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>
20 #include "cpufreq_ondemand.h"
22 /* On-demand governor macros */
23 #define DEF_FREQUENCY_UP_THRESHOLD (80)
24 #define DEF_SAMPLING_DOWN_FACTOR (1)
25 #define MAX_SAMPLING_DOWN_FACTOR (100000)
26 #define MICRO_FREQUENCY_UP_THRESHOLD (95)
27 #define MICRO_FREQUENCY_MIN_SAMPLE_RATE (10000)
28 #define MIN_FREQUENCY_UP_THRESHOLD (11)
29 #define MAX_FREQUENCY_UP_THRESHOLD (100)
31 static struct od_ops od_ops;
33 static unsigned int default_powersave_bias;
36 * Not all CPUs want IO time to be accounted as busy; this depends on how
37 * efficient idling at a higher frequency/voltage is.
38 * Pavel Machek says this is not so for various generations of AMD and old
40 * Mike Chan (android.com) claims this is also not true for ARM.
41 * Because of this, whitelist specific known (series) of CPUs by default, and
42 * leave all others up to the user.
44 static int should_io_be_busy(void)
46 #if defined(CONFIG_X86)
48 * For Intel, Core 2 (model 15) and later have an efficient idle.
50 if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
51 boot_cpu_data.x86 == 6 &&
52 boot_cpu_data.x86_model >= 15)
59 * Find right freq to be set now with powersave_bias on.
60 * Returns the freq_hi to be used right now and will set freq_hi_delay_us,
61 * freq_lo, and freq_lo_delay_us in percpu area for averaging freqs.
63 static unsigned int generic_powersave_bias_target(struct cpufreq_policy *policy,
64 unsigned int freq_next, unsigned int relation)
66 unsigned int freq_req, freq_reduc, freq_avg;
67 unsigned int freq_hi, freq_lo;
69 unsigned int delay_hi_us;
70 struct policy_dbs_info *policy_dbs = policy->governor_data;
71 struct od_policy_dbs_info *dbs_info = to_dbs_info(policy_dbs);
72 struct dbs_data *dbs_data = policy_dbs->dbs_data;
73 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
74 struct cpufreq_frequency_table *freq_table = policy->freq_table;
77 dbs_info->freq_lo = 0;
78 dbs_info->freq_lo_delay_us = 0;
82 index = cpufreq_frequency_table_target(policy, freq_next, relation);
83 freq_req = freq_table[index].frequency;
84 freq_reduc = freq_req * od_tuners->powersave_bias / 1000;
85 freq_avg = freq_req - freq_reduc;
87 /* Find freq bounds for freq_avg in freq_table */
88 index = cpufreq_frequency_table_target(policy, freq_avg,
90 freq_lo = freq_table[index].frequency;
91 index = cpufreq_frequency_table_target(policy, freq_avg,
93 freq_hi = freq_table[index].frequency;
95 /* Find out how long we have to be in hi and lo freqs */
96 if (freq_hi == freq_lo) {
97 dbs_info->freq_lo = 0;
98 dbs_info->freq_lo_delay_us = 0;
101 delay_hi_us = (freq_avg - freq_lo) * dbs_data->sampling_rate;
102 delay_hi_us += (freq_hi - freq_lo) / 2;
103 delay_hi_us /= freq_hi - freq_lo;
104 dbs_info->freq_hi_delay_us = delay_hi_us;
105 dbs_info->freq_lo = freq_lo;
106 dbs_info->freq_lo_delay_us = dbs_data->sampling_rate - delay_hi_us;
110 static void ondemand_powersave_bias_init(struct cpufreq_policy *policy)
112 struct od_policy_dbs_info *dbs_info = to_dbs_info(policy->governor_data);
114 dbs_info->freq_lo = 0;
117 static void dbs_freq_increase(struct cpufreq_policy *policy, unsigned int freq)
119 struct policy_dbs_info *policy_dbs = policy->governor_data;
120 struct dbs_data *dbs_data = policy_dbs->dbs_data;
121 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
123 if (od_tuners->powersave_bias)
124 freq = od_ops.powersave_bias_target(policy, freq,
126 else if (policy->cur == policy->max)
129 __cpufreq_driver_target(policy, freq, od_tuners->powersave_bias ?
130 CPUFREQ_RELATION_L : CPUFREQ_RELATION_H);
134 * Every sampling_rate, we check, if current idle time is less than 20%
135 * (default), then we try to increase frequency. Else, we adjust the frequency
136 * proportional to load.
138 static void od_update(struct cpufreq_policy *policy)
140 struct policy_dbs_info *policy_dbs = policy->governor_data;
141 struct od_policy_dbs_info *dbs_info = to_dbs_info(policy_dbs);
142 struct dbs_data *dbs_data = policy_dbs->dbs_data;
143 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
144 unsigned int load = dbs_update(policy);
146 dbs_info->freq_lo = 0;
148 /* Check for frequency increase */
149 if (load > dbs_data->up_threshold) {
150 /* If switching to max speed, apply sampling_down_factor */
151 if (policy->cur < policy->max)
152 policy_dbs->rate_mult = dbs_data->sampling_down_factor;
153 dbs_freq_increase(policy, policy->max);
155 /* Calculate the next frequency proportional to load */
156 unsigned int freq_next, min_f, max_f;
158 min_f = policy->cpuinfo.min_freq;
159 max_f = policy->cpuinfo.max_freq;
160 freq_next = min_f + load * (max_f - min_f) / 100;
162 /* No longer fully busy, reset rate_mult */
163 policy_dbs->rate_mult = 1;
165 if (od_tuners->powersave_bias)
166 freq_next = od_ops.powersave_bias_target(policy,
170 __cpufreq_driver_target(policy, freq_next, CPUFREQ_RELATION_C);
174 static unsigned int od_dbs_timer(struct cpufreq_policy *policy)
176 struct policy_dbs_info *policy_dbs = policy->governor_data;
177 struct dbs_data *dbs_data = policy_dbs->dbs_data;
178 struct od_policy_dbs_info *dbs_info = to_dbs_info(policy_dbs);
179 int sample_type = dbs_info->sample_type;
181 /* Common NORMAL_SAMPLE setup */
182 dbs_info->sample_type = OD_NORMAL_SAMPLE;
184 * OD_SUB_SAMPLE doesn't make sense if sample_delay_ns is 0, so ignore
187 if (sample_type == OD_SUB_SAMPLE && policy_dbs->sample_delay_ns > 0) {
188 __cpufreq_driver_target(policy, dbs_info->freq_lo,
190 return dbs_info->freq_lo_delay_us;
195 if (dbs_info->freq_lo) {
196 /* Setup timer for SUB_SAMPLE */
197 dbs_info->sample_type = OD_SUB_SAMPLE;
198 return dbs_info->freq_hi_delay_us;
201 return dbs_data->sampling_rate * policy_dbs->rate_mult;
204 /************************** sysfs interface ************************/
205 static struct dbs_governor od_dbs_gov;
207 static ssize_t store_io_is_busy(struct gov_attr_set *attr_set, const char *buf,
210 struct dbs_data *dbs_data = to_dbs_data(attr_set);
214 ret = sscanf(buf, "%u", &input);
217 dbs_data->io_is_busy = !!input;
219 /* we need to re-evaluate prev_cpu_idle */
220 gov_update_cpu_data(dbs_data);
225 static ssize_t store_up_threshold(struct gov_attr_set *attr_set,
226 const char *buf, size_t count)
228 struct dbs_data *dbs_data = to_dbs_data(attr_set);
231 ret = sscanf(buf, "%u", &input);
233 if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
234 input < MIN_FREQUENCY_UP_THRESHOLD) {
238 dbs_data->up_threshold = input;
242 static ssize_t store_sampling_down_factor(struct gov_attr_set *attr_set,
243 const char *buf, size_t count)
245 struct dbs_data *dbs_data = to_dbs_data(attr_set);
246 struct policy_dbs_info *policy_dbs;
249 ret = sscanf(buf, "%u", &input);
251 if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
254 dbs_data->sampling_down_factor = input;
256 /* Reset down sampling multiplier in case it was active */
257 list_for_each_entry(policy_dbs, &attr_set->policy_list, list) {
259 * Doing this without locking might lead to using different
260 * rate_mult values in od_update() and od_dbs_timer().
262 mutex_lock(&policy_dbs->timer_mutex);
263 policy_dbs->rate_mult = 1;
264 mutex_unlock(&policy_dbs->timer_mutex);
270 static ssize_t store_ignore_nice_load(struct gov_attr_set *attr_set,
271 const char *buf, size_t count)
273 struct dbs_data *dbs_data = to_dbs_data(attr_set);
277 ret = sscanf(buf, "%u", &input);
284 if (input == dbs_data->ignore_nice_load) { /* nothing to do */
287 dbs_data->ignore_nice_load = input;
289 /* we need to re-evaluate prev_cpu_idle */
290 gov_update_cpu_data(dbs_data);
295 static ssize_t store_powersave_bias(struct gov_attr_set *attr_set,
296 const char *buf, size_t count)
298 struct dbs_data *dbs_data = to_dbs_data(attr_set);
299 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
300 struct policy_dbs_info *policy_dbs;
303 ret = sscanf(buf, "%u", &input);
311 od_tuners->powersave_bias = input;
313 list_for_each_entry(policy_dbs, &attr_set->policy_list, list)
314 ondemand_powersave_bias_init(policy_dbs->policy);
319 gov_show_one_common(sampling_rate);
320 gov_show_one_common(up_threshold);
321 gov_show_one_common(sampling_down_factor);
322 gov_show_one_common(ignore_nice_load);
323 gov_show_one_common(min_sampling_rate);
324 gov_show_one_common(io_is_busy);
325 gov_show_one(od, powersave_bias);
327 gov_attr_rw(sampling_rate);
328 gov_attr_rw(io_is_busy);
329 gov_attr_rw(up_threshold);
330 gov_attr_rw(sampling_down_factor);
331 gov_attr_rw(ignore_nice_load);
332 gov_attr_rw(powersave_bias);
333 gov_attr_ro(min_sampling_rate);
335 static struct attribute *od_attributes[] = {
336 &min_sampling_rate.attr,
339 &sampling_down_factor.attr,
340 &ignore_nice_load.attr,
341 &powersave_bias.attr,
346 /************************** sysfs end ************************/
348 static struct policy_dbs_info *od_alloc(void)
350 struct od_policy_dbs_info *dbs_info;
352 dbs_info = kzalloc(sizeof(*dbs_info), GFP_KERNEL);
353 return dbs_info ? &dbs_info->policy_dbs : NULL;
356 static void od_free(struct policy_dbs_info *policy_dbs)
358 kfree(to_dbs_info(policy_dbs));
361 static int od_init(struct dbs_data *dbs_data)
363 struct od_dbs_tuners *tuners;
367 tuners = kzalloc(sizeof(*tuners), GFP_KERNEL);
372 idle_time = get_cpu_idle_time_us(cpu, NULL);
374 if (idle_time != -1ULL) {
375 /* Idle micro accounting is supported. Use finer thresholds */
376 dbs_data->up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
378 * In nohz/micro accounting case we set the minimum frequency
379 * not depending on HZ, but fixed (very low). The deferred
380 * timer might skip some samples if idle/sleeping as needed.
382 dbs_data->min_sampling_rate = MICRO_FREQUENCY_MIN_SAMPLE_RATE;
384 dbs_data->up_threshold = DEF_FREQUENCY_UP_THRESHOLD;
386 /* For correct statistics, we need 10 ticks for each measure */
387 dbs_data->min_sampling_rate = MIN_SAMPLING_RATE_RATIO *
388 jiffies_to_usecs(10);
391 dbs_data->sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR;
392 dbs_data->ignore_nice_load = 0;
393 tuners->powersave_bias = default_powersave_bias;
394 dbs_data->io_is_busy = should_io_be_busy();
396 dbs_data->tuners = tuners;
400 static void od_exit(struct dbs_data *dbs_data)
402 kfree(dbs_data->tuners);
405 static void od_start(struct cpufreq_policy *policy)
407 struct od_policy_dbs_info *dbs_info = to_dbs_info(policy->governor_data);
409 dbs_info->sample_type = OD_NORMAL_SAMPLE;
410 ondemand_powersave_bias_init(policy);
413 static struct od_ops od_ops = {
414 .powersave_bias_target = generic_powersave_bias_target,
417 static struct dbs_governor od_dbs_gov = {
418 .gov = CPUFREQ_DBS_GOVERNOR_INITIALIZER("ondemand"),
419 .kobj_type = { .default_attrs = od_attributes },
420 .gov_dbs_timer = od_dbs_timer,
428 #define CPU_FREQ_GOV_ONDEMAND (&od_dbs_gov.gov)
430 static void od_set_powersave_bias(unsigned int powersave_bias)
435 default_powersave_bias = powersave_bias;
436 cpumask_clear(&done);
439 for_each_online_cpu(cpu) {
440 struct cpufreq_policy *policy;
441 struct policy_dbs_info *policy_dbs;
442 struct dbs_data *dbs_data;
443 struct od_dbs_tuners *od_tuners;
445 if (cpumask_test_cpu(cpu, &done))
448 policy = cpufreq_cpu_get_raw(cpu);
449 if (!policy || policy->governor != CPU_FREQ_GOV_ONDEMAND)
452 policy_dbs = policy->governor_data;
456 cpumask_or(&done, &done, policy->cpus);
458 dbs_data = policy_dbs->dbs_data;
459 od_tuners = dbs_data->tuners;
460 od_tuners->powersave_bias = default_powersave_bias;
465 void od_register_powersave_bias_handler(unsigned int (*f)
466 (struct cpufreq_policy *, unsigned int, unsigned int),
467 unsigned int powersave_bias)
469 od_ops.powersave_bias_target = f;
470 od_set_powersave_bias(powersave_bias);
472 EXPORT_SYMBOL_GPL(od_register_powersave_bias_handler);
474 void od_unregister_powersave_bias_handler(void)
476 od_ops.powersave_bias_target = generic_powersave_bias_target;
477 od_set_powersave_bias(0);
479 EXPORT_SYMBOL_GPL(od_unregister_powersave_bias_handler);
481 static int __init cpufreq_gov_dbs_init(void)
483 return cpufreq_register_governor(CPU_FREQ_GOV_ONDEMAND);
486 static void __exit cpufreq_gov_dbs_exit(void)
488 cpufreq_unregister_governor(CPU_FREQ_GOV_ONDEMAND);
491 MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
492 MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
493 MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
494 "Low Latency Frequency Transition capable processors");
495 MODULE_LICENSE("GPL");
497 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
498 struct cpufreq_governor *cpufreq_default_governor(void)
500 return CPU_FREQ_GOV_ONDEMAND;
503 fs_initcall(cpufreq_gov_dbs_init);
505 module_init(cpufreq_gov_dbs_init);
507 module_exit(cpufreq_gov_dbs_exit);