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 workqueue_struct *kconservative_wq;
86 static struct dbs_tuners {
87 unsigned int sampling_rate;
88 unsigned int sampling_down_factor;
89 unsigned int up_threshold;
90 unsigned int down_threshold;
91 unsigned int ignore_nice;
92 unsigned int freq_step;
94 .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
95 .down_threshold = DEF_FREQUENCY_DOWN_THRESHOLD,
96 .sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR,
101 static inline cputime64_t get_cpu_idle_time_jiffy(unsigned int cpu,
104 cputime64_t idle_time;
105 cputime64_t cur_wall_time;
106 cputime64_t busy_time;
108 cur_wall_time = jiffies64_to_cputime64(get_jiffies_64());
109 busy_time = cputime64_add(kstat_cpu(cpu).cpustat.user,
110 kstat_cpu(cpu).cpustat.system);
112 busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.irq);
113 busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.softirq);
114 busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.steal);
115 busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.nice);
117 idle_time = cputime64_sub(cur_wall_time, busy_time);
119 *wall = (cputime64_t)jiffies_to_usecs(cur_wall_time);
121 return (cputime64_t)jiffies_to_usecs(idle_time);;
124 static inline cputime64_t get_cpu_idle_time(unsigned int cpu, cputime64_t *wall)
126 u64 idle_time = get_cpu_idle_time_us(cpu, wall);
128 if (idle_time == -1ULL)
129 return get_cpu_idle_time_jiffy(cpu, wall);
134 /* keep track of frequency transitions */
136 dbs_cpufreq_notifier(struct notifier_block *nb, unsigned long val,
139 struct cpufreq_freqs *freq = data;
140 struct cpu_dbs_info_s *this_dbs_info = &per_cpu(cs_cpu_dbs_info,
143 struct cpufreq_policy *policy;
145 if (!this_dbs_info->enable)
148 policy = this_dbs_info->cur_policy;
151 * we only care if our internally tracked freq moves outside
152 * the 'valid' ranges of freqency available to us otherwise
153 * we do not change it
155 if (this_dbs_info->requested_freq > policy->max
156 || this_dbs_info->requested_freq < policy->min)
157 this_dbs_info->requested_freq = freq->new;
162 static struct notifier_block dbs_cpufreq_notifier_block = {
163 .notifier_call = dbs_cpufreq_notifier
166 /************************** sysfs interface ************************/
167 static ssize_t show_sampling_rate_max(struct kobject *kobj,
168 struct attribute *attr, char *buf)
170 printk_once(KERN_INFO "CPUFREQ: conservative sampling_rate_max "
171 "sysfs file is deprecated - used by: %s\n", current->comm);
172 return sprintf(buf, "%u\n", -1U);
175 static ssize_t show_sampling_rate_min(struct kobject *kobj,
176 struct attribute *attr, char *buf)
178 return sprintf(buf, "%u\n", min_sampling_rate);
181 define_one_global_ro(sampling_rate_max);
182 define_one_global_ro(sampling_rate_min);
184 /* cpufreq_conservative Governor Tunables */
185 #define show_one(file_name, object) \
186 static ssize_t show_##file_name \
187 (struct kobject *kobj, struct attribute *attr, char *buf) \
189 return sprintf(buf, "%u\n", dbs_tuners_ins.object); \
191 show_one(sampling_rate, sampling_rate);
192 show_one(sampling_down_factor, sampling_down_factor);
193 show_one(up_threshold, up_threshold);
194 show_one(down_threshold, down_threshold);
195 show_one(ignore_nice_load, ignore_nice);
196 show_one(freq_step, freq_step);
198 /*** delete after deprecation time ***/
199 #define DEPRECATION_MSG(file_name) \
200 printk_once(KERN_INFO "CPUFREQ: Per core conservative sysfs " \
201 "interface is deprecated - " #file_name "\n");
203 #define show_one_old(file_name) \
204 static ssize_t show_##file_name##_old \
205 (struct cpufreq_policy *unused, char *buf) \
207 printk_once(KERN_INFO "CPUFREQ: Per core conservative sysfs " \
208 "interface is deprecated - " #file_name "\n"); \
209 return show_##file_name(NULL, NULL, buf); \
211 show_one_old(sampling_rate);
212 show_one_old(sampling_down_factor);
213 show_one_old(up_threshold);
214 show_one_old(down_threshold);
215 show_one_old(ignore_nice_load);
216 show_one_old(freq_step);
217 show_one_old(sampling_rate_min);
218 show_one_old(sampling_rate_max);
220 cpufreq_freq_attr_ro_old(sampling_rate_min);
221 cpufreq_freq_attr_ro_old(sampling_rate_max);
223 /*** delete after deprecation time ***/
225 static ssize_t store_sampling_down_factor(struct kobject *a,
227 const char *buf, size_t count)
231 ret = sscanf(buf, "%u", &input);
233 if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
236 mutex_lock(&dbs_mutex);
237 dbs_tuners_ins.sampling_down_factor = input;
238 mutex_unlock(&dbs_mutex);
243 static ssize_t store_sampling_rate(struct kobject *a, struct attribute *b,
244 const char *buf, size_t count)
248 ret = sscanf(buf, "%u", &input);
253 mutex_lock(&dbs_mutex);
254 dbs_tuners_ins.sampling_rate = max(input, min_sampling_rate);
255 mutex_unlock(&dbs_mutex);
260 static ssize_t store_up_threshold(struct kobject *a, struct attribute *b,
261 const char *buf, size_t count)
265 ret = sscanf(buf, "%u", &input);
267 mutex_lock(&dbs_mutex);
268 if (ret != 1 || input > 100 ||
269 input <= dbs_tuners_ins.down_threshold) {
270 mutex_unlock(&dbs_mutex);
274 dbs_tuners_ins.up_threshold = input;
275 mutex_unlock(&dbs_mutex);
280 static ssize_t store_down_threshold(struct kobject *a, struct attribute *b,
281 const char *buf, size_t count)
285 ret = sscanf(buf, "%u", &input);
287 mutex_lock(&dbs_mutex);
288 /* cannot be lower than 11 otherwise freq will not fall */
289 if (ret != 1 || input < 11 || input > 100 ||
290 input >= dbs_tuners_ins.up_threshold) {
291 mutex_unlock(&dbs_mutex);
295 dbs_tuners_ins.down_threshold = input;
296 mutex_unlock(&dbs_mutex);
301 static ssize_t store_ignore_nice_load(struct kobject *a, struct attribute *b,
302 const char *buf, size_t count)
309 ret = sscanf(buf, "%u", &input);
316 mutex_lock(&dbs_mutex);
317 if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */
318 mutex_unlock(&dbs_mutex);
321 dbs_tuners_ins.ignore_nice = input;
323 /* we need to re-evaluate prev_cpu_idle */
324 for_each_online_cpu(j) {
325 struct cpu_dbs_info_s *dbs_info;
326 dbs_info = &per_cpu(cs_cpu_dbs_info, j);
327 dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
328 &dbs_info->prev_cpu_wall);
329 if (dbs_tuners_ins.ignore_nice)
330 dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
332 mutex_unlock(&dbs_mutex);
337 static ssize_t store_freq_step(struct kobject *a, struct attribute *b,
338 const char *buf, size_t count)
342 ret = sscanf(buf, "%u", &input);
350 /* no need to test here if freq_step is zero as the user might actually
351 * want this, they would be crazy though :) */
352 mutex_lock(&dbs_mutex);
353 dbs_tuners_ins.freq_step = input;
354 mutex_unlock(&dbs_mutex);
359 define_one_global_rw(sampling_rate);
360 define_one_global_rw(sampling_down_factor);
361 define_one_global_rw(up_threshold);
362 define_one_global_rw(down_threshold);
363 define_one_global_rw(ignore_nice_load);
364 define_one_global_rw(freq_step);
366 static struct attribute *dbs_attributes[] = {
367 &sampling_rate_max.attr,
368 &sampling_rate_min.attr,
370 &sampling_down_factor.attr,
372 &down_threshold.attr,
373 &ignore_nice_load.attr,
378 static struct attribute_group dbs_attr_group = {
379 .attrs = dbs_attributes,
380 .name = "conservative",
383 /*** delete after deprecation time ***/
385 #define write_one_old(file_name) \
386 static ssize_t store_##file_name##_old \
387 (struct cpufreq_policy *unused, const char *buf, size_t count) \
389 printk_once(KERN_INFO "CPUFREQ: Per core conservative sysfs " \
390 "interface is deprecated - " #file_name "\n"); \
391 return store_##file_name(NULL, NULL, buf, count); \
393 write_one_old(sampling_rate);
394 write_one_old(sampling_down_factor);
395 write_one_old(up_threshold);
396 write_one_old(down_threshold);
397 write_one_old(ignore_nice_load);
398 write_one_old(freq_step);
400 cpufreq_freq_attr_rw_old(sampling_rate);
401 cpufreq_freq_attr_rw_old(sampling_down_factor);
402 cpufreq_freq_attr_rw_old(up_threshold);
403 cpufreq_freq_attr_rw_old(down_threshold);
404 cpufreq_freq_attr_rw_old(ignore_nice_load);
405 cpufreq_freq_attr_rw_old(freq_step);
407 static struct attribute *dbs_attributes_old[] = {
408 &sampling_rate_max_old.attr,
409 &sampling_rate_min_old.attr,
410 &sampling_rate_old.attr,
411 &sampling_down_factor_old.attr,
412 &up_threshold_old.attr,
413 &down_threshold_old.attr,
414 &ignore_nice_load_old.attr,
419 static struct attribute_group dbs_attr_group_old = {
420 .attrs = dbs_attributes_old,
421 .name = "conservative",
424 /*** delete after deprecation time ***/
426 /************************** sysfs end ************************/
428 static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
430 unsigned int load = 0;
431 unsigned int max_load = 0;
432 unsigned int freq_target;
434 struct cpufreq_policy *policy;
437 policy = this_dbs_info->cur_policy;
440 * Every sampling_rate, we check, if current idle time is less
441 * than 20% (default), then we try to increase frequency
442 * Every sampling_rate*sampling_down_factor, we check, if current
443 * idle time is more than 80%, then we try to decrease frequency
445 * Any frequency increase takes it to the maximum frequency.
446 * Frequency reduction happens at minimum steps of
447 * 5% (default) of maximum frequency
450 /* Get Absolute Load */
451 for_each_cpu(j, policy->cpus) {
452 struct cpu_dbs_info_s *j_dbs_info;
453 cputime64_t cur_wall_time, cur_idle_time;
454 unsigned int idle_time, wall_time;
456 j_dbs_info = &per_cpu(cs_cpu_dbs_info, j);
458 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time);
460 wall_time = (unsigned int) cputime64_sub(cur_wall_time,
461 j_dbs_info->prev_cpu_wall);
462 j_dbs_info->prev_cpu_wall = cur_wall_time;
464 idle_time = (unsigned int) cputime64_sub(cur_idle_time,
465 j_dbs_info->prev_cpu_idle);
466 j_dbs_info->prev_cpu_idle = cur_idle_time;
468 if (dbs_tuners_ins.ignore_nice) {
469 cputime64_t cur_nice;
470 unsigned long cur_nice_jiffies;
472 cur_nice = cputime64_sub(kstat_cpu(j).cpustat.nice,
473 j_dbs_info->prev_cpu_nice);
475 * Assumption: nice time between sampling periods will
476 * be less than 2^32 jiffies for 32 bit sys
478 cur_nice_jiffies = (unsigned long)
479 cputime64_to_jiffies64(cur_nice);
481 j_dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
482 idle_time += jiffies_to_usecs(cur_nice_jiffies);
485 if (unlikely(!wall_time || wall_time < idle_time))
488 load = 100 * (wall_time - idle_time) / wall_time;
495 * break out if we 'cannot' reduce the speed as the user might
496 * want freq_step to be zero
498 if (dbs_tuners_ins.freq_step == 0)
501 /* Check for frequency increase */
502 if (max_load > dbs_tuners_ins.up_threshold) {
503 this_dbs_info->down_skip = 0;
505 /* if we are already at full speed then break out early */
506 if (this_dbs_info->requested_freq == policy->max)
509 freq_target = (dbs_tuners_ins.freq_step * policy->max) / 100;
511 /* max freq cannot be less than 100. But who knows.... */
512 if (unlikely(freq_target == 0))
515 this_dbs_info->requested_freq += freq_target;
516 if (this_dbs_info->requested_freq > policy->max)
517 this_dbs_info->requested_freq = policy->max;
519 __cpufreq_driver_target(policy, this_dbs_info->requested_freq,
525 * The optimal frequency is the frequency that is the lowest that
526 * can support the current CPU usage without triggering the up
527 * policy. To be safe, we focus 10 points under the threshold.
529 if (max_load < (dbs_tuners_ins.down_threshold - 10)) {
530 freq_target = (dbs_tuners_ins.freq_step * policy->max) / 100;
532 this_dbs_info->requested_freq -= freq_target;
533 if (this_dbs_info->requested_freq < policy->min)
534 this_dbs_info->requested_freq = policy->min;
537 * if we cannot reduce the frequency anymore, break out early
539 if (policy->cur == policy->min)
542 __cpufreq_driver_target(policy, this_dbs_info->requested_freq,
548 static void do_dbs_timer(struct work_struct *work)
550 struct cpu_dbs_info_s *dbs_info =
551 container_of(work, struct cpu_dbs_info_s, work.work);
552 unsigned int cpu = dbs_info->cpu;
554 /* We want all CPUs to do sampling nearly on same jiffy */
555 int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
557 delay -= jiffies % delay;
559 mutex_lock(&dbs_info->timer_mutex);
561 dbs_check_cpu(dbs_info);
563 queue_delayed_work_on(cpu, kconservative_wq, &dbs_info->work, delay);
564 mutex_unlock(&dbs_info->timer_mutex);
567 static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
569 /* We want all CPUs to do sampling nearly on same jiffy */
570 int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
571 delay -= jiffies % delay;
573 dbs_info->enable = 1;
574 INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
575 queue_delayed_work_on(dbs_info->cpu, kconservative_wq, &dbs_info->work,
579 static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
581 dbs_info->enable = 0;
582 cancel_delayed_work_sync(&dbs_info->work);
585 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
588 unsigned int cpu = policy->cpu;
589 struct cpu_dbs_info_s *this_dbs_info;
593 this_dbs_info = &per_cpu(cs_cpu_dbs_info, cpu);
596 case CPUFREQ_GOV_START:
597 if ((!cpu_online(cpu)) || (!policy->cur))
600 mutex_lock(&dbs_mutex);
602 rc = sysfs_create_group(&policy->kobj, &dbs_attr_group_old);
604 mutex_unlock(&dbs_mutex);
608 for_each_cpu(j, policy->cpus) {
609 struct cpu_dbs_info_s *j_dbs_info;
610 j_dbs_info = &per_cpu(cs_cpu_dbs_info, j);
611 j_dbs_info->cur_policy = policy;
613 j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
614 &j_dbs_info->prev_cpu_wall);
615 if (dbs_tuners_ins.ignore_nice) {
616 j_dbs_info->prev_cpu_nice =
617 kstat_cpu(j).cpustat.nice;
620 this_dbs_info->down_skip = 0;
621 this_dbs_info->requested_freq = policy->cur;
623 mutex_init(&this_dbs_info->timer_mutex);
626 * Start the timerschedule work, when this governor
627 * is used for first time
629 if (dbs_enable == 1) {
630 unsigned int latency;
631 /* policy latency is in nS. Convert it to uS first */
632 latency = policy->cpuinfo.transition_latency / 1000;
636 rc = sysfs_create_group(cpufreq_global_kobject,
639 mutex_unlock(&dbs_mutex);
644 * conservative does not implement micro like ondemand
645 * governor, thus we are bound to jiffes/HZ
648 MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10);
649 /* Bring kernel and HW constraints together */
650 min_sampling_rate = max(min_sampling_rate,
651 MIN_LATENCY_MULTIPLIER * latency);
652 dbs_tuners_ins.sampling_rate =
653 max(min_sampling_rate,
654 latency * LATENCY_MULTIPLIER);
656 cpufreq_register_notifier(
657 &dbs_cpufreq_notifier_block,
658 CPUFREQ_TRANSITION_NOTIFIER);
660 mutex_unlock(&dbs_mutex);
662 dbs_timer_init(this_dbs_info);
666 case CPUFREQ_GOV_STOP:
667 dbs_timer_exit(this_dbs_info);
669 mutex_lock(&dbs_mutex);
670 sysfs_remove_group(&policy->kobj, &dbs_attr_group_old);
672 mutex_destroy(&this_dbs_info->timer_mutex);
675 * Stop the timerschedule work, when this governor
676 * is used for first time
679 cpufreq_unregister_notifier(
680 &dbs_cpufreq_notifier_block,
681 CPUFREQ_TRANSITION_NOTIFIER);
683 mutex_unlock(&dbs_mutex);
685 sysfs_remove_group(cpufreq_global_kobject,
690 case CPUFREQ_GOV_LIMITS:
691 mutex_lock(&this_dbs_info->timer_mutex);
692 if (policy->max < this_dbs_info->cur_policy->cur)
693 __cpufreq_driver_target(
694 this_dbs_info->cur_policy,
695 policy->max, CPUFREQ_RELATION_H);
696 else if (policy->min > this_dbs_info->cur_policy->cur)
697 __cpufreq_driver_target(
698 this_dbs_info->cur_policy,
699 policy->min, CPUFREQ_RELATION_L);
700 mutex_unlock(&this_dbs_info->timer_mutex);
707 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_CONSERVATIVE
710 struct cpufreq_governor cpufreq_gov_conservative = {
711 .name = "conservative",
712 .governor = cpufreq_governor_dbs,
713 .max_transition_latency = TRANSITION_LATENCY_LIMIT,
714 .owner = THIS_MODULE,
717 static int __init cpufreq_gov_dbs_init(void)
721 kconservative_wq = create_workqueue("kconservative");
722 if (!kconservative_wq) {
723 printk(KERN_ERR "Creation of kconservative failed\n");
727 err = cpufreq_register_governor(&cpufreq_gov_conservative);
729 destroy_workqueue(kconservative_wq);
734 static void __exit cpufreq_gov_dbs_exit(void)
736 cpufreq_unregister_governor(&cpufreq_gov_conservative);
737 destroy_workqueue(kconservative_wq);
741 MODULE_AUTHOR("Alexander Clouter <alex@digriz.org.uk>");
742 MODULE_DESCRIPTION("'cpufreq_conservative' - A dynamic cpufreq governor for "
743 "Low Latency Frequency Transition capable processors "
744 "optimised for use in a battery environment");
745 MODULE_LICENSE("GPL");
747 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_CONSERVATIVE
748 fs_initcall(cpufreq_gov_dbs_init);
750 module_init(cpufreq_gov_dbs_init);
752 module_exit(cpufreq_gov_dbs_exit);