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/cpu.h>
16 #include <linux/percpu-defs.h>
17 #include <linux/slab.h>
18 #include <linux/tick.h>
19 #include "cpufreq_governor.h"
21 /* On-demand governor macros */
22 #define DEF_FREQUENCY_UP_THRESHOLD (80)
23 #define DEF_SAMPLING_DOWN_FACTOR (1)
24 #define MAX_SAMPLING_DOWN_FACTOR (100000)
25 #define MICRO_FREQUENCY_UP_THRESHOLD (95)
26 #define MICRO_FREQUENCY_MIN_SAMPLE_RATE (10000)
27 #define MIN_FREQUENCY_UP_THRESHOLD (11)
28 #define MAX_FREQUENCY_UP_THRESHOLD (100)
30 static DEFINE_PER_CPU(struct od_cpu_dbs_info_s, od_cpu_dbs_info);
32 static struct od_ops od_ops;
34 static struct cpufreq_governor cpufreq_gov_ondemand;
36 static unsigned int default_powersave_bias;
38 static void ondemand_powersave_bias_init_cpu(int cpu)
40 struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
42 dbs_info->freq_table = cpufreq_frequency_get_table(cpu);
43 dbs_info->freq_lo = 0;
47 * Not all CPUs want IO time to be accounted as busy; this depends on how
48 * efficient idling at a higher frequency/voltage is.
49 * Pavel Machek says this is not so for various generations of AMD and old
51 * Mike Chan (android.com) claims this is also not true for ARM.
52 * Because of this, whitelist specific known (series) of CPUs by default, and
53 * leave all others up to the user.
55 static int should_io_be_busy(void)
57 #if defined(CONFIG_X86)
59 * For Intel, Core 2 (model 15) and later have an efficient idle.
61 if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
62 boot_cpu_data.x86 == 6 &&
63 boot_cpu_data.x86_model >= 15)
70 * Find right freq to be set now with powersave_bias on.
71 * Returns the freq_hi to be used right now and will set freq_hi_jiffies,
72 * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs.
74 static unsigned int generic_powersave_bias_target(struct cpufreq_policy *policy,
75 unsigned int freq_next, unsigned int relation)
77 unsigned int freq_req, freq_reduc, freq_avg;
78 unsigned int freq_hi, freq_lo;
79 unsigned int index = 0;
80 unsigned int jiffies_total, jiffies_hi, jiffies_lo;
81 struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
83 struct dbs_data *dbs_data = policy->governor_data;
84 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
86 if (!dbs_info->freq_table) {
87 dbs_info->freq_lo = 0;
88 dbs_info->freq_lo_jiffies = 0;
92 cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next,
94 freq_req = dbs_info->freq_table[index].frequency;
95 freq_reduc = freq_req * od_tuners->powersave_bias / 1000;
96 freq_avg = freq_req - freq_reduc;
98 /* Find freq bounds for freq_avg in freq_table */
100 cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
101 CPUFREQ_RELATION_H, &index);
102 freq_lo = dbs_info->freq_table[index].frequency;
104 cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
105 CPUFREQ_RELATION_L, &index);
106 freq_hi = dbs_info->freq_table[index].frequency;
108 /* Find out how long we have to be in hi and lo freqs */
109 if (freq_hi == freq_lo) {
110 dbs_info->freq_lo = 0;
111 dbs_info->freq_lo_jiffies = 0;
114 jiffies_total = usecs_to_jiffies(od_tuners->sampling_rate);
115 jiffies_hi = (freq_avg - freq_lo) * jiffies_total;
116 jiffies_hi += ((freq_hi - freq_lo) / 2);
117 jiffies_hi /= (freq_hi - freq_lo);
118 jiffies_lo = jiffies_total - jiffies_hi;
119 dbs_info->freq_lo = freq_lo;
120 dbs_info->freq_lo_jiffies = jiffies_lo;
121 dbs_info->freq_hi_jiffies = jiffies_hi;
125 static void ondemand_powersave_bias_init(void)
128 for_each_online_cpu(i) {
129 ondemand_powersave_bias_init_cpu(i);
133 static void dbs_freq_increase(struct cpufreq_policy *policy, unsigned int freq)
135 struct dbs_data *dbs_data = policy->governor_data;
136 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
138 if (od_tuners->powersave_bias)
139 freq = od_ops.powersave_bias_target(policy, freq,
141 else if (policy->cur == policy->max)
144 __cpufreq_driver_target(policy, freq, od_tuners->powersave_bias ?
145 CPUFREQ_RELATION_L : CPUFREQ_RELATION_H);
149 * Every sampling_rate, we check, if current idle time is less than 20%
150 * (default), then we try to increase frequency. Else, we adjust the frequency
151 * proportional to load.
153 static void od_check_cpu(int cpu, unsigned int load)
155 struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
156 struct cpufreq_policy *policy = dbs_info->cdbs.shared->policy;
157 struct dbs_data *dbs_data = policy->governor_data;
158 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
160 dbs_info->freq_lo = 0;
162 /* Check for frequency increase */
163 if (load > od_tuners->up_threshold) {
164 /* If switching to max speed, apply sampling_down_factor */
165 if (policy->cur < policy->max)
166 dbs_info->rate_mult =
167 od_tuners->sampling_down_factor;
168 dbs_freq_increase(policy, policy->max);
170 /* Calculate the next frequency proportional to load */
171 unsigned int freq_next, min_f, max_f;
173 min_f = policy->cpuinfo.min_freq;
174 max_f = policy->cpuinfo.max_freq;
175 freq_next = min_f + load * (max_f - min_f) / 100;
177 /* No longer fully busy, reset rate_mult */
178 dbs_info->rate_mult = 1;
180 if (!od_tuners->powersave_bias) {
181 __cpufreq_driver_target(policy, freq_next,
186 freq_next = od_ops.powersave_bias_target(policy, freq_next,
188 __cpufreq_driver_target(policy, freq_next, CPUFREQ_RELATION_C);
192 static unsigned int od_dbs_timer(struct cpufreq_policy *policy, bool modify_all)
194 struct dbs_data *dbs_data = policy->governor_data;
195 unsigned int cpu = policy->cpu;
196 struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
198 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
199 int delay = 0, sample_type = dbs_info->sample_type;
204 /* Common NORMAL_SAMPLE setup */
205 dbs_info->sample_type = OD_NORMAL_SAMPLE;
206 if (sample_type == OD_SUB_SAMPLE) {
207 delay = dbs_info->freq_lo_jiffies;
208 __cpufreq_driver_target(policy, dbs_info->freq_lo,
211 dbs_check_cpu(dbs_data, cpu);
212 if (dbs_info->freq_lo) {
213 /* Setup timer for SUB_SAMPLE */
214 dbs_info->sample_type = OD_SUB_SAMPLE;
215 delay = dbs_info->freq_hi_jiffies;
221 delay = delay_for_sampling_rate(od_tuners->sampling_rate
222 * dbs_info->rate_mult);
227 /************************** sysfs interface ************************/
228 static struct common_dbs_data od_dbs_cdata;
231 * update_sampling_rate - update sampling rate effective immediately if needed.
232 * @new_rate: new sampling rate
234 * If new rate is smaller than the old, simply updating
235 * dbs_tuners_int.sampling_rate might not be appropriate. For example, if the
236 * original sampling_rate was 1 second and the requested new sampling rate is 10
237 * ms because the user needs immediate reaction from ondemand governor, but not
238 * sure if higher frequency will be required or not, then, the governor may
239 * change the sampling rate too late; up to 1 second later. Thus, if we are
240 * reducing the sampling rate, we need to make the new value effective
243 static void update_sampling_rate(struct dbs_data *dbs_data,
244 unsigned int new_rate)
246 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
247 struct cpumask cpumask;
250 od_tuners->sampling_rate = new_rate = max(new_rate,
251 dbs_data->min_sampling_rate);
254 * Lock governor so that governor start/stop can't execute in parallel.
256 mutex_lock(&od_dbs_cdata.mutex);
258 cpumask_copy(&cpumask, cpu_online_mask);
260 for_each_cpu(cpu, &cpumask) {
261 struct cpufreq_policy *policy;
262 struct od_cpu_dbs_info_s *dbs_info;
263 struct cpu_dbs_info *cdbs;
264 struct cpu_common_dbs_info *shared;
265 unsigned long next_sampling, appointed_at;
267 dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
268 cdbs = &dbs_info->cdbs;
269 shared = cdbs->shared;
272 * A valid shared and shared->policy means governor hasn't
273 * stopped or exited yet.
275 if (!shared || !shared->policy)
278 policy = shared->policy;
280 /* clear all CPUs of this policy */
281 cpumask_andnot(&cpumask, &cpumask, policy->cpus);
284 * Update sampling rate for CPUs whose policy is governed by
285 * dbs_data. In case of governor_per_policy, only a single
286 * policy will be governed by dbs_data, otherwise there can be
287 * multiple policies that are governed by the same dbs_data.
289 if (dbs_data != policy->governor_data)
293 * Checking this for any CPU should be fine, timers for all of
294 * them are scheduled together.
296 next_sampling = jiffies + usecs_to_jiffies(new_rate);
297 appointed_at = dbs_info->cdbs.timer.expires;
299 if (time_before(next_sampling, appointed_at)) {
300 gov_cancel_work(shared);
301 gov_add_timers(policy, usecs_to_jiffies(new_rate));
306 mutex_unlock(&od_dbs_cdata.mutex);
309 static ssize_t store_sampling_rate(struct dbs_data *dbs_data, const char *buf,
314 ret = sscanf(buf, "%u", &input);
318 update_sampling_rate(dbs_data, input);
322 static ssize_t store_io_is_busy(struct dbs_data *dbs_data, const char *buf,
325 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
330 ret = sscanf(buf, "%u", &input);
333 od_tuners->io_is_busy = !!input;
335 /* we need to re-evaluate prev_cpu_idle */
336 for_each_online_cpu(j) {
337 struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
339 dbs_info->cdbs.prev_cpu_idle = get_cpu_idle_time(j,
340 &dbs_info->cdbs.prev_cpu_wall, od_tuners->io_is_busy);
345 static ssize_t store_up_threshold(struct dbs_data *dbs_data, const char *buf,
348 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
351 ret = sscanf(buf, "%u", &input);
353 if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
354 input < MIN_FREQUENCY_UP_THRESHOLD) {
358 od_tuners->up_threshold = input;
362 static ssize_t store_sampling_down_factor(struct dbs_data *dbs_data,
363 const char *buf, size_t count)
365 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
366 unsigned int input, j;
368 ret = sscanf(buf, "%u", &input);
370 if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
372 od_tuners->sampling_down_factor = input;
374 /* Reset down sampling multiplier in case it was active */
375 for_each_online_cpu(j) {
376 struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
378 dbs_info->rate_mult = 1;
383 static ssize_t store_ignore_nice_load(struct dbs_data *dbs_data,
384 const char *buf, size_t count)
386 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
392 ret = sscanf(buf, "%u", &input);
399 if (input == od_tuners->ignore_nice_load) { /* nothing to do */
402 od_tuners->ignore_nice_load = input;
404 /* we need to re-evaluate prev_cpu_idle */
405 for_each_online_cpu(j) {
406 struct od_cpu_dbs_info_s *dbs_info;
407 dbs_info = &per_cpu(od_cpu_dbs_info, j);
408 dbs_info->cdbs.prev_cpu_idle = get_cpu_idle_time(j,
409 &dbs_info->cdbs.prev_cpu_wall, od_tuners->io_is_busy);
410 if (od_tuners->ignore_nice_load)
411 dbs_info->cdbs.prev_cpu_nice =
412 kcpustat_cpu(j).cpustat[CPUTIME_NICE];
418 static ssize_t store_powersave_bias(struct dbs_data *dbs_data, const char *buf,
421 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
424 ret = sscanf(buf, "%u", &input);
432 od_tuners->powersave_bias = input;
433 ondemand_powersave_bias_init();
437 show_store_one(od, sampling_rate);
438 show_store_one(od, io_is_busy);
439 show_store_one(od, up_threshold);
440 show_store_one(od, sampling_down_factor);
441 show_store_one(od, ignore_nice_load);
442 show_store_one(od, powersave_bias);
443 declare_show_sampling_rate_min(od);
445 gov_sys_pol_attr_rw(sampling_rate);
446 gov_sys_pol_attr_rw(io_is_busy);
447 gov_sys_pol_attr_rw(up_threshold);
448 gov_sys_pol_attr_rw(sampling_down_factor);
449 gov_sys_pol_attr_rw(ignore_nice_load);
450 gov_sys_pol_attr_rw(powersave_bias);
451 gov_sys_pol_attr_ro(sampling_rate_min);
453 static struct attribute *dbs_attributes_gov_sys[] = {
454 &sampling_rate_min_gov_sys.attr,
455 &sampling_rate_gov_sys.attr,
456 &up_threshold_gov_sys.attr,
457 &sampling_down_factor_gov_sys.attr,
458 &ignore_nice_load_gov_sys.attr,
459 &powersave_bias_gov_sys.attr,
460 &io_is_busy_gov_sys.attr,
464 static struct attribute_group od_attr_group_gov_sys = {
465 .attrs = dbs_attributes_gov_sys,
469 static struct attribute *dbs_attributes_gov_pol[] = {
470 &sampling_rate_min_gov_pol.attr,
471 &sampling_rate_gov_pol.attr,
472 &up_threshold_gov_pol.attr,
473 &sampling_down_factor_gov_pol.attr,
474 &ignore_nice_load_gov_pol.attr,
475 &powersave_bias_gov_pol.attr,
476 &io_is_busy_gov_pol.attr,
480 static struct attribute_group od_attr_group_gov_pol = {
481 .attrs = dbs_attributes_gov_pol,
485 /************************** sysfs end ************************/
487 static int od_init(struct dbs_data *dbs_data, bool notify)
489 struct od_dbs_tuners *tuners;
493 tuners = kzalloc(sizeof(*tuners), GFP_KERNEL);
495 pr_err("%s: kzalloc failed\n", __func__);
500 idle_time = get_cpu_idle_time_us(cpu, NULL);
502 if (idle_time != -1ULL) {
503 /* Idle micro accounting is supported. Use finer thresholds */
504 tuners->up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
506 * In nohz/micro accounting case we set the minimum frequency
507 * not depending on HZ, but fixed (very low). The deferred
508 * timer might skip some samples if idle/sleeping as needed.
510 dbs_data->min_sampling_rate = MICRO_FREQUENCY_MIN_SAMPLE_RATE;
512 tuners->up_threshold = DEF_FREQUENCY_UP_THRESHOLD;
514 /* For correct statistics, we need 10 ticks for each measure */
515 dbs_data->min_sampling_rate = MIN_SAMPLING_RATE_RATIO *
516 jiffies_to_usecs(10);
519 tuners->sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR;
520 tuners->ignore_nice_load = 0;
521 tuners->powersave_bias = default_powersave_bias;
522 tuners->io_is_busy = should_io_be_busy();
524 dbs_data->tuners = tuners;
528 static void od_exit(struct dbs_data *dbs_data, bool notify)
530 kfree(dbs_data->tuners);
533 define_get_cpu_dbs_routines(od_cpu_dbs_info);
535 static struct od_ops od_ops = {
536 .powersave_bias_init_cpu = ondemand_powersave_bias_init_cpu,
537 .powersave_bias_target = generic_powersave_bias_target,
538 .freq_increase = dbs_freq_increase,
541 static struct common_dbs_data od_dbs_cdata = {
542 .governor = GOV_ONDEMAND,
543 .attr_group_gov_sys = &od_attr_group_gov_sys,
544 .attr_group_gov_pol = &od_attr_group_gov_pol,
545 .get_cpu_cdbs = get_cpu_cdbs,
546 .get_cpu_dbs_info_s = get_cpu_dbs_info_s,
547 .gov_dbs_timer = od_dbs_timer,
548 .gov_check_cpu = od_check_cpu,
552 .mutex = __MUTEX_INITIALIZER(od_dbs_cdata.mutex),
555 static int od_cpufreq_governor_dbs(struct cpufreq_policy *policy,
558 return cpufreq_governor_dbs(policy, &od_dbs_cdata, event);
561 static struct cpufreq_governor cpufreq_gov_ondemand = {
563 .governor = od_cpufreq_governor_dbs,
564 .max_transition_latency = TRANSITION_LATENCY_LIMIT,
565 .owner = THIS_MODULE,
568 static void od_set_powersave_bias(unsigned int powersave_bias)
570 struct cpufreq_policy *policy;
571 struct dbs_data *dbs_data;
572 struct od_dbs_tuners *od_tuners;
576 default_powersave_bias = powersave_bias;
577 cpumask_clear(&done);
580 for_each_online_cpu(cpu) {
581 struct cpu_common_dbs_info *shared;
583 if (cpumask_test_cpu(cpu, &done))
586 shared = per_cpu(od_cpu_dbs_info, cpu).cdbs.shared;
590 policy = shared->policy;
591 cpumask_or(&done, &done, policy->cpus);
593 if (policy->governor != &cpufreq_gov_ondemand)
596 dbs_data = policy->governor_data;
597 od_tuners = dbs_data->tuners;
598 od_tuners->powersave_bias = default_powersave_bias;
603 void od_register_powersave_bias_handler(unsigned int (*f)
604 (struct cpufreq_policy *, unsigned int, unsigned int),
605 unsigned int powersave_bias)
607 od_ops.powersave_bias_target = f;
608 od_set_powersave_bias(powersave_bias);
610 EXPORT_SYMBOL_GPL(od_register_powersave_bias_handler);
612 void od_unregister_powersave_bias_handler(void)
614 od_ops.powersave_bias_target = generic_powersave_bias_target;
615 od_set_powersave_bias(0);
617 EXPORT_SYMBOL_GPL(od_unregister_powersave_bias_handler);
619 static int __init cpufreq_gov_dbs_init(void)
621 return cpufreq_register_governor(&cpufreq_gov_ondemand);
624 static void __exit cpufreq_gov_dbs_exit(void)
626 cpufreq_unregister_governor(&cpufreq_gov_ondemand);
629 MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
630 MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
631 MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
632 "Low Latency Frequency Transition capable processors");
633 MODULE_LICENSE("GPL");
635 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
636 struct cpufreq_governor *cpufreq_default_governor(void)
638 return &cpufreq_gov_ondemand;
641 fs_initcall(cpufreq_gov_dbs_init);
643 module_init(cpufreq_gov_dbs_init);
645 module_exit(cpufreq_gov_dbs_exit);