2 * drivers/cpufreq/cpufreq_ondemand.c
4 * Copyright (C) 2001 Russell King
5 * (C) 2003 Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>.
6 * Jun Nakajima <jun.nakajima@intel.com>
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 as
10 * published by the Free Software Foundation.
13 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
15 #include <linux/cpufreq.h>
16 #include <linux/init.h>
17 #include <linux/kernel.h>
18 #include <linux/kernel_stat.h>
19 #include <linux/kobject.h>
20 #include <linux/module.h>
21 #include <linux/mutex.h>
22 #include <linux/percpu-defs.h>
23 #include <linux/slab.h>
24 #include <linux/sysfs.h>
25 #include <linux/tick.h>
26 #include <linux/types.h>
27 #include <linux/cpu.h>
29 #include "cpufreq_governor.h"
31 /* On-demand governor macros */
32 #define DEF_FREQUENCY_DOWN_DIFFERENTIAL (10)
33 #define DEF_FREQUENCY_UP_THRESHOLD (80)
34 #define DEF_SAMPLING_DOWN_FACTOR (1)
35 #define MAX_SAMPLING_DOWN_FACTOR (100000)
36 #define MICRO_FREQUENCY_DOWN_DIFFERENTIAL (3)
37 #define MICRO_FREQUENCY_UP_THRESHOLD (95)
38 #define MICRO_FREQUENCY_MIN_SAMPLE_RATE (10000)
39 #define MIN_FREQUENCY_UP_THRESHOLD (11)
40 #define MAX_FREQUENCY_UP_THRESHOLD (100)
42 static DEFINE_PER_CPU(struct od_cpu_dbs_info_s, od_cpu_dbs_info);
44 static struct od_ops od_ops;
46 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
47 static struct cpufreq_governor cpufreq_gov_ondemand;
50 static void ondemand_powersave_bias_init_cpu(int cpu)
52 struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
54 dbs_info->freq_table = cpufreq_frequency_get_table(cpu);
55 dbs_info->freq_lo = 0;
59 * Not all CPUs want IO time to be accounted as busy; this depends on how
60 * efficient idling at a higher frequency/voltage is.
61 * Pavel Machek says this is not so for various generations of AMD and old
63 * Mike Chan (android.com) claims this is also not true for ARM.
64 * Because of this, whitelist specific known (series) of CPUs by default, and
65 * leave all others up to the user.
67 static int should_io_be_busy(void)
69 #if defined(CONFIG_X86)
71 * For Intel, Core 2 (model 15) and later have an efficient idle.
73 if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
74 boot_cpu_data.x86 == 6 &&
75 boot_cpu_data.x86_model >= 15)
82 * Find right freq to be set now with powersave_bias on.
83 * Returns the freq_hi to be used right now and will set freq_hi_jiffies,
84 * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs.
86 static unsigned int generic_powersave_bias_target(struct cpufreq_policy *policy,
87 unsigned int freq_next, unsigned int relation)
89 unsigned int freq_req, freq_reduc, freq_avg;
90 unsigned int freq_hi, freq_lo;
91 unsigned int index = 0;
92 unsigned int jiffies_total, jiffies_hi, jiffies_lo;
93 struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
95 struct dbs_data *dbs_data = policy->governor_data;
96 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
98 if (!dbs_info->freq_table) {
99 dbs_info->freq_lo = 0;
100 dbs_info->freq_lo_jiffies = 0;
104 cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next,
106 freq_req = dbs_info->freq_table[index].frequency;
107 freq_reduc = freq_req * od_tuners->powersave_bias / 1000;
108 freq_avg = freq_req - freq_reduc;
110 /* Find freq bounds for freq_avg in freq_table */
112 cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
113 CPUFREQ_RELATION_H, &index);
114 freq_lo = dbs_info->freq_table[index].frequency;
116 cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
117 CPUFREQ_RELATION_L, &index);
118 freq_hi = dbs_info->freq_table[index].frequency;
120 /* Find out how long we have to be in hi and lo freqs */
121 if (freq_hi == freq_lo) {
122 dbs_info->freq_lo = 0;
123 dbs_info->freq_lo_jiffies = 0;
126 jiffies_total = usecs_to_jiffies(od_tuners->sampling_rate);
127 jiffies_hi = (freq_avg - freq_lo) * jiffies_total;
128 jiffies_hi += ((freq_hi - freq_lo) / 2);
129 jiffies_hi /= (freq_hi - freq_lo);
130 jiffies_lo = jiffies_total - jiffies_hi;
131 dbs_info->freq_lo = freq_lo;
132 dbs_info->freq_lo_jiffies = jiffies_lo;
133 dbs_info->freq_hi_jiffies = jiffies_hi;
137 static void ondemand_powersave_bias_init(void)
140 for_each_online_cpu(i) {
141 ondemand_powersave_bias_init_cpu(i);
145 static void dbs_freq_increase(struct cpufreq_policy *p, unsigned int freq)
147 struct dbs_data *dbs_data = p->governor_data;
148 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
150 if (od_tuners->powersave_bias)
151 freq = od_ops.powersave_bias_target(p, freq,
153 else if (p->cur == p->max)
156 __cpufreq_driver_target(p, freq, od_tuners->powersave_bias ?
157 CPUFREQ_RELATION_L : CPUFREQ_RELATION_H);
161 * Every sampling_rate, we check, if current idle time is less than 20%
162 * (default), then we try to increase frequency. Every sampling_rate, we look
163 * for the lowest frequency which can sustain the load while keeping idle time
164 * over 30%. If such a frequency exist, we try to decrease to this frequency.
166 * Any frequency increase takes it to the maximum frequency. Frequency reduction
167 * happens at minimum steps of 5% (default) of current frequency
169 static void od_check_cpu(int cpu, unsigned int load_freq)
171 struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
172 struct cpufreq_policy *policy = dbs_info->cdbs.cur_policy;
173 struct dbs_data *dbs_data = policy->governor_data;
174 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
176 dbs_info->freq_lo = 0;
178 /* Check for frequency increase */
179 if (load_freq > od_tuners->up_threshold * policy->cur) {
180 /* If switching to max speed, apply sampling_down_factor */
181 if (policy->cur < policy->max)
182 dbs_info->rate_mult =
183 od_tuners->sampling_down_factor;
184 dbs_freq_increase(policy, policy->max);
188 /* Check for frequency decrease */
189 /* if we cannot reduce the frequency anymore, break out early */
190 if (policy->cur == policy->min)
194 * The optimal frequency is the frequency that is the lowest that can
195 * support the current CPU usage without triggering the up policy. To be
196 * safe, we focus 10 points under the threshold.
198 if (load_freq < od_tuners->adj_up_threshold
200 unsigned int freq_next;
201 freq_next = load_freq / od_tuners->adj_up_threshold;
203 /* No longer fully busy, reset rate_mult */
204 dbs_info->rate_mult = 1;
206 if (freq_next < policy->min)
207 freq_next = policy->min;
209 if (!od_tuners->powersave_bias) {
210 __cpufreq_driver_target(policy, freq_next,
215 freq_next = od_ops.powersave_bias_target(policy, freq_next,
217 __cpufreq_driver_target(policy, freq_next, CPUFREQ_RELATION_L);
221 static void od_dbs_timer(struct work_struct *work)
223 struct od_cpu_dbs_info_s *dbs_info =
224 container_of(work, struct od_cpu_dbs_info_s, cdbs.work.work);
225 unsigned int cpu = dbs_info->cdbs.cur_policy->cpu;
226 struct od_cpu_dbs_info_s *core_dbs_info = &per_cpu(od_cpu_dbs_info,
228 struct dbs_data *dbs_data = dbs_info->cdbs.cur_policy->governor_data;
229 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
230 int delay = 0, sample_type = core_dbs_info->sample_type;
231 bool modify_all = true;
233 mutex_lock(&core_dbs_info->cdbs.timer_mutex);
234 if (!need_load_eval(&core_dbs_info->cdbs, od_tuners->sampling_rate)) {
239 /* Common NORMAL_SAMPLE setup */
240 core_dbs_info->sample_type = OD_NORMAL_SAMPLE;
241 if (sample_type == OD_SUB_SAMPLE) {
242 delay = core_dbs_info->freq_lo_jiffies;
243 __cpufreq_driver_target(core_dbs_info->cdbs.cur_policy,
244 core_dbs_info->freq_lo, CPUFREQ_RELATION_H);
246 dbs_check_cpu(dbs_data, cpu);
247 if (core_dbs_info->freq_lo) {
248 /* Setup timer for SUB_SAMPLE */
249 core_dbs_info->sample_type = OD_SUB_SAMPLE;
250 delay = core_dbs_info->freq_hi_jiffies;
256 delay = delay_for_sampling_rate(od_tuners->sampling_rate
257 * core_dbs_info->rate_mult);
259 gov_queue_work(dbs_data, dbs_info->cdbs.cur_policy, delay, modify_all);
260 mutex_unlock(&core_dbs_info->cdbs.timer_mutex);
263 /************************** sysfs interface ************************/
264 static struct common_dbs_data od_dbs_cdata;
267 * update_sampling_rate - update sampling rate effective immediately if needed.
268 * @new_rate: new sampling rate
270 * If new rate is smaller than the old, simply updating
271 * dbs_tuners_int.sampling_rate might not be appropriate. For example, if the
272 * original sampling_rate was 1 second and the requested new sampling rate is 10
273 * ms because the user needs immediate reaction from ondemand governor, but not
274 * sure if higher frequency will be required or not, then, the governor may
275 * change the sampling rate too late; up to 1 second later. Thus, if we are
276 * reducing the sampling rate, we need to make the new value effective
279 static void update_sampling_rate(struct dbs_data *dbs_data,
280 unsigned int new_rate)
282 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
285 od_tuners->sampling_rate = new_rate = max(new_rate,
286 dbs_data->min_sampling_rate);
288 for_each_online_cpu(cpu) {
289 struct cpufreq_policy *policy;
290 struct od_cpu_dbs_info_s *dbs_info;
291 unsigned long next_sampling, appointed_at;
293 policy = cpufreq_cpu_get(cpu);
296 if (policy->governor != &cpufreq_gov_ondemand) {
297 cpufreq_cpu_put(policy);
300 dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
301 cpufreq_cpu_put(policy);
303 mutex_lock(&dbs_info->cdbs.timer_mutex);
305 if (!delayed_work_pending(&dbs_info->cdbs.work)) {
306 mutex_unlock(&dbs_info->cdbs.timer_mutex);
310 next_sampling = jiffies + usecs_to_jiffies(new_rate);
311 appointed_at = dbs_info->cdbs.work.timer.expires;
313 if (time_before(next_sampling, appointed_at)) {
315 mutex_unlock(&dbs_info->cdbs.timer_mutex);
316 cancel_delayed_work_sync(&dbs_info->cdbs.work);
317 mutex_lock(&dbs_info->cdbs.timer_mutex);
319 gov_queue_work(dbs_data, dbs_info->cdbs.cur_policy,
320 usecs_to_jiffies(new_rate), true);
323 mutex_unlock(&dbs_info->cdbs.timer_mutex);
327 static ssize_t store_sampling_rate(struct dbs_data *dbs_data, const char *buf,
332 ret = sscanf(buf, "%u", &input);
336 update_sampling_rate(dbs_data, input);
340 static ssize_t store_io_is_busy(struct dbs_data *dbs_data, const char *buf,
343 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
348 ret = sscanf(buf, "%u", &input);
351 od_tuners->io_is_busy = !!input;
353 /* we need to re-evaluate prev_cpu_idle */
354 for_each_online_cpu(j) {
355 struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
357 dbs_info->cdbs.prev_cpu_idle = get_cpu_idle_time(j,
358 &dbs_info->cdbs.prev_cpu_wall, od_tuners->io_is_busy);
363 static ssize_t store_up_threshold(struct dbs_data *dbs_data, const char *buf,
366 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
369 ret = sscanf(buf, "%u", &input);
371 if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
372 input < MIN_FREQUENCY_UP_THRESHOLD) {
375 /* Calculate the new adj_up_threshold */
376 od_tuners->adj_up_threshold += input;
377 od_tuners->adj_up_threshold -= od_tuners->up_threshold;
379 od_tuners->up_threshold = input;
383 static ssize_t store_sampling_down_factor(struct dbs_data *dbs_data,
384 const char *buf, size_t count)
386 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
387 unsigned int input, j;
389 ret = sscanf(buf, "%u", &input);
391 if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
393 od_tuners->sampling_down_factor = input;
395 /* Reset down sampling multiplier in case it was active */
396 for_each_online_cpu(j) {
397 struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
399 dbs_info->rate_mult = 1;
404 static ssize_t store_ignore_nice(struct dbs_data *dbs_data, const char *buf,
407 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
413 ret = sscanf(buf, "%u", &input);
420 if (input == od_tuners->ignore_nice) { /* nothing to do */
423 od_tuners->ignore_nice = input;
425 /* we need to re-evaluate prev_cpu_idle */
426 for_each_online_cpu(j) {
427 struct od_cpu_dbs_info_s *dbs_info;
428 dbs_info = &per_cpu(od_cpu_dbs_info, j);
429 dbs_info->cdbs.prev_cpu_idle = get_cpu_idle_time(j,
430 &dbs_info->cdbs.prev_cpu_wall, od_tuners->io_is_busy);
431 if (od_tuners->ignore_nice)
432 dbs_info->cdbs.prev_cpu_nice =
433 kcpustat_cpu(j).cpustat[CPUTIME_NICE];
439 static ssize_t store_powersave_bias(struct dbs_data *dbs_data, const char *buf,
442 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
445 ret = sscanf(buf, "%u", &input);
453 od_tuners->powersave_bias = input;
454 ondemand_powersave_bias_init();
458 show_store_one(od, sampling_rate);
459 show_store_one(od, io_is_busy);
460 show_store_one(od, up_threshold);
461 show_store_one(od, sampling_down_factor);
462 show_store_one(od, ignore_nice);
463 show_store_one(od, powersave_bias);
464 declare_show_sampling_rate_min(od);
466 gov_sys_pol_attr_rw(sampling_rate);
467 gov_sys_pol_attr_rw(io_is_busy);
468 gov_sys_pol_attr_rw(up_threshold);
469 gov_sys_pol_attr_rw(sampling_down_factor);
470 gov_sys_pol_attr_rw(ignore_nice);
471 gov_sys_pol_attr_rw(powersave_bias);
472 gov_sys_pol_attr_ro(sampling_rate_min);
474 static struct attribute *dbs_attributes_gov_sys[] = {
475 &sampling_rate_min_gov_sys.attr,
476 &sampling_rate_gov_sys.attr,
477 &up_threshold_gov_sys.attr,
478 &sampling_down_factor_gov_sys.attr,
479 &ignore_nice_gov_sys.attr,
480 &powersave_bias_gov_sys.attr,
481 &io_is_busy_gov_sys.attr,
485 static struct attribute_group od_attr_group_gov_sys = {
486 .attrs = dbs_attributes_gov_sys,
490 static struct attribute *dbs_attributes_gov_pol[] = {
491 &sampling_rate_min_gov_pol.attr,
492 &sampling_rate_gov_pol.attr,
493 &up_threshold_gov_pol.attr,
494 &sampling_down_factor_gov_pol.attr,
495 &ignore_nice_gov_pol.attr,
496 &powersave_bias_gov_pol.attr,
497 &io_is_busy_gov_pol.attr,
501 static struct attribute_group od_attr_group_gov_pol = {
502 .attrs = dbs_attributes_gov_pol,
506 /************************** sysfs end ************************/
508 static int od_init(struct dbs_data *dbs_data)
510 struct od_dbs_tuners *tuners;
514 tuners = kzalloc(sizeof(struct od_dbs_tuners), GFP_KERNEL);
516 pr_err("%s: kzalloc failed\n", __func__);
521 idle_time = get_cpu_idle_time_us(cpu, NULL);
523 if (idle_time != -1ULL) {
524 /* Idle micro accounting is supported. Use finer thresholds */
525 tuners->up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
526 tuners->adj_up_threshold = MICRO_FREQUENCY_UP_THRESHOLD -
527 MICRO_FREQUENCY_DOWN_DIFFERENTIAL;
529 * In nohz/micro accounting case we set the minimum frequency
530 * not depending on HZ, but fixed (very low). The deferred
531 * timer might skip some samples if idle/sleeping as needed.
533 dbs_data->min_sampling_rate = MICRO_FREQUENCY_MIN_SAMPLE_RATE;
535 tuners->up_threshold = DEF_FREQUENCY_UP_THRESHOLD;
536 tuners->adj_up_threshold = DEF_FREQUENCY_UP_THRESHOLD -
537 DEF_FREQUENCY_DOWN_DIFFERENTIAL;
539 /* For correct statistics, we need 10 ticks for each measure */
540 dbs_data->min_sampling_rate = MIN_SAMPLING_RATE_RATIO *
541 jiffies_to_usecs(10);
544 tuners->sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR;
545 tuners->ignore_nice = 0;
546 tuners->powersave_bias = 0;
547 tuners->io_is_busy = should_io_be_busy();
549 dbs_data->tuners = tuners;
550 pr_info("%s: tuners %p\n", __func__, tuners);
551 mutex_init(&dbs_data->mutex);
555 static void od_exit(struct dbs_data *dbs_data)
557 kfree(dbs_data->tuners);
560 define_get_cpu_dbs_routines(od_cpu_dbs_info);
562 static struct od_ops od_ops = {
563 .powersave_bias_init_cpu = ondemand_powersave_bias_init_cpu,
564 .powersave_bias_target = generic_powersave_bias_target,
565 .freq_increase = dbs_freq_increase,
568 static struct common_dbs_data od_dbs_cdata = {
569 .governor = GOV_ONDEMAND,
570 .attr_group_gov_sys = &od_attr_group_gov_sys,
571 .attr_group_gov_pol = &od_attr_group_gov_pol,
572 .get_cpu_cdbs = get_cpu_cdbs,
573 .get_cpu_dbs_info_s = get_cpu_dbs_info_s,
574 .gov_dbs_timer = od_dbs_timer,
575 .gov_check_cpu = od_check_cpu,
581 static void od_set_powersave_bias(unsigned int powersave_bias)
583 struct cpufreq_policy *policy;
584 struct dbs_data *dbs_data;
585 struct od_dbs_tuners *od_tuners;
589 cpumask_clear(&done);
592 for_each_online_cpu(cpu) {
593 if (cpumask_test_cpu(cpu, &done))
596 policy = per_cpu(od_cpu_dbs_info, cpu).cdbs.cur_policy;
597 dbs_data = policy->governor_data;
598 od_tuners = dbs_data->tuners;
599 od_tuners->powersave_bias = powersave_bias;
601 cpumask_or(&done, &done, policy->cpus);
606 void od_register_powersave_bias_handler(unsigned int (*f)
607 (struct cpufreq_policy *, unsigned int, unsigned int),
608 unsigned int powersave_bias)
610 od_ops.powersave_bias_target = f;
611 od_set_powersave_bias(powersave_bias);
613 EXPORT_SYMBOL_GPL(od_register_powersave_bias_handler);
615 void od_unregister_powersave_bias_handler(void)
617 od_ops.powersave_bias_target = generic_powersave_bias_target;
618 od_set_powersave_bias(0);
620 EXPORT_SYMBOL_GPL(od_unregister_powersave_bias_handler);
622 static int od_cpufreq_governor_dbs(struct cpufreq_policy *policy,
625 return cpufreq_governor_dbs(policy, &od_dbs_cdata, event);
628 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
631 struct cpufreq_governor cpufreq_gov_ondemand = {
633 .governor = od_cpufreq_governor_dbs,
634 .max_transition_latency = TRANSITION_LATENCY_LIMIT,
635 .owner = THIS_MODULE,
638 static int __init cpufreq_gov_dbs_init(void)
640 return cpufreq_register_governor(&cpufreq_gov_ondemand);
643 static void __exit cpufreq_gov_dbs_exit(void)
645 cpufreq_unregister_governor(&cpufreq_gov_ondemand);
648 MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
649 MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
650 MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
651 "Low Latency Frequency Transition capable processors");
652 MODULE_LICENSE("GPL");
654 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
655 fs_initcall(cpufreq_gov_dbs_init);
657 module_init(cpufreq_gov_dbs_init);
659 module_exit(cpufreq_gov_dbs_exit);