= cpu->prev_mperf;
520 intel_pstate_calc_busy(cpu, &cpu->samples[cpu->sample_ptr]);
522 cpu->prev_aperf = aperf;
523 cpu->prev_mperf = mperf;
526 static inline void intel_pstate_set_sample_time(struct cpudata *cpu)
528 int sample_time, delay;
530 sample_time = pid_params.sample_rate_ms;
531 delay = msecs_to_jiffies(sample_time);
532 mod_timer_pinned(&cpu->timer, jiffies + delay);
535 static inline int32_t intel_pstate_get_scaled_busy(struct cpudata *cpu)
537 int32_t core_busy, max_pstate, current_pstate;
539 core_busy = cpu->samples[cpu->sample_ptr].core_pct_busy;
540 max_pstate = int_tofp(cpu->pstate.max_pstate);
541 current_pstate = int_tofp(cpu->pstate.current_pstate);
542 return mul_fp(core_busy, div_fp(max_pstate, current_pstate));
545 static inline void intel_pstate_adjust_busy_pstate(struct cpudata *cpu)
553 busy_scaled = intel_pstate_get_scaled_busy(cpu);
555 ctl = pid_calc(pid, busy_scaled);
559 intel_pstate_pstate_increase(cpu, steps);
561 intel_pstate_pstate_decrease(cpu, steps);
564 static void intel_pstate_timer_func(unsigned long __data)
566 struct cpudata *cpu = (struct cpudata *) __data;
568 intel_pstate_sample(cpu);
569 intel_pstate_adjust_busy_pstate(cpu);
571 if (cpu->pstate.current_pstate == cpu->pstate.min_pstate) {
572 cpu->min_pstate_count++;
573 if (!(cpu->min_pstate_count % 5)) {
574 intel_pstate_set_pstate(cpu, cpu->pstate.max_pstate);
577 cpu->min_pstate_count = 0;
579 intel_pstate_set_sample_time(cpu);
582 #define ICPU(model, policy) \
583 { X86_VENDOR_INTEL, 6, model, X86_FEATURE_ANY, (unsigned long)&policy }
585 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
586 ICPU(0x2a, core_params),
587 ICPU(0x2d, core_params),
588 ICPU(0x37, byt_params),
589 ICPU(0x3a, core_params),
590 ICPU(0x3c, core_params),
591 ICPU(0x3e, core_params),
592 ICPU(0x3f, core_params),
593 ICPU(0x45, core_params),
594 ICPU(0x46, core_params),
597 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
599 static int intel_pstate_init_cpu(unsigned int cpunum)
602 const struct x86_cpu_id *id;
605 id = x86_match_cpu(intel_pstate_cpu_ids);
609 all_cpu_data[cpunum] = kzalloc(sizeof(struct cpudata), GFP_KERNEL);
610 if (!all_cpu_data[cpunum])
613 cpu = all_cpu_data[cpunum];
615 intel_pstate_get_cpu_pstates(cpu);
619 init_timer_deferrable(&cpu->timer);
620 cpu->timer.function = intel_pstate_timer_func;
623 cpu->timer.expires = jiffies + HZ/100;
624 intel_pstate_busy_pid_reset(cpu);
625 intel_pstate_sample(cpu);
626 intel_pstate_set_pstate(cpu, cpu->pstate.max_pstate);
628 add_timer_on(&cpu->timer, cpunum);
630 pr_info("Intel pstate controlling: cpu %d\n", cpunum);
635 static unsigned int intel_pstate_get(unsigned int cpu_num)
637 struct sample *sample;
640 cpu = all_cpu_data[cpu_num];
643 sample = &cpu->samples[cpu->sample_ptr];
647 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
651 cpu = all_cpu_data[policy->cpu];
653 if (!policy->cpuinfo.max_freq)
656 if (policy->policy == CPUFREQ_POLICY_PERFORMANCE) {
657 limits.min_perf_pct = 100;
658 limits.min_perf = int_tofp(1);
659 limits.max_perf_pct = 100;
660 limits.max_perf = int_tofp(1);
664 limits.min_perf_pct = (policy->min * 100) / policy->cpuinfo.max_freq;
665 limits.min_perf_pct = clamp_t(int, limits.min_perf_pct, 0 , 100);
666 limits.min_perf = div_fp(int_tofp(limits.min_perf_pct), int_tofp(100));
668 limits.max_policy_pct = policy->max * 100 / policy->cpuinfo.max_freq;
669 limits.max_policy_pct = clamp_t(int, limits.max_policy_pct, 0 , 100);
670 limits.max_perf_pct = min(limits.max_policy_pct, limits.max_sysfs_pct);
671 limits.max_perf = div_fp(int_tofp(limits.max_perf_pct), int_tofp(100));
676 static int intel_pstate_verify_policy(struct cpufreq_policy *policy)
678 cpufreq_verify_within_cpu_limits(policy);
680 if ((policy->policy != CPUFREQ_POLICY_POWERSAVE) &&
681 (policy->policy != CPUFREQ_POLICY_PERFORMANCE))
687 static int intel_pstate_cpu_exit(struct cpufreq_policy *policy)
689 int cpu = policy->cpu;
691 del_timer(&all_cpu_data[cpu]->timer);
692 kfree(all_cpu_data[cpu]);
693 all_cpu_data[cpu] = NULL;
697 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
702 rc = intel_pstate_init_cpu(policy->cpu);
706 cpu = all_cpu_data[policy->cpu];
708 if (!limits.no_turbo &&
709 limits.min_perf_pct == 100 && limits.max_perf_pct == 100)
710 policy->policy = CPUFREQ_POLICY_PERFORMANCE;
712 policy->policy = CPUFREQ_POLICY_POWERSAVE;
714 policy->min = cpu->pstate.min_pstate * 100000;
715 policy->max = cpu->pstate.turbo_pstate * 100000;
717 /* cpuinfo and default policy values */
718 policy->cpuinfo.min_freq = cpu->pstate.min_pstate * 100000;
719 policy->cpuinfo.max_freq = cpu->pstate.turbo_pstate * 100000;
720 policy->cpuinfo.transition_latency = CPUFREQ_ETERNAL;
721 cpumask_set_cpu(policy->cpu, policy->cpus);
726 static struct cpufreq_driver intel_pstate_driver = {
727 .flags = CPUFREQ_CONST_LOOPS,
728 .verify = intel_pstate_verify_policy,
729 .setpolicy = intel_pstate_set_policy,
730 .get = intel_pstate_get,
731 .init = intel_pstate_cpu_init,
732 .exit = intel_pstate_cpu_exit,
733 .name = "intel_pstate",
736 static int __initdata no_load;
738 static int intel_pstate_msrs_not_valid(void)
740 /* Check that all the msr's we are using are valid. */
741 u64 aperf, mperf, tmp;
743 rdmsrl(MSR_IA32_APERF, aperf);
744 rdmsrl(MSR_IA32_MPERF, mperf);
746 if (!pstate_funcs.get_max() ||
747 !pstate_funcs.get_min() ||
748 !pstate_funcs.get_turbo())
751 rdmsrl(MSR_IA32_APERF, tmp);
755 rdmsrl(MSR_IA32_MPERF, tmp);
762 void copy_pid_params(struct pstate_adjust_policy *policy)
764 pid_params.sample_rate_ms = policy->sample_rate_ms;
765 pid_params.p_gain_pct = policy->p_gain_pct;
766 pid_params.i_gain_pct = policy->i_gain_pct;
767 pid_params.d_gain_pct = policy->d_gain_pct;
768 pid_params.deadband = policy->deadband;
769 pid_params.setpoint = policy->setpoint;
772 void copy_cpu_funcs(struct pstate_funcs *funcs)
774 pstate_funcs.get_max = funcs->get_max;
775 pstate_funcs.get_min = funcs->get_min;
776 pstate_funcs.get_turbo = funcs->get_turbo;
777 pstate_funcs.set = funcs->set;
780 static int __init intel_pstate_init(void)
783 const struct x86_cpu_id *id;
784 struct cpu_defaults *cpu_info;
789 id = x86_match_cpu(intel_pstate_cpu_ids);
793 cpu_info = (struct cpu_defaults *)id->driver_data;
795 copy_pid_params(&cpu_info->pid_policy);
796 copy_cpu_funcs(&cpu_info->funcs);
798 if (intel_pstate_msrs_not_valid())
801 pr_info("Intel P-state driver initializing.\n");
803 all_cpu_data = vzalloc(sizeof(void *) * num_possible_cpus());
807 rc = cpufreq_register_driver(&intel_pstate_driver);
811 intel_pstate_debug_expose_params();
812 intel_pstate_sysfs_expose_params();
816 for_each_online_cpu(cpu) {
817 if (all_cpu_data[cpu]) {
818 del_timer_sync(&all_cpu_data[cpu]->timer);
819 kfree(all_cpu_data[cpu]);
827 device_initcall(intel_pstate_init);
829 static int __init intel_pstate_setup(char *str)
834 if (!strcmp(str, "disable"))
838 early_param("intel_pstate", intel_pstate_setup);
840 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
841 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
842 MODULE_LICENSE("GPL");