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 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);
382 if (pcpu->target_freq >= tunables->hispeed_freq &&
383 new_freq > pcpu->target_freq &&
384 now - pcpu->hispeed_validate_time <
385 freq_to_above_hispeed_delay(tunables, pcpu->target_freq)) {
386 trace_cpufreq_interactive_notyet(
387 data, cpu_load, pcpu->target_freq,
388 pcpu->policy->cur, new_freq);
389 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
393 pcpu->hispeed_validate_time = now;
395 if (cpufreq_frequency_table_target(pcpu->policy, pcpu->freq_table,
396 new_freq, CPUFREQ_RELATION_L,
398 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
402 new_freq = pcpu->freq_table[index].frequency;
405 * Do not scale below floor_freq unless we have been at or above the
406 * floor frequency for the minimum sample time since last validated.
408 if (new_freq < pcpu->floor_freq) {
409 if (now - pcpu->floor_validate_time <
410 tunables->min_sample_time) {
411 trace_cpufreq_interactive_notyet(
412 data, cpu_load, pcpu->target_freq,
413 pcpu->policy->cur, new_freq);
414 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
420 * Update the timestamp for checking whether speed has been held at
421 * or above the selected frequency for a minimum of min_sample_time,
422 * if not boosted to hispeed_freq. If boosted to hispeed_freq then we
423 * allow the speed to drop as soon as the boostpulse duration expires
424 * (or the indefinite boost is turned off).
427 if (!boosted || new_freq > tunables->hispeed_freq) {
428 pcpu->floor_freq = new_freq;
429 pcpu->floor_validate_time = now;
432 if (pcpu->target_freq == new_freq) {
433 trace_cpufreq_interactive_already(
434 data, cpu_load, pcpu->target_freq,
435 pcpu->policy->cur, new_freq);
436 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
437 goto rearm_if_notmax;
440 trace_cpufreq_interactive_target(data, cpu_load, pcpu->target_freq,
441 pcpu->policy->cur, new_freq);
443 pcpu->target_freq = new_freq;
444 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
445 spin_lock_irqsave(&speedchange_cpumask_lock, flags);
446 cpumask_set_cpu(data, &speedchange_cpumask);
447 spin_unlock_irqrestore(&speedchange_cpumask_lock, flags);
448 wake_up_process(speedchange_task);
452 * Already set max speed and don't see a need to change that,
453 * wait until next idle to re-evaluate, don't need timer.
455 if (pcpu->target_freq == pcpu->policy->max)
459 if (!timer_pending(&pcpu->cpu_timer))
460 cpufreq_interactive_timer_resched(pcpu);
463 up_read(&pcpu->enable_sem);
467 static void cpufreq_interactive_idle_start(void)
469 struct cpufreq_interactive_cpuinfo *pcpu =
470 &per_cpu(cpuinfo, smp_processor_id());
473 if (!down_read_trylock(&pcpu->enable_sem))
475 if (!pcpu->governor_enabled) {
476 up_read(&pcpu->enable_sem);
480 pending = timer_pending(&pcpu->cpu_timer);
482 if (pcpu->target_freq != pcpu->policy->min) {
484 * Entering idle while not at lowest speed. On some
485 * platforms this can hold the other CPU(s) at that speed
486 * even though the CPU is idle. Set a timer to re-evaluate
487 * speed so this idle CPU doesn't hold the other CPUs above
488 * min indefinitely. This should probably be a quirk of
489 * the CPUFreq driver.
492 cpufreq_interactive_timer_resched(pcpu);
495 up_read(&pcpu->enable_sem);
498 static void cpufreq_interactive_idle_end(void)
500 struct cpufreq_interactive_cpuinfo *pcpu =
501 &per_cpu(cpuinfo, smp_processor_id());
503 if (!down_read_trylock(&pcpu->enable_sem))
505 if (!pcpu->governor_enabled) {
506 up_read(&pcpu->enable_sem);
510 /* Arm the timer for 1-2 ticks later if not already. */
511 if (!timer_pending(&pcpu->cpu_timer)) {
512 cpufreq_interactive_timer_resched(pcpu);
513 } else if (time_after_eq(jiffies, pcpu->cpu_timer.expires)) {
514 del_timer(&pcpu->cpu_timer);
515 del_timer(&pcpu->cpu_slack_timer);
516 cpufreq_interactive_timer(smp_processor_id());
519 up_read(&pcpu->enable_sem);
522 static int cpufreq_interactive_speedchange_task(void *data)
527 struct cpufreq_interactive_cpuinfo *pcpu;
530 set_current_state(TASK_INTERRUPTIBLE);
531 spin_lock_irqsave(&speedchange_cpumask_lock, flags);
533 if (cpumask_empty(&speedchange_cpumask)) {
534 spin_unlock_irqrestore(&speedchange_cpumask_lock,
538 if (kthread_should_stop())
541 spin_lock_irqsave(&speedchange_cpumask_lock, flags);
544 set_current_state(TASK_RUNNING);
545 tmp_mask = speedchange_cpumask;
546 cpumask_clear(&speedchange_cpumask);
547 spin_unlock_irqrestore(&speedchange_cpumask_lock, flags);
549 for_each_cpu(cpu, &tmp_mask) {
551 unsigned int max_freq = 0;
553 pcpu = &per_cpu(cpuinfo, cpu);
554 if (!down_read_trylock(&pcpu->enable_sem))
556 if (!pcpu->governor_enabled) {
557 up_read(&pcpu->enable_sem);
561 for_each_cpu(j, pcpu->policy->cpus) {
562 struct cpufreq_interactive_cpuinfo *pjcpu =
563 &per_cpu(cpuinfo, j);
565 if (pjcpu->target_freq > max_freq)
566 max_freq = pjcpu->target_freq;
569 if (max_freq != pcpu->policy->cur)
570 __cpufreq_driver_target(pcpu->policy,
573 trace_cpufreq_interactive_setspeed(cpu,
577 up_read(&pcpu->enable_sem);
584 static void cpufreq_interactive_boost(void)
588 unsigned long flags[2];
589 struct cpufreq_interactive_cpuinfo *pcpu;
590 struct cpufreq_interactive_tunables *tunables;
592 spin_lock_irqsave(&speedchange_cpumask_lock, flags[0]);
594 for_each_online_cpu(i) {
595 pcpu = &per_cpu(cpuinfo, i);
596 tunables = pcpu->policy->governor_data;
598 spin_lock_irqsave(&pcpu->target_freq_lock, flags[1]);
599 if (pcpu->target_freq < tunables->hispeed_freq) {
600 pcpu->target_freq = tunables->hispeed_freq;
601 cpumask_set_cpu(i, &speedchange_cpumask);
602 pcpu->hispeed_validate_time =
603 ktime_to_us(ktime_get());
608 * Set floor freq and (re)start timer for when last
612 pcpu->floor_freq = tunables->hispeed_freq;
613 pcpu->floor_validate_time = ktime_to_us(ktime_get());
614 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags[1]);
617 spin_unlock_irqrestore(&speedchange_cpumask_lock, flags[0]);
620 wake_up_process(speedchange_task);
623 static int cpufreq_interactive_notifier(
624 struct notifier_block *nb, unsigned long val, void *data)
626 struct cpufreq_freqs *freq = data;
627 struct cpufreq_interactive_cpuinfo *pcpu;
631 if (val == CPUFREQ_POSTCHANGE) {
632 pcpu = &per_cpu(cpuinfo, freq->cpu);
633 if (!down_read_trylock(&pcpu->enable_sem))
635 if (!pcpu->governor_enabled) {
636 up_read(&pcpu->enable_sem);
640 for_each_cpu(cpu, pcpu->policy->cpus) {
641 struct cpufreq_interactive_cpuinfo *pjcpu =
642 &per_cpu(cpuinfo, cpu);
643 if (cpu != freq->cpu) {
644 if (!down_read_trylock(&pjcpu->enable_sem))
646 if (!pjcpu->governor_enabled) {
647 up_read(&pjcpu->enable_sem);
651 spin_lock_irqsave(&pjcpu->load_lock, flags);
653 spin_unlock_irqrestore(&pjcpu->load_lock, flags);
654 if (cpu != freq->cpu)
655 up_read(&pjcpu->enable_sem);
658 up_read(&pcpu->enable_sem);
663 static struct notifier_block cpufreq_notifier_block = {
664 .notifier_call = cpufreq_interactive_notifier,
667 static unsigned int *get_tokenized_data(const char *buf, int *num_tokens)
672 unsigned int *tokenized_data;
676 while ((cp = strpbrk(cp + 1, " :")))
679 if (!(ntokens & 0x1))
682 tokenized_data = kmalloc(ntokens * sizeof(unsigned int), GFP_KERNEL);
683 if (!tokenized_data) {
690 while (i < ntokens) {
691 if (sscanf(cp, "%u", &tokenized_data[i++]) != 1)
694 cp = strpbrk(cp, " :");
703 *num_tokens = ntokens;
704 return tokenized_data;
707 kfree(tokenized_data);
712 static ssize_t show_target_loads(
713 struct cpufreq_interactive_tunables *tunables,
720 spin_lock_irqsave(&tunables->target_loads_lock, flags);
722 for (i = 0; i < tunables->ntarget_loads; i++)
723 ret += sprintf(buf + ret, "%u%s", tunables->target_loads[i],
724 i & 0x1 ? ":" : " ");
726 sprintf(buf + ret - 1, "\n");
727 spin_unlock_irqrestore(&tunables->target_loads_lock, flags);
731 static ssize_t store_target_loads(
732 struct cpufreq_interactive_tunables *tunables,
733 const char *buf, size_t count)
736 unsigned int *new_target_loads = NULL;
739 new_target_loads = get_tokenized_data(buf, &ntokens);
740 if (IS_ERR(new_target_loads))
741 return PTR_RET(new_target_loads);
743 spin_lock_irqsave(&tunables->target_loads_lock, flags);
744 if (tunables->target_loads != default_target_loads)
745 kfree(tunables->target_loads);
746 tunables->target_loads = new_target_loads;
747 tunables->ntarget_loads = ntokens;
748 spin_unlock_irqrestore(&tunables->target_loads_lock, flags);
752 static ssize_t show_above_hispeed_delay(
753 struct cpufreq_interactive_tunables *tunables, char *buf)
759 spin_lock_irqsave(&tunables->above_hispeed_delay_lock, flags);
761 for (i = 0; i < tunables->nabove_hispeed_delay; i++)
762 ret += sprintf(buf + ret, "%u%s",
763 tunables->above_hispeed_delay[i],
764 i & 0x1 ? ":" : " ");
766 sprintf(buf + ret - 1, "\n");
767 spin_unlock_irqrestore(&tunables->above_hispeed_delay_lock, flags);
771 static ssize_t store_above_hispeed_delay(
772 struct cpufreq_interactive_tunables *tunables,
773 const char *buf, size_t count)
776 unsigned int *new_above_hispeed_delay = NULL;
779 new_above_hispeed_delay = get_tokenized_data(buf, &ntokens);
780 if (IS_ERR(new_above_hispeed_delay))
781 return PTR_RET(new_above_hispeed_delay);
783 spin_lock_irqsave(&tunables->above_hispeed_delay_lock, flags);
784 if (tunables->above_hispeed_delay != default_above_hispeed_delay)
785 kfree(tunables->above_hispeed_delay);
786 tunables->above_hispeed_delay = new_above_hispeed_delay;
787 tunables->nabove_hispeed_delay = ntokens;
788 spin_unlock_irqrestore(&tunables->above_hispeed_delay_lock, flags);
793 static ssize_t show_hispeed_freq(struct cpufreq_interactive_tunables *tunables,
796 return sprintf(buf, "%u\n", tunables->hispeed_freq);
799 static ssize_t store_hispeed_freq(struct cpufreq_interactive_tunables *tunables,
800 const char *buf, size_t count)
803 long unsigned int val;
805 ret = strict_strtoul(buf, 0, &val);
808 tunables->hispeed_freq = val;
812 static ssize_t show_go_hispeed_load(struct cpufreq_interactive_tunables
813 *tunables, char *buf)
815 return sprintf(buf, "%lu\n", tunables->go_hispeed_load);
818 static ssize_t store_go_hispeed_load(struct cpufreq_interactive_tunables
819 *tunables, const char *buf, size_t count)
824 ret = strict_strtoul(buf, 0, &val);
827 tunables->go_hispeed_load = val;
831 static ssize_t show_min_sample_time(struct cpufreq_interactive_tunables
832 *tunables, char *buf)
834 return sprintf(buf, "%lu\n", tunables->min_sample_time);
837 static ssize_t store_min_sample_time(struct cpufreq_interactive_tunables
838 *tunables, const char *buf, size_t count)
843 ret = strict_strtoul(buf, 0, &val);
846 tunables->min_sample_time = val;
850 static ssize_t show_timer_rate(struct cpufreq_interactive_tunables *tunables,
853 return sprintf(buf, "%lu\n", tunables->timer_rate);
856 static ssize_t store_timer_rate(struct cpufreq_interactive_tunables *tunables,
857 const char *buf, size_t count)
862 ret = strict_strtoul(buf, 0, &val);
865 tunables->timer_rate = val;
869 static ssize_t show_timer_slack(struct cpufreq_interactive_tunables *tunables,
872 return sprintf(buf, "%d\n", tunables->timer_slack_val);
875 static ssize_t store_timer_slack(struct cpufreq_interactive_tunables *tunables,
876 const char *buf, size_t count)
881 ret = kstrtol(buf, 10, &val);
885 tunables->timer_slack_val = val;
889 static ssize_t show_boost(struct cpufreq_interactive_tunables *tunables,
892 return sprintf(buf, "%d\n", tunables->boost_val);
895 static ssize_t store_boost(struct cpufreq_interactive_tunables *tunables,
896 const char *buf, size_t count)
901 ret = kstrtoul(buf, 0, &val);
905 tunables->boost_val = val;
907 if (tunables->boost_val) {
908 trace_cpufreq_interactive_boost("on");
909 cpufreq_interactive_boost();
911 tunables->boostpulse_endtime = ktime_to_us(ktime_get());
912 trace_cpufreq_interactive_unboost("off");
918 static ssize_t store_boostpulse(struct cpufreq_interactive_tunables *tunables,
919 const char *buf, size_t count)
924 ret = kstrtoul(buf, 0, &val);
928 tunables->boostpulse_endtime = ktime_to_us(ktime_get()) +
929 tunables->boostpulse_duration_val;
930 trace_cpufreq_interactive_boost("pulse");
931 cpufreq_interactive_boost();
935 static ssize_t show_boostpulse_duration(struct cpufreq_interactive_tunables
936 *tunables, char *buf)
938 return sprintf(buf, "%d\n", tunables->boostpulse_duration_val);
941 static ssize_t store_boostpulse_duration(struct cpufreq_interactive_tunables
942 *tunables, const char *buf, size_t count)
947 ret = kstrtoul(buf, 0, &val);
951 tunables->boostpulse_duration_val = val;
955 static ssize_t show_io_is_busy(struct cpufreq_interactive_tunables *tunables,
958 return sprintf(buf, "%u\n", tunables->io_is_busy);
961 static ssize_t store_io_is_busy(struct cpufreq_interactive_tunables *tunables,
962 const char *buf, size_t count)
967 ret = kstrtoul(buf, 0, &val);
970 tunables->io_is_busy = val;
975 * Create show/store routines
976 * - sys: One governor instance for complete SYSTEM
977 * - pol: One governor instance per struct cpufreq_policy
979 #define show_gov_pol_sys(file_name) \
980 static ssize_t show_##file_name##_gov_sys \
981 (struct kobject *kobj, struct attribute *attr, char *buf) \
983 return show_##file_name(common_tunables, buf); \
986 static ssize_t show_##file_name##_gov_pol \
987 (struct cpufreq_policy *policy, char *buf) \
989 return show_##file_name(policy->governor_data, buf); \
992 #define store_gov_pol_sys(file_name) \
993 static ssize_t store_##file_name##_gov_sys \
994 (struct kobject *kobj, struct attribute *attr, const char *buf, \
997 return store_##file_name(common_tunables, buf, count); \
1000 static ssize_t store_##file_name##_gov_pol \
1001 (struct cpufreq_policy *policy, const char *buf, size_t count) \
1003 return store_##file_name(policy->governor_data, buf, count); \
1006 #define show_store_gov_pol_sys(file_name) \
1007 show_gov_pol_sys(file_name); \
1008 store_gov_pol_sys(file_name)
1010 show_store_gov_pol_sys(target_loads);
1011 show_store_gov_pol_sys(above_hispeed_delay);
1012 show_store_gov_pol_sys(hispeed_freq);
1013 show_store_gov_pol_sys(go_hispeed_load);
1014 show_store_gov_pol_sys(min_sample_time);
1015 show_store_gov_pol_sys(timer_rate);
1016 show_store_gov_pol_sys(timer_slack);
1017 show_store_gov_pol_sys(boost);
1018 store_gov_pol_sys(boostpulse);
1019 show_store_gov_pol_sys(boostpulse_duration);
1020 show_store_gov_pol_sys(io_is_busy);
1022 #define gov_sys_attr_rw(_name) \
1023 static struct global_attr _name##_gov_sys = \
1024 __ATTR(_name, 0644, show_##_name##_gov_sys, store_##_name##_gov_sys)
1026 #define gov_pol_attr_rw(_name) \
1027 static struct freq_attr _name##_gov_pol = \
1028 __ATTR(_name, 0644, show_##_name##_gov_pol, store_##_name##_gov_pol)
1030 #define gov_sys_pol_attr_rw(_name) \
1031 gov_sys_attr_rw(_name); \
1032 gov_pol_attr_rw(_name)
1034 gov_sys_pol_attr_rw(target_loads);
1035 gov_sys_pol_attr_rw(above_hispeed_delay);
1036 gov_sys_pol_attr_rw(hispeed_freq);
1037 gov_sys_pol_attr_rw(go_hispeed_load);
1038 gov_sys_pol_attr_rw(min_sample_time);
1039 gov_sys_pol_attr_rw(timer_rate);
1040 gov_sys_pol_attr_rw(timer_slack);
1041 gov_sys_pol_attr_rw(boost);
1042 gov_sys_pol_attr_rw(boostpulse_duration);
1043 gov_sys_pol_attr_rw(io_is_busy);
1045 static struct global_attr boostpulse_gov_sys =
1046 __ATTR(boostpulse, 0200, NULL, store_boostpulse_gov_sys);
1048 static struct freq_attr boostpulse_gov_pol =
1049 __ATTR(boostpulse, 0200, NULL, store_boostpulse_gov_pol);
1051 /* One Governor instance for entire system */
1052 static struct attribute *interactive_attributes_gov_sys[] = {
1053 &target_loads_gov_sys.attr,
1054 &above_hispeed_delay_gov_sys.attr,
1055 &hispeed_freq_gov_sys.attr,
1056 &go_hispeed_load_gov_sys.attr,
1057 &min_sample_time_gov_sys.attr,
1058 &timer_rate_gov_sys.attr,
1059 &timer_slack_gov_sys.attr,
1060 &boost_gov_sys.attr,
1061 &boostpulse_gov_sys.attr,
1062 &boostpulse_duration_gov_sys.attr,
1063 &io_is_busy_gov_sys.attr,
1067 static struct attribute_group interactive_attr_group_gov_sys = {
1068 .attrs = interactive_attributes_gov_sys,
1069 .name = "interactive",
1072 /* Per policy governor instance */
1073 static struct attribute *interactive_attributes_gov_pol[] = {
1074 &target_loads_gov_pol.attr,
1075 &above_hispeed_delay_gov_pol.attr,
1076 &hispeed_freq_gov_pol.attr,
1077 &go_hispeed_load_gov_pol.attr,
1078 &min_sample_time_gov_pol.attr,
1079 &timer_rate_gov_pol.attr,
1080 &timer_slack_gov_pol.attr,
1081 &boost_gov_pol.attr,
1082 &boostpulse_gov_pol.attr,
1083 &boostpulse_duration_gov_pol.attr,
1084 &io_is_busy_gov_pol.attr,
1088 static struct attribute_group interactive_attr_group_gov_pol = {
1089 .attrs = interactive_attributes_gov_pol,
1090 .name = "interactive",
1093 static struct attribute_group *get_sysfs_attr(void)
1095 if (have_governor_per_policy())
1096 return &interactive_attr_group_gov_pol;
1098 return &interactive_attr_group_gov_sys;
1101 static int cpufreq_interactive_idle_notifier(struct notifier_block *nb,
1107 cpufreq_interactive_idle_start();
1110 cpufreq_interactive_idle_end();
1117 static struct notifier_block cpufreq_interactive_idle_nb = {
1118 .notifier_call = cpufreq_interactive_idle_notifier,
1121 static int cpufreq_governor_interactive(struct cpufreq_policy *policy,
1126 struct cpufreq_interactive_cpuinfo *pcpu;
1127 struct cpufreq_frequency_table *freq_table;
1128 struct cpufreq_interactive_tunables *tunables;
1129 unsigned long flags;
1131 if (have_governor_per_policy())
1132 tunables = policy->governor_data;
1134 tunables = common_tunables;
1136 WARN_ON(!tunables && (event != CPUFREQ_GOV_POLICY_INIT));
1139 case CPUFREQ_GOV_POLICY_INIT:
1140 if (have_governor_per_policy()) {
1142 } else if (tunables) {
1143 tunables->usage_count++;
1144 policy->governor_data = tunables;
1148 tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
1150 pr_err("%s: POLICY_INIT: kzalloc failed\n", __func__);
1154 tunables->usage_count = 1;
1155 tunables->above_hispeed_delay = default_above_hispeed_delay;
1156 tunables->nabove_hispeed_delay =
1157 ARRAY_SIZE(default_above_hispeed_delay);
1158 tunables->go_hispeed_load = DEFAULT_GO_HISPEED_LOAD;
1159 tunables->target_loads = default_target_loads;
1160 tunables->ntarget_loads = ARRAY_SIZE(default_target_loads);
1161 tunables->min_sample_time = DEFAULT_MIN_SAMPLE_TIME;
1162 tunables->timer_rate = DEFAULT_TIMER_RATE;
1163 tunables->boostpulse_duration_val = DEFAULT_MIN_SAMPLE_TIME;
1164 tunables->timer_slack_val = DEFAULT_TIMER_SLACK;
1166 spin_lock_init(&tunables->target_loads_lock);
1167 spin_lock_init(&tunables->above_hispeed_delay_lock);
1169 policy->governor_data = tunables;
1170 if (!have_governor_per_policy()) {
1171 common_tunables = tunables;
1172 WARN_ON(cpufreq_get_global_kobject());
1175 rc = sysfs_create_group(get_governor_parent_kobj(policy),
1179 policy->governor_data = NULL;
1180 if (!have_governor_per_policy())
1181 common_tunables = NULL;
1185 if (!policy->governor->initialized) {
1186 idle_notifier_register(&cpufreq_interactive_idle_nb);
1187 cpufreq_register_notifier(&cpufreq_notifier_block,
1188 CPUFREQ_TRANSITION_NOTIFIER);
1193 case CPUFREQ_GOV_POLICY_EXIT:
1194 if (!--tunables->usage_count) {
1195 if (policy->governor->initialized == 1) {
1196 cpufreq_unregister_notifier(&cpufreq_notifier_block,
1197 CPUFREQ_TRANSITION_NOTIFIER);
1198 idle_notifier_unregister(&cpufreq_interactive_idle_nb);
1201 sysfs_remove_group(get_governor_parent_kobj(policy),
1204 if (!have_governor_per_policy())
1205 cpufreq_put_global_kobject();
1208 common_tunables = NULL;
1211 policy->governor_data = NULL;
1214 case CPUFREQ_GOV_START:
1215 mutex_lock(&gov_lock);
1217 freq_table = cpufreq_frequency_get_table(policy->cpu);
1218 if (!tunables->hispeed_freq)
1219 tunables->hispeed_freq = policy->max;
1221 for_each_cpu(j, policy->cpus) {
1222 pcpu = &per_cpu(cpuinfo, j);
1223 pcpu->policy = policy;
1224 pcpu->target_freq = policy->cur;
1225 pcpu->freq_table = freq_table;
1226 pcpu->floor_freq = pcpu->target_freq;
1227 pcpu->floor_validate_time =
1228 ktime_to_us(ktime_get());
1229 pcpu->hispeed_validate_time =
1230 pcpu->floor_validate_time;
1231 pcpu->max_freq = policy->max;
1232 down_write(&pcpu->enable_sem);
1233 del_timer_sync(&pcpu->cpu_timer);
1234 del_timer_sync(&pcpu->cpu_slack_timer);
1235 cpufreq_interactive_timer_start(tunables, j);
1236 pcpu->governor_enabled = 1;
1237 up_write(&pcpu->enable_sem);
1240 mutex_unlock(&gov_lock);
1243 case CPUFREQ_GOV_STOP:
1244 mutex_lock(&gov_lock);
1245 for_each_cpu(j, policy->cpus) {
1246 pcpu = &per_cpu(cpuinfo, j);
1247 down_write(&pcpu->enable_sem);
1248 pcpu->governor_enabled = 0;
1249 del_timer_sync(&pcpu->cpu_timer);
1250 del_timer_sync(&pcpu->cpu_slack_timer);
1251 up_write(&pcpu->enable_sem);
1254 mutex_unlock(&gov_lock);
1257 case CPUFREQ_GOV_LIMITS:
1258 if (policy->max < policy->cur)
1259 __cpufreq_driver_target(policy,
1260 policy->max, CPUFREQ_RELATION_H);
1261 else if (policy->min > policy->cur)
1262 __cpufreq_driver_target(policy,
1263 policy->min, CPUFREQ_RELATION_L);
1264 for_each_cpu(j, policy->cpus) {
1265 pcpu = &per_cpu(cpuinfo, j);
1267 down_read(&pcpu->enable_sem);
1268 if (pcpu->governor_enabled == 0) {
1269 up_read(&pcpu->enable_sem);
1273 spin_lock_irqsave(&pcpu->target_freq_lock, flags);
1274 if (policy->max < pcpu->target_freq)
1275 pcpu->target_freq = policy->max;
1276 else if (policy->min > pcpu->target_freq)
1277 pcpu->target_freq = policy->min;
1279 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
1280 up_read(&pcpu->enable_sem);
1282 /* Reschedule timer only if policy->max is raised.
1283 * Delete the timers, else the timer callback may
1284 * return without re-arm the timer when failed
1285 * acquire the semaphore. This race may cause timer
1286 * stopped unexpectedly.
1289 if (policy->max > pcpu->max_freq) {
1290 down_write(&pcpu->enable_sem);
1291 del_timer_sync(&pcpu->cpu_timer);
1292 del_timer_sync(&pcpu->cpu_slack_timer);
1293 cpufreq_interactive_timer_start(tunables, j);
1294 up_write(&pcpu->enable_sem);
1297 pcpu->max_freq = policy->max;
1304 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_INTERACTIVE
1307 struct cpufreq_governor cpufreq_gov_interactive = {
1308 .name = "interactive",
1309 .governor = cpufreq_governor_interactive,
1310 .max_transition_latency = 10000000,
1311 .owner = THIS_MODULE,
1314 static void cpufreq_interactive_nop_timer(unsigned long data)
1318 static int __init cpufreq_interactive_init(void)
1321 struct cpufreq_interactive_cpuinfo *pcpu;
1322 struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
1324 /* Initalize per-cpu timers */
1325 for_each_possible_cpu(i) {
1326 pcpu = &per_cpu(cpuinfo, i);
1327 init_timer_deferrable(&pcpu->cpu_timer);
1328 pcpu->cpu_timer.function = cpufreq_interactive_timer;
1329 pcpu->cpu_timer.data = i;
1330 init_timer(&pcpu->cpu_slack_timer);
1331 pcpu->cpu_slack_timer.function = cpufreq_interactive_nop_timer;
1332 spin_lock_init(&pcpu->load_lock);
1333 spin_lock_init(&pcpu->target_freq_lock);
1334 init_rwsem(&pcpu->enable_sem);
1337 spin_lock_init(&speedchange_cpumask_lock);
1338 mutex_init(&gov_lock);
1340 kthread_create(cpufreq_interactive_speedchange_task, NULL,
1342 if (IS_ERR(speedchange_task))
1343 return PTR_ERR(speedchange_task);
1345 sched_setscheduler_nocheck(speedchange_task, SCHED_FIFO, ¶m);
1346 get_task_struct(speedchange_task);
1348 /* NB: wake up so the thread does not look hung to the freezer */
1349 wake_up_process(speedchange_task);
1351 return cpufreq_register_governor(&cpufreq_gov_interactive);
1354 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_INTERACTIVE
1355 fs_initcall(cpufreq_interactive_init);
1357 module_init(cpufreq_interactive_init);
1360 static void __exit cpufreq_interactive_exit(void)
1362 cpufreq_unregister_governor(&cpufreq_gov_interactive);
1363 kthread_stop(speedchange_task);
1364 put_task_struct(speedchange_task);
1367 module_exit(cpufreq_interactive_exit);
1369 MODULE_AUTHOR("Mike Chan <mike@android.com>");
1370 MODULE_DESCRIPTION("'cpufreq_interactive' - A cpufreq governor for "
1371 "Latency sensitive workloads");
1372 MODULE_LICENSE("GPL");