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;
67 * percpu mutex that serializes governor limit change with
68 * do_dbs_timer invocation. We do not want do_dbs_timer to run
69 * when user is changing the governor or limits.
71 struct mutex timer_mutex;
73 static DEFINE_PER_CPU(struct cpu_dbs_info_s, cpu_dbs_info);
75 static unsigned int dbs_enable; /* number of CPUs using this policy */
78 * dbs_mutex protects data in dbs_tuners_ins from concurrent changes on
79 * different CPUs. It protects dbs_enable in governor start/stop.
81 static DEFINE_MUTEX(dbs_mutex);
83 static struct workqueue_struct *kconservative_wq;
85 static struct dbs_tuners {
86 unsigned int sampling_rate;
87 unsigned int sampling_down_factor;
88 unsigned int up_threshold;
89 unsigned int down_threshold;
90 unsigned int ignore_nice;
91 unsigned int freq_step;
93 .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
94 .down_threshold = DEF_FREQUENCY_DOWN_THRESHOLD,
95 .sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR,
100 static inline cputime64_t get_cpu_idle_time_jiffy(unsigned int cpu,
103 cputime64_t idle_time;
104 cputime64_t cur_wall_time;
105 cputime64_t busy_time;
107 cur_wall_time = jiffies64_to_cputime64(get_jiffies_64());
108 busy_time = cputime64_add(kstat_cpu(cpu).cpustat.user,
109 kstat_cpu(cpu).cpustat.system);
111 busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.irq);
112 busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.softirq);
113 busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.steal);
114 busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.nice);
116 idle_time = cputime64_sub(cur_wall_time, busy_time);
118 *wall = cur_wall_time;
123 static inline cputime64_t get_cpu_idle_time(unsigned int cpu, cputime64_t *wall)
125 u64 idle_time = get_cpu_idle_time_us(cpu, wall);
127 if (idle_time == -1ULL)
128 return get_cpu_idle_time_jiffy(cpu, wall);
133 /* keep track of frequency transitions */
135 dbs_cpufreq_notifier(struct notifier_block *nb, unsigned long val,
138 struct cpufreq_freqs *freq = data;
139 struct cpu_dbs_info_s *this_dbs_info = &per_cpu(cpu_dbs_info,
142 struct cpufreq_policy *policy;
144 policy = this_dbs_info->cur_policy;
147 * we only care if our internally tracked freq moves outside
148 * the 'valid' ranges of freqency available to us otherwise
149 * we do not change it
151 if (this_dbs_info->requested_freq > policy->max
152 || this_dbs_info->requested_freq < policy->min)
153 this_dbs_info->requested_freq = freq->new;
158 static struct notifier_block dbs_cpufreq_notifier_block = {
159 .notifier_call = dbs_cpufreq_notifier
162 /************************** sysfs interface ************************/
163 static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf)
165 printk_once(KERN_INFO "CPUFREQ: conservative sampling_rate_max "
166 "sysfs file is deprecated - used by: %s\n", current->comm);
167 return sprintf(buf, "%u\n", -1U);
170 static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf)
172 return sprintf(buf, "%u\n", min_sampling_rate);
175 #define define_one_ro(_name) \
176 static struct freq_attr _name = \
177 __ATTR(_name, 0444, show_##_name, NULL)
179 define_one_ro(sampling_rate_max);
180 define_one_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 cpufreq_policy *unused, 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 static ssize_t store_sampling_down_factor(struct cpufreq_policy *unused,
197 const char *buf, size_t count)
201 ret = sscanf(buf, "%u", &input);
203 if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
206 mutex_lock(&dbs_mutex);
207 dbs_tuners_ins.sampling_down_factor = input;
208 mutex_unlock(&dbs_mutex);
213 static ssize_t store_sampling_rate(struct cpufreq_policy *unused,
214 const char *buf, size_t count)
218 ret = sscanf(buf, "%u", &input);
223 mutex_lock(&dbs_mutex);
224 dbs_tuners_ins.sampling_rate = max(input, min_sampling_rate);
225 mutex_unlock(&dbs_mutex);
230 static ssize_t store_up_threshold(struct cpufreq_policy *unused,
231 const char *buf, size_t count)
235 ret = sscanf(buf, "%u", &input);
237 mutex_lock(&dbs_mutex);
238 if (ret != 1 || input > 100 ||
239 input <= dbs_tuners_ins.down_threshold) {
240 mutex_unlock(&dbs_mutex);
244 dbs_tuners_ins.up_threshold = input;
245 mutex_unlock(&dbs_mutex);
250 static ssize_t store_down_threshold(struct cpufreq_policy *unused,
251 const char *buf, size_t count)
255 ret = sscanf(buf, "%u", &input);
257 mutex_lock(&dbs_mutex);
258 /* cannot be lower than 11 otherwise freq will not fall */
259 if (ret != 1 || input < 11 || input > 100 ||
260 input >= dbs_tuners_ins.up_threshold) {
261 mutex_unlock(&dbs_mutex);
265 dbs_tuners_ins.down_threshold = input;
266 mutex_unlock(&dbs_mutex);
271 static ssize_t store_ignore_nice_load(struct cpufreq_policy *policy,
272 const char *buf, size_t count)
279 ret = sscanf(buf, "%u", &input);
286 mutex_lock(&dbs_mutex);
287 if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */
288 mutex_unlock(&dbs_mutex);
291 dbs_tuners_ins.ignore_nice = input;
293 /* we need to re-evaluate prev_cpu_idle */
294 for_each_online_cpu(j) {
295 struct cpu_dbs_info_s *dbs_info;
296 dbs_info = &per_cpu(cpu_dbs_info, j);
297 dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
298 &dbs_info->prev_cpu_wall);
299 if (dbs_tuners_ins.ignore_nice)
300 dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
302 mutex_unlock(&dbs_mutex);
307 static ssize_t store_freq_step(struct cpufreq_policy *policy,
308 const char *buf, size_t count)
312 ret = sscanf(buf, "%u", &input);
320 /* no need to test here if freq_step is zero as the user might actually
321 * want this, they would be crazy though :) */
322 mutex_lock(&dbs_mutex);
323 dbs_tuners_ins.freq_step = input;
324 mutex_unlock(&dbs_mutex);
329 #define define_one_rw(_name) \
330 static struct freq_attr _name = \
331 __ATTR(_name, 0644, show_##_name, store_##_name)
333 define_one_rw(sampling_rate);
334 define_one_rw(sampling_down_factor);
335 define_one_rw(up_threshold);
336 define_one_rw(down_threshold);
337 define_one_rw(ignore_nice_load);
338 define_one_rw(freq_step);
340 static struct attribute *dbs_attributes[] = {
341 &sampling_rate_max.attr,
342 &sampling_rate_min.attr,
344 &sampling_down_factor.attr,
346 &down_threshold.attr,
347 &ignore_nice_load.attr,
352 static struct attribute_group dbs_attr_group = {
353 .attrs = dbs_attributes,
354 .name = "conservative",
357 /************************** sysfs end ************************/
359 static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
361 unsigned int load = 0;
362 unsigned int freq_target;
364 struct cpufreq_policy *policy;
367 policy = this_dbs_info->cur_policy;
370 * Every sampling_rate, we check, if current idle time is less
371 * than 20% (default), then we try to increase frequency
372 * Every sampling_rate*sampling_down_factor, we check, if current
373 * idle time is more than 80%, then we try to decrease frequency
375 * Any frequency increase takes it to the maximum frequency.
376 * Frequency reduction happens at minimum steps of
377 * 5% (default) of maximum frequency
380 /* Get Absolute Load */
381 for_each_cpu(j, policy->cpus) {
382 struct cpu_dbs_info_s *j_dbs_info;
383 cputime64_t cur_wall_time, cur_idle_time;
384 unsigned int idle_time, wall_time;
386 j_dbs_info = &per_cpu(cpu_dbs_info, j);
388 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time);
390 wall_time = (unsigned int) cputime64_sub(cur_wall_time,
391 j_dbs_info->prev_cpu_wall);
392 j_dbs_info->prev_cpu_wall = cur_wall_time;
394 idle_time = (unsigned int) cputime64_sub(cur_idle_time,
395 j_dbs_info->prev_cpu_idle);
396 j_dbs_info->prev_cpu_idle = cur_idle_time;
398 if (dbs_tuners_ins.ignore_nice) {
399 cputime64_t cur_nice;
400 unsigned long cur_nice_jiffies;
402 cur_nice = cputime64_sub(kstat_cpu(j).cpustat.nice,
403 j_dbs_info->prev_cpu_nice);
405 * Assumption: nice time between sampling periods will
406 * be less than 2^32 jiffies for 32 bit sys
408 cur_nice_jiffies = (unsigned long)
409 cputime64_to_jiffies64(cur_nice);
411 j_dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
412 idle_time += jiffies_to_usecs(cur_nice_jiffies);
415 if (unlikely(!wall_time || wall_time < idle_time))
418 load = 100 * (wall_time - idle_time) / wall_time;
422 * break out if we 'cannot' reduce the speed as the user might
423 * want freq_step to be zero
425 if (dbs_tuners_ins.freq_step == 0)
428 /* Check for frequency increase */
429 if (load > dbs_tuners_ins.up_threshold) {
430 this_dbs_info->down_skip = 0;
432 /* if we are already at full speed then break out early */
433 if (this_dbs_info->requested_freq == policy->max)
436 freq_target = (dbs_tuners_ins.freq_step * policy->max) / 100;
438 /* max freq cannot be less than 100. But who knows.... */
439 if (unlikely(freq_target == 0))
442 this_dbs_info->requested_freq += freq_target;
443 if (this_dbs_info->requested_freq > policy->max)
444 this_dbs_info->requested_freq = policy->max;
446 __cpufreq_driver_target(policy, this_dbs_info->requested_freq,
452 * The optimal frequency is the frequency that is the lowest that
453 * can support the current CPU usage without triggering the up
454 * policy. To be safe, we focus 10 points under the threshold.
456 if (load < (dbs_tuners_ins.down_threshold - 10)) {
457 freq_target = (dbs_tuners_ins.freq_step * policy->max) / 100;
459 this_dbs_info->requested_freq -= freq_target;
460 if (this_dbs_info->requested_freq < policy->min)
461 this_dbs_info->requested_freq = policy->min;
464 * if we cannot reduce the frequency anymore, break out early
466 if (policy->cur == policy->min)
469 __cpufreq_driver_target(policy, this_dbs_info->requested_freq,
475 static void do_dbs_timer(struct work_struct *work)
477 struct cpu_dbs_info_s *dbs_info =
478 container_of(work, struct cpu_dbs_info_s, work.work);
479 unsigned int cpu = dbs_info->cpu;
481 /* We want all CPUs to do sampling nearly on same jiffy */
482 int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
484 delay -= jiffies % delay;
486 mutex_lock(&dbs_info->timer_mutex);
488 dbs_check_cpu(dbs_info);
490 queue_delayed_work_on(cpu, kconservative_wq, &dbs_info->work, delay);
491 mutex_unlock(&dbs_info->timer_mutex);
494 static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
496 /* We want all CPUs to do sampling nearly on same jiffy */
497 int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
498 delay -= jiffies % delay;
500 INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
501 queue_delayed_work_on(dbs_info->cpu, kconservative_wq, &dbs_info->work,
505 static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
507 cancel_delayed_work_sync(&dbs_info->work);
510 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
513 unsigned int cpu = policy->cpu;
514 struct cpu_dbs_info_s *this_dbs_info;
518 this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
521 case CPUFREQ_GOV_START:
522 if ((!cpu_online(cpu)) || (!policy->cur))
525 mutex_lock(&dbs_mutex);
527 rc = sysfs_create_group(&policy->kobj, &dbs_attr_group);
529 mutex_unlock(&dbs_mutex);
533 for_each_cpu(j, policy->cpus) {
534 struct cpu_dbs_info_s *j_dbs_info;
535 j_dbs_info = &per_cpu(cpu_dbs_info, j);
536 j_dbs_info->cur_policy = policy;
538 j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
539 &j_dbs_info->prev_cpu_wall);
540 if (dbs_tuners_ins.ignore_nice) {
541 j_dbs_info->prev_cpu_nice =
542 kstat_cpu(j).cpustat.nice;
545 this_dbs_info->down_skip = 0;
546 this_dbs_info->requested_freq = policy->cur;
548 mutex_init(&this_dbs_info->timer_mutex);
551 * Start the timerschedule work, when this governor
552 * is used for first time
554 if (dbs_enable == 1) {
555 unsigned int latency;
556 /* policy latency is in nS. Convert it to uS first */
557 latency = policy->cpuinfo.transition_latency / 1000;
562 * conservative does not implement micro like ondemand
563 * governor, thus we are bound to jiffes/HZ
566 MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10);
567 /* Bring kernel and HW constraints together */
568 min_sampling_rate = max(min_sampling_rate,
569 MIN_LATENCY_MULTIPLIER * latency);
570 dbs_tuners_ins.sampling_rate =
571 max(min_sampling_rate,
572 latency * LATENCY_MULTIPLIER);
574 cpufreq_register_notifier(
575 &dbs_cpufreq_notifier_block,
576 CPUFREQ_TRANSITION_NOTIFIER);
578 mutex_unlock(&dbs_mutex);
580 dbs_timer_init(this_dbs_info);
584 case CPUFREQ_GOV_STOP:
585 dbs_timer_exit(this_dbs_info);
587 mutex_lock(&dbs_mutex);
588 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
590 mutex_destroy(&this_dbs_info->timer_mutex);
593 * Stop the timerschedule work, when this governor
594 * is used for first time
597 cpufreq_unregister_notifier(
598 &dbs_cpufreq_notifier_block,
599 CPUFREQ_TRANSITION_NOTIFIER);
601 mutex_unlock(&dbs_mutex);
605 case CPUFREQ_GOV_LIMITS:
606 mutex_lock(&this_dbs_info->timer_mutex);
607 if (policy->max < this_dbs_info->cur_policy->cur)
608 __cpufreq_driver_target(
609 this_dbs_info->cur_policy,
610 policy->max, CPUFREQ_RELATION_H);
611 else if (policy->min > this_dbs_info->cur_policy->cur)
612 __cpufreq_driver_target(
613 this_dbs_info->cur_policy,
614 policy->min, CPUFREQ_RELATION_L);
615 mutex_unlock(&this_dbs_info->timer_mutex);
622 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_CONSERVATIVE
625 struct cpufreq_governor cpufreq_gov_conservative = {
626 .name = "conservative",
627 .governor = cpufreq_governor_dbs,
628 .max_transition_latency = TRANSITION_LATENCY_LIMIT,
629 .owner = THIS_MODULE,
632 static int __init cpufreq_gov_dbs_init(void)
636 kconservative_wq = create_workqueue("kconservative");
637 if (!kconservative_wq) {
638 printk(KERN_ERR "Creation of kconservative failed\n");
642 err = cpufreq_register_governor(&cpufreq_gov_conservative);
644 destroy_workqueue(kconservative_wq);
649 static void __exit cpufreq_gov_dbs_exit(void)
651 cpufreq_unregister_governor(&cpufreq_gov_conservative);
652 destroy_workqueue(kconservative_wq);
656 MODULE_AUTHOR("Alexander Clouter <alex@digriz.org.uk>");
657 MODULE_DESCRIPTION("'cpufreq_conservative' - A dynamic cpufreq governor for "
658 "Low Latency Frequency Transition capable processors "
659 "optimised for use in a battery environment");
660 MODULE_LICENSE("GPL");
662 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_CONSERVATIVE
663 fs_initcall(cpufreq_gov_dbs_init);
665 module_init(cpufreq_gov_dbs_init);
667 module_exit(cpufreq_gov_dbs_exit);