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 #include <linux/kernel.h>
14 #include <linux/module.h>
15 #include <linux/init.h>
16 #include <linux/cpufreq.h>
17 #include <linux/cpu.h>
18 #include <linux/jiffies.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/mutex.h>
23 * dbs is used in this file as a shortform for demandbased switching
24 * It helps to keep variable names smaller, simpler
27 #define DEF_FREQUENCY_UP_THRESHOLD (80)
28 #define MIN_FREQUENCY_UP_THRESHOLD (11)
29 #define MAX_FREQUENCY_UP_THRESHOLD (100)
32 * The polling frequency of this governor depends on the capability of
33 * the processor. Default polling frequency is 1000 times the transition
34 * latency of the processor. The governor will work on any processor with
35 * transition latency <= 10mS, using appropriate sampling
37 * For CPUs with transition latency > 10mS (mostly drivers with CPUFREQ_ETERNAL)
38 * this governor will not work.
39 * All times here are in uS.
41 static unsigned int def_sampling_rate;
42 #define MIN_SAMPLING_RATE_RATIO (2)
43 /* for correct statistics, we need at least 10 ticks between each measure */
44 #define MIN_STAT_SAMPLING_RATE (MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10))
45 #define MIN_SAMPLING_RATE (def_sampling_rate / MIN_SAMPLING_RATE_RATIO)
46 #define MAX_SAMPLING_RATE (500 * def_sampling_rate)
47 #define DEF_SAMPLING_RATE_LATENCY_MULTIPLIER (1000)
48 #define TRANSITION_LATENCY_LIMIT (10 * 1000)
50 static void do_dbs_timer(struct work_struct *work);
53 enum dbs_sample {DBS_NORMAL_SAMPLE, DBS_SUB_SAMPLE};
55 struct cpu_dbs_info_s {
56 cputime64_t prev_cpu_idle;
57 cputime64_t prev_cpu_wall;
58 struct cpufreq_policy *cur_policy;
59 struct delayed_work work;
60 enum dbs_sample sample_type;
62 struct cpufreq_frequency_table *freq_table;
64 unsigned int freq_lo_jiffies;
65 unsigned int freq_hi_jiffies;
67 static DEFINE_PER_CPU(struct cpu_dbs_info_s, cpu_dbs_info);
69 static unsigned int dbs_enable; /* number of CPUs using this policy */
72 * DEADLOCK ALERT! There is a ordering requirement between cpu_hotplug
73 * lock and dbs_mutex. cpu_hotplug lock should always be held before
74 * dbs_mutex. If any function that can potentially take cpu_hotplug lock
75 * (like __cpufreq_driver_target()) is being called with dbs_mutex taken, then
76 * cpu_hotplug lock should be taken before that. Note that cpu_hotplug lock
77 * is recursive for the same process. -Venki
79 static DEFINE_MUTEX(dbs_mutex);
81 static struct workqueue_struct *kondemand_wq;
83 static struct dbs_tuners {
84 unsigned int sampling_rate;
85 unsigned int up_threshold;
86 unsigned int ignore_nice;
87 unsigned int powersave_bias;
89 .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
94 static inline cputime64_t get_cpu_idle_time(unsigned int cpu)
98 retval = cputime64_add(kstat_cpu(cpu).cpustat.idle,
99 kstat_cpu(cpu).cpustat.iowait);
101 if (dbs_tuners_ins.ignore_nice)
102 retval = cputime64_add(retval, kstat_cpu(cpu).cpustat.nice);
108 * Find right freq to be set now with powersave_bias on.
109 * Returns the freq_hi to be used right now and will set freq_hi_jiffies,
110 * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs.
112 static unsigned int powersave_bias_target(struct cpufreq_policy *policy,
113 unsigned int freq_next,
114 unsigned int relation)
116 unsigned int freq_req, freq_reduc, freq_avg;
117 unsigned int freq_hi, freq_lo;
118 unsigned int index = 0;
119 unsigned int jiffies_total, jiffies_hi, jiffies_lo;
120 struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, policy->cpu);
122 if (!dbs_info->freq_table) {
123 dbs_info->freq_lo = 0;
124 dbs_info->freq_lo_jiffies = 0;
128 cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next,
130 freq_req = dbs_info->freq_table[index].frequency;
131 freq_reduc = freq_req * dbs_tuners_ins.powersave_bias / 1000;
132 freq_avg = freq_req - freq_reduc;
134 /* Find freq bounds for freq_avg in freq_table */
136 cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
137 CPUFREQ_RELATION_H, &index);
138 freq_lo = dbs_info->freq_table[index].frequency;
140 cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
141 CPUFREQ_RELATION_L, &index);
142 freq_hi = dbs_info->freq_table[index].frequency;
144 /* Find out how long we have to be in hi and lo freqs */
145 if (freq_hi == freq_lo) {
146 dbs_info->freq_lo = 0;
147 dbs_info->freq_lo_jiffies = 0;
150 jiffies_total = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
151 jiffies_hi = (freq_avg - freq_lo) * jiffies_total;
152 jiffies_hi += ((freq_hi - freq_lo) / 2);
153 jiffies_hi /= (freq_hi - freq_lo);
154 jiffies_lo = jiffies_total - jiffies_hi;
155 dbs_info->freq_lo = freq_lo;
156 dbs_info->freq_lo_jiffies = jiffies_lo;
157 dbs_info->freq_hi_jiffies = jiffies_hi;
161 static void ondemand_powersave_bias_init(void)
164 for_each_online_cpu(i) {
165 struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, i);
166 dbs_info->freq_table = cpufreq_frequency_get_table(i);
167 dbs_info->freq_lo = 0;
171 /************************** sysfs interface ************************/
172 static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf)
174 return sprintf (buf, "%u\n", MAX_SAMPLING_RATE);
177 static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf)
179 return sprintf (buf, "%u\n", MIN_SAMPLING_RATE);
182 #define define_one_ro(_name) \
183 static struct freq_attr _name = \
184 __ATTR(_name, 0444, show_##_name, NULL)
186 define_one_ro(sampling_rate_max);
187 define_one_ro(sampling_rate_min);
189 /* cpufreq_ondemand Governor Tunables */
190 #define show_one(file_name, object) \
191 static ssize_t show_##file_name \
192 (struct cpufreq_policy *unused, char *buf) \
194 return sprintf(buf, "%u\n", dbs_tuners_ins.object); \
196 show_one(sampling_rate, sampling_rate);
197 show_one(up_threshold, up_threshold);
198 show_one(ignore_nice_load, ignore_nice);
199 show_one(powersave_bias, powersave_bias);
201 static ssize_t store_sampling_rate(struct cpufreq_policy *unused,
202 const char *buf, size_t count)
206 ret = sscanf(buf, "%u", &input);
208 mutex_lock(&dbs_mutex);
209 if (ret != 1 || input > MAX_SAMPLING_RATE || input < MIN_SAMPLING_RATE) {
210 mutex_unlock(&dbs_mutex);
214 dbs_tuners_ins.sampling_rate = input;
215 mutex_unlock(&dbs_mutex);
220 static ssize_t store_up_threshold(struct cpufreq_policy *unused,
221 const char *buf, size_t count)
225 ret = sscanf(buf, "%u", &input);
227 mutex_lock(&dbs_mutex);
228 if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
229 input < MIN_FREQUENCY_UP_THRESHOLD) {
230 mutex_unlock(&dbs_mutex);
234 dbs_tuners_ins.up_threshold = input;
235 mutex_unlock(&dbs_mutex);
240 static ssize_t store_ignore_nice_load(struct cpufreq_policy *policy,
241 const char *buf, size_t count)
248 ret = sscanf(buf, "%u", &input);
255 mutex_lock(&dbs_mutex);
256 if ( input == dbs_tuners_ins.ignore_nice ) { /* nothing to do */
257 mutex_unlock(&dbs_mutex);
260 dbs_tuners_ins.ignore_nice = input;
262 /* we need to re-evaluate prev_cpu_idle */
263 for_each_online_cpu(j) {
264 struct cpu_dbs_info_s *dbs_info;
265 dbs_info = &per_cpu(cpu_dbs_info, j);
266 dbs_info->prev_cpu_idle = get_cpu_idle_time(j);
267 dbs_info->prev_cpu_wall = get_jiffies_64();
269 mutex_unlock(&dbs_mutex);
274 static ssize_t store_powersave_bias(struct cpufreq_policy *unused,
275 const char *buf, size_t count)
279 ret = sscanf(buf, "%u", &input);
287 mutex_lock(&dbs_mutex);
288 dbs_tuners_ins.powersave_bias = input;
289 ondemand_powersave_bias_init();
290 mutex_unlock(&dbs_mutex);
295 #define define_one_rw(_name) \
296 static struct freq_attr _name = \
297 __ATTR(_name, 0644, show_##_name, store_##_name)
299 define_one_rw(sampling_rate);
300 define_one_rw(up_threshold);
301 define_one_rw(ignore_nice_load);
302 define_one_rw(powersave_bias);
304 static struct attribute * dbs_attributes[] = {
305 &sampling_rate_max.attr,
306 &sampling_rate_min.attr,
309 &ignore_nice_load.attr,
310 &powersave_bias.attr,
314 static struct attribute_group dbs_attr_group = {
315 .attrs = dbs_attributes,
319 /************************** sysfs end ************************/
321 static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
323 unsigned int idle_ticks, total_ticks;
325 cputime64_t cur_jiffies;
327 struct cpufreq_policy *policy;
330 if (!this_dbs_info->enable)
333 this_dbs_info->freq_lo = 0;
334 policy = this_dbs_info->cur_policy;
335 cur_jiffies = jiffies64_to_cputime64(get_jiffies_64());
336 total_ticks = (unsigned int) cputime64_sub(cur_jiffies,
337 this_dbs_info->prev_cpu_wall);
338 this_dbs_info->prev_cpu_wall = cur_jiffies;
342 * Every sampling_rate, we check, if current idle time is less
343 * than 20% (default), then we try to increase frequency
344 * Every sampling_rate, we look for a the lowest
345 * frequency which can sustain the load while keeping idle time over
346 * 30%. If such a frequency exist, we try to decrease to this frequency.
348 * Any frequency increase takes it to the maximum frequency.
349 * Frequency reduction happens at minimum steps of
350 * 5% (default) of current frequency
354 idle_ticks = UINT_MAX;
355 for_each_cpu_mask(j, policy->cpus) {
356 cputime64_t total_idle_ticks;
357 unsigned int tmp_idle_ticks;
358 struct cpu_dbs_info_s *j_dbs_info;
360 j_dbs_info = &per_cpu(cpu_dbs_info, j);
361 total_idle_ticks = get_cpu_idle_time(j);
362 tmp_idle_ticks = (unsigned int) cputime64_sub(total_idle_ticks,
363 j_dbs_info->prev_cpu_idle);
364 j_dbs_info->prev_cpu_idle = total_idle_ticks;
366 if (tmp_idle_ticks < idle_ticks)
367 idle_ticks = tmp_idle_ticks;
369 load = (100 * (total_ticks - idle_ticks)) / total_ticks;
371 /* Check for frequency increase */
372 if (load > dbs_tuners_ins.up_threshold) {
373 /* if we are already at full speed then break out early */
374 if (!dbs_tuners_ins.powersave_bias) {
375 if (policy->cur == policy->max)
378 __cpufreq_driver_target(policy, policy->max,
381 int freq = powersave_bias_target(policy, policy->max,
383 __cpufreq_driver_target(policy, freq,
389 /* Check for frequency decrease */
390 /* if we cannot reduce the frequency anymore, break out early */
391 if (policy->cur == policy->min)
395 * The optimal frequency is the frequency that is the lowest that
396 * can support the current CPU usage without triggering the up
397 * policy. To be safe, we focus 10 points under the threshold.
399 if (load < (dbs_tuners_ins.up_threshold - 10)) {
400 unsigned int freq_next = (policy->cur * load) /
401 (dbs_tuners_ins.up_threshold - 10);
402 if (!dbs_tuners_ins.powersave_bias) {
403 __cpufreq_driver_target(policy, freq_next,
406 int freq = powersave_bias_target(policy, freq_next,
408 __cpufreq_driver_target(policy, freq,
414 static void do_dbs_timer(struct work_struct *work)
416 unsigned int cpu = smp_processor_id();
417 struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, cpu);
418 enum dbs_sample sample_type = dbs_info->sample_type;
419 /* We want all CPUs to do sampling nearly on same jiffy */
420 int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
422 /* Permit rescheduling of this work item */
425 delay -= jiffies % delay;
427 if (!dbs_info->enable)
429 /* Common NORMAL_SAMPLE setup */
430 dbs_info->sample_type = DBS_NORMAL_SAMPLE;
431 if (!dbs_tuners_ins.powersave_bias ||
432 sample_type == DBS_NORMAL_SAMPLE) {
434 dbs_check_cpu(dbs_info);
435 unlock_cpu_hotplug();
436 if (dbs_info->freq_lo) {
437 /* Setup timer for SUB_SAMPLE */
438 dbs_info->sample_type = DBS_SUB_SAMPLE;
439 delay = dbs_info->freq_hi_jiffies;
442 __cpufreq_driver_target(dbs_info->cur_policy,
446 queue_delayed_work_on(cpu, kondemand_wq, &dbs_info->work, delay);
449 static inline void dbs_timer_init(unsigned int cpu)
451 struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, cpu);
452 /* We want all CPUs to do sampling nearly on same jiffy */
453 int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
454 delay -= jiffies % delay;
456 ondemand_powersave_bias_init();
457 INIT_DELAYED_WORK_NAR(&dbs_info->work, do_dbs_timer);
458 dbs_info->sample_type = DBS_NORMAL_SAMPLE;
459 queue_delayed_work_on(cpu, kondemand_wq, &dbs_info->work, delay);
462 static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
464 dbs_info->enable = 0;
465 cancel_delayed_work(&dbs_info->work);
466 flush_workqueue(kondemand_wq);
469 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
472 unsigned int cpu = policy->cpu;
473 struct cpu_dbs_info_s *this_dbs_info;
476 this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
479 case CPUFREQ_GOV_START:
480 if ((!cpu_online(cpu)) || (!policy->cur))
483 if (policy->cpuinfo.transition_latency >
484 (TRANSITION_LATENCY_LIMIT * 1000)) {
485 printk(KERN_WARNING "ondemand governor failed to load "
486 "due to too long transition latency\n");
489 if (this_dbs_info->enable) /* Already enabled */
492 mutex_lock(&dbs_mutex);
494 if (dbs_enable == 1) {
495 kondemand_wq = create_workqueue("kondemand");
497 printk(KERN_ERR "Creation of kondemand failed\n");
499 mutex_unlock(&dbs_mutex);
503 for_each_cpu_mask(j, policy->cpus) {
504 struct cpu_dbs_info_s *j_dbs_info;
505 j_dbs_info = &per_cpu(cpu_dbs_info, j);
506 j_dbs_info->cur_policy = policy;
508 j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j);
509 j_dbs_info->prev_cpu_wall = get_jiffies_64();
511 this_dbs_info->enable = 1;
512 sysfs_create_group(&policy->kobj, &dbs_attr_group);
514 * Start the timerschedule work, when this governor
515 * is used for first time
517 if (dbs_enable == 1) {
518 unsigned int latency;
519 /* policy latency is in nS. Convert it to uS first */
520 latency = policy->cpuinfo.transition_latency / 1000;
524 def_sampling_rate = latency *
525 DEF_SAMPLING_RATE_LATENCY_MULTIPLIER;
527 if (def_sampling_rate < MIN_STAT_SAMPLING_RATE)
528 def_sampling_rate = MIN_STAT_SAMPLING_RATE;
530 dbs_tuners_ins.sampling_rate = def_sampling_rate;
532 dbs_timer_init(policy->cpu);
534 mutex_unlock(&dbs_mutex);
537 case CPUFREQ_GOV_STOP:
538 mutex_lock(&dbs_mutex);
539 dbs_timer_exit(this_dbs_info);
540 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
543 destroy_workqueue(kondemand_wq);
545 mutex_unlock(&dbs_mutex);
549 case CPUFREQ_GOV_LIMITS:
550 mutex_lock(&dbs_mutex);
551 if (policy->max < this_dbs_info->cur_policy->cur)
552 __cpufreq_driver_target(this_dbs_info->cur_policy,
555 else if (policy->min > this_dbs_info->cur_policy->cur)
556 __cpufreq_driver_target(this_dbs_info->cur_policy,
559 mutex_unlock(&dbs_mutex);
565 static struct cpufreq_governor cpufreq_gov_dbs = {
567 .governor = cpufreq_governor_dbs,
568 .owner = THIS_MODULE,
571 static int __init cpufreq_gov_dbs_init(void)
573 return cpufreq_register_governor(&cpufreq_gov_dbs);
576 static void __exit cpufreq_gov_dbs_exit(void)
578 cpufreq_unregister_governor(&cpufreq_gov_dbs);
582 MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
583 MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
584 MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
585 "Low Latency Frequency Transition capable processors");
586 MODULE_LICENSE("GPL");
588 module_init(cpufreq_gov_dbs_init);
589 module_exit(cpufreq_gov_dbs_exit);