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) 2004 Alexander Clouter <alex-kernel@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/smp.h>
17 #include <linux/init.h>
18 #include <linux/interrupt.h>
19 #include <linux/ctype.h>
20 #include <linux/cpufreq.h>
21 #include <linux/sysctl.h>
22 #include <linux/types.h>
24 #include <linux/sysfs.h>
25 #include <linux/cpu.h>
26 #include <linux/kmod.h>
27 #include <linux/workqueue.h>
28 #include <linux/jiffies.h>
29 #include <linux/kernel_stat.h>
30 #include <linux/percpu.h>
31 #include <linux/mutex.h>
33 * dbs is used in this file as a shortform for demandbased switching
34 * It helps to keep variable names smaller, simpler
37 #define DEF_FREQUENCY_UP_THRESHOLD (80)
38 #define DEF_FREQUENCY_DOWN_THRESHOLD (20)
41 * The polling frequency of this governor depends on the capability of
42 * the processor. Default polling frequency is 1000 times the transition
43 * latency of the processor. The governor will work on any processor with
44 * transition latency <= 10mS, using appropriate sampling
46 * For CPUs with transition latency > 10mS (mostly drivers
47 * with CPUFREQ_ETERNAL), this governor will not work.
48 * All times here are in uS.
50 static unsigned int def_sampling_rate;
51 #define MIN_SAMPLING_RATE_RATIO (2)
52 /* for correct statistics, we need at least 10 ticks between each measure */
53 #define MIN_STAT_SAMPLING_RATE \
54 (MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10))
55 #define MIN_SAMPLING_RATE \
56 (def_sampling_rate / MIN_SAMPLING_RATE_RATIO)
57 #define MAX_SAMPLING_RATE (500 * def_sampling_rate)
58 #define DEF_SAMPLING_RATE_LATENCY_MULTIPLIER (1000)
59 #define DEF_SAMPLING_DOWN_FACTOR (1)
60 #define MAX_SAMPLING_DOWN_FACTOR (10)
61 #define TRANSITION_LATENCY_LIMIT (10 * 1000 * 1000)
63 static void do_dbs_timer(struct work_struct *work);
65 struct cpu_dbs_info_s {
66 struct cpufreq_policy *cur_policy;
67 unsigned int prev_cpu_idle_up;
68 unsigned int prev_cpu_idle_down;
70 unsigned int down_skip;
71 unsigned int requested_freq;
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 * DEADLOCK ALERT! There is a ordering requirement between cpu_hotplug
79 * lock and dbs_mutex. cpu_hotplug lock should always be held before
80 * dbs_mutex. If any function that can potentially take cpu_hotplug lock
81 * (like __cpufreq_driver_target()) is being called with dbs_mutex taken, then
82 * cpu_hotplug lock should be taken before that. Note that cpu_hotplug lock
83 * is recursive for the same process. -Venki
85 static DEFINE_MUTEX(dbs_mutex);
86 static DECLARE_DELAYED_WORK(dbs_work, do_dbs_timer);
89 unsigned int sampling_rate;
90 unsigned int sampling_down_factor;
91 unsigned int up_threshold;
92 unsigned int down_threshold;
93 unsigned int ignore_nice;
94 unsigned int freq_step;
97 static struct dbs_tuners dbs_tuners_ins = {
98 .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
99 .down_threshold = DEF_FREQUENCY_DOWN_THRESHOLD,
100 .sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR,
105 static inline unsigned int get_cpu_idle_time(unsigned int cpu)
107 unsigned int add_nice = 0, ret;
109 if (dbs_tuners_ins.ignore_nice)
110 add_nice = kstat_cpu(cpu).cpustat.nice;
112 ret = kstat_cpu(cpu).cpustat.idle +
113 kstat_cpu(cpu).cpustat.iowait +
119 /* keep track of frequency transitions */
121 dbs_cpufreq_notifier(struct notifier_block *nb, unsigned long val,
124 struct cpufreq_freqs *freq = data;
125 struct cpu_dbs_info_s *this_dbs_info = &per_cpu(cpu_dbs_info,
128 if (!this_dbs_info->enable)
131 this_dbs_info->requested_freq = freq->new;
136 static struct notifier_block dbs_cpufreq_notifier_block = {
137 .notifier_call = dbs_cpufreq_notifier
140 /************************** sysfs interface ************************/
141 static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf)
143 static int print_once;
146 printk(KERN_INFO "CPUFREQ: conservative sampling_rate_max "
147 "sysfs file is deprecated - used by: %s\n",
151 return sprintf(buf, "%u\n", MAX_SAMPLING_RATE);
154 static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf)
156 static int print_once;
159 printk(KERN_INFO "CPUFREQ: conservative sampling_rate_max "
160 "sysfs file is deprecated - used by: %s\n", current->comm);
163 return sprintf(buf, "%u\n", MIN_SAMPLING_RATE);
166 #define define_one_ro(_name) \
167 static struct freq_attr _name = \
168 __ATTR(_name, 0444, show_##_name, NULL)
170 define_one_ro(sampling_rate_max);
171 define_one_ro(sampling_rate_min);
173 /* cpufreq_conservative Governor Tunables */
174 #define show_one(file_name, object) \
175 static ssize_t show_##file_name \
176 (struct cpufreq_policy *unused, char *buf) \
178 return sprintf(buf, "%u\n", dbs_tuners_ins.object); \
180 show_one(sampling_rate, sampling_rate);
181 show_one(sampling_down_factor, sampling_down_factor);
182 show_one(up_threshold, up_threshold);
183 show_one(down_threshold, down_threshold);
184 show_one(ignore_nice_load, ignore_nice);
185 show_one(freq_step, freq_step);
187 static ssize_t store_sampling_down_factor(struct cpufreq_policy *unused,
188 const char *buf, size_t count)
192 ret = sscanf(buf, "%u", &input);
193 if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
196 mutex_lock(&dbs_mutex);
197 dbs_tuners_ins.sampling_down_factor = input;
198 mutex_unlock(&dbs_mutex);
203 static ssize_t store_sampling_rate(struct cpufreq_policy *unused,
204 const char *buf, size_t count)
208 ret = sscanf(buf, "%u", &input);
210 mutex_lock(&dbs_mutex);
211 if (ret != 1 || input > MAX_SAMPLING_RATE ||
212 input < MIN_SAMPLING_RATE) {
213 mutex_unlock(&dbs_mutex);
217 dbs_tuners_ins.sampling_rate = input;
218 mutex_unlock(&dbs_mutex);
223 static ssize_t store_up_threshold(struct cpufreq_policy *unused,
224 const char *buf, size_t count)
228 ret = sscanf(buf, "%u", &input);
230 mutex_lock(&dbs_mutex);
231 if (ret != 1 || input > 100 ||
232 input <= dbs_tuners_ins.down_threshold) {
233 mutex_unlock(&dbs_mutex);
237 dbs_tuners_ins.up_threshold = input;
238 mutex_unlock(&dbs_mutex);
243 static ssize_t store_down_threshold(struct cpufreq_policy *unused,
244 const char *buf, size_t count)
248 ret = sscanf(buf, "%u", &input);
250 mutex_lock(&dbs_mutex);
251 if (ret != 1 || input > 100 || input >= dbs_tuners_ins.up_threshold) {
252 mutex_unlock(&dbs_mutex);
256 dbs_tuners_ins.down_threshold = input;
257 mutex_unlock(&dbs_mutex);
262 static ssize_t store_ignore_nice_load(struct cpufreq_policy *policy,
263 const char *buf, size_t count)
270 ret = sscanf(buf, "%u", &input);
277 mutex_lock(&dbs_mutex);
278 if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */
279 mutex_unlock(&dbs_mutex);
282 dbs_tuners_ins.ignore_nice = input;
284 /* we need to re-evaluate prev_cpu_idle_up and prev_cpu_idle_down */
285 for_each_online_cpu(j) {
286 struct cpu_dbs_info_s *j_dbs_info;
287 j_dbs_info = &per_cpu(cpu_dbs_info, j);
288 j_dbs_info->prev_cpu_idle_up = get_cpu_idle_time(j);
289 j_dbs_info->prev_cpu_idle_down = j_dbs_info->prev_cpu_idle_up;
291 mutex_unlock(&dbs_mutex);
296 static ssize_t store_freq_step(struct cpufreq_policy *policy,
297 const char *buf, size_t count)
302 ret = sscanf(buf, "%u", &input);
310 /* no need to test here if freq_step is zero as the user might actually
311 * want this, they would be crazy though :) */
312 mutex_lock(&dbs_mutex);
313 dbs_tuners_ins.freq_step = input;
314 mutex_unlock(&dbs_mutex);
319 #define define_one_rw(_name) \
320 static struct freq_attr _name = \
321 __ATTR(_name, 0644, show_##_name, store_##_name)
323 define_one_rw(sampling_rate);
324 define_one_rw(sampling_down_factor);
325 define_one_rw(up_threshold);
326 define_one_rw(down_threshold);
327 define_one_rw(ignore_nice_load);
328 define_one_rw(freq_step);
330 static struct attribute *dbs_attributes[] = {
331 &sampling_rate_max.attr,
332 &sampling_rate_min.attr,
334 &sampling_down_factor.attr,
336 &down_threshold.attr,
337 &ignore_nice_load.attr,
342 static struct attribute_group dbs_attr_group = {
343 .attrs = dbs_attributes,
344 .name = "conservative",
347 /************************** sysfs end ************************/
349 static void dbs_check_cpu(int cpu)
351 unsigned int idle_ticks, up_idle_ticks, down_idle_ticks;
352 unsigned int tmp_idle_ticks, total_idle_ticks;
353 unsigned int freq_target;
354 unsigned int freq_down_sampling_rate;
355 struct cpu_dbs_info_s *this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
356 struct cpufreq_policy *policy;
358 if (!this_dbs_info->enable)
361 policy = this_dbs_info->cur_policy;
364 * The default safe range is 20% to 80%
365 * Every sampling_rate, we check
366 * - If current idle time is less than 20%, then we try to
368 * Every sampling_rate*sampling_down_factor, we check
369 * - If current idle time is more than 80%, then we try to
372 * Any frequency increase takes it to the maximum frequency.
373 * Frequency reduction happens at minimum steps of
374 * 5% (default) of max_frequency
377 /* Check for frequency increase */
378 idle_ticks = UINT_MAX;
380 /* Check for frequency increase */
381 total_idle_ticks = get_cpu_idle_time(cpu);
382 tmp_idle_ticks = total_idle_ticks -
383 this_dbs_info->prev_cpu_idle_up;
384 this_dbs_info->prev_cpu_idle_up = total_idle_ticks;
386 if (tmp_idle_ticks < idle_ticks)
387 idle_ticks = tmp_idle_ticks;
389 /* Scale idle ticks by 100 and compare with up and down ticks */
391 up_idle_ticks = (100 - dbs_tuners_ins.up_threshold) *
392 usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
394 if (idle_ticks < up_idle_ticks) {
395 this_dbs_info->down_skip = 0;
396 this_dbs_info->prev_cpu_idle_down =
397 this_dbs_info->prev_cpu_idle_up;
399 /* if we are already at full speed then break out early */
400 if (this_dbs_info->requested_freq == policy->max)
403 freq_target = (dbs_tuners_ins.freq_step * policy->max) / 100;
405 /* max freq cannot be less than 100. But who knows.... */
406 if (unlikely(freq_target == 0))
409 this_dbs_info->requested_freq += freq_target;
410 if (this_dbs_info->requested_freq > policy->max)
411 this_dbs_info->requested_freq = policy->max;
413 __cpufreq_driver_target(policy, this_dbs_info->requested_freq,
418 /* Check for frequency decrease */
419 this_dbs_info->down_skip++;
420 if (this_dbs_info->down_skip < dbs_tuners_ins.sampling_down_factor)
423 /* Check for frequency decrease */
424 total_idle_ticks = this_dbs_info->prev_cpu_idle_up;
425 tmp_idle_ticks = total_idle_ticks -
426 this_dbs_info->prev_cpu_idle_down;
427 this_dbs_info->prev_cpu_idle_down = total_idle_ticks;
429 if (tmp_idle_ticks < idle_ticks)
430 idle_ticks = tmp_idle_ticks;
432 /* Scale idle ticks by 100 and compare with up and down ticks */
434 this_dbs_info->down_skip = 0;
436 freq_down_sampling_rate = dbs_tuners_ins.sampling_rate *
437 dbs_tuners_ins.sampling_down_factor;
438 down_idle_ticks = (100 - dbs_tuners_ins.down_threshold) *
439 usecs_to_jiffies(freq_down_sampling_rate);
441 if (idle_ticks > down_idle_ticks) {
443 * if we are already at the lowest speed then break out early
444 * or if we 'cannot' reduce the speed as the user might want
445 * freq_target to be zero
447 if (this_dbs_info->requested_freq == policy->min
448 || dbs_tuners_ins.freq_step == 0)
451 freq_target = (dbs_tuners_ins.freq_step * policy->max) / 100;
453 /* max freq cannot be less than 100. But who knows.... */
454 if (unlikely(freq_target == 0))
457 this_dbs_info->requested_freq -= freq_target;
458 if (this_dbs_info->requested_freq < policy->min)
459 this_dbs_info->requested_freq = policy->min;
461 __cpufreq_driver_target(policy, this_dbs_info->requested_freq,
467 static void do_dbs_timer(struct work_struct *work)
470 mutex_lock(&dbs_mutex);
471 for_each_online_cpu(i)
473 schedule_delayed_work(&dbs_work,
474 usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
475 mutex_unlock(&dbs_mutex);
478 static inline void dbs_timer_init(void)
480 init_timer_deferrable(&dbs_work.timer);
481 schedule_delayed_work(&dbs_work,
482 usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
486 static inline void dbs_timer_exit(void)
488 cancel_delayed_work(&dbs_work);
492 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
495 unsigned int cpu = policy->cpu;
496 struct cpu_dbs_info_s *this_dbs_info;
500 this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
503 case CPUFREQ_GOV_START:
504 if ((!cpu_online(cpu)) || (!policy->cur))
507 if (this_dbs_info->enable) /* Already enabled */
510 mutex_lock(&dbs_mutex);
512 rc = sysfs_create_group(&policy->kobj, &dbs_attr_group);
514 mutex_unlock(&dbs_mutex);
518 for_each_cpu(j, policy->cpus) {
519 struct cpu_dbs_info_s *j_dbs_info;
520 j_dbs_info = &per_cpu(cpu_dbs_info, j);
521 j_dbs_info->cur_policy = policy;
523 j_dbs_info->prev_cpu_idle_up = get_cpu_idle_time(cpu);
524 j_dbs_info->prev_cpu_idle_down
525 = j_dbs_info->prev_cpu_idle_up;
527 this_dbs_info->enable = 1;
528 this_dbs_info->down_skip = 0;
529 this_dbs_info->requested_freq = policy->cur;
533 * Start the timerschedule work, when this governor
534 * is used for first time
536 if (dbs_enable == 1) {
537 unsigned int latency;
538 /* policy latency is in nS. Convert it to uS first */
539 latency = policy->cpuinfo.transition_latency / 1000;
543 def_sampling_rate = 10 * latency *
544 DEF_SAMPLING_RATE_LATENCY_MULTIPLIER;
546 if (def_sampling_rate < MIN_STAT_SAMPLING_RATE)
547 def_sampling_rate = MIN_STAT_SAMPLING_RATE;
549 dbs_tuners_ins.sampling_rate = def_sampling_rate;
552 cpufreq_register_notifier(
553 &dbs_cpufreq_notifier_block,
554 CPUFREQ_TRANSITION_NOTIFIER);
557 mutex_unlock(&dbs_mutex);
560 case CPUFREQ_GOV_STOP:
561 mutex_lock(&dbs_mutex);
562 this_dbs_info->enable = 0;
563 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
566 * Stop the timerschedule work, when this governor
567 * is used for first time
569 if (dbs_enable == 0) {
571 cpufreq_unregister_notifier(
572 &dbs_cpufreq_notifier_block,
573 CPUFREQ_TRANSITION_NOTIFIER);
576 mutex_unlock(&dbs_mutex);
580 case CPUFREQ_GOV_LIMITS:
581 mutex_lock(&dbs_mutex);
582 if (policy->max < this_dbs_info->cur_policy->cur)
583 __cpufreq_driver_target(
584 this_dbs_info->cur_policy,
585 policy->max, CPUFREQ_RELATION_H);
586 else if (policy->min > this_dbs_info->cur_policy->cur)
587 __cpufreq_driver_target(
588 this_dbs_info->cur_policy,
589 policy->min, CPUFREQ_RELATION_L);
590 mutex_unlock(&dbs_mutex);
596 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_CONSERVATIVE
599 struct cpufreq_governor cpufreq_gov_conservative = {
600 .name = "conservative",
601 .governor = cpufreq_governor_dbs,
602 .max_transition_latency = TRANSITION_LATENCY_LIMIT,
603 .owner = THIS_MODULE,
606 static int __init cpufreq_gov_dbs_init(void)
608 return cpufreq_register_governor(&cpufreq_gov_conservative);
611 static void __exit cpufreq_gov_dbs_exit(void)
613 /* Make sure that the scheduled work is indeed not running */
614 flush_scheduled_work();
616 cpufreq_unregister_governor(&cpufreq_gov_conservative);
620 MODULE_AUTHOR("Alexander Clouter <alex-kernel@digriz.org.uk>");
621 MODULE_DESCRIPTION("'cpufreq_conservative' - A dynamic cpufreq governor for "
622 "Low Latency Frequency Transition capable processors "
623 "optimised for use in a battery environment");
624 MODULE_LICENSE("GPL");
626 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_CONSERVATIVE
627 fs_initcall(cpufreq_gov_dbs_init);
629 module_init(cpufreq_gov_dbs_init);
631 module_exit(cpufreq_gov_dbs_exit);