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>
34 #define CREATE_TRACE_POINTS
35 #include <trace/events/cpufreq_interactive.h>
37 struct cpufreq_interactive_cpuinfo {
38 struct timer_list cpu_timer;
39 struct timer_list cpu_slack_timer;
40 spinlock_t load_lock; /* protects the next 4 fields */
42 u64 time_in_idle_timestamp;
44 u64 cputime_speedadj_timestamp;
45 struct cpufreq_policy *policy;
46 struct cpufreq_frequency_table *freq_table;
47 spinlock_t target_freq_lock; /*protects target freq */
48 unsigned int target_freq;
49 unsigned int floor_freq;
50 unsigned int max_freq;
51 u64 floor_validate_time;
52 u64 hispeed_validate_time;
53 struct rw_semaphore enable_sem;
57 static DEFINE_PER_CPU(struct cpufreq_interactive_cpuinfo, cpuinfo);
59 /* realtime thread handles frequency scaling */
60 static struct task_struct *speedchange_task;
61 static cpumask_t speedchange_cpumask;
62 static spinlock_t speedchange_cpumask_lock;
63 static struct mutex gov_lock;
65 /* Target load. Lower values result in higher CPU speeds. */
66 #define DEFAULT_TARGET_LOAD 90
67 static unsigned int default_target_loads[] = {DEFAULT_TARGET_LOAD};
69 #define DEFAULT_TIMER_RATE (20 * USEC_PER_MSEC)
70 #define DEFAULT_ABOVE_HISPEED_DELAY DEFAULT_TIMER_RATE
71 static unsigned int default_above_hispeed_delay[] = {
72 DEFAULT_ABOVE_HISPEED_DELAY };
74 struct cpufreq_interactive_tunables {
76 /* Hi speed to bump to from lo speed when load burst (default max) */
77 unsigned int hispeed_freq;
78 /* Go to hi speed when CPU load at or above this value. */
79 #define DEFAULT_GO_HISPEED_LOAD 99
80 unsigned long go_hispeed_load;
81 /* Target load. Lower values result in higher CPU speeds. */
82 spinlock_t target_loads_lock;
83 unsigned int *target_loads;
86 * The minimum amount of time to spend at a frequency before we can ramp
89 #define DEFAULT_MIN_SAMPLE_TIME (80 * USEC_PER_MSEC)
90 unsigned long min_sample_time;
92 * The sample rate of the timer used to increase frequency
94 unsigned long timer_rate;
96 * Wait this long before raising speed above hispeed, by default a
97 * single timer interval.
99 spinlock_t above_hispeed_delay_lock;
100 unsigned int *above_hispeed_delay;
101 int nabove_hispeed_delay;
102 /* Non-zero means indefinite speed boost active */
104 /* Duration of a boot pulse in usecs */
105 int boostpulse_duration_val;
106 /* End time of boost pulse in ktime converted to usecs */
107 u64 boostpulse_endtime;
109 * Max additional time to wait in idle, beyond timer_rate, at speeds
110 * above minimum before wakeup to reduce speed, or -1 if unnecessary.
112 #define DEFAULT_TIMER_SLACK (4 * DEFAULT_TIMER_RATE)
117 /* For cases where we have single governor instance for system */
118 static struct cpufreq_interactive_tunables *common_tunables;
120 static struct attribute_group *get_sysfs_attr(void);
122 static void cpufreq_interactive_timer_resched(
123 struct cpufreq_interactive_cpuinfo *pcpu)
125 struct cpufreq_interactive_tunables *tunables =
126 pcpu->policy->governor_data;
127 unsigned long expires;
130 spin_lock_irqsave(&pcpu->load_lock, flags);
132 get_cpu_idle_time(smp_processor_id(),
133 &pcpu->time_in_idle_timestamp,
134 tunables->io_is_busy);
135 pcpu->cputime_speedadj = 0;
136 pcpu->cputime_speedadj_timestamp = pcpu->time_in_idle_timestamp;
137 expires = jiffies + usecs_to_jiffies(tunables->timer_rate);
138 mod_timer_pinned(&pcpu->cpu_timer, expires);
140 if (tunables->timer_slack_val >= 0 &&
141 pcpu->target_freq > pcpu->policy->min) {
142 expires += usecs_to_jiffies(tunables->timer_slack_val);
143 mod_timer_pinned(&pcpu->cpu_slack_timer, expires);
146 spin_unlock_irqrestore(&pcpu->load_lock, flags);
149 /* The caller shall take enable_sem write semaphore to avoid any timer race.
150 * The cpu_timer and cpu_slack_timer must be deactivated when calling this
153 static void cpufreq_interactive_timer_start(
154 struct cpufreq_interactive_tunables *tunables, int cpu)
156 struct cpufreq_interactive_cpuinfo *pcpu = &per_cpu(cpuinfo, cpu);
157 unsigned long expires = jiffies +
158 usecs_to_jiffies(tunables->timer_rate);
161 pcpu->cpu_timer.expires = expires;
162 add_timer_on(&pcpu->cpu_timer, cpu);
163 if (tunables->timer_slack_val >= 0 &&
164 pcpu->target_freq > pcpu->policy->min) {
165 expires += usecs_to_jiffies(tunables->timer_slack_val);
166 pcpu->cpu_slack_timer.expires = expires;
167 add_timer_on(&pcpu->cpu_slack_timer, cpu);
170 spin_lock_irqsave(&pcpu->load_lock, flags);
172 get_cpu_idle_time(cpu, &pcpu->time_in_idle_timestamp,
173 tunables->io_is_busy);
174 pcpu->cputime_speedadj = 0;
175 pcpu->cputime_speedadj_timestamp = pcpu->time_in_idle_timestamp;
176 spin_unlock_irqrestore(&pcpu->load_lock, flags);
179 static unsigned int freq_to_above_hispeed_delay(
180 struct cpufreq_interactive_tunables *tunables,
187 spin_lock_irqsave(&tunables->above_hispeed_delay_lock, flags);
189 for (i = 0; i < tunables->nabove_hispeed_delay - 1 &&
190 freq >= tunables->above_hispeed_delay[i+1]; i += 2)
193 ret = tunables->above_hispeed_delay[i];
194 spin_unlock_irqrestore(&tunables->above_hispeed_delay_lock, flags);
198 static unsigned int freq_to_targetload(
199 struct cpufreq_interactive_tunables *tunables, unsigned int freq)
205 spin_lock_irqsave(&tunables->target_loads_lock, flags);
207 for (i = 0; i < tunables->ntarget_loads - 1 &&
208 freq >= tunables->target_loads[i+1]; i += 2)
211 ret = tunables->target_loads[i];
212 spin_unlock_irqrestore(&tunables->target_loads_lock, flags);
217 * If increasing frequencies never map to a lower target load then
218 * choose_freq() will find the minimum frequency that does not exceed its
219 * target load given the current load.
221 static unsigned int choose_freq(struct cpufreq_interactive_cpuinfo *pcpu,
222 unsigned int loadadjfreq)
224 unsigned int freq = pcpu->policy->cur;
225 unsigned int prevfreq, freqmin, freqmax;
234 tl = freq_to_targetload(pcpu->policy->governor_data, freq);
237 * Find the lowest frequency where the computed load is less
238 * than or equal to the target load.
241 if (cpufreq_frequency_table_target(
242 pcpu->policy, pcpu->freq_table, loadadjfreq / tl,
243 CPUFREQ_RELATION_L, &index))
245 freq = pcpu->freq_table[index].frequency;
247 if (freq > prevfreq) {
248 /* The previous frequency is too low. */
251 if (freq >= freqmax) {
253 * Find the highest frequency that is less
256 if (cpufreq_frequency_table_target(
257 pcpu->policy, pcpu->freq_table,
258 freqmax - 1, CPUFREQ_RELATION_H,
261 freq = pcpu->freq_table[index].frequency;
263 if (freq == freqmin) {
265 * The first frequency below freqmax
266 * has already been found to be too
267 * low. freqmax is the lowest speed
268 * we found that is fast enough.
274 } else if (freq < prevfreq) {
275 /* The previous frequency is high enough. */
278 if (freq <= freqmin) {
280 * Find the lowest frequency that is higher
283 if (cpufreq_frequency_table_target(
284 pcpu->policy, pcpu->freq_table,
285 freqmin + 1, CPUFREQ_RELATION_L,
288 freq = pcpu->freq_table[index].frequency;
291 * If freqmax is the first frequency above
292 * freqmin then we have already found that
293 * this speed is fast enough.
300 /* If same frequency chosen as previous then done. */
301 } while (freq != prevfreq);
306 static u64 update_load(int cpu)
308 struct cpufreq_interactive_cpuinfo *pcpu = &per_cpu(cpuinfo, cpu);
309 struct cpufreq_interactive_tunables *tunables =
310 pcpu->policy->governor_data;
313 unsigned int delta_idle;
314 unsigned int delta_time;
317 now_idle = get_cpu_idle_time(cpu, &now, tunables->io_is_busy);
318 delta_idle = (unsigned int)(now_idle - pcpu->time_in_idle);
319 delta_time = (unsigned int)(now - pcpu->time_in_idle_timestamp);
321 if (delta_time <= delta_idle)
324 active_time = delta_time - delta_idle;
326 pcpu->cputime_speedadj += active_time * pcpu->policy->cur;
328 pcpu->time_in_idle = now_idle;
329 pcpu->time_in_idle_timestamp = now;
333 static void cpufreq_interactive_timer(unsigned long data)
336 unsigned int delta_time;
337 u64 cputime_speedadj;
339 struct cpufreq_interactive_cpuinfo *pcpu =
340 &per_cpu(cpuinfo, data);
341 struct cpufreq_interactive_tunables *tunables =
342 pcpu->policy->governor_data;
343 unsigned int new_freq;
344 unsigned int loadadjfreq;
349 if (!down_read_trylock(&pcpu->enable_sem))
351 if (!pcpu->governor_enabled)
354 spin_lock_irqsave(&pcpu->load_lock, flags);
355 now = update_load(data);
356 delta_time = (unsigned int)(now - pcpu->cputime_speedadj_timestamp);
357 cputime_speedadj = pcpu->cputime_speedadj;
358 spin_unlock_irqrestore(&pcpu->load_lock, flags);
360 if (WARN_ON_ONCE(!delta_time))
363 spin_lock_irqsave(&pcpu->target_freq_lock, flags);
364 do_div(cputime_speedadj, delta_time);
365 loadadjfreq = (unsigned int)cputime_speedadj * 100;
366 cpu_load = loadadjfreq / pcpu->target_freq;
367 boosted = tunables->boost_val || now < tunables->boostpulse_endtime;
369 if (cpu_load >= tunables->go_hispeed_load || boosted) {
370 if (pcpu->target_freq < tunables->hispeed_freq) {
371 new_freq = tunables->hispeed_freq;
373 new_freq = choose_freq(pcpu, loadadjfreq);
375 if (new_freq < tunables->hispeed_freq)
376 new_freq = tunables->hispeed_freq;
379 new_freq = choose_freq(pcpu, loadadjfreq);
380 if (new_freq > tunables->hispeed_freq &&
381 pcpu->target_freq < tunables->hispeed_freq)
382 new_freq = tunables->hispeed_freq;
385 if (pcpu->target_freq >= tunables->hispeed_freq &&
386 new_freq > pcpu->target_freq &&
387 now - pcpu->hispeed_validate_time <
388 freq_to_above_hispeed_delay(tunables, pcpu->target_freq)) {
389 trace_cpufreq_interactive_notyet(
390 data, cpu_load, pcpu->target_freq,
391 pcpu->policy->cur, new_freq);
392 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
396 pcpu->hispeed_validate_time = now;
398 if (cpufreq_frequency_table_target(pcpu->policy, pcpu->freq_table,
399 new_freq, CPUFREQ_RELATION_L,
401 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
405 new_freq = pcpu->freq_table[index].frequency;
408 * Do not scale below floor_freq unless we have been at or above the
409 * floor frequency for the minimum sample time since last validated.
411 if (new_freq < pcpu->floor_freq) {
412 if (now - pcpu->floor_validate_time <
413 tunables->min_sample_time) {
414 trace_cpufreq_interactive_notyet(
415 data, cpu_load, pcpu->target_freq,
416 pcpu->policy->cur, new_freq);
417 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
423 * Update the timestamp for checking whether speed has been held at
424 * or above the selected frequency for a minimum of min_sample_time,
425 * if not boosted to hispeed_freq. If boosted to hispeed_freq then we
426 * allow the speed to drop as soon as the boostpulse duration expires
427 * (or the indefinite boost is turned off).
430 if (!boosted || new_freq > tunables->hispeed_freq) {
431 pcpu->floor_freq = new_freq;
432 pcpu->floor_validate_time = now;
435 if (pcpu->target_freq == new_freq &&
436 pcpu->target_freq <= pcpu->policy->cur) {
437 trace_cpufreq_interactive_already(
438 data, cpu_load, pcpu->target_freq,
439 pcpu->policy->cur, new_freq);
440 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
441 goto rearm_if_notmax;
444 trace_cpufreq_interactive_target(data, cpu_load, pcpu->target_freq,
445 pcpu->policy->cur, new_freq);
447 pcpu->target_freq = new_freq;
448 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
449 spin_lock_irqsave(&speedchange_cpumask_lock, flags);
450 cpumask_set_cpu(data, &speedchange_cpumask);
451 spin_unlock_irqrestore(&speedchange_cpumask_lock, flags);
452 wake_up_process(speedchange_task);
456 * Already set max speed and don't see a need to change that,
457 * wait until next idle to re-evaluate, don't need timer.
459 if (pcpu->target_freq == pcpu->policy->max)
463 if (!timer_pending(&pcpu->cpu_timer))
464 cpufreq_interactive_timer_resched(pcpu);
467 up_read(&pcpu->enable_sem);
471 static void cpufreq_interactive_idle_start(void)
473 struct cpufreq_interactive_cpuinfo *pcpu =
474 &per_cpu(cpuinfo, smp_processor_id());
477 if (!down_read_trylock(&pcpu->enable_sem))
479 if (!pcpu->governor_enabled) {
480 up_read(&pcpu->enable_sem);
484 pending = timer_pending(&pcpu->cpu_timer);
486 if (pcpu->target_freq != pcpu->policy->min) {
488 * Entering idle while not at lowest speed. On some
489 * platforms this can hold the other CPU(s) at that speed
490 * even though the CPU is idle. Set a timer to re-evaluate
491 * speed so this idle CPU doesn't hold the other CPUs above
492 * min indefinitely. This should probably be a quirk of
493 * the CPUFreq driver.
496 cpufreq_interactive_timer_resched(pcpu);
499 up_read(&pcpu->enable_sem);
502 static void cpufreq_interactive_idle_end(void)
504 struct cpufreq_interactive_cpuinfo *pcpu =
505 &per_cpu(cpuinfo, smp_processor_id());
507 if (!down_read_trylock(&pcpu->enable_sem))
509 if (!pcpu->governor_enabled) {
510 up_read(&pcpu->enable_sem);
514 /* Arm the timer for 1-2 ticks later if not already. */
515 if (!timer_pending(&pcpu->cpu_timer)) {
516 cpufreq_interactive_timer_resched(pcpu);
517 } else if (time_after_eq(jiffies, pcpu->cpu_timer.expires)) {
518 del_timer(&pcpu->cpu_timer);
519 del_timer(&pcpu->cpu_slack_timer);
520 cpufreq_interactive_timer(smp_processor_id());
523 up_read(&pcpu->enable_sem);
526 static int cpufreq_interactive_speedchange_task(void *data)
531 struct cpufreq_interactive_cpuinfo *pcpu;
534 set_current_state(TASK_INTERRUPTIBLE);
535 spin_lock_irqsave(&speedchange_cpumask_lock, flags);
537 if (cpumask_empty(&speedchange_cpumask)) {
538 spin_unlock_irqrestore(&speedchange_cpumask_lock,
542 if (kthread_should_stop())
545 spin_lock_irqsave(&speedchange_cpumask_lock, flags);
548 set_current_state(TASK_RUNNING);
549 tmp_mask = speedchange_cpumask;
550 cpumask_clear(&speedchange_cpumask);
551 spin_unlock_irqrestore(&speedchange_cpumask_lock, flags);
553 for_each_cpu(cpu, &tmp_mask) {
555 unsigned int max_freq = 0;
557 pcpu = &per_cpu(cpuinfo, cpu);
558 if (!down_read_trylock(&pcpu->enable_sem))
560 if (!pcpu->governor_enabled) {
561 up_read(&pcpu->enable_sem);
565 for_each_cpu(j, pcpu->policy->cpus) {
566 struct cpufreq_interactive_cpuinfo *pjcpu =
567 &per_cpu(cpuinfo, j);
569 if (pjcpu->target_freq > max_freq)
570 max_freq = pjcpu->target_freq;
573 if (max_freq != pcpu->policy->cur)
574 __cpufreq_driver_target(pcpu->policy,
577 trace_cpufreq_interactive_setspeed(cpu,
581 up_read(&pcpu->enable_sem);
588 static void cpufreq_interactive_boost(void)
592 unsigned long flags[2];
593 struct cpufreq_interactive_cpuinfo *pcpu;
594 struct cpufreq_interactive_tunables *tunables;
596 spin_lock_irqsave(&speedchange_cpumask_lock, flags[0]);
598 for_each_online_cpu(i) {
599 pcpu = &per_cpu(cpuinfo, i);
600 tunables = pcpu->policy->governor_data;
602 spin_lock_irqsave(&pcpu->target_freq_lock, flags[1]);
603 if (pcpu->target_freq < tunables->hispeed_freq) {
604 pcpu->target_freq = tunables->hispeed_freq;
605 cpumask_set_cpu(i, &speedchange_cpumask);
606 pcpu->hispeed_validate_time =
607 ktime_to_us(ktime_get());
612 * Set floor freq and (re)start timer for when last
616 pcpu->floor_freq = tunables->hispeed_freq;
617 pcpu->floor_validate_time = ktime_to_us(ktime_get());
618 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags[1]);
621 spin_unlock_irqrestore(&speedchange_cpumask_lock, flags[0]);
624 wake_up_process(speedchange_task);
627 static int cpufreq_interactive_notifier(
628 struct notifier_block *nb, unsigned long val, void *data)
630 struct cpufreq_freqs *freq = data;
631 struct cpufreq_interactive_cpuinfo *pcpu;
635 if (val == CPUFREQ_POSTCHANGE) {
636 pcpu = &per_cpu(cpuinfo, freq->cpu);
637 if (!down_read_trylock(&pcpu->enable_sem))
639 if (!pcpu->governor_enabled) {
640 up_read(&pcpu->enable_sem);
644 for_each_cpu(cpu, pcpu->policy->cpus) {
645 struct cpufreq_interactive_cpuinfo *pjcpu =
646 &per_cpu(cpuinfo, cpu);
647 if (cpu != freq->cpu) {
648 if (!down_read_trylock(&pjcpu->enable_sem))
650 if (!pjcpu->governor_enabled) {
651 up_read(&pjcpu->enable_sem);
655 spin_lock_irqsave(&pjcpu->load_lock, flags);
657 spin_unlock_irqrestore(&pjcpu->load_lock, flags);
658 if (cpu != freq->cpu)
659 up_read(&pjcpu->enable_sem);
662 up_read(&pcpu->enable_sem);
667 static struct notifier_block cpufreq_notifier_block = {
668 .notifier_call = cpufreq_interactive_notifier,
671 static unsigned int *get_tokenized_data(const char *buf, int *num_tokens)
676 unsigned int *tokenized_data;
680 while ((cp = strpbrk(cp + 1, " :")))
683 if (!(ntokens & 0x1))
686 tokenized_data = kmalloc(ntokens * sizeof(unsigned int), GFP_KERNEL);
687 if (!tokenized_data) {
694 while (i < ntokens) {
695 if (sscanf(cp, "%u", &tokenized_data[i++]) != 1)
698 cp = strpbrk(cp, " :");
707 *num_tokens = ntokens;
708 return tokenized_data;
711 kfree(tokenized_data);
716 static ssize_t show_target_loads(
717 struct cpufreq_interactive_tunables *tunables,
724 spin_lock_irqsave(&tunables->target_loads_lock, flags);
726 for (i = 0; i < tunables->ntarget_loads; i++)
727 ret += sprintf(buf + ret, "%u%s", tunables->target_loads[i],
728 i & 0x1 ? ":" : " ");
730 sprintf(buf + ret - 1, "\n");
731 spin_unlock_irqrestore(&tunables->target_loads_lock, flags);
735 static ssize_t store_target_loads(
736 struct cpufreq_interactive_tunables *tunables,
737 const char *buf, size_t count)
740 unsigned int *new_target_loads = NULL;
743 new_target_loads = get_tokenized_data(buf, &ntokens);
744 if (IS_ERR(new_target_loads))
745 return PTR_RET(new_target_loads);
747 spin_lock_irqsave(&tunables->target_loads_lock, flags);
748 if (tunables->target_loads != default_target_loads)
749 kfree(tunables->target_loads);
750 tunables->target_loads = new_target_loads;
751 tunables->ntarget_loads = ntokens;
752 spin_unlock_irqrestore(&tunables->target_loads_lock, flags);
756 static ssize_t show_above_hispeed_delay(
757 struct cpufreq_interactive_tunables *tunables, char *buf)
763 spin_lock_irqsave(&tunables->above_hispeed_delay_lock, flags);
765 for (i = 0; i < tunables->nabove_hispeed_delay; i++)
766 ret += sprintf(buf + ret, "%u%s",
767 tunables->above_hispeed_delay[i],
768 i & 0x1 ? ":" : " ");
770 sprintf(buf + ret - 1, "\n");
771 spin_unlock_irqrestore(&tunables->above_hispeed_delay_lock, flags);
775 static ssize_t store_above_hispeed_delay(
776 struct cpufreq_interactive_tunables *tunables,
777 const char *buf, size_t count)
780 unsigned int *new_above_hispeed_delay = NULL;
783 new_above_hispeed_delay = get_tokenized_data(buf, &ntokens);
784 if (IS_ERR(new_above_hispeed_delay))
785 return PTR_RET(new_above_hispeed_delay);
787 spin_lock_irqsave(&tunables->above_hispeed_delay_lock, flags);
788 if (tunables->above_hispeed_delay != default_above_hispeed_delay)
789 kfree(tunables->above_hispeed_delay);
790 tunables->above_hispeed_delay = new_above_hispeed_delay;
791 tunables->nabove_hispeed_delay = ntokens;
792 spin_unlock_irqrestore(&tunables->above_hispeed_delay_lock, flags);
797 static ssize_t show_hispeed_freq(struct cpufreq_interactive_tunables *tunables,
800 return sprintf(buf, "%u\n", tunables->hispeed_freq);
803 static ssize_t store_hispeed_freq(struct cpufreq_interactive_tunables *tunables,
804 const char *buf, size_t count)
807 long unsigned int val;
809 ret = strict_strtoul(buf, 0, &val);
812 tunables->hispeed_freq = val;
816 static ssize_t show_go_hispeed_load(struct cpufreq_interactive_tunables
817 *tunables, char *buf)
819 return sprintf(buf, "%lu\n", tunables->go_hispeed_load);
822 static ssize_t store_go_hispeed_load(struct cpufreq_interactive_tunables
823 *tunables, const char *buf, size_t count)
828 ret = strict_strtoul(buf, 0, &val);
831 tunables->go_hispeed_load = val;
835 static ssize_t show_min_sample_time(struct cpufreq_interactive_tunables
836 *tunables, char *buf)
838 return sprintf(buf, "%lu\n", tunables->min_sample_time);
841 static ssize_t store_min_sample_time(struct cpufreq_interactive_tunables
842 *tunables, const char *buf, size_t count)
847 ret = strict_strtoul(buf, 0, &val);
850 tunables->min_sample_time = val;
854 static ssize_t show_timer_rate(struct cpufreq_interactive_tunables *tunables,
857 return sprintf(buf, "%lu\n", tunables->timer_rate);
860 static ssize_t store_timer_rate(struct cpufreq_interactive_tunables *tunables,
861 const char *buf, size_t count)
866 ret = strict_strtoul(buf, 0, &val);
869 tunables->timer_rate = val;
873 static ssize_t show_timer_slack(struct cpufreq_interactive_tunables *tunables,
876 return sprintf(buf, "%d\n", tunables->timer_slack_val);
879 static ssize_t store_timer_slack(struct cpufreq_interactive_tunables *tunables,
880 const char *buf, size_t count)
885 ret = kstrtol(buf, 10, &val);
889 tunables->timer_slack_val = val;
893 static ssize_t show_boost(struct cpufreq_interactive_tunables *tunables,
896 return sprintf(buf, "%d\n", tunables->boost_val);
899 static ssize_t store_boost(struct cpufreq_interactive_tunables *tunables,
900 const char *buf, size_t count)
905 ret = kstrtoul(buf, 0, &val);
909 tunables->boost_val = val;
911 if (tunables->boost_val) {
912 trace_cpufreq_interactive_boost("on");
913 cpufreq_interactive_boost();
915 tunables->boostpulse_endtime = ktime_to_us(ktime_get());
916 trace_cpufreq_interactive_unboost("off");
922 static ssize_t store_boostpulse(struct cpufreq_interactive_tunables *tunables,
923 const char *buf, size_t count)
928 ret = kstrtoul(buf, 0, &val);
932 tunables->boostpulse_endtime = ktime_to_us(ktime_get()) +
933 tunables->boostpulse_duration_val;
934 trace_cpufreq_interactive_boost("pulse");
935 cpufreq_interactive_boost();
939 static ssize_t show_boostpulse_duration(struct cpufreq_interactive_tunables
940 *tunables, char *buf)
942 return sprintf(buf, "%d\n", tunables->boostpulse_duration_val);
945 static ssize_t store_boostpulse_duration(struct cpufreq_interactive_tunables
946 *tunables, const char *buf, size_t count)
951 ret = kstrtoul(buf, 0, &val);
955 tunables->boostpulse_duration_val = val;
959 static ssize_t show_io_is_busy(struct cpufreq_interactive_tunables *tunables,
962 return sprintf(buf, "%u\n", tunables->io_is_busy);
965 static ssize_t store_io_is_busy(struct cpufreq_interactive_tunables *tunables,
966 const char *buf, size_t count)
971 ret = kstrtoul(buf, 0, &val);
974 tunables->io_is_busy = val;
979 * Create show/store routines
980 * - sys: One governor instance for complete SYSTEM
981 * - pol: One governor instance per struct cpufreq_policy
983 #define show_gov_pol_sys(file_name) \
984 static ssize_t show_##file_name##_gov_sys \
985 (struct kobject *kobj, struct attribute *attr, char *buf) \
987 return show_##file_name(common_tunables, buf); \
990 static ssize_t show_##file_name##_gov_pol \
991 (struct cpufreq_policy *policy, char *buf) \
993 return show_##file_name(policy->governor_data, buf); \
996 #define store_gov_pol_sys(file_name) \
997 static ssize_t store_##file_name##_gov_sys \
998 (struct kobject *kobj, struct attribute *attr, const char *buf, \
1001 return store_##file_name(common_tunables, buf, count); \
1004 static ssize_t store_##file_name##_gov_pol \
1005 (struct cpufreq_policy *policy, const char *buf, size_t count) \
1007 return store_##file_name(policy->governor_data, buf, count); \
1010 #define show_store_gov_pol_sys(file_name) \
1011 show_gov_pol_sys(file_name); \
1012 store_gov_pol_sys(file_name)
1014 show_store_gov_pol_sys(target_loads);
1015 show_store_gov_pol_sys(above_hispeed_delay);
1016 show_store_gov_pol_sys(hispeed_freq);
1017 show_store_gov_pol_sys(go_hispeed_load);
1018 show_store_gov_pol_sys(min_sample_time);
1019 show_store_gov_pol_sys(timer_rate);
1020 show_store_gov_pol_sys(timer_slack);
1021 show_store_gov_pol_sys(boost);
1022 store_gov_pol_sys(boostpulse);
1023 show_store_gov_pol_sys(boostpulse_duration);
1024 show_store_gov_pol_sys(io_is_busy);
1026 #define gov_sys_attr_rw(_name) \
1027 static struct global_attr _name##_gov_sys = \
1028 __ATTR(_name, 0644, show_##_name##_gov_sys, store_##_name##_gov_sys)
1030 #define gov_pol_attr_rw(_name) \
1031 static struct freq_attr _name##_gov_pol = \
1032 __ATTR(_name, 0644, show_##_name##_gov_pol, store_##_name##_gov_pol)
1034 #define gov_sys_pol_attr_rw(_name) \
1035 gov_sys_attr_rw(_name); \
1036 gov_pol_attr_rw(_name)
1038 gov_sys_pol_attr_rw(target_loads);
1039 gov_sys_pol_attr_rw(above_hispeed_delay);
1040 gov_sys_pol_attr_rw(hispeed_freq);
1041 gov_sys_pol_attr_rw(go_hispeed_load);
1042 gov_sys_pol_attr_rw(min_sample_time);
1043 gov_sys_pol_attr_rw(timer_rate);
1044 gov_sys_pol_attr_rw(timer_slack);
1045 gov_sys_pol_attr_rw(boost);
1046 gov_sys_pol_attr_rw(boostpulse_duration);
1047 gov_sys_pol_attr_rw(io_is_busy);
1049 static struct global_attr boostpulse_gov_sys =
1050 __ATTR(boostpulse, 0200, NULL, store_boostpulse_gov_sys);
1052 static struct freq_attr boostpulse_gov_pol =
1053 __ATTR(boostpulse, 0200, NULL, store_boostpulse_gov_pol);
1055 /* One Governor instance for entire system */
1056 static struct attribute *interactive_attributes_gov_sys[] = {
1057 &target_loads_gov_sys.attr,
1058 &above_hispeed_delay_gov_sys.attr,
1059 &hispeed_freq_gov_sys.attr,
1060 &go_hispeed_load_gov_sys.attr,
1061 &min_sample_time_gov_sys.attr,
1062 &timer_rate_gov_sys.attr,
1063 &timer_slack_gov_sys.attr,
1064 &boost_gov_sys.attr,
1065 &boostpulse_gov_sys.attr,
1066 &boostpulse_duration_gov_sys.attr,
1067 &io_is_busy_gov_sys.attr,
1071 static struct attribute_group interactive_attr_group_gov_sys = {
1072 .attrs = interactive_attributes_gov_sys,
1073 .name = "interactive",
1076 /* Per policy governor instance */
1077 static struct attribute *interactive_attributes_gov_pol[] = {
1078 &target_loads_gov_pol.attr,
1079 &above_hispeed_delay_gov_pol.attr,
1080 &hispeed_freq_gov_pol.attr,
1081 &go_hispeed_load_gov_pol.attr,
1082 &min_sample_time_gov_pol.attr,
1083 &timer_rate_gov_pol.attr,
1084 &timer_slack_gov_pol.attr,
1085 &boost_gov_pol.attr,
1086 &boostpulse_gov_pol.attr,
1087 &boostpulse_duration_gov_pol.attr,
1088 &io_is_busy_gov_pol.attr,
1092 static struct attribute_group interactive_attr_group_gov_pol = {
1093 .attrs = interactive_attributes_gov_pol,
1094 .name = "interactive",
1097 static struct attribute_group *get_sysfs_attr(void)
1099 if (have_governor_per_policy())
1100 return &interactive_attr_group_gov_pol;
1102 return &interactive_attr_group_gov_sys;
1105 static int cpufreq_interactive_idle_notifier(struct notifier_block *nb,
1111 cpufreq_interactive_idle_start();
1114 cpufreq_interactive_idle_end();
1121 static struct notifier_block cpufreq_interactive_idle_nb = {
1122 .notifier_call = cpufreq_interactive_idle_notifier,
1125 static int cpufreq_governor_interactive(struct cpufreq_policy *policy,
1130 struct cpufreq_interactive_cpuinfo *pcpu;
1131 struct cpufreq_frequency_table *freq_table;
1132 struct cpufreq_interactive_tunables *tunables;
1133 unsigned long flags;
1135 if (have_governor_per_policy())
1136 tunables = policy->governor_data;
1138 tunables = common_tunables;
1140 WARN_ON(!tunables && (event != CPUFREQ_GOV_POLICY_INIT));
1143 case CPUFREQ_GOV_POLICY_INIT:
1144 if (have_governor_per_policy()) {
1146 } else if (tunables) {
1147 tunables->usage_count++;
1148 policy->governor_data = tunables;
1152 tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
1154 pr_err("%s: POLICY_INIT: kzalloc failed\n", __func__);
1158 tunables->usage_count = 1;
1159 tunables->above_hispeed_delay = default_above_hispeed_delay;
1160 tunables->nabove_hispeed_delay =
1161 ARRAY_SIZE(default_above_hispeed_delay);
1162 tunables->go_hispeed_load = DEFAULT_GO_HISPEED_LOAD;
1163 tunables->target_loads = default_target_loads;
1164 tunables->ntarget_loads = ARRAY_SIZE(default_target_loads);
1165 tunables->min_sample_time = DEFAULT_MIN_SAMPLE_TIME;
1166 tunables->timer_rate = DEFAULT_TIMER_RATE;
1167 tunables->boostpulse_duration_val = DEFAULT_MIN_SAMPLE_TIME;
1168 tunables->timer_slack_val = DEFAULT_TIMER_SLACK;
1170 spin_lock_init(&tunables->target_loads_lock);
1171 spin_lock_init(&tunables->above_hispeed_delay_lock);
1173 policy->governor_data = tunables;
1174 if (!have_governor_per_policy()) {
1175 common_tunables = tunables;
1176 WARN_ON(cpufreq_get_global_kobject());
1179 rc = sysfs_create_group(get_governor_parent_kobj(policy),
1183 policy->governor_data = NULL;
1184 if (!have_governor_per_policy())
1185 common_tunables = NULL;
1189 if (!policy->governor->initialized) {
1190 idle_notifier_register(&cpufreq_interactive_idle_nb);
1191 cpufreq_register_notifier(&cpufreq_notifier_block,
1192 CPUFREQ_TRANSITION_NOTIFIER);
1197 case CPUFREQ_GOV_POLICY_EXIT:
1198 if (!--tunables->usage_count) {
1199 if (policy->governor->initialized == 1) {
1200 cpufreq_unregister_notifier(&cpufreq_notifier_block,
1201 CPUFREQ_TRANSITION_NOTIFIER);
1202 idle_notifier_unregister(&cpufreq_interactive_idle_nb);
1205 sysfs_remove_group(get_governor_parent_kobj(policy),
1208 if (!have_governor_per_policy())
1209 cpufreq_put_global_kobject();
1212 common_tunables = NULL;
1215 policy->governor_data = NULL;
1218 case CPUFREQ_GOV_START:
1219 mutex_lock(&gov_lock);
1221 freq_table = cpufreq_frequency_get_table(policy->cpu);
1222 if (!tunables->hispeed_freq)
1223 tunables->hispeed_freq = policy->max;
1225 for_each_cpu(j, policy->cpus) {
1226 pcpu = &per_cpu(cpuinfo, j);
1227 pcpu->policy = policy;
1228 pcpu->target_freq = policy->cur;
1229 pcpu->freq_table = freq_table;
1230 pcpu->floor_freq = pcpu->target_freq;
1231 pcpu->floor_validate_time =
1232 ktime_to_us(ktime_get());
1233 pcpu->hispeed_validate_time =
1234 pcpu->floor_validate_time;
1235 pcpu->max_freq = policy->max;
1236 down_write(&pcpu->enable_sem);
1237 del_timer_sync(&pcpu->cpu_timer);
1238 del_timer_sync(&pcpu->cpu_slack_timer);
1239 cpufreq_interactive_timer_start(tunables, j);
1240 pcpu->governor_enabled = 1;
1241 up_write(&pcpu->enable_sem);
1244 mutex_unlock(&gov_lock);
1247 case CPUFREQ_GOV_STOP:
1248 mutex_lock(&gov_lock);
1249 for_each_cpu(j, policy->cpus) {
1250 pcpu = &per_cpu(cpuinfo, j);
1251 down_write(&pcpu->enable_sem);
1252 pcpu->governor_enabled = 0;
1253 del_timer_sync(&pcpu->cpu_timer);
1254 del_timer_sync(&pcpu->cpu_slack_timer);
1255 up_write(&pcpu->enable_sem);
1258 mutex_unlock(&gov_lock);
1261 case CPUFREQ_GOV_LIMITS:
1262 if (policy->max < policy->cur)
1263 __cpufreq_driver_target(policy,
1264 policy->max, CPUFREQ_RELATION_H);
1265 else if (policy->min > policy->cur)
1266 __cpufreq_driver_target(policy,
1267 policy->min, CPUFREQ_RELATION_L);
1268 for_each_cpu(j, policy->cpus) {
1269 pcpu = &per_cpu(cpuinfo, j);
1271 down_read(&pcpu->enable_sem);
1272 if (pcpu->governor_enabled == 0) {
1273 up_read(&pcpu->enable_sem);
1277 spin_lock_irqsave(&pcpu->target_freq_lock, flags);
1278 if (policy->max < pcpu->target_freq)
1279 pcpu->target_freq = policy->max;
1280 else if (policy->min > pcpu->target_freq)
1281 pcpu->target_freq = policy->min;
1283 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
1284 up_read(&pcpu->enable_sem);
1286 /* Reschedule timer only if policy->max is raised.
1287 * Delete the timers, else the timer callback may
1288 * return without re-arm the timer when failed
1289 * acquire the semaphore. This race may cause timer
1290 * stopped unexpectedly.
1293 if (policy->max > pcpu->max_freq) {
1294 down_write(&pcpu->enable_sem);
1295 del_timer_sync(&pcpu->cpu_timer);
1296 del_timer_sync(&pcpu->cpu_slack_timer);
1297 cpufreq_interactive_timer_start(tunables, j);
1298 up_write(&pcpu->enable_sem);
1301 pcpu->max_freq = policy->max;
1308 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_INTERACTIVE
1311 struct cpufreq_governor cpufreq_gov_interactive = {
1312 .name = "interactive",
1313 .governor = cpufreq_governor_interactive,
1314 .max_transition_latency = 10000000,
1315 .owner = THIS_MODULE,
1318 static void cpufreq_interactive_nop_timer(unsigned long data)
1322 static int __init cpufreq_interactive_init(void)
1325 struct cpufreq_interactive_cpuinfo *pcpu;
1326 struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
1328 /* Initalize per-cpu timers */
1329 for_each_possible_cpu(i) {
1330 pcpu = &per_cpu(cpuinfo, i);
1331 init_timer_deferrable(&pcpu->cpu_timer);
1332 pcpu->cpu_timer.function = cpufreq_interactive_timer;
1333 pcpu->cpu_timer.data = i;
1334 init_timer(&pcpu->cpu_slack_timer);
1335 pcpu->cpu_slack_timer.function = cpufreq_interactive_nop_timer;
1336 spin_lock_init(&pcpu->load_lock);
1337 spin_lock_init(&pcpu->target_freq_lock);
1338 init_rwsem(&pcpu->enable_sem);
1341 spin_lock_init(&speedchange_cpumask_lock);
1342 mutex_init(&gov_lock);
1344 kthread_create(cpufreq_interactive_speedchange_task, NULL,
1346 if (IS_ERR(speedchange_task))
1347 return PTR_ERR(speedchange_task);
1349 sched_setscheduler_nocheck(speedchange_task, SCHED_FIFO, ¶m);
1350 get_task_struct(speedchange_task);
1352 /* NB: wake up so the thread does not look hung to the freezer */
1353 wake_up_process(speedchange_task);
1355 return cpufreq_register_governor(&cpufreq_gov_interactive);
1358 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_INTERACTIVE
1359 fs_initcall(cpufreq_interactive_init);
1361 module_init(cpufreq_interactive_init);
1364 static void __exit cpufreq_interactive_exit(void)
1366 cpufreq_unregister_governor(&cpufreq_gov_interactive);
1367 kthread_stop(speedchange_task);
1368 put_task_struct(speedchange_task);
1371 module_exit(cpufreq_interactive_exit);
1373 MODULE_AUTHOR("Mike Chan <mike@android.com>");
1374 MODULE_DESCRIPTION("'cpufreq_interactive' - A cpufreq governor for "
1375 "Latency sensitive workloads");
1376 MODULE_LICENSE("GPL");