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/sysfs.h>
24 #include <linux/tick.h>
25 #include <linux/types.h>
27 #include "cpufreq_governor.h"
29 /* On-demand governor macors */
30 #define DEF_FREQUENCY_DOWN_DIFFERENTIAL (10)
31 #define DEF_FREQUENCY_UP_THRESHOLD (80)
32 #define DEF_SAMPLING_DOWN_FACTOR (1)
33 #define MAX_SAMPLING_DOWN_FACTOR (100000)
34 #define MICRO_FREQUENCY_DOWN_DIFFERENTIAL (3)
35 #define MICRO_FREQUENCY_UP_THRESHOLD (95)
36 #define MICRO_FREQUENCY_MIN_SAMPLE_RATE (10000)
37 #define MIN_FREQUENCY_UP_THRESHOLD (11)
38 #define MAX_FREQUENCY_UP_THRESHOLD (100)
40 static struct dbs_data od_dbs_data;
41 static DEFINE_PER_CPU(struct od_cpu_dbs_info_s, od_cpu_dbs_info);
43 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
44 static struct cpufreq_governor cpufreq_gov_ondemand;
47 static struct od_dbs_tuners od_tuners = {
48 .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
49 .sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR,
50 .down_differential = DEF_FREQUENCY_DOWN_DIFFERENTIAL,
55 static void ondemand_powersave_bias_init_cpu(int cpu)
57 struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
59 dbs_info->freq_table = cpufreq_frequency_get_table(cpu);
60 dbs_info->freq_lo = 0;
64 * Not all CPUs want IO time to be accounted as busy; this depends on how
65 * efficient idling at a higher frequency/voltage is.
66 * Pavel Machek says this is not so for various generations of AMD and old
68 * Mike Chan (androidlcom) calis this is also not true for ARM.
69 * Because of this, whitelist specific known (series) of CPUs by default, and
70 * leave all others up to the user.
72 static int should_io_be_busy(void)
74 #if defined(CONFIG_X86)
76 * For Intel, Core 2 (model 15) andl later have an efficient idle.
78 if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
79 boot_cpu_data.x86 == 6 &&
80 boot_cpu_data.x86_model >= 15)
87 * Find right freq to be set now with powersave_bias on.
88 * Returns the freq_hi to be used right now and will set freq_hi_jiffies,
89 * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs.
91 static unsigned int powersave_bias_target(struct cpufreq_policy *policy,
92 unsigned int freq_next, unsigned int relation)
94 unsigned int freq_req, freq_reduc, freq_avg;
95 unsigned int freq_hi, freq_lo;
96 unsigned int index = 0;
97 unsigned int jiffies_total, jiffies_hi, jiffies_lo;
98 struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
101 if (!dbs_info->freq_table) {
102 dbs_info->freq_lo = 0;
103 dbs_info->freq_lo_jiffies = 0;
107 cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next,
109 freq_req = dbs_info->freq_table[index].frequency;
110 freq_reduc = freq_req * od_tuners.powersave_bias / 1000;
111 freq_avg = freq_req - freq_reduc;
113 /* Find freq bounds for freq_avg in freq_table */
115 cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
116 CPUFREQ_RELATION_H, &index);
117 freq_lo = dbs_info->freq_table[index].frequency;
119 cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
120 CPUFREQ_RELATION_L, &index);
121 freq_hi = dbs_info->freq_table[index].frequency;
123 /* Find out how long we have to be in hi and lo freqs */
124 if (freq_hi == freq_lo) {
125 dbs_info->freq_lo = 0;
126 dbs_info->freq_lo_jiffies = 0;
129 jiffies_total = usecs_to_jiffies(od_tuners.sampling_rate);
130 jiffies_hi = (freq_avg - freq_lo) * jiffies_total;
131 jiffies_hi += ((freq_hi - freq_lo) / 2);
132 jiffies_hi /= (freq_hi - freq_lo);
133 jiffies_lo = jiffies_total - jiffies_hi;
134 dbs_info->freq_lo = freq_lo;
135 dbs_info->freq_lo_jiffies = jiffies_lo;
136 dbs_info->freq_hi_jiffies = jiffies_hi;
140 static void ondemand_powersave_bias_init(void)
143 for_each_online_cpu(i) {
144 ondemand_powersave_bias_init_cpu(i);
148 static void dbs_freq_increase(struct cpufreq_policy *p, unsigned int freq)
150 if (od_tuners.powersave_bias)
151 freq = powersave_bias_target(p, freq, CPUFREQ_RELATION_H);
152 else if (p->cur == p->max)
155 __cpufreq_driver_target(p, freq, od_tuners.powersave_bias ?
156 CPUFREQ_RELATION_L : CPUFREQ_RELATION_H);
160 * Every sampling_rate, we check, if current idle time is less than 20%
161 * (default), then we try to increase frequency Every sampling_rate, we look for
162 * a the lowest frequency which can sustain the load while keeping idle time
163 * over 30%. If such a frequency exist, we try to decrease to this frequency.
165 * Any frequency increase takes it to the maximum frequency. Frequency reduction
166 * happens at minimum steps of 5% (default) of current frequency
168 static void od_check_cpu(int cpu, unsigned int load_freq)
170 struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
171 struct cpufreq_policy *policy = dbs_info->cdbs.cur_policy;
173 dbs_info->freq_lo = 0;
175 /* Check for frequency increase */
176 if (load_freq > od_tuners.up_threshold * policy->cur) {
177 /* If switching to max speed, apply sampling_down_factor */
178 if (policy->cur < policy->max)
179 dbs_info->rate_mult =
180 od_tuners.sampling_down_factor;
181 dbs_freq_increase(policy, policy->max);
185 /* Check for frequency decrease */
186 /* if we cannot reduce the frequency anymore, break out early */
187 if (policy->cur == policy->min)
191 * The optimal frequency is the frequency that is the lowest that can
192 * support the current CPU usage without triggering the up policy. To be
193 * safe, we focus 10 points under the threshold.
195 if (load_freq < (od_tuners.up_threshold - od_tuners.down_differential) *
197 unsigned int freq_next;
198 freq_next = load_freq / (od_tuners.up_threshold -
199 od_tuners.down_differential);
201 /* No longer fully busy, reset rate_mult */
202 dbs_info->rate_mult = 1;
204 if (freq_next < policy->min)
205 freq_next = policy->min;
207 if (!od_tuners.powersave_bias) {
208 __cpufreq_driver_target(policy, freq_next,
211 int freq = powersave_bias_target(policy, freq_next,
213 __cpufreq_driver_target(policy, freq,
219 static void od_timer_update(struct od_cpu_dbs_info_s *dbs_info, bool sample,
220 struct delayed_work *dw)
222 unsigned int cpu = dbs_info->cdbs.cpu;
223 int delay, sample_type = dbs_info->sample_type;
225 /* Common NORMAL_SAMPLE setup */
226 dbs_info->sample_type = OD_NORMAL_SAMPLE;
227 if (sample_type == OD_SUB_SAMPLE) {
228 delay = dbs_info->freq_lo_jiffies;
230 __cpufreq_driver_target(dbs_info->cdbs.cur_policy,
235 dbs_check_cpu(&od_dbs_data, cpu);
236 if (dbs_info->freq_lo) {
237 /* Setup timer for SUB_SAMPLE */
238 dbs_info->sample_type = OD_SUB_SAMPLE;
239 delay = dbs_info->freq_hi_jiffies;
241 delay = delay_for_sampling_rate(od_tuners.sampling_rate
242 * dbs_info->rate_mult);
246 schedule_delayed_work_on(smp_processor_id(), dw, delay);
249 static void od_timer_coordinated(struct od_cpu_dbs_info_s *dbs_info_local,
250 struct delayed_work *dw)
252 struct od_cpu_dbs_info_s *dbs_info;
257 /* use leader CPU's dbs_info */
258 dbs_info = &per_cpu(od_cpu_dbs_info, dbs_info_local->cdbs.cpu);
259 mutex_lock(&dbs_info->cdbs.timer_mutex);
261 time_now = ktime_get();
262 delta_us = ktime_us_delta(time_now, dbs_info->cdbs.time_stamp);
264 /* Do nothing if we recently have sampled */
265 if (delta_us < (s64)(od_tuners.sampling_rate / 2))
268 dbs_info->cdbs.time_stamp = time_now;
270 od_timer_update(dbs_info, sample, dw);
271 mutex_unlock(&dbs_info->cdbs.timer_mutex);
274 static void od_dbs_timer(struct work_struct *work)
276 struct delayed_work *dw = to_delayed_work(work);
277 struct od_cpu_dbs_info_s *dbs_info =
278 container_of(work, struct od_cpu_dbs_info_s, cdbs.work.work);
280 if (dbs_sw_coordinated_cpus(&dbs_info->cdbs)) {
281 od_timer_coordinated(dbs_info, dw);
283 mutex_lock(&dbs_info->cdbs.timer_mutex);
284 od_timer_update(dbs_info, true, dw);
285 mutex_unlock(&dbs_info->cdbs.timer_mutex);
289 /************************** sysfs interface ************************/
291 static ssize_t show_sampling_rate_min(struct kobject *kobj,
292 struct attribute *attr, char *buf)
294 return sprintf(buf, "%u\n", od_dbs_data.min_sampling_rate);
298 * update_sampling_rate - update sampling rate effective immediately if needed.
299 * @new_rate: new sampling rate
301 * If new rate is smaller than the old, simply updaing
302 * dbs_tuners_int.sampling_rate might not be appropriate. For example, if the
303 * original sampling_rate was 1 second and the requested new sampling rate is 10
304 * ms because the user needs immediate reaction from ondemand governor, but not
305 * sure if higher frequency will be required or not, then, the governor may
306 * change the sampling rate too late; up to 1 second later. Thus, if we are
307 * reducing the sampling rate, we need to make the new value effective
310 static void update_sampling_rate(unsigned int new_rate)
314 od_tuners.sampling_rate = new_rate = max(new_rate,
315 od_dbs_data.min_sampling_rate);
317 for_each_online_cpu(cpu) {
318 struct cpufreq_policy *policy;
319 struct od_cpu_dbs_info_s *dbs_info;
320 unsigned long next_sampling, appointed_at;
322 policy = cpufreq_cpu_get(cpu);
325 if (policy->governor != &cpufreq_gov_ondemand) {
326 cpufreq_cpu_put(policy);
329 dbs_info = &per_cpu(od_cpu_dbs_info, policy->cpu);
330 cpufreq_cpu_put(policy);
332 mutex_lock(&dbs_info->cdbs.timer_mutex);
334 if (!delayed_work_pending(&dbs_info->cdbs.work)) {
335 mutex_unlock(&dbs_info->cdbs.timer_mutex);
339 next_sampling = jiffies + usecs_to_jiffies(new_rate);
340 appointed_at = dbs_info->cdbs.work.timer.expires;
342 if (time_before(next_sampling, appointed_at)) {
344 mutex_unlock(&dbs_info->cdbs.timer_mutex);
345 cancel_delayed_work_sync(&dbs_info->cdbs.work);
346 mutex_lock(&dbs_info->cdbs.timer_mutex);
348 schedule_delayed_work_on(dbs_info->cdbs.cpu,
349 &dbs_info->cdbs.work,
350 usecs_to_jiffies(new_rate));
353 mutex_unlock(&dbs_info->cdbs.timer_mutex);
357 static ssize_t store_sampling_rate(struct kobject *a, struct attribute *b,
358 const char *buf, size_t count)
362 ret = sscanf(buf, "%u", &input);
365 update_sampling_rate(input);
369 static ssize_t store_io_is_busy(struct kobject *a, struct attribute *b,
370 const char *buf, size_t count)
375 ret = sscanf(buf, "%u", &input);
378 od_tuners.io_is_busy = !!input;
382 static ssize_t store_up_threshold(struct kobject *a, struct attribute *b,
383 const char *buf, size_t count)
387 ret = sscanf(buf, "%u", &input);
389 if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
390 input < MIN_FREQUENCY_UP_THRESHOLD) {
393 od_tuners.up_threshold = input;
397 static ssize_t store_sampling_down_factor(struct kobject *a,
398 struct attribute *b, const char *buf, size_t count)
400 unsigned int input, j;
402 ret = sscanf(buf, "%u", &input);
404 if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
406 od_tuners.sampling_down_factor = input;
408 /* Reset down sampling multiplier in case it was active */
409 for_each_online_cpu(j) {
410 struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
412 dbs_info->rate_mult = 1;
417 static ssize_t store_ignore_nice_load(struct kobject *a, struct attribute *b,
418 const char *buf, size_t count)
425 ret = sscanf(buf, "%u", &input);
432 if (input == od_tuners.ignore_nice) { /* nothing to do */
435 od_tuners.ignore_nice = input;
437 /* we need to re-evaluate prev_cpu_idle */
438 for_each_online_cpu(j) {
439 struct od_cpu_dbs_info_s *dbs_info;
440 dbs_info = &per_cpu(od_cpu_dbs_info, j);
441 dbs_info->cdbs.prev_cpu_idle = get_cpu_idle_time(j,
442 &dbs_info->cdbs.prev_cpu_wall);
443 if (od_tuners.ignore_nice)
444 dbs_info->cdbs.prev_cpu_nice =
445 kcpustat_cpu(j).cpustat[CPUTIME_NICE];
451 static ssize_t store_powersave_bias(struct kobject *a, struct attribute *b,
452 const char *buf, size_t count)
456 ret = sscanf(buf, "%u", &input);
464 od_tuners.powersave_bias = input;
465 ondemand_powersave_bias_init();
469 show_one(od, sampling_rate, sampling_rate);
470 show_one(od, io_is_busy, io_is_busy);
471 show_one(od, up_threshold, up_threshold);
472 show_one(od, sampling_down_factor, sampling_down_factor);
473 show_one(od, ignore_nice_load, ignore_nice);
474 show_one(od, powersave_bias, powersave_bias);
476 define_one_global_rw(sampling_rate);
477 define_one_global_rw(io_is_busy);
478 define_one_global_rw(up_threshold);
479 define_one_global_rw(sampling_down_factor);
480 define_one_global_rw(ignore_nice_load);
481 define_one_global_rw(powersave_bias);
482 define_one_global_ro(sampling_rate_min);
484 static struct attribute *dbs_attributes[] = {
485 &sampling_rate_min.attr,
488 &sampling_down_factor.attr,
489 &ignore_nice_load.attr,
490 &powersave_bias.attr,
495 static struct attribute_group od_attr_group = {
496 .attrs = dbs_attributes,
500 /************************** sysfs end ************************/
502 define_get_cpu_dbs_routines(od_cpu_dbs_info);
504 static struct od_ops od_ops = {
505 .io_busy = should_io_be_busy,
506 .powersave_bias_init_cpu = ondemand_powersave_bias_init_cpu,
507 .powersave_bias_target = powersave_bias_target,
508 .freq_increase = dbs_freq_increase,
511 static struct dbs_data od_dbs_data = {
512 .governor = GOV_ONDEMAND,
513 .attr_group = &od_attr_group,
514 .tuners = &od_tuners,
515 .get_cpu_cdbs = get_cpu_cdbs,
516 .get_cpu_dbs_info_s = get_cpu_dbs_info_s,
517 .gov_dbs_timer = od_dbs_timer,
518 .gov_check_cpu = od_check_cpu,
522 static int od_cpufreq_governor_dbs(struct cpufreq_policy *policy,
525 return cpufreq_governor_dbs(&od_dbs_data, policy, event);
528 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
531 struct cpufreq_governor cpufreq_gov_ondemand = {
533 .governor = od_cpufreq_governor_dbs,
534 .max_transition_latency = TRANSITION_LATENCY_LIMIT,
535 .owner = THIS_MODULE,
538 static int __init cpufreq_gov_dbs_init(void)
543 mutex_init(&od_dbs_data.mutex);
544 idle_time = get_cpu_idle_time_us(cpu, NULL);
546 if (idle_time != -1ULL) {
547 /* Idle micro accounting is supported. Use finer thresholds */
548 od_tuners.up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
549 od_tuners.down_differential = MICRO_FREQUENCY_DOWN_DIFFERENTIAL;
551 * In nohz/micro accounting case we set the minimum frequency
552 * not depending on HZ, but fixed (very low). The deferred
553 * timer might skip some samples if idle/sleeping as needed.
555 od_dbs_data.min_sampling_rate = MICRO_FREQUENCY_MIN_SAMPLE_RATE;
557 /* For correct statistics, we need 10 ticks for each measure */
558 od_dbs_data.min_sampling_rate = MIN_SAMPLING_RATE_RATIO *
559 jiffies_to_usecs(10);
562 return cpufreq_register_governor(&cpufreq_gov_ondemand);
565 static void __exit cpufreq_gov_dbs_exit(void)
567 cpufreq_unregister_governor(&cpufreq_gov_ondemand);
570 MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
571 MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
572 MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
573 "Low Latency Frequency Transition capable processors");
574 MODULE_LICENSE("GPL");
576 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
577 fs_initcall(cpufreq_gov_dbs_init);
579 module_init(cpufreq_gov_dbs_init);
581 module_exit(cpufreq_gov_dbs_exit);