2 * drivers/cpufreq/cpufreq_conservative.c
4 * Copyright (C) 2001 Russell King
5 * (C) 2003 Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>.
6 * Jun Nakajima <jun.nakajima@intel.com>
7 * (C) 2009 Alexander Clouter <alex@digriz.org.uk>
9 * This program is free software; you can redistribute it and/or modify
10 * it under the terms of the GNU General Public License version 2 as
11 * published by the Free Software Foundation.
14 #include <linux/kernel.h>
15 #include <linux/module.h>
16 #include <linux/init.h>
17 #include <linux/cpufreq.h>
18 #include <linux/cpu.h>
19 #include <linux/jiffies.h>
20 #include <linux/kernel_stat.h>
21 #include <linux/mutex.h>
22 #include <linux/hrtimer.h>
23 #include <linux/tick.h>
24 #include <linux/ktime.h>
25 #include <linux/sched.h>
28 * dbs is used in this file as a shortform for demandbased switching
29 * It helps to keep variable names smaller, simpler
32 #define DEF_FREQUENCY_UP_THRESHOLD (80)
33 #define DEF_FREQUENCY_DOWN_THRESHOLD (20)
36 * The polling frequency of this governor depends on the capability of
37 * the processor. Default polling frequency is 1000 times the transition
38 * latency of the processor. The governor will work on any processor with
39 * transition latency <= 10mS, using appropriate sampling
41 * For CPUs with transition latency > 10mS (mostly drivers with CPUFREQ_ETERNAL)
42 * this governor will not work.
43 * All times here are in uS.
45 #define MIN_SAMPLING_RATE_RATIO (2)
47 static unsigned int min_sampling_rate;
49 #define LATENCY_MULTIPLIER (1000)
50 #define MIN_LATENCY_MULTIPLIER (100)
51 #define DEF_SAMPLING_DOWN_FACTOR (1)
52 #define MAX_SAMPLING_DOWN_FACTOR (10)
53 #define TRANSITION_LATENCY_LIMIT (10 * 1000 * 1000)
55 static void do_dbs_timer(struct work_struct *work);
57 struct cpu_dbs_info_s {
58 cputime64_t prev_cpu_idle;
59 cputime64_t prev_cpu_wall;
60 cputime64_t prev_cpu_nice;
61 struct cpufreq_policy *cur_policy;
62 struct delayed_work work;
63 unsigned int down_skip;
64 unsigned int requested_freq;
66 unsigned int enable:1;
68 * percpu mutex that serializes governor limit change with
69 * do_dbs_timer invocation. We do not want do_dbs_timer to run
70 * when user is changing the governor or limits.
72 struct mutex timer_mutex;
74 static DEFINE_PER_CPU(struct cpu_dbs_info_s, cs_cpu_dbs_info);
76 static unsigned int dbs_enable; /* number of CPUs using this policy */
79 * dbs_mutex protects data in dbs_tuners_ins from concurrent changes on
80 * different CPUs. It protects dbs_enable in governor start/stop.
82 static DEFINE_MUTEX(dbs_mutex);
84 static struct dbs_tuners {
85 unsigned int sampling_rate;
86 unsigned int sampling_down_factor;
87 unsigned int up_threshold;
88 unsigned int down_threshold;
89 unsigned int ignore_nice;
90 unsigned int freq_step;
92 .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
93 .down_threshold = DEF_FREQUENCY_DOWN_THRESHOLD,
94 .sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR,
99 static inline cputime64_t get_cpu_idle_time_jiffy(unsigned int cpu,
102 cputime64_t idle_time;
103 cputime64_t cur_wall_time;
104 cputime64_t busy_time;
106 cur_wall_time = jiffies64_to_cputime64(get_jiffies_64());
107 busy_time = cputime64_add(kstat_cpu(cpu).cpustat.user,
108 kstat_cpu(cpu).cpustat.system);
110 busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.irq);
111 busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.softirq);
112 busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.steal);
113 busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.nice);
115 idle_time = cputime64_sub(cur_wall_time, busy_time);
117 *wall = (cputime64_t)jiffies_to_usecs(cur_wall_time);
119 return (cputime64_t)jiffies_to_usecs(idle_time);;
122 static inline cputime64_t get_cpu_idle_time(unsigned int cpu, cputime64_t *wall)
124 u64 idle_time = get_cpu_idle_time_us(cpu, wall);
126 if (idle_time == -1ULL)
127 return get_cpu_idle_time_jiffy(cpu, wall);
132 /* keep track of frequency transitions */
134 dbs_cpufreq_notifier(struct notifier_block *nb, unsigned long val,
137 struct cpufreq_freqs *freq = data;
138 struct cpu_dbs_info_s *this_dbs_info = &per_cpu(cs_cpu_dbs_info,
141 struct cpufreq_policy *policy;
143 if (!this_dbs_info->enable)
146 policy = this_dbs_info->cur_policy;
149 * we only care if our internally tracked freq moves outside
150 * the 'valid' ranges of freqency available to us otherwise
151 * we do not change it
153 if (this_dbs_info->requested_freq > policy->max
154 || this_dbs_info->requested_freq < policy->min)
155 this_dbs_info->requested_freq = freq->new;
160 static struct notifier_block dbs_cpufreq_notifier_block = {
161 .notifier_call = dbs_cpufreq_notifier
164 /************************** sysfs interface ************************/
165 static ssize_t show_sampling_rate_max(struct kobject *kobj,
166 struct attribute *attr, char *buf)
168 printk_once(KERN_INFO "CPUFREQ: conservative sampling_rate_max "
169 "sysfs file is deprecated - used by: %s\n", current->comm);
170 return sprintf(buf, "%u\n", -1U);
173 static ssize_t show_sampling_rate_min(struct kobject *kobj,
174 struct attribute *attr, char *buf)
176 return sprintf(buf, "%u\n", min_sampling_rate);
179 define_one_global_ro(sampling_rate_max);
180 define_one_global_ro(sampling_rate_min);
182 /* cpufreq_conservative Governor Tunables */
183 #define show_one(file_name, object) \
184 static ssize_t show_##file_name \
185 (struct kobject *kobj, struct attribute *attr, char *buf) \
187 return sprintf(buf, "%u\n", dbs_tuners_ins.object); \
189 show_one(sampling_rate, sampling_rate);
190 show_one(sampling_down_factor, sampling_down_factor);
191 show_one(up_threshold, up_threshold);
192 show_one(down_threshold, down_threshold);
193 show_one(ignore_nice_load, ignore_nice);
194 show_one(freq_step, freq_step);
196 /*** delete after deprecation time ***/
197 #define DEPRECATION_MSG(file_name) \
198 printk_once(KERN_INFO "CPUFREQ: Per core conservative sysfs " \
199 "interface is deprecated - " #file_name "\n");
201 #define show_one_old(file_name) \
202 static ssize_t show_##file_name##_old \
203 (struct cpufreq_policy *unused, char *buf) \
205 printk_once(KERN_INFO "CPUFREQ: Per core conservative sysfs " \
206 "interface is deprecated - " #file_name "\n"); \
207 return show_##file_name(NULL, NULL, buf); \
209 show_one_old(sampling_rate);
210 show_one_old(sampling_down_factor);
211 show_one_old(up_threshold);
212 show_one_old(down_threshold);
213 show_one_old(ignore_nice_load);
214 show_one_old(freq_step);
215 show_one_old(sampling_rate_min);
216 show_one_old(sampling_rate_max);
218 cpufreq_freq_attr_ro_old(sampling_rate_min);
219 cpufreq_freq_attr_ro_old(sampling_rate_max);
221 /*** delete after deprecation time ***/
223 static ssize_t store_sampling_down_factor(struct kobject *a,
225 const char *buf, size_t count)
229 ret = sscanf(buf, "%u", &input);
231 if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
234 mutex_lock(&dbs_mutex);
235 dbs_tuners_ins.sampling_down_factor = input;
236 mutex_unlock(&dbs_mutex);
241 static ssize_t store_sampling_rate(struct kobject *a, struct attribute *b,
242 const char *buf, size_t count)
246 ret = sscanf(buf, "%u", &input);
251 mutex_lock(&dbs_mutex);
252 dbs_tuners_ins.sampling_rate = max(input, min_sampling_rate);
253 mutex_unlock(&dbs_mutex);
258 static ssize_t store_up_threshold(struct kobject *a, struct attribute *b,
259 const char *buf, size_t count)
263 ret = sscanf(buf, "%u", &input);
265 mutex_lock(&dbs_mutex);
266 if (ret != 1 || input > 100 ||
267 input <= dbs_tuners_ins.down_threshold) {
268 mutex_unlock(&dbs_mutex);
272 dbs_tuners_ins.up_threshold = input;
273 mutex_unlock(&dbs_mutex);
278 static ssize_t store_down_threshold(struct kobject *a, struct attribute *b,
279 const char *buf, size_t count)
283 ret = sscanf(buf, "%u", &input);
285 mutex_lock(&dbs_mutex);
286 /* cannot be lower than 11 otherwise freq will not fall */
287 if (ret != 1 || input < 11 || input > 100 ||
288 input >= dbs_tuners_ins.up_threshold) {
289 mutex_unlock(&dbs_mutex);
293 dbs_tuners_ins.down_threshold = input;
294 mutex_unlock(&dbs_mutex);
299 static ssize_t store_ignore_nice_load(struct kobject *a, struct attribute *b,
300 const char *buf, size_t count)
307 ret = sscanf(buf, "%u", &input);
314 mutex_lock(&dbs_mutex);
315 if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */
316 mutex_unlock(&dbs_mutex);
319 dbs_tuners_ins.ignore_nice = input;
321 /* we need to re-evaluate prev_cpu_idle */
322 for_each_online_cpu(j) {
323 struct cpu_dbs_info_s *dbs_info;
324 dbs_info = &per_cpu(cs_cpu_dbs_info, j);
325 dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
326 &dbs_info->prev_cpu_wall);
327 if (dbs_tuners_ins.ignore_nice)
328 dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
330 mutex_unlock(&dbs_mutex);
335 static ssize_t store_freq_step(struct kobject *a, struct attribute *b,
336 const char *buf, size_t count)
340 ret = sscanf(buf, "%u", &input);
348 /* no need to test here if freq_step is zero as the user might actually
349 * want this, they would be crazy though :) */
350 mutex_lock(&dbs_mutex);
351 dbs_tuners_ins.freq_step = input;
352 mutex_unlock(&dbs_mutex);
357 define_one_global_rw(sampling_rate);
358 define_one_global_rw(sampling_down_factor);
359 define_one_global_rw(up_threshold);
360 define_one_global_rw(down_threshold);
361 define_one_global_rw(ignore_nice_load);
362 define_one_global_rw(freq_step);
364 static struct attribute *dbs_attributes[] = {
365 &sampling_rate_max.attr,
366 &sampling_rate_min.attr,
368 &sampling_down_factor.attr,
370 &down_threshold.attr,
371 &ignore_nice_load.attr,
376 static struct attribute_group dbs_attr_group = {
377 .attrs = dbs_attributes,
378 .name = "conservative",
381 /*** delete after deprecation time ***/
383 #define write_one_old(file_name) \
384 static ssize_t store_##file_name##_old \
385 (struct cpufreq_policy *unused, const char *buf, size_t count) \
387 printk_once(KERN_INFO "CPUFREQ: Per core conservative sysfs " \
388 "interface is deprecated - " #file_name "\n"); \
389 return store_##file_name(NULL, NULL, buf, count); \
391 write_one_old(sampling_rate);
392 write_one_old(sampling_down_factor);
393 write_one_old(up_threshold);
394 write_one_old(down_threshold);
395 write_one_old(ignore_nice_load);
396 write_one_old(freq_step);
398 cpufreq_freq_attr_rw_old(sampling_rate);
399 cpufreq_freq_attr_rw_old(sampling_down_factor);
400 cpufreq_freq_attr_rw_old(up_threshold);
401 cpufreq_freq_attr_rw_old(down_threshold);
402 cpufreq_freq_attr_rw_old(ignore_nice_load);
403 cpufreq_freq_attr_rw_old(freq_step);
405 static struct attribute *dbs_attributes_old[] = {
406 &sampling_rate_max_old.attr,
407 &sampling_rate_min_old.attr,
408 &sampling_rate_old.attr,
409 &sampling_down_factor_old.attr,
410 &up_threshold_old.attr,
411 &down_threshold_old.attr,
412 &ignore_nice_load_old.attr,
417 static struct attribute_group dbs_attr_group_old = {
418 .attrs = dbs_attributes_old,
419 .name = "conservative",
422 /*** delete after deprecation time ***/
424 /************************** sysfs end ************************/
426 static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
428 unsigned int load = 0;
429 unsigned int max_load = 0;
430 unsigned int freq_target;
432 struct cpufreq_policy *policy;
435 policy = this_dbs_info->cur_policy;
438 * Every sampling_rate, we check, if current idle time is less
439 * than 20% (default), then we try to increase frequency
440 * Every sampling_rate*sampling_down_factor, we check, if current
441 * idle time is more than 80%, then we try to decrease frequency
443 * Any frequency increase takes it to the maximum frequency.
444 * Frequency reduction happens at minimum steps of
445 * 5% (default) of maximum frequency
448 /* Get Absolute Load */
449 for_each_cpu(j, policy->cpus) {
450 struct cpu_dbs_info_s *j_dbs_info;
451 cputime64_t cur_wall_time, cur_idle_time;
452 unsigned int idle_time, wall_time;
454 j_dbs_info = &per_cpu(cs_cpu_dbs_info, j);
456 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time);
458 wall_time = (unsigned int) cputime64_sub(cur_wall_time,
459 j_dbs_info->prev_cpu_wall);
460 j_dbs_info->prev_cpu_wall = cur_wall_time;
462 idle_time = (unsigned int) cputime64_sub(cur_idle_time,
463 j_dbs_info->prev_cpu_idle);
464 j_dbs_info->prev_cpu_idle = cur_idle_time;
466 if (dbs_tuners_ins.ignore_nice) {
467 cputime64_t cur_nice;
468 unsigned long cur_nice_jiffies;
470 cur_nice = cputime64_sub(kstat_cpu(j).cpustat.nice,
471 j_dbs_info->prev_cpu_nice);
473 * Assumption: nice time between sampling periods will
474 * be less than 2^32 jiffies for 32 bit sys
476 cur_nice_jiffies = (unsigned long)
477 cputime64_to_jiffies64(cur_nice);
479 j_dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
480 idle_time += jiffies_to_usecs(cur_nice_jiffies);
483 if (unlikely(!wall_time || wall_time < idle_time))
486 load = 100 * (wall_time - idle_time) / wall_time;
493 * break out if we 'cannot' reduce the speed as the user might
494 * want freq_step to be zero
496 if (dbs_tuners_ins.freq_step == 0)
499 /* Check for frequency increase */
500 if (max_load > dbs_tuners_ins.up_threshold) {
501 this_dbs_info->down_skip = 0;
503 /* if we are already at full speed then break out early */
504 if (this_dbs_info->requested_freq == policy->max)
507 freq_target = (dbs_tuners_ins.freq_step * policy->max) / 100;
509 /* max freq cannot be less than 100. But who knows.... */
510 if (unlikely(freq_target == 0))
513 this_dbs_info->requested_freq += freq_target;
514 if (this_dbs_info->requested_freq > policy->max)
515 this_dbs_info->requested_freq = policy->max;
517 __cpufreq_driver_target(policy, this_dbs_info->requested_freq,
523 * The optimal frequency is the frequency that is the lowest that
524 * can support the current CPU usage without triggering the up
525 * policy. To be safe, we focus 10 points under the threshold.
527 if (max_load < (dbs_tuners_ins.down_threshold - 10)) {
528 freq_target = (dbs_tuners_ins.freq_step * policy->max) / 100;
530 this_dbs_info->requested_freq -= freq_target;
531 if (this_dbs_info->requested_freq < policy->min)
532 this_dbs_info->requested_freq = policy->min;
535 * if we cannot reduce the frequency anymore, break out early
537 if (policy->cur == policy->min)
540 __cpufreq_driver_target(policy, this_dbs_info->requested_freq,
546 static void do_dbs_timer(struct work_struct *work)
548 struct cpu_dbs_info_s *dbs_info =
549 container_of(work, struct cpu_dbs_info_s, work.work);
550 unsigned int cpu = dbs_info->cpu;
552 /* We want all CPUs to do sampling nearly on same jiffy */
553 int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
555 delay -= jiffies % delay;
557 mutex_lock(&dbs_info->timer_mutex);
559 dbs_check_cpu(dbs_info);
561 schedule_delayed_work_on(cpu, &dbs_info->work, delay);
562 mutex_unlock(&dbs_info->timer_mutex);
565 static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
567 /* We want all CPUs to do sampling nearly on same jiffy */
568 int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
569 delay -= jiffies % delay;
571 dbs_info->enable = 1;
572 INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
573 schedule_delayed_work_on(dbs_info->cpu, &dbs_info->work, delay);
576 static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
578 dbs_info->enable = 0;
579 cancel_delayed_work_sync(&dbs_info->work);
582 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
585 unsigned int cpu = policy->cpu;
586 struct cpu_dbs_info_s *this_dbs_info;
590 this_dbs_info = &per_cpu(cs_cpu_dbs_info, cpu);
593 case CPUFREQ_GOV_START:
594 if ((!cpu_online(cpu)) || (!policy->cur))
597 mutex_lock(&dbs_mutex);
599 rc = sysfs_create_group(&policy->kobj, &dbs_attr_group_old);
601 mutex_unlock(&dbs_mutex);
605 for_each_cpu(j, policy->cpus) {
606 struct cpu_dbs_info_s *j_dbs_info;
607 j_dbs_info = &per_cpu(cs_cpu_dbs_info, j);
608 j_dbs_info->cur_policy = policy;
610 j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
611 &j_dbs_info->prev_cpu_wall);
612 if (dbs_tuners_ins.ignore_nice) {
613 j_dbs_info->prev_cpu_nice =
614 kstat_cpu(j).cpustat.nice;
617 this_dbs_info->down_skip = 0;
618 this_dbs_info->requested_freq = policy->cur;
620 mutex_init(&this_dbs_info->timer_mutex);
623 * Start the timerschedule work, when this governor
624 * is used for first time
626 if (dbs_enable == 1) {
627 unsigned int latency;
628 /* policy latency is in nS. Convert it to uS first */
629 latency = policy->cpuinfo.transition_latency / 1000;
633 rc = sysfs_create_group(cpufreq_global_kobject,
636 mutex_unlock(&dbs_mutex);
641 * conservative does not implement micro like ondemand
642 * governor, thus we are bound to jiffes/HZ
645 MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10);
646 /* Bring kernel and HW constraints together */
647 min_sampling_rate = max(min_sampling_rate,
648 MIN_LATENCY_MULTIPLIER * latency);
649 dbs_tuners_ins.sampling_rate =
650 max(min_sampling_rate,
651 latency * LATENCY_MULTIPLIER);
653 cpufreq_register_notifier(
654 &dbs_cpufreq_notifier_block,
655 CPUFREQ_TRANSITION_NOTIFIER);
657 mutex_unlock(&dbs_mutex);
659 dbs_timer_init(this_dbs_info);
663 case CPUFREQ_GOV_STOP:
664 dbs_timer_exit(this_dbs_info);
666 mutex_lock(&dbs_mutex);
667 sysfs_remove_group(&policy->kobj, &dbs_attr_group_old);
669 mutex_destroy(&this_dbs_info->timer_mutex);
672 * Stop the timerschedule work, when this governor
673 * is used for first time
676 cpufreq_unregister_notifier(
677 &dbs_cpufreq_notifier_block,
678 CPUFREQ_TRANSITION_NOTIFIER);
680 mutex_unlock(&dbs_mutex);
682 sysfs_remove_group(cpufreq_global_kobject,
687 case CPUFREQ_GOV_LIMITS:
688 mutex_lock(&this_dbs_info->timer_mutex);
689 if (policy->max < this_dbs_info->cur_policy->cur)
690 __cpufreq_driver_target(
691 this_dbs_info->cur_policy,
692 policy->max, CPUFREQ_RELATION_H);
693 else if (policy->min > this_dbs_info->cur_policy->cur)
694 __cpufreq_driver_target(
695 this_dbs_info->cur_policy,
696 policy->min, CPUFREQ_RELATION_L);
697 mutex_unlock(&this_dbs_info->timer_mutex);
704 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_CONSERVATIVE
707 struct cpufreq_governor cpufreq_gov_conservative = {
708 .name = "conservative",
709 .governor = cpufreq_governor_dbs,
710 .max_transition_latency = TRANSITION_LATENCY_LIMIT,
711 .owner = THIS_MODULE,
714 static int __init cpufreq_gov_dbs_init(void)
716 return cpufreq_register_governor(&cpufreq_gov_conservative);
719 static void __exit cpufreq_gov_dbs_exit(void)
721 cpufreq_unregister_governor(&cpufreq_gov_conservative);
725 MODULE_AUTHOR("Alexander Clouter <alex@digriz.org.uk>");
726 MODULE_DESCRIPTION("'cpufreq_conservative' - A dynamic cpufreq governor for "
727 "Low Latency Frequency Transition capable processors "
728 "optimised for use in a battery environment");
729 MODULE_LICENSE("GPL");
731 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_CONSERVATIVE
732 fs_initcall(cpufreq_gov_dbs_init);
734 module_init(cpufreq_gov_dbs_init);
736 module_exit(cpufreq_gov_dbs_exit);