2 * drivers/cpufreq/cpufreq_interactive.c
4 * Copyright (C) 2010 Google, Inc.
6 * This software is licensed under the terms of the GNU General Public
7 * License version 2, as published by the Free Software Foundation, and
8 * may be copied, distributed, and modified under those terms.
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
15 * Author: Mike Chan (mike@android.com)
19 #include <linux/cpu.h>
20 #include <linux/cpumask.h>
21 #include <linux/cpufreq.h>
22 #include <linux/module.h>
23 #include <linux/moduleparam.h>
24 #include <linux/rwsem.h>
25 #include <linux/sched.h>
26 #include <linux/sched/rt.h>
27 #include <linux/tick.h>
28 #include <linux/time.h>
29 #include <linux/timer.h>
30 #include <linux/workqueue.h>
31 #include <linux/kthread.h>
32 #include <linux/slab.h>
33 #include "cpufreq_governor.h"
35 #define CREATE_TRACE_POINTS
36 #include <trace/events/cpufreq_interactive.h>
38 struct cpufreq_interactive_cpuinfo {
39 struct timer_list cpu_timer;
40 struct timer_list cpu_slack_timer;
41 spinlock_t load_lock; /* protects the next 4 fields */
43 u64 time_in_idle_timestamp;
45 u64 cputime_speedadj_timestamp;
46 struct cpufreq_policy *policy;
47 struct cpufreq_frequency_table *freq_table;
48 spinlock_t target_freq_lock; /*protects target freq */
49 unsigned int target_freq;
50 unsigned int floor_freq;
51 unsigned int max_freq;
52 u64 floor_validate_time;
53 u64 hispeed_validate_time;
54 struct rw_semaphore enable_sem;
58 static DEFINE_PER_CPU(struct cpufreq_interactive_cpuinfo, cpuinfo);
60 /* realtime thread handles frequency scaling */
61 static struct task_struct *speedchange_task;
62 static cpumask_t speedchange_cpumask;
63 static spinlock_t speedchange_cpumask_lock;
64 static struct mutex gov_lock;
66 /* Target load. Lower values result in higher CPU speeds. */
67 #define DEFAULT_TARGET_LOAD 90
68 static unsigned int default_target_loads[] = {DEFAULT_TARGET_LOAD};
70 #define DEFAULT_TIMER_RATE (20 * USEC_PER_MSEC)
71 #define DEFAULT_ABOVE_HISPEED_DELAY DEFAULT_TIMER_RATE
72 static unsigned int default_above_hispeed_delay[] = {
73 DEFAULT_ABOVE_HISPEED_DELAY };
75 struct cpufreq_interactive_tunables {
77 /* Hi speed to bump to from lo speed when load burst (default max) */
78 unsigned int hispeed_freq;
79 /* Go to hi speed when CPU load at or above this value. */
80 #define DEFAULT_GO_HISPEED_LOAD 99
81 unsigned long go_hispeed_load;
82 /* Target load. Lower values result in higher CPU speeds. */
83 spinlock_t target_loads_lock;
84 unsigned int *target_loads;
87 * The minimum amount of time to spend at a frequency before we can ramp
90 #define DEFAULT_MIN_SAMPLE_TIME (80 * USEC_PER_MSEC)
91 unsigned long min_sample_time;
93 * The sample rate of the timer used to increase frequency
95 unsigned long timer_rate;
97 * Wait this long before raising speed above hispeed, by default a
98 * single timer interval.
100 spinlock_t above_hispeed_delay_lock;
101 unsigned int *above_hispeed_delay;
102 int nabove_hispeed_delay;
103 /* Non-zero means indefinite speed boost active */
105 /* Duration of a boot pulse in usecs */
106 int boostpulse_duration_val;
107 /* End time of boost pulse in ktime converted to usecs */
108 u64 boostpulse_endtime;
111 * Max additional time to wait in idle, beyond timer_rate, at speeds
112 * above minimum before wakeup to reduce speed, or -1 if unnecessary.
114 #define DEFAULT_TIMER_SLACK (4 * DEFAULT_TIMER_RATE)
119 /* For cases where we have single governor instance for system */
120 static struct cpufreq_interactive_tunables *common_tunables;
122 static struct attribute_group *get_sysfs_attr(void);
124 static void cpufreq_interactive_timer_resched(
125 struct cpufreq_interactive_cpuinfo *pcpu)
127 struct cpufreq_interactive_tunables *tunables =
128 pcpu->policy->governor_data;
129 unsigned long expires;
132 spin_lock_irqsave(&pcpu->load_lock, flags);
134 get_cpu_idle_time(smp_processor_id(),
135 &pcpu->time_in_idle_timestamp,
136 tunables->io_is_busy);
137 pcpu->cputime_speedadj = 0;
138 pcpu->cputime_speedadj_timestamp = pcpu->time_in_idle_timestamp;
139 expires = jiffies + usecs_to_jiffies(tunables->timer_rate);
140 mod_timer_pinned(&pcpu->cpu_timer, expires);
142 if (tunables->timer_slack_val >= 0 &&
143 pcpu->target_freq > pcpu->policy->min) {
144 expires += usecs_to_jiffies(tunables->timer_slack_val);
145 mod_timer_pinned(&pcpu->cpu_slack_timer, expires);
148 spin_unlock_irqrestore(&pcpu->load_lock, flags);
151 /* The caller shall take enable_sem write semaphore to avoid any timer race.
152 * The cpu_timer and cpu_slack_timer must be deactivated when calling this
155 static void cpufreq_interactive_timer_start(
156 struct cpufreq_interactive_tunables *tunables, int cpu)
158 struct cpufreq_interactive_cpuinfo *pcpu = &per_cpu(cpuinfo, cpu);
159 unsigned long expires = jiffies +
160 usecs_to_jiffies(tunables->timer_rate);
163 pcpu->cpu_timer.expires = expires;
164 add_timer_on(&pcpu->cpu_timer, cpu);
165 if (tunables->timer_slack_val >= 0 &&
166 pcpu->target_freq > pcpu->policy->min) {
167 expires += usecs_to_jiffies(tunables->timer_slack_val);
168 pcpu->cpu_slack_timer.expires = expires;
169 add_timer_on(&pcpu->cpu_slack_timer, cpu);
172 spin_lock_irqsave(&pcpu->load_lock, flags);
174 get_cpu_idle_time(cpu, &pcpu->time_in_idle_timestamp,
175 tunables->io_is_busy);
176 pcpu->cputime_speedadj = 0;
177 pcpu->cputime_speedadj_timestamp = pcpu->time_in_idle_timestamp;
178 spin_unlock_irqrestore(&pcpu->load_lock, flags);
181 static unsigned int freq_to_above_hispeed_delay(
182 struct cpufreq_interactive_tunables *tunables,
189 spin_lock_irqsave(&tunables->above_hispeed_delay_lock, flags);
191 for (i = 0; i < tunables->nabove_hispeed_delay - 1 &&
192 freq >= tunables->above_hispeed_delay[i+1]; i += 2)
195 ret = tunables->above_hispeed_delay[i];
196 spin_unlock_irqrestore(&tunables->above_hispeed_delay_lock, flags);
200 static unsigned int freq_to_targetload(
201 struct cpufreq_interactive_tunables *tunables, unsigned int freq)
207 spin_lock_irqsave(&tunables->target_loads_lock, flags);
209 for (i = 0; i < tunables->ntarget_loads - 1 &&
210 freq >= tunables->target_loads[i+1]; i += 2)
213 ret = tunables->target_loads[i];
214 spin_unlock_irqrestore(&tunables->target_loads_lock, flags);
219 * If increasing frequencies never map to a lower target load then
220 * choose_freq() will find the minimum frequency that does not exceed its
221 * target load given the current load.
223 static unsigned int choose_freq(struct cpufreq_interactive_cpuinfo *pcpu,
224 unsigned int loadadjfreq)
226 unsigned int freq = pcpu->policy->cur;
227 unsigned int prevfreq, freqmin, freqmax;
236 tl = freq_to_targetload(pcpu->policy->governor_data, freq);
239 * Find the lowest frequency where the computed load is less
240 * than or equal to the target load.
243 if (cpufreq_frequency_table_target(
244 pcpu->policy, pcpu->freq_table, loadadjfreq / tl,
245 CPUFREQ_RELATION_L, &index))
247 freq = pcpu->freq_table[index].frequency;
249 if (freq > prevfreq) {
250 /* The previous frequency is too low. */
253 if (freq >= freqmax) {
255 * Find the highest frequency that is less
258 if (cpufreq_frequency_table_target(
259 pcpu->policy, pcpu->freq_table,
260 freqmax - 1, CPUFREQ_RELATION_H,
263 freq = pcpu->freq_table[index].frequency;
265 if (freq == freqmin) {
267 * The first frequency below freqmax
268 * has already been found to be too
269 * low. freqmax is the lowest speed
270 * we found that is fast enough.
276 } else if (freq < prevfreq) {
277 /* The previous frequency is high enough. */
280 if (freq <= freqmin) {
282 * Find the lowest frequency that is higher
285 if (cpufreq_frequency_table_target(
286 pcpu->policy, pcpu->freq_table,
287 freqmin + 1, CPUFREQ_RELATION_L,
290 freq = pcpu->freq_table[index].frequency;
293 * If freqmax is the first frequency above
294 * freqmin then we have already found that
295 * this speed is fast enough.
302 /* If same frequency chosen as previous then done. */
303 } while (freq != prevfreq);
308 static u64 update_load(int cpu)
310 struct cpufreq_interactive_cpuinfo *pcpu = &per_cpu(cpuinfo, cpu);
311 struct cpufreq_interactive_tunables *tunables =
312 pcpu->policy->governor_data;
319 now_idle = get_cpu_idle_time(cpu, &now, tunables->io_is_busy);
320 delta_idle = (now_idle - pcpu->time_in_idle);
321 delta_time = (now - pcpu->time_in_idle_timestamp);
323 if (delta_time <= delta_idle)
326 active_time = delta_time - delta_idle;
328 pcpu->cputime_speedadj += active_time * pcpu->policy->cur;
330 pcpu->time_in_idle = now_idle;
331 pcpu->time_in_idle_timestamp = now;
335 static void cpufreq_interactive_timer(unsigned long data)
338 unsigned int delta_time;
339 u64 cputime_speedadj;
341 struct cpufreq_interactive_cpuinfo *pcpu =
342 &per_cpu(cpuinfo, data);
343 struct cpufreq_interactive_tunables *tunables =
344 pcpu->policy->governor_data;
345 unsigned int new_freq;
346 unsigned int loadadjfreq;
350 if (!down_read_trylock(&pcpu->enable_sem))
352 if (!pcpu->governor_enabled)
355 spin_lock_irqsave(&pcpu->load_lock, flags);
356 now = update_load(data);
357 delta_time = (unsigned int)(now - pcpu->cputime_speedadj_timestamp);
358 cputime_speedadj = pcpu->cputime_speedadj;
359 spin_unlock_irqrestore(&pcpu->load_lock, flags);
361 if (WARN_ON_ONCE(!delta_time))
364 spin_lock_irqsave(&pcpu->target_freq_lock, flags);
365 do_div(cputime_speedadj, delta_time);
366 loadadjfreq = (unsigned int)cputime_speedadj * 100;
367 cpu_load = loadadjfreq / pcpu->target_freq;
368 tunables->boosted = tunables->boost_val || now < tunables->boostpulse_endtime;
370 if (cpu_load >= tunables->go_hispeed_load || tunables->boosted) {
371 if (pcpu->target_freq < tunables->hispeed_freq) {
372 new_freq = tunables->hispeed_freq;
374 new_freq = choose_freq(pcpu, loadadjfreq);
376 if (new_freq < tunables->hispeed_freq)
377 new_freq = tunables->hispeed_freq;
380 new_freq = choose_freq(pcpu, loadadjfreq);
381 if (new_freq > tunables->hispeed_freq &&
382 pcpu->target_freq < tunables->hispeed_freq)
383 new_freq = tunables->hispeed_freq;
386 if (pcpu->target_freq >= tunables->hispeed_freq &&
387 new_freq > pcpu->target_freq &&
388 now - pcpu->hispeed_validate_time <
389 freq_to_above_hispeed_delay(tunables, pcpu->target_freq)) {
390 trace_cpufreq_interactive_notyet(
391 data, cpu_load, pcpu->target_freq,
392 pcpu->policy->cur, new_freq);
393 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
397 pcpu->hispeed_validate_time = now;
399 if (cpufreq_frequency_table_target(pcpu->policy, pcpu->freq_table,
400 new_freq, CPUFREQ_RELATION_L,
402 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
406 new_freq = pcpu->freq_table[index].frequency;
409 * Do not scale below floor_freq unless we have been at or above the
410 * floor frequency for the minimum sample time since last validated.
412 if (new_freq < pcpu->floor_freq) {
413 if (now - pcpu->floor_validate_time <
414 tunables->min_sample_time) {
415 trace_cpufreq_interactive_notyet(
416 data, cpu_load, pcpu->target_freq,
417 pcpu->policy->cur, new_freq);
418 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
424 * Update the timestamp for checking whether speed has been held at
425 * or above the selected frequency for a minimum of min_sample_time,
426 * if not boosted to hispeed_freq. If boosted to hispeed_freq then we
427 * allow the speed to drop as soon as the boostpulse duration expires
428 * (or the indefinite boost is turned off).
431 if (!tunables->boosted || new_freq > tunables->hispeed_freq) {
432 pcpu->floor_freq = new_freq;
433 pcpu->floor_validate_time = now;
436 if (pcpu->target_freq == new_freq &&
437 pcpu->target_freq <= pcpu->policy->cur) {
438 trace_cpufreq_interactive_already(
439 data, cpu_load, pcpu->target_freq,
440 pcpu->policy->cur, new_freq);
441 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
442 goto rearm_if_notmax;
445 trace_cpufreq_interactive_target(data, cpu_load, pcpu->target_freq,
446 pcpu->policy->cur, new_freq);
448 pcpu->target_freq = new_freq;
449 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
450 spin_lock_irqsave(&speedchange_cpumask_lock, flags);
451 cpumask_set_cpu(data, &speedchange_cpumask);
452 spin_unlock_irqrestore(&speedchange_cpumask_lock, flags);
453 wake_up_process(speedchange_task);
457 * Already set max speed and don't see a need to change that,
458 * wait until next idle to re-evaluate, don't need timer.
460 if (pcpu->target_freq == pcpu->policy->max)
464 if (!timer_pending(&pcpu->cpu_timer))
465 cpufreq_interactive_timer_resched(pcpu);
468 up_read(&pcpu->enable_sem);
472 static void cpufreq_interactive_idle_start(void)
474 struct cpufreq_interactive_cpuinfo *pcpu =
475 &per_cpu(cpuinfo, smp_processor_id());
478 if (!down_read_trylock(&pcpu->enable_sem))
480 if (!pcpu->governor_enabled) {
481 up_read(&pcpu->enable_sem);
485 pending = timer_pending(&pcpu->cpu_timer);
487 if (pcpu->target_freq != pcpu->policy->min) {
489 * Entering idle while not at lowest speed. On some
490 * platforms this can hold the other CPU(s) at that speed
491 * even though the CPU is idle. Set a timer to re-evaluate
492 * speed so this idle CPU doesn't hold the other CPUs above
493 * min indefinitely. This should probably be a quirk of
494 * the CPUFreq driver.
497 cpufreq_interactive_timer_resched(pcpu);
500 up_read(&pcpu->enable_sem);
503 static void cpufreq_interactive_idle_end(void)
505 struct cpufreq_interactive_cpuinfo *pcpu =
506 &per_cpu(cpuinfo, smp_processor_id());
508 if (!down_read_trylock(&pcpu->enable_sem))
510 if (!pcpu->governor_enabled) {
511 up_read(&pcpu->enable_sem);
515 /* Arm the timer for 1-2 ticks later if not already. */
516 if (!timer_pending(&pcpu->cpu_timer)) {
517 cpufreq_interactive_timer_resched(pcpu);
518 } else if (time_after_eq(jiffies, pcpu->cpu_timer.expires)) {
519 del_timer(&pcpu->cpu_timer);
520 del_timer(&pcpu->cpu_slack_timer);
521 cpufreq_interactive_timer(smp_processor_id());
524 up_read(&pcpu->enable_sem);
527 static int cpufreq_interactive_speedchange_task(void *data)
532 struct cpufreq_interactive_cpuinfo *pcpu;
535 set_current_state(TASK_INTERRUPTIBLE);
536 spin_lock_irqsave(&speedchange_cpumask_lock, flags);
538 if (cpumask_empty(&speedchange_cpumask)) {
539 spin_unlock_irqrestore(&speedchange_cpumask_lock,
543 if (kthread_should_stop())
546 spin_lock_irqsave(&speedchange_cpumask_lock, flags);
549 set_current_state(TASK_RUNNING);
550 tmp_mask = speedchange_cpumask;
551 cpumask_clear(&speedchange_cpumask);
552 spin_unlock_irqrestore(&speedchange_cpumask_lock, flags);
554 for_each_cpu(cpu, &tmp_mask) {
556 unsigned int max_freq = 0;
558 pcpu = &per_cpu(cpuinfo, cpu);
559 if (!down_read_trylock(&pcpu->enable_sem))
561 if (!pcpu->governor_enabled) {
562 up_read(&pcpu->enable_sem);
566 for_each_cpu(j, pcpu->policy->cpus) {
567 struct cpufreq_interactive_cpuinfo *pjcpu =
568 &per_cpu(cpuinfo, j);
570 if (pjcpu->target_freq > max_freq)
571 max_freq = pjcpu->target_freq;
574 if (max_freq != pcpu->policy->cur)
575 __cpufreq_driver_target(pcpu->policy,
578 trace_cpufreq_interactive_setspeed(cpu,
582 up_read(&pcpu->enable_sem);
589 static void cpufreq_interactive_boost(struct cpufreq_interactive_tunables *tunables)
593 unsigned long flags[2];
594 struct cpufreq_interactive_cpuinfo *pcpu;
596 tunables->boosted = true;
598 spin_lock_irqsave(&speedchange_cpumask_lock, flags[0]);
600 for_each_online_cpu(i) {
601 pcpu = &per_cpu(cpuinfo, i);
602 if (tunables != pcpu->policy->governor_data)
605 spin_lock_irqsave(&pcpu->target_freq_lock, flags[1]);
606 if (pcpu->target_freq < tunables->hispeed_freq) {
607 pcpu->target_freq = tunables->hispeed_freq;
608 cpumask_set_cpu(i, &speedchange_cpumask);
609 pcpu->hispeed_validate_time =
610 ktime_to_us(ktime_get());
615 * Set floor freq and (re)start timer for when last
619 pcpu->floor_freq = tunables->hispeed_freq;
620 pcpu->floor_validate_time = ktime_to_us(ktime_get());
621 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags[1]);
624 spin_unlock_irqrestore(&speedchange_cpumask_lock, flags[0]);
627 wake_up_process(speedchange_task);
630 static int cpufreq_interactive_notifier(
631 struct notifier_block *nb, unsigned long val, void *data)
633 struct cpufreq_freqs *freq = data;
634 struct cpufreq_interactive_cpuinfo *pcpu;
638 if (val == CPUFREQ_POSTCHANGE) {
639 pcpu = &per_cpu(cpuinfo, freq->cpu);
640 if (!down_read_trylock(&pcpu->enable_sem))
642 if (!pcpu->governor_enabled) {
643 up_read(&pcpu->enable_sem);
647 for_each_cpu(cpu, pcpu->policy->cpus) {
648 struct cpufreq_interactive_cpuinfo *pjcpu =
649 &per_cpu(cpuinfo, cpu);
650 if (cpu != freq->cpu) {
651 if (!down_read_trylock(&pjcpu->enable_sem))
653 if (!pjcpu->governor_enabled) {
654 up_read(&pjcpu->enable_sem);
658 spin_lock_irqsave(&pjcpu->load_lock, flags);
660 spin_unlock_irqrestore(&pjcpu->load_lock, flags);
661 if (cpu != freq->cpu)
662 up_read(&pjcpu->enable_sem);
665 up_read(&pcpu->enable_sem);
670 static struct notifier_block cpufreq_notifier_block = {
671 .notifier_call = cpufreq_interactive_notifier,
674 static unsigned int *get_tokenized_data(const char *buf, int *num_tokens)
679 unsigned int *tokenized_data;
683 while ((cp = strpbrk(cp + 1, " :")))
686 if (!(ntokens & 0x1))
689 tokenized_data = kmalloc(ntokens * sizeof(unsigned int), GFP_KERNEL);
690 if (!tokenized_data) {
697 while (i < ntokens) {
698 if (sscanf(cp, "%u", &tokenized_data[i++]) != 1)
701 cp = strpbrk(cp, " :");
710 *num_tokens = ntokens;
711 return tokenized_data;
714 kfree(tokenized_data);
719 static ssize_t show_target_loads(
720 struct cpufreq_interactive_tunables *tunables,
727 spin_lock_irqsave(&tunables->target_loads_lock, flags);
729 for (i = 0; i < tunables->ntarget_loads; i++)
730 ret += sprintf(buf + ret, "%u%s", tunables->target_loads[i],
731 i & 0x1 ? ":" : " ");
733 sprintf(buf + ret - 1, "\n");
734 spin_unlock_irqrestore(&tunables->target_loads_lock, flags);
738 static ssize_t store_target_loads(
739 struct cpufreq_interactive_tunables *tunables,
740 const char *buf, size_t count)
743 unsigned int *new_target_loads = NULL;
746 new_target_loads = get_tokenized_data(buf, &ntokens);
747 if (IS_ERR(new_target_loads))
748 return PTR_RET(new_target_loads);
750 spin_lock_irqsave(&tunables->target_loads_lock, flags);
751 if (tunables->target_loads != default_target_loads)
752 kfree(tunables->target_loads);
753 tunables->target_loads = new_target_loads;
754 tunables->ntarget_loads = ntokens;
755 spin_unlock_irqrestore(&tunables->target_loads_lock, flags);
759 static ssize_t show_above_hispeed_delay(
760 struct cpufreq_interactive_tunables *tunables, char *buf)
766 spin_lock_irqsave(&tunables->above_hispeed_delay_lock, flags);
768 for (i = 0; i < tunables->nabove_hispeed_delay; i++)
769 ret += sprintf(buf + ret, "%u%s",
770 tunables->above_hispeed_delay[i],
771 i & 0x1 ? ":" : " ");
773 sprintf(buf + ret - 1, "\n");
774 spin_unlock_irqrestore(&tunables->above_hispeed_delay_lock, flags);
778 static ssize_t store_above_hispeed_delay(
779 struct cpufreq_interactive_tunables *tunables,
780 const char *buf, size_t count)
783 unsigned int *new_above_hispeed_delay = NULL;
786 new_above_hispeed_delay = get_tokenized_data(buf, &ntokens);
787 if (IS_ERR(new_above_hispeed_delay))
788 return PTR_RET(new_above_hispeed_delay);
790 spin_lock_irqsave(&tunables->above_hispeed_delay_lock, flags);
791 if (tunables->above_hispeed_delay != default_above_hispeed_delay)
792 kfree(tunables->above_hispeed_delay);
793 tunables->above_hispeed_delay = new_above_hispeed_delay;
794 tunables->nabove_hispeed_delay = ntokens;
795 spin_unlock_irqrestore(&tunables->above_hispeed_delay_lock, flags);
800 static ssize_t show_hispeed_freq(struct cpufreq_interactive_tunables *tunables,
803 return sprintf(buf, "%u\n", tunables->hispeed_freq);
806 static ssize_t store_hispeed_freq(struct cpufreq_interactive_tunables *tunables,
807 const char *buf, size_t count)
810 long unsigned int val;
812 ret = strict_strtoul(buf, 0, &val);
815 tunables->hispeed_freq = val;
819 static ssize_t show_go_hispeed_load(struct cpufreq_interactive_tunables
820 *tunables, char *buf)
822 return sprintf(buf, "%lu\n", tunables->go_hispeed_load);
825 static ssize_t store_go_hispeed_load(struct cpufreq_interactive_tunables
826 *tunables, const char *buf, size_t count)
831 ret = strict_strtoul(buf, 0, &val);
834 tunables->go_hispeed_load = val;
838 static ssize_t show_min_sample_time(struct cpufreq_interactive_tunables
839 *tunables, char *buf)
841 return sprintf(buf, "%lu\n", tunables->min_sample_time);
844 static ssize_t store_min_sample_time(struct cpufreq_interactive_tunables
845 *tunables, const char *buf, size_t count)
850 ret = strict_strtoul(buf, 0, &val);
853 tunables->min_sample_time = val;
857 static ssize_t show_timer_rate(struct cpufreq_interactive_tunables *tunables,
860 return sprintf(buf, "%lu\n", tunables->timer_rate);
863 static ssize_t store_timer_rate(struct cpufreq_interactive_tunables *tunables,
864 const char *buf, size_t count)
869 ret = strict_strtoul(buf, 0, &val);
872 tunables->timer_rate = val;
876 static ssize_t show_timer_slack(struct cpufreq_interactive_tunables *tunables,
879 return sprintf(buf, "%d\n", tunables->timer_slack_val);
882 static ssize_t store_timer_slack(struct cpufreq_interactive_tunables *tunables,
883 const char *buf, size_t count)
888 ret = kstrtol(buf, 10, &val);
892 tunables->timer_slack_val = val;
896 static ssize_t show_boost(struct cpufreq_interactive_tunables *tunables,
899 return sprintf(buf, "%d\n", tunables->boost_val);
902 static ssize_t store_boost(struct cpufreq_interactive_tunables *tunables,
903 const char *buf, size_t count)
908 ret = kstrtoul(buf, 0, &val);
912 tunables->boost_val = val;
914 if (tunables->boost_val) {
915 trace_cpufreq_interactive_boost("on");
916 if (!tunables->boosted)
917 cpufreq_interactive_boost(tunables);
919 tunables->boostpulse_endtime = ktime_to_us(ktime_get());
920 trace_cpufreq_interactive_unboost("off");
926 static ssize_t store_boostpulse(struct cpufreq_interactive_tunables *tunables,
927 const char *buf, size_t count)
932 ret = kstrtoul(buf, 0, &val);
936 tunables->boostpulse_endtime = ktime_to_us(ktime_get()) +
937 tunables->boostpulse_duration_val;
938 trace_cpufreq_interactive_boost("pulse");
939 if (!tunables->boosted)
940 cpufreq_interactive_boost(tunables);
944 static ssize_t show_boostpulse_duration(struct cpufreq_interactive_tunables
945 *tunables, char *buf)
947 return sprintf(buf, "%d\n", tunables->boostpulse_duration_val);
950 static ssize_t store_boostpulse_duration(struct cpufreq_interactive_tunables
951 *tunables, const char *buf, size_t count)
956 ret = kstrtoul(buf, 0, &val);
960 tunables->boostpulse_duration_val = val;
964 static ssize_t show_io_is_busy(struct cpufreq_interactive_tunables *tunables,
967 return sprintf(buf, "%u\n", tunables->io_is_busy);
970 static ssize_t store_io_is_busy(struct cpufreq_interactive_tunables *tunables,
971 const char *buf, size_t count)
976 ret = kstrtoul(buf, 0, &val);
979 tunables->io_is_busy = val;
984 * Create show/store routines
985 * - sys: One governor instance for complete SYSTEM
986 * - pol: One governor instance per struct cpufreq_policy
988 #define show_gov_pol_sys(file_name) \
989 static ssize_t show_##file_name##_gov_sys \
990 (struct kobject *kobj, struct attribute *attr, char *buf) \
992 return show_##file_name(common_tunables, buf); \
995 static ssize_t show_##file_name##_gov_pol \
996 (struct cpufreq_policy *policy, char *buf) \
998 return show_##file_name(policy->governor_data, buf); \
1001 #define store_gov_pol_sys(file_name) \
1002 static ssize_t store_##file_name##_gov_sys \
1003 (struct kobject *kobj, struct attribute *attr, const char *buf, \
1006 return store_##file_name(common_tunables, buf, count); \
1009 static ssize_t store_##file_name##_gov_pol \
1010 (struct cpufreq_policy *policy, const char *buf, size_t count) \
1012 return store_##file_name(policy->governor_data, buf, count); \
1015 #define show_store_gov_pol_sys(file_name) \
1016 show_gov_pol_sys(file_name); \
1017 store_gov_pol_sys(file_name)
1019 show_store_gov_pol_sys(target_loads);
1020 show_store_gov_pol_sys(above_hispeed_delay);
1021 show_store_gov_pol_sys(hispeed_freq);
1022 show_store_gov_pol_sys(go_hispeed_load);
1023 show_store_gov_pol_sys(min_sample_time);
1024 show_store_gov_pol_sys(timer_rate);
1025 show_store_gov_pol_sys(timer_slack);
1026 show_store_gov_pol_sys(boost);
1027 store_gov_pol_sys(boostpulse);
1028 show_store_gov_pol_sys(boostpulse_duration);
1029 show_store_gov_pol_sys(io_is_busy);
1031 #define gov_sys_attr_rw(_name) \
1032 static struct global_attr _name##_gov_sys = \
1033 __ATTR(_name, 0644, show_##_name##_gov_sys, store_##_name##_gov_sys)
1035 #define gov_pol_attr_rw(_name) \
1036 static struct freq_attr _name##_gov_pol = \
1037 __ATTR(_name, 0644, show_##_name##_gov_pol, store_##_name##_gov_pol)
1039 #define gov_sys_pol_attr_rw(_name) \
1040 gov_sys_attr_rw(_name); \
1041 gov_pol_attr_rw(_name)
1043 gov_sys_pol_attr_rw(target_loads);
1044 gov_sys_pol_attr_rw(above_hispeed_delay);
1045 gov_sys_pol_attr_rw(hispeed_freq);
1046 gov_sys_pol_attr_rw(go_hispeed_load);
1047 gov_sys_pol_attr_rw(min_sample_time);
1048 gov_sys_pol_attr_rw(timer_rate);
1049 gov_sys_pol_attr_rw(timer_slack);
1050 gov_sys_pol_attr_rw(boost);
1051 gov_sys_pol_attr_rw(boostpulse_duration);
1052 gov_sys_pol_attr_rw(io_is_busy);
1054 static struct global_attr boostpulse_gov_sys =
1055 __ATTR(boostpulse, 0200, NULL, store_boostpulse_gov_sys);
1057 static struct freq_attr boostpulse_gov_pol =
1058 __ATTR(boostpulse, 0200, NULL, store_boostpulse_gov_pol);
1060 /* One Governor instance for entire system */
1061 static struct attribute *interactive_attributes_gov_sys[] = {
1062 &target_loads_gov_sys.attr,
1063 &above_hispeed_delay_gov_sys.attr,
1064 &hispeed_freq_gov_sys.attr,
1065 &go_hispeed_load_gov_sys.attr,
1066 &min_sample_time_gov_sys.attr,
1067 &timer_rate_gov_sys.attr,
1068 &timer_slack_gov_sys.attr,
1069 &boost_gov_sys.attr,
1070 &boostpulse_gov_sys.attr,
1071 &boostpulse_duration_gov_sys.attr,
1072 &io_is_busy_gov_sys.attr,
1076 static struct attribute_group interactive_attr_group_gov_sys = {
1077 .attrs = interactive_attributes_gov_sys,
1078 .name = "interactive",
1081 /* Per policy governor instance */
1082 static struct attribute *interactive_attributes_gov_pol[] = {
1083 &target_loads_gov_pol.attr,
1084 &above_hispeed_delay_gov_pol.attr,
1085 &hispeed_freq_gov_pol.attr,
1086 &go_hispeed_load_gov_pol.attr,
1087 &min_sample_time_gov_pol.attr,
1088 &timer_rate_gov_pol.attr,
1089 &timer_slack_gov_pol.attr,
1090 &boost_gov_pol.attr,
1091 &boostpulse_gov_pol.attr,
1092 &boostpulse_duration_gov_pol.attr,
1093 &io_is_busy_gov_pol.attr,
1097 static struct attribute_group interactive_attr_group_gov_pol = {
1098 .attrs = interactive_attributes_gov_pol,
1099 .name = "interactive",
1102 static struct attribute_group *get_sysfs_attr(void)
1104 if (have_governor_per_policy())
1105 return &interactive_attr_group_gov_pol;
1107 return &interactive_attr_group_gov_sys;
1110 static int cpufreq_interactive_idle_notifier(struct notifier_block *nb,
1116 cpufreq_interactive_idle_start();
1119 cpufreq_interactive_idle_end();
1126 static struct notifier_block cpufreq_interactive_idle_nb = {
1127 .notifier_call = cpufreq_interactive_idle_notifier,
1130 static int cpufreq_governor_interactive(struct cpufreq_policy *policy,
1135 struct cpufreq_interactive_cpuinfo *pcpu;
1136 struct cpufreq_frequency_table *freq_table;
1137 struct cpufreq_interactive_tunables *tunables;
1138 unsigned long flags;
1140 if (have_governor_per_policy())
1141 tunables = policy->governor_data;
1143 tunables = common_tunables;
1145 WARN_ON(!tunables && (event != CPUFREQ_GOV_POLICY_INIT));
1148 case CPUFREQ_GOV_POLICY_INIT:
1149 if (have_governor_per_policy()) {
1151 } else if (tunables) {
1152 tunables->usage_count++;
1153 policy->governor_data = tunables;
1157 tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
1159 pr_err("%s: POLICY_INIT: kzalloc failed\n", __func__);
1163 tunables->usage_count = 1;
1164 tunables->above_hispeed_delay = default_above_hispeed_delay;
1165 tunables->nabove_hispeed_delay =
1166 ARRAY_SIZE(default_above_hispeed_delay);
1167 tunables->go_hispeed_load = DEFAULT_GO_HISPEED_LOAD;
1168 tunables->target_loads = default_target_loads;
1169 tunables->ntarget_loads = ARRAY_SIZE(default_target_loads);
1170 tunables->min_sample_time = DEFAULT_MIN_SAMPLE_TIME;
1171 tunables->timer_rate = DEFAULT_TIMER_RATE;
1172 tunables->boostpulse_duration_val = DEFAULT_MIN_SAMPLE_TIME;
1173 tunables->timer_slack_val = DEFAULT_TIMER_SLACK;
1175 spin_lock_init(&tunables->target_loads_lock);
1176 spin_lock_init(&tunables->above_hispeed_delay_lock);
1178 policy->governor_data = tunables;
1179 if (!have_governor_per_policy())
1180 common_tunables = tunables;
1182 rc = sysfs_create_group(get_governor_parent_kobj(policy),
1186 policy->governor_data = NULL;
1187 if (!have_governor_per_policy())
1188 common_tunables = NULL;
1192 if (!policy->governor->initialized) {
1193 idle_notifier_register(&cpufreq_interactive_idle_nb);
1194 cpufreq_register_notifier(&cpufreq_notifier_block,
1195 CPUFREQ_TRANSITION_NOTIFIER);
1200 case CPUFREQ_GOV_POLICY_EXIT:
1201 if (!--tunables->usage_count) {
1202 if (policy->governor->initialized == 1) {
1203 cpufreq_unregister_notifier(&cpufreq_notifier_block,
1204 CPUFREQ_TRANSITION_NOTIFIER);
1205 idle_notifier_unregister(&cpufreq_interactive_idle_nb);
1208 sysfs_remove_group(get_governor_parent_kobj(policy),
1211 common_tunables = NULL;
1214 policy->governor_data = NULL;
1217 case CPUFREQ_GOV_START:
1218 mutex_lock(&gov_lock);
1220 freq_table = cpufreq_frequency_get_table(policy->cpu);
1221 if (!tunables->hispeed_freq)
1222 tunables->hispeed_freq = policy->max;
1224 for_each_cpu(j, policy->cpus) {
1225 pcpu = &per_cpu(cpuinfo, j);
1226 pcpu->policy = policy;
1227 pcpu->target_freq = policy->cur;
1228 pcpu->freq_table = freq_table;
1229 pcpu->floor_freq = pcpu->target_freq;
1230 pcpu->floor_validate_time =
1231 ktime_to_us(ktime_get());
1232 pcpu->hispeed_validate_time =
1233 pcpu->floor_validate_time;
1234 pcpu->max_freq = policy->max;
1235 down_write(&pcpu->enable_sem);
1236 del_timer_sync(&pcpu->cpu_timer);
1237 del_timer_sync(&pcpu->cpu_slack_timer);
1238 cpufreq_interactive_timer_start(tunables, j);
1239 pcpu->governor_enabled = 1;
1240 up_write(&pcpu->enable_sem);
1243 mutex_unlock(&gov_lock);
1246 case CPUFREQ_GOV_STOP:
1247 mutex_lock(&gov_lock);
1248 for_each_cpu(j, policy->cpus) {
1249 pcpu = &per_cpu(cpuinfo, j);
1250 down_write(&pcpu->enable_sem);
1251 pcpu->governor_enabled = 0;
1252 del_timer_sync(&pcpu->cpu_timer);
1253 del_timer_sync(&pcpu->cpu_slack_timer);
1254 up_write(&pcpu->enable_sem);
1257 mutex_unlock(&gov_lock);
1260 case CPUFREQ_GOV_LIMITS:
1261 if (policy->max < policy->cur)
1262 __cpufreq_driver_target(policy,
1263 policy->max, CPUFREQ_RELATION_H);
1264 else if (policy->min > policy->cur)
1265 __cpufreq_driver_target(policy,
1266 policy->min, CPUFREQ_RELATION_L);
1267 for_each_cpu(j, policy->cpus) {
1268 pcpu = &per_cpu(cpuinfo, j);
1270 down_read(&pcpu->enable_sem);
1271 if (pcpu->governor_enabled == 0) {
1272 up_read(&pcpu->enable_sem);
1276 spin_lock_irqsave(&pcpu->target_freq_lock, flags);
1277 if (policy->max < pcpu->target_freq)
1278 pcpu->target_freq = policy->max;
1279 else if (policy->min > pcpu->target_freq)
1280 pcpu->target_freq = policy->min;
1282 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
1283 up_read(&pcpu->enable_sem);
1285 /* Reschedule timer only if policy->max is raised.
1286 * Delete the timers, else the timer callback may
1287 * return without re-arm the timer when failed
1288 * acquire the semaphore. This race may cause timer
1289 * stopped unexpectedly.
1292 if (policy->max > pcpu->max_freq) {
1293 down_write(&pcpu->enable_sem);
1294 del_timer_sync(&pcpu->cpu_timer);
1295 del_timer_sync(&pcpu->cpu_slack_timer);
1296 cpufreq_interactive_timer_start(tunables, j);
1297 up_write(&pcpu->enable_sem);
1300 pcpu->max_freq = policy->max;
1307 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_INTERACTIVE
1310 struct cpufreq_governor cpufreq_gov_interactive = {
1311 .name = "interactive",
1312 .governor = cpufreq_governor_interactive,
1313 .max_transition_latency = 10000000,
1314 .owner = THIS_MODULE,
1317 static void cpufreq_interactive_nop_timer(unsigned long data)
1321 static int __init cpufreq_interactive_init(void)
1324 struct cpufreq_interactive_cpuinfo *pcpu;
1325 struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
1327 /* Initalize per-cpu timers */
1328 for_each_possible_cpu(i) {
1329 pcpu = &per_cpu(cpuinfo, i);
1330 init_timer_deferrable(&pcpu->cpu_timer);
1331 pcpu->cpu_timer.function = cpufreq_interactive_timer;
1332 pcpu->cpu_timer.data = i;
1333 init_timer(&pcpu->cpu_slack_timer);
1334 pcpu->cpu_slack_timer.function = cpufreq_interactive_nop_timer;
1335 spin_lock_init(&pcpu->load_lock);
1336 spin_lock_init(&pcpu->target_freq_lock);
1337 init_rwsem(&pcpu->enable_sem);
1340 spin_lock_init(&speedchange_cpumask_lock);
1341 mutex_init(&gov_lock);
1343 kthread_create(cpufreq_interactive_speedchange_task, NULL,
1345 if (IS_ERR(speedchange_task))
1346 return PTR_ERR(speedchange_task);
1348 sched_setscheduler_nocheck(speedchange_task, SCHED_FIFO, ¶m);
1349 get_task_struct(speedchange_task);
1351 /* NB: wake up so the thread does not look hung to the freezer */
1352 wake_up_process(speedchange_task);
1354 return cpufreq_register_governor(&cpufreq_gov_interactive);
1357 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_INTERACTIVE
1358 fs_initcall(cpufreq_interactive_init);
1360 module_init(cpufreq_interactive_init);
1363 static void __exit cpufreq_interactive_exit(void)
1365 cpufreq_unregister_governor(&cpufreq_gov_interactive);
1366 kthread_stop(speedchange_task);
1367 put_task_struct(speedchange_task);
1370 module_exit(cpufreq_interactive_exit);
1372 MODULE_AUTHOR("Mike Chan <mike@android.com>");
1373 MODULE_DESCRIPTION("'cpufreq_interactive' - A cpufreq governor for "
1374 "Latency sensitive workloads");
1375 MODULE_LICENSE("GPL");