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 static unsigned int def_sampling_rate;
46 #define MIN_SAMPLING_RATE_RATIO (2)
47 /* for correct statistics, we need at least 10 ticks between each measure */
48 #define MIN_STAT_SAMPLING_RATE \
49 (MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10))
50 #define MIN_SAMPLING_RATE \
51 (def_sampling_rate / MIN_SAMPLING_RATE_RATIO)
52 /* Above MIN_SAMPLING_RATE will vanish with its sysfs file soon
53 * Define the minimal settable sampling rate to the greater of:
54 * - "HW transition latency" * 100 (same as default sampling / 10)
55 * - MIN_STAT_SAMPLING_RATE
56 * To avoid that userspace shoots itself.
58 static unsigned int minimum_sampling_rate(void)
60 return max(def_sampling_rate / 10, MIN_STAT_SAMPLING_RATE);
63 /* This will also vanish soon with removing sampling_rate_max */
64 #define MAX_SAMPLING_RATE (500 * def_sampling_rate)
65 #define LATENCY_MULTIPLIER (1000)
66 #define DEF_SAMPLING_DOWN_FACTOR (1)
67 #define MAX_SAMPLING_DOWN_FACTOR (10)
68 #define TRANSITION_LATENCY_LIMIT (10 * 1000 * 1000)
70 static void do_dbs_timer(struct work_struct *work);
72 struct cpu_dbs_info_s {
73 cputime64_t prev_cpu_idle;
74 cputime64_t prev_cpu_wall;
75 cputime64_t prev_cpu_nice;
76 struct cpufreq_policy *cur_policy;
77 struct delayed_work work;
78 unsigned int down_skip;
79 unsigned int requested_freq;
81 unsigned int enable:1;
83 static DEFINE_PER_CPU(struct cpu_dbs_info_s, cpu_dbs_info);
85 static unsigned int dbs_enable; /* number of CPUs using this policy */
88 * DEADLOCK ALERT! There is a ordering requirement between cpu_hotplug
89 * lock and dbs_mutex. cpu_hotplug lock should always be held before
90 * dbs_mutex. If any function that can potentially take cpu_hotplug lock
91 * (like __cpufreq_driver_target()) is being called with dbs_mutex taken, then
92 * cpu_hotplug lock should be taken before that. Note that cpu_hotplug lock
93 * is recursive for the same process. -Venki
94 * DEADLOCK ALERT! (2) : do_dbs_timer() must not take the dbs_mutex, because it
95 * would deadlock with cancel_delayed_work_sync(), which is needed for proper
96 * raceless workqueue teardown.
98 static DEFINE_MUTEX(dbs_mutex);
100 static struct workqueue_struct *kconservative_wq;
102 static struct dbs_tuners {
103 unsigned int sampling_rate;
104 unsigned int sampling_down_factor;
105 unsigned int up_threshold;
106 unsigned int down_threshold;
107 unsigned int ignore_nice;
108 unsigned int freq_step;
110 .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
111 .down_threshold = DEF_FREQUENCY_DOWN_THRESHOLD,
112 .sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR,
117 static inline cputime64_t get_cpu_idle_time_jiffy(unsigned int cpu,
120 cputime64_t idle_time;
121 cputime64_t cur_wall_time;
122 cputime64_t busy_time;
124 cur_wall_time = jiffies64_to_cputime64(get_jiffies_64());
125 busy_time = cputime64_add(kstat_cpu(cpu).cpustat.user,
126 kstat_cpu(cpu).cpustat.system);
128 busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.irq);
129 busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.softirq);
130 busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.steal);
131 busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.nice);
133 idle_time = cputime64_sub(cur_wall_time, busy_time);
135 *wall = cur_wall_time;
140 static inline cputime64_t get_cpu_idle_time(unsigned int cpu, cputime64_t *wall)
142 u64 idle_time = get_cpu_idle_time_us(cpu, wall);
144 if (idle_time == -1ULL)
145 return get_cpu_idle_time_jiffy(cpu, wall);
150 /* keep track of frequency transitions */
152 dbs_cpufreq_notifier(struct notifier_block *nb, unsigned long val,
155 struct cpufreq_freqs *freq = data;
156 struct cpu_dbs_info_s *this_dbs_info = &per_cpu(cpu_dbs_info,
159 struct cpufreq_policy *policy;
161 if (!this_dbs_info->enable)
164 policy = this_dbs_info->cur_policy;
167 * we only care if our internally tracked freq moves outside
168 * the 'valid' ranges of freqency available to us otherwise
169 * we do not change it
171 if (this_dbs_info->requested_freq > policy->max
172 || this_dbs_info->requested_freq < policy->min)
173 this_dbs_info->requested_freq = freq->new;
178 static struct notifier_block dbs_cpufreq_notifier_block = {
179 .notifier_call = dbs_cpufreq_notifier
182 /************************** sysfs interface ************************/
183 static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf)
185 static int print_once;
188 printk(KERN_INFO "CPUFREQ: conservative sampling_rate_max "
189 "sysfs file is deprecated - used by: %s\n",
193 return sprintf(buf, "%u\n", MAX_SAMPLING_RATE);
196 static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf)
198 static int print_once;
201 printk(KERN_INFO "CPUFREQ: conservative sampling_rate_max "
202 "sysfs file is deprecated - used by: %s\n", current->comm);
205 return sprintf(buf, "%u\n", MIN_SAMPLING_RATE);
208 #define define_one_ro(_name) \
209 static struct freq_attr _name = \
210 __ATTR(_name, 0444, show_##_name, NULL)
212 define_one_ro(sampling_rate_max);
213 define_one_ro(sampling_rate_min);
215 /* cpufreq_conservative Governor Tunables */
216 #define show_one(file_name, object) \
217 static ssize_t show_##file_name \
218 (struct cpufreq_policy *unused, char *buf) \
220 return sprintf(buf, "%u\n", dbs_tuners_ins.object); \
222 show_one(sampling_rate, sampling_rate);
223 show_one(sampling_down_factor, sampling_down_factor);
224 show_one(up_threshold, up_threshold);
225 show_one(down_threshold, down_threshold);
226 show_one(ignore_nice_load, ignore_nice);
227 show_one(freq_step, freq_step);
229 static ssize_t store_sampling_down_factor(struct cpufreq_policy *unused,
230 const char *buf, size_t count)
234 ret = sscanf(buf, "%u", &input);
236 if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
239 mutex_lock(&dbs_mutex);
240 dbs_tuners_ins.sampling_down_factor = input;
241 mutex_unlock(&dbs_mutex);
246 static ssize_t store_sampling_rate(struct cpufreq_policy *unused,
247 const char *buf, size_t count)
251 ret = sscanf(buf, "%u", &input);
256 mutex_lock(&dbs_mutex);
257 dbs_tuners_ins.sampling_rate = max(input, minimum_sampling_rate());
258 mutex_unlock(&dbs_mutex);
263 static ssize_t store_up_threshold(struct cpufreq_policy *unused,
264 const char *buf, size_t count)
268 ret = sscanf(buf, "%u", &input);
270 mutex_lock(&dbs_mutex);
271 if (ret != 1 || input > 100 ||
272 input <= dbs_tuners_ins.down_threshold) {
273 mutex_unlock(&dbs_mutex);
277 dbs_tuners_ins.up_threshold = input;
278 mutex_unlock(&dbs_mutex);
283 static ssize_t store_down_threshold(struct cpufreq_policy *unused,
284 const char *buf, size_t count)
288 ret = sscanf(buf, "%u", &input);
290 mutex_lock(&dbs_mutex);
291 /* cannot be lower than 11 otherwise freq will not fall */
292 if (ret != 1 || input < 11 || input > 100 ||
293 input >= dbs_tuners_ins.up_threshold) {
294 mutex_unlock(&dbs_mutex);
298 dbs_tuners_ins.down_threshold = input;
299 mutex_unlock(&dbs_mutex);
304 static ssize_t store_ignore_nice_load(struct cpufreq_policy *policy,
305 const char *buf, size_t count)
312 ret = sscanf(buf, "%u", &input);
319 mutex_lock(&dbs_mutex);
320 if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */
321 mutex_unlock(&dbs_mutex);
324 dbs_tuners_ins.ignore_nice = input;
326 /* we need to re-evaluate prev_cpu_idle */
327 for_each_online_cpu(j) {
328 struct cpu_dbs_info_s *dbs_info;
329 dbs_info = &per_cpu(cpu_dbs_info, j);
330 dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
331 &dbs_info->prev_cpu_wall);
332 if (dbs_tuners_ins.ignore_nice)
333 dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
335 mutex_unlock(&dbs_mutex);
340 static ssize_t store_freq_step(struct cpufreq_policy *policy,
341 const char *buf, size_t count)
345 ret = sscanf(buf, "%u", &input);
353 /* no need to test here if freq_step is zero as the user might actually
354 * want this, they would be crazy though :) */
355 mutex_lock(&dbs_mutex);
356 dbs_tuners_ins.freq_step = input;
357 mutex_unlock(&dbs_mutex);
362 #define define_one_rw(_name) \
363 static struct freq_attr _name = \
364 __ATTR(_name, 0644, show_##_name, store_##_name)
366 define_one_rw(sampling_rate);
367 define_one_rw(sampling_down_factor);
368 define_one_rw(up_threshold);
369 define_one_rw(down_threshold);
370 define_one_rw(ignore_nice_load);
371 define_one_rw(freq_step);
373 static struct attribute *dbs_attributes[] = {
374 &sampling_rate_max.attr,
375 &sampling_rate_min.attr,
377 &sampling_down_factor.attr,
379 &down_threshold.attr,
380 &ignore_nice_load.attr,
385 static struct attribute_group dbs_attr_group = {
386 .attrs = dbs_attributes,
387 .name = "conservative",
390 /************************** sysfs end ************************/
392 static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
394 unsigned int load = 0;
395 unsigned int freq_target;
397 struct cpufreq_policy *policy;
400 policy = this_dbs_info->cur_policy;
403 * Every sampling_rate, we check, if current idle time is less
404 * than 20% (default), then we try to increase frequency
405 * Every sampling_rate*sampling_down_factor, we check, if current
406 * idle time is more than 80%, then we try to decrease frequency
408 * Any frequency increase takes it to the maximum frequency.
409 * Frequency reduction happens at minimum steps of
410 * 5% (default) of maximum frequency
413 /* Get Absolute Load */
414 for_each_cpu(j, policy->cpus) {
415 struct cpu_dbs_info_s *j_dbs_info;
416 cputime64_t cur_wall_time, cur_idle_time;
417 unsigned int idle_time, wall_time;
419 j_dbs_info = &per_cpu(cpu_dbs_info, j);
421 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time);
423 wall_time = (unsigned int) cputime64_sub(cur_wall_time,
424 j_dbs_info->prev_cpu_wall);
425 j_dbs_info->prev_cpu_wall = cur_wall_time;
427 idle_time = (unsigned int) cputime64_sub(cur_idle_time,
428 j_dbs_info->prev_cpu_idle);
429 j_dbs_info->prev_cpu_idle = cur_idle_time;
431 if (dbs_tuners_ins.ignore_nice) {
432 cputime64_t cur_nice;
433 unsigned long cur_nice_jiffies;
435 cur_nice = cputime64_sub(kstat_cpu(j).cpustat.nice,
436 j_dbs_info->prev_cpu_nice);
438 * Assumption: nice time between sampling periods will
439 * be less than 2^32 jiffies for 32 bit sys
441 cur_nice_jiffies = (unsigned long)
442 cputime64_to_jiffies64(cur_nice);
444 j_dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
445 idle_time += jiffies_to_usecs(cur_nice_jiffies);
448 if (unlikely(!wall_time || wall_time < idle_time))
451 load = 100 * (wall_time - idle_time) / wall_time;
455 * break out if we 'cannot' reduce the speed as the user might
456 * want freq_step to be zero
458 if (dbs_tuners_ins.freq_step == 0)
461 /* Check for frequency increase */
462 if (load > dbs_tuners_ins.up_threshold) {
463 this_dbs_info->down_skip = 0;
465 /* if we are already at full speed then break out early */
466 if (this_dbs_info->requested_freq == policy->max)
469 freq_target = (dbs_tuners_ins.freq_step * policy->max) / 100;
471 /* max freq cannot be less than 100. But who knows.... */
472 if (unlikely(freq_target == 0))
475 this_dbs_info->requested_freq += freq_target;
476 if (this_dbs_info->requested_freq > policy->max)
477 this_dbs_info->requested_freq = policy->max;
479 __cpufreq_driver_target(policy, this_dbs_info->requested_freq,
485 * The optimal frequency is the frequency that is the lowest that
486 * can support the current CPU usage without triggering the up
487 * policy. To be safe, we focus 10 points under the threshold.
489 if (load < (dbs_tuners_ins.down_threshold - 10)) {
490 freq_target = (dbs_tuners_ins.freq_step * policy->max) / 100;
492 this_dbs_info->requested_freq -= freq_target;
493 if (this_dbs_info->requested_freq < policy->min)
494 this_dbs_info->requested_freq = policy->min;
497 * if we cannot reduce the frequency anymore, break out early
499 if (policy->cur == policy->min)
502 __cpufreq_driver_target(policy, this_dbs_info->requested_freq,
508 static void do_dbs_timer(struct work_struct *work)
510 struct cpu_dbs_info_s *dbs_info =
511 container_of(work, struct cpu_dbs_info_s, work.work);
512 unsigned int cpu = dbs_info->cpu;
514 /* We want all CPUs to do sampling nearly on same jiffy */
515 int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
517 delay -= jiffies % delay;
519 if (lock_policy_rwsem_write(cpu) < 0)
522 if (!dbs_info->enable) {
523 unlock_policy_rwsem_write(cpu);
527 dbs_check_cpu(dbs_info);
529 queue_delayed_work_on(cpu, kconservative_wq, &dbs_info->work, delay);
530 unlock_policy_rwsem_write(cpu);
533 static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
535 /* We want all CPUs to do sampling nearly on same jiffy */
536 int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
537 delay -= jiffies % delay;
539 dbs_info->enable = 1;
540 INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
541 queue_delayed_work_on(dbs_info->cpu, kconservative_wq, &dbs_info->work,
545 static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
547 dbs_info->enable = 0;
548 cancel_delayed_work_sync(&dbs_info->work);
551 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
554 unsigned int cpu = policy->cpu;
555 struct cpu_dbs_info_s *this_dbs_info;
559 this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
562 case CPUFREQ_GOV_START:
563 if ((!cpu_online(cpu)) || (!policy->cur))
566 if (this_dbs_info->enable) /* Already enabled */
569 mutex_lock(&dbs_mutex);
571 rc = sysfs_create_group(&policy->kobj, &dbs_attr_group);
573 mutex_unlock(&dbs_mutex);
577 for_each_cpu(j, policy->cpus) {
578 struct cpu_dbs_info_s *j_dbs_info;
579 j_dbs_info = &per_cpu(cpu_dbs_info, j);
580 j_dbs_info->cur_policy = policy;
582 j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
583 &j_dbs_info->prev_cpu_wall);
584 if (dbs_tuners_ins.ignore_nice) {
585 j_dbs_info->prev_cpu_nice =
586 kstat_cpu(j).cpustat.nice;
589 this_dbs_info->down_skip = 0;
590 this_dbs_info->requested_freq = policy->cur;
594 * Start the timerschedule work, when this governor
595 * is used for first time
597 if (dbs_enable == 1) {
598 unsigned int latency;
599 /* policy latency is in nS. Convert it to uS first */
600 latency = policy->cpuinfo.transition_latency / 1000;
605 max(latency * LATENCY_MULTIPLIER,
606 MIN_STAT_SAMPLING_RATE);
608 dbs_tuners_ins.sampling_rate = def_sampling_rate;
610 cpufreq_register_notifier(
611 &dbs_cpufreq_notifier_block,
612 CPUFREQ_TRANSITION_NOTIFIER);
614 dbs_timer_init(this_dbs_info);
616 mutex_unlock(&dbs_mutex);
620 case CPUFREQ_GOV_STOP:
621 mutex_lock(&dbs_mutex);
622 dbs_timer_exit(this_dbs_info);
623 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
627 * Stop the timerschedule work, when this governor
628 * is used for first time
631 cpufreq_unregister_notifier(
632 &dbs_cpufreq_notifier_block,
633 CPUFREQ_TRANSITION_NOTIFIER);
635 mutex_unlock(&dbs_mutex);
639 case CPUFREQ_GOV_LIMITS:
640 mutex_lock(&dbs_mutex);
641 if (policy->max < this_dbs_info->cur_policy->cur)
642 __cpufreq_driver_target(
643 this_dbs_info->cur_policy,
644 policy->max, CPUFREQ_RELATION_H);
645 else if (policy->min > this_dbs_info->cur_policy->cur)
646 __cpufreq_driver_target(
647 this_dbs_info->cur_policy,
648 policy->min, CPUFREQ_RELATION_L);
649 mutex_unlock(&dbs_mutex);
656 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_CONSERVATIVE
659 struct cpufreq_governor cpufreq_gov_conservative = {
660 .name = "conservative",
661 .governor = cpufreq_governor_dbs,
662 .max_transition_latency = TRANSITION_LATENCY_LIMIT,
663 .owner = THIS_MODULE,
666 static int __init cpufreq_gov_dbs_init(void)
670 kconservative_wq = create_workqueue("kconservative");
671 if (!kconservative_wq) {
672 printk(KERN_ERR "Creation of kconservative failed\n");
676 err = cpufreq_register_governor(&cpufreq_gov_conservative);
678 destroy_workqueue(kconservative_wq);
683 static void __exit cpufreq_gov_dbs_exit(void)
685 cpufreq_unregister_governor(&cpufreq_gov_conservative);
686 destroy_workqueue(kconservative_wq);
690 MODULE_AUTHOR("Alexander Clouter <alex@digriz.org.uk>");
691 MODULE_DESCRIPTION("'cpufreq_conservative' - A dynamic cpufreq governor for "
692 "Low Latency Frequency Transition capable processors "
693 "optimised for use in a battery environment");
694 MODULE_LICENSE("GPL");
696 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_CONSERVATIVE
697 fs_initcall(cpufreq_gov_dbs_init);
699 module_init(cpufreq_gov_dbs_init);
701 module_exit(cpufreq_gov_dbs_exit);