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 u64 pol_floor_val_time; /* policy floor_validate_time */
51 u64 loc_floor_val_time; /* per-cpu floor_validate_time */
52 u64 pol_hispeed_val_time; /* policy hispeed_validate_time */
53 u64 loc_hispeed_val_time; /* per-cpu 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;
315 unsigned int delta_idle;
316 unsigned int delta_time;
319 now_idle = get_cpu_idle_time(cpu, &now, tunables->io_is_busy);
320 delta_idle = (unsigned int)(now_idle - pcpu->time_in_idle);
321 delta_time = (unsigned int)(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;
351 if (!down_read_trylock(&pcpu->enable_sem))
353 if (!pcpu->governor_enabled)
356 spin_lock_irqsave(&pcpu->load_lock, flags);
357 now = update_load(data);
358 delta_time = (unsigned int)(now - pcpu->cputime_speedadj_timestamp);
359 cputime_speedadj = pcpu->cputime_speedadj;
360 spin_unlock_irqrestore(&pcpu->load_lock, flags);
362 if (WARN_ON_ONCE(!delta_time))
365 spin_lock_irqsave(&pcpu->target_freq_lock, flags);
366 do_div(cputime_speedadj, delta_time);
367 loadadjfreq = (unsigned int)cputime_speedadj * 100;
368 cpu_load = loadadjfreq / pcpu->policy->cur;
369 tunables->boosted = tunables->boost_val || now < tunables->boostpulse_endtime;
371 if (cpu_load >= tunables->go_hispeed_load || tunables->boosted) {
372 if (pcpu->policy->cur < tunables->hispeed_freq) {
373 new_freq = tunables->hispeed_freq;
375 new_freq = choose_freq(pcpu, loadadjfreq);
377 if (new_freq < tunables->hispeed_freq)
378 new_freq = tunables->hispeed_freq;
381 new_freq = choose_freq(pcpu, loadadjfreq);
382 if (new_freq > tunables->hispeed_freq &&
383 pcpu->policy->cur < tunables->hispeed_freq)
384 new_freq = tunables->hispeed_freq;
387 if (pcpu->policy->cur >= tunables->hispeed_freq &&
388 new_freq > pcpu->policy->cur &&
389 now - pcpu->pol_hispeed_val_time <
390 freq_to_above_hispeed_delay(tunables, pcpu->policy->cur)) {
391 trace_cpufreq_interactive_notyet(
392 data, cpu_load, pcpu->target_freq,
393 pcpu->policy->cur, new_freq);
394 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
398 pcpu->loc_hispeed_val_time = now;
400 if (cpufreq_frequency_table_target(pcpu->policy, pcpu->freq_table,
401 new_freq, CPUFREQ_RELATION_L,
403 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
407 new_freq = pcpu->freq_table[index].frequency;
410 * Do not scale below floor_freq unless we have been at or above the
411 * floor frequency for the minimum sample time since last validated.
413 max_fvtime = max(pcpu->pol_floor_val_time, pcpu->loc_floor_val_time);
414 if (new_freq < pcpu->floor_freq &&
415 pcpu->target_freq >= pcpu->policy->cur) {
416 if (now - max_fvtime < tunables->min_sample_time) {
417 trace_cpufreq_interactive_notyet(
418 data, cpu_load, pcpu->target_freq,
419 pcpu->policy->cur, new_freq);
420 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
426 * Update the timestamp for checking whether speed has been held at
427 * or above the selected frequency for a minimum of min_sample_time,
428 * if not boosted to hispeed_freq. If boosted to hispeed_freq then we
429 * allow the speed to drop as soon as the boostpulse duration expires
430 * (or the indefinite boost is turned off).
433 if (!tunables->boosted || new_freq > tunables->hispeed_freq) {
434 pcpu->floor_freq = new_freq;
435 if (pcpu->target_freq >= pcpu->policy->cur ||
436 new_freq >= pcpu->policy->cur)
437 pcpu->loc_floor_val_time = now;
440 if (pcpu->target_freq == new_freq &&
441 pcpu->target_freq <= pcpu->policy->cur) {
442 trace_cpufreq_interactive_already(
443 data, cpu_load, pcpu->target_freq,
444 pcpu->policy->cur, new_freq);
445 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
449 trace_cpufreq_interactive_target(data, cpu_load, pcpu->target_freq,
450 pcpu->policy->cur, new_freq);
452 pcpu->target_freq = new_freq;
453 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
454 spin_lock_irqsave(&speedchange_cpumask_lock, flags);
455 cpumask_set_cpu(data, &speedchange_cpumask);
456 spin_unlock_irqrestore(&speedchange_cpumask_lock, flags);
457 wake_up_process(speedchange_task);
460 if (!timer_pending(&pcpu->cpu_timer))
461 cpufreq_interactive_timer_resched(pcpu);
464 up_read(&pcpu->enable_sem);
468 static void cpufreq_interactive_idle_end(void)
470 struct cpufreq_interactive_cpuinfo *pcpu =
471 &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 /* Arm the timer for 1-2 ticks later if not already. */
481 if (!timer_pending(&pcpu->cpu_timer)) {
482 cpufreq_interactive_timer_resched(pcpu);
483 } else if (time_after_eq(jiffies, pcpu->cpu_timer.expires)) {
484 del_timer(&pcpu->cpu_timer);
485 del_timer(&pcpu->cpu_slack_timer);
486 cpufreq_interactive_timer(smp_processor_id());
489 up_read(&pcpu->enable_sem);
492 static void cpufreq_interactive_get_policy_info(struct cpufreq_policy *policy,
493 unsigned int *pmax_freq,
494 u64 *phvt, u64 *pfvt)
496 struct cpufreq_interactive_cpuinfo *pcpu;
497 unsigned int max_freq = 0;
498 u64 hvt = ~0ULL, fvt = 0;
501 for_each_cpu(i, policy->cpus) {
502 pcpu = &per_cpu(cpuinfo, i);
504 fvt = max(fvt, pcpu->loc_floor_val_time);
505 if (pcpu->target_freq > max_freq) {
506 max_freq = pcpu->target_freq;
507 hvt = pcpu->loc_hispeed_val_time;
508 } else if (pcpu->target_freq == max_freq) {
509 hvt = min(hvt, pcpu->loc_hispeed_val_time);
513 *pmax_freq = max_freq;
518 static void cpufreq_interactive_adjust_cpu(unsigned int cpu,
519 struct cpufreq_policy *policy)
521 struct cpufreq_interactive_cpuinfo *pcpu;
523 unsigned int max_freq;
526 cpufreq_interactive_get_policy_info(policy, &max_freq, &hvt, &fvt);
528 for_each_cpu(i, policy->cpus) {
529 pcpu = &per_cpu(cpuinfo, i);
530 pcpu->pol_floor_val_time = fvt;
533 if (max_freq != policy->cur) {
534 __cpufreq_driver_target(policy, max_freq, CPUFREQ_RELATION_H);
535 for_each_cpu(i, policy->cpus) {
536 pcpu = &per_cpu(cpuinfo, i);
537 pcpu->pol_hispeed_val_time = hvt;
541 trace_cpufreq_interactive_setspeed(cpu, max_freq, policy->cur);
544 static int cpufreq_interactive_speedchange_task(void *data)
549 struct cpufreq_interactive_cpuinfo *pcpu;
552 set_current_state(TASK_INTERRUPTIBLE);
553 spin_lock_irqsave(&speedchange_cpumask_lock, flags);
555 if (cpumask_empty(&speedchange_cpumask)) {
556 spin_unlock_irqrestore(&speedchange_cpumask_lock,
560 if (kthread_should_stop())
563 spin_lock_irqsave(&speedchange_cpumask_lock, flags);
566 set_current_state(TASK_RUNNING);
567 tmp_mask = speedchange_cpumask;
568 cpumask_clear(&speedchange_cpumask);
569 spin_unlock_irqrestore(&speedchange_cpumask_lock, flags);
571 for_each_cpu(cpu, &tmp_mask) {
572 pcpu = &per_cpu(cpuinfo, cpu);
574 down_write(&pcpu->policy->rwsem);
576 if (likely(down_read_trylock(&pcpu->enable_sem))) {
577 if (likely(pcpu->governor_enabled))
578 cpufreq_interactive_adjust_cpu(cpu,
580 up_read(&pcpu->enable_sem);
583 up_write(&pcpu->policy->rwsem);
590 static void cpufreq_interactive_boost(struct cpufreq_interactive_tunables *tunables)
594 unsigned long flags[2];
595 struct cpufreq_interactive_cpuinfo *pcpu;
597 tunables->boosted = true;
599 spin_lock_irqsave(&speedchange_cpumask_lock, flags[0]);
601 for_each_online_cpu(i) {
602 pcpu = &per_cpu(cpuinfo, i);
604 if (!down_read_trylock(&pcpu->enable_sem))
607 if (!pcpu->governor_enabled) {
608 up_read(&pcpu->enable_sem);
612 if (tunables != pcpu->policy->governor_data) {
613 up_read(&pcpu->enable_sem);
617 spin_lock_irqsave(&pcpu->target_freq_lock, flags[1]);
618 if (pcpu->target_freq < tunables->hispeed_freq) {
619 pcpu->target_freq = tunables->hispeed_freq;
620 cpumask_set_cpu(i, &speedchange_cpumask);
621 pcpu->pol_hispeed_val_time =
622 ktime_to_us(ktime_get());
625 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags[1]);
627 up_read(&pcpu->enable_sem);
630 spin_unlock_irqrestore(&speedchange_cpumask_lock, flags[0]);
633 wake_up_process(speedchange_task);
636 static int cpufreq_interactive_notifier(
637 struct notifier_block *nb, unsigned long val, void *data)
639 struct cpufreq_freqs *freq = data;
640 struct cpufreq_interactive_cpuinfo *pcpu;
644 if (val == CPUFREQ_POSTCHANGE) {
645 pcpu = &per_cpu(cpuinfo, freq->cpu);
646 if (!down_read_trylock(&pcpu->enable_sem))
648 if (!pcpu->governor_enabled) {
649 up_read(&pcpu->enable_sem);
653 for_each_cpu(cpu, pcpu->policy->cpus) {
654 struct cpufreq_interactive_cpuinfo *pjcpu =
655 &per_cpu(cpuinfo, cpu);
656 if (cpu != freq->cpu) {
657 if (!down_read_trylock(&pjcpu->enable_sem))
659 if (!pjcpu->governor_enabled) {
660 up_read(&pjcpu->enable_sem);
664 spin_lock_irqsave(&pjcpu->load_lock, flags);
666 spin_unlock_irqrestore(&pjcpu->load_lock, flags);
667 if (cpu != freq->cpu)
668 up_read(&pjcpu->enable_sem);
671 up_read(&pcpu->enable_sem);
676 static struct notifier_block cpufreq_notifier_block = {
677 .notifier_call = cpufreq_interactive_notifier,
680 static unsigned int *get_tokenized_data(const char *buf, int *num_tokens)
685 unsigned int *tokenized_data;
689 while ((cp = strpbrk(cp + 1, " :")))
692 if (!(ntokens & 0x1))
695 tokenized_data = kmalloc(ntokens * sizeof(unsigned int), GFP_KERNEL);
696 if (!tokenized_data) {
703 while (i < ntokens) {
704 if (sscanf(cp, "%u", &tokenized_data[i++]) != 1)
707 cp = strpbrk(cp, " :");
716 *num_tokens = ntokens;
717 return tokenized_data;
720 kfree(tokenized_data);
725 static ssize_t show_target_loads(
726 struct cpufreq_interactive_tunables *tunables,
733 spin_lock_irqsave(&tunables->target_loads_lock, flags);
735 for (i = 0; i < tunables->ntarget_loads; i++)
736 ret += sprintf(buf + ret, "%u%s", tunables->target_loads[i],
737 i & 0x1 ? ":" : " ");
739 sprintf(buf + ret - 1, "\n");
740 spin_unlock_irqrestore(&tunables->target_loads_lock, flags);
744 static ssize_t store_target_loads(
745 struct cpufreq_interactive_tunables *tunables,
746 const char *buf, size_t count)
749 unsigned int *new_target_loads = NULL;
752 new_target_loads = get_tokenized_data(buf, &ntokens);
753 if (IS_ERR(new_target_loads))
754 return PTR_RET(new_target_loads);
756 spin_lock_irqsave(&tunables->target_loads_lock, flags);
757 if (tunables->target_loads != default_target_loads)
758 kfree(tunables->target_loads);
759 tunables->target_loads = new_target_loads;
760 tunables->ntarget_loads = ntokens;
761 spin_unlock_irqrestore(&tunables->target_loads_lock, flags);
765 static ssize_t show_above_hispeed_delay(
766 struct cpufreq_interactive_tunables *tunables, char *buf)
772 spin_lock_irqsave(&tunables->above_hispeed_delay_lock, flags);
774 for (i = 0; i < tunables->nabove_hispeed_delay; i++)
775 ret += sprintf(buf + ret, "%u%s",
776 tunables->above_hispeed_delay[i],
777 i & 0x1 ? ":" : " ");
779 sprintf(buf + ret - 1, "\n");
780 spin_unlock_irqrestore(&tunables->above_hispeed_delay_lock, flags);
784 static ssize_t store_above_hispeed_delay(
785 struct cpufreq_interactive_tunables *tunables,
786 const char *buf, size_t count)
789 unsigned int *new_above_hispeed_delay = NULL;
792 new_above_hispeed_delay = get_tokenized_data(buf, &ntokens);
793 if (IS_ERR(new_above_hispeed_delay))
794 return PTR_RET(new_above_hispeed_delay);
796 spin_lock_irqsave(&tunables->above_hispeed_delay_lock, flags);
797 if (tunables->above_hispeed_delay != default_above_hispeed_delay)
798 kfree(tunables->above_hispeed_delay);
799 tunables->above_hispeed_delay = new_above_hispeed_delay;
800 tunables->nabove_hispeed_delay = ntokens;
801 spin_unlock_irqrestore(&tunables->above_hispeed_delay_lock, flags);
806 static ssize_t show_hispeed_freq(struct cpufreq_interactive_tunables *tunables,
809 return sprintf(buf, "%u\n", tunables->hispeed_freq);
812 static ssize_t store_hispeed_freq(struct cpufreq_interactive_tunables *tunables,
813 const char *buf, size_t count)
816 long unsigned int val;
818 ret = kstrtoul(buf, 0, &val);
821 tunables->hispeed_freq = val;
825 static ssize_t show_go_hispeed_load(struct cpufreq_interactive_tunables
826 *tunables, char *buf)
828 return sprintf(buf, "%lu\n", tunables->go_hispeed_load);
831 static ssize_t store_go_hispeed_load(struct cpufreq_interactive_tunables
832 *tunables, const char *buf, size_t count)
837 ret = kstrtoul(buf, 0, &val);
840 tunables->go_hispeed_load = val;
844 static ssize_t show_min_sample_time(struct cpufreq_interactive_tunables
845 *tunables, char *buf)
847 return sprintf(buf, "%lu\n", tunables->min_sample_time);
850 static ssize_t store_min_sample_time(struct cpufreq_interactive_tunables
851 *tunables, const char *buf, size_t count)
856 ret = kstrtoul(buf, 0, &val);
859 tunables->min_sample_time = val;
863 static ssize_t show_timer_rate(struct cpufreq_interactive_tunables *tunables,
866 return sprintf(buf, "%lu\n", tunables->timer_rate);
869 static ssize_t store_timer_rate(struct cpufreq_interactive_tunables *tunables,
870 const char *buf, size_t count)
873 unsigned long val, val_round;
875 ret = kstrtoul(buf, 0, &val);
879 val_round = jiffies_to_usecs(usecs_to_jiffies(val));
880 if (val != val_round)
881 pr_warn("timer_rate not aligned to jiffy. Rounded up to %lu\n",
884 tunables->timer_rate = val_round;
888 static ssize_t show_timer_slack(struct cpufreq_interactive_tunables *tunables,
891 return sprintf(buf, "%d\n", tunables->timer_slack_val);
894 static ssize_t store_timer_slack(struct cpufreq_interactive_tunables *tunables,
895 const char *buf, size_t count)
900 ret = kstrtol(buf, 10, &val);
904 tunables->timer_slack_val = val;
908 static ssize_t show_boost(struct cpufreq_interactive_tunables *tunables,
911 return sprintf(buf, "%d\n", tunables->boost_val);
914 static ssize_t store_boost(struct cpufreq_interactive_tunables *tunables,
915 const char *buf, size_t count)
920 ret = kstrtoul(buf, 0, &val);
924 tunables->boost_val = val;
926 if (tunables->boost_val) {
927 trace_cpufreq_interactive_boost("on");
928 if (!tunables->boosted)
929 cpufreq_interactive_boost(tunables);
931 tunables->boostpulse_endtime = ktime_to_us(ktime_get());
932 trace_cpufreq_interactive_unboost("off");
938 static ssize_t store_boostpulse(struct cpufreq_interactive_tunables *tunables,
939 const char *buf, size_t count)
944 ret = kstrtoul(buf, 0, &val);
948 tunables->boostpulse_endtime = ktime_to_us(ktime_get()) +
949 tunables->boostpulse_duration_val;
950 trace_cpufreq_interactive_boost("pulse");
951 if (!tunables->boosted)
952 cpufreq_interactive_boost(tunables);
956 static ssize_t show_boostpulse_duration(struct cpufreq_interactive_tunables
957 *tunables, char *buf)
959 return sprintf(buf, "%d\n", tunables->boostpulse_duration_val);
962 static ssize_t store_boostpulse_duration(struct cpufreq_interactive_tunables
963 *tunables, const char *buf, size_t count)
968 ret = kstrtoul(buf, 0, &val);
972 tunables->boostpulse_duration_val = val;
976 static ssize_t show_io_is_busy(struct cpufreq_interactive_tunables *tunables,
979 return sprintf(buf, "%u\n", tunables->io_is_busy);
982 static ssize_t store_io_is_busy(struct cpufreq_interactive_tunables *tunables,
983 const char *buf, size_t count)
988 ret = kstrtoul(buf, 0, &val);
991 tunables->io_is_busy = val;
996 * Create show/store routines
997 * - sys: One governor instance for complete SYSTEM
998 * - pol: One governor instance per struct cpufreq_policy
1000 #define show_gov_pol_sys(file_name) \
1001 static ssize_t show_##file_name##_gov_sys \
1002 (struct kobject *kobj, struct attribute *attr, char *buf) \
1004 return show_##file_name(common_tunables, buf); \
1007 static ssize_t show_##file_name##_gov_pol \
1008 (struct cpufreq_policy *policy, char *buf) \
1010 return show_##file_name(policy->governor_data, buf); \
1013 #define store_gov_pol_sys(file_name) \
1014 static ssize_t store_##file_name##_gov_sys \
1015 (struct kobject *kobj, struct attribute *attr, const char *buf, \
1018 return store_##file_name(common_tunables, buf, count); \
1021 static ssize_t store_##file_name##_gov_pol \
1022 (struct cpufreq_policy *policy, const char *buf, size_t count) \
1024 return store_##file_name(policy->governor_data, buf, count); \
1027 #define show_store_gov_pol_sys(file_name) \
1028 show_gov_pol_sys(file_name); \
1029 store_gov_pol_sys(file_name)
1031 show_store_gov_pol_sys(target_loads);
1032 show_store_gov_pol_sys(above_hispeed_delay);
1033 show_store_gov_pol_sys(hispeed_freq);
1034 show_store_gov_pol_sys(go_hispeed_load);
1035 show_store_gov_pol_sys(min_sample_time);
1036 show_store_gov_pol_sys(timer_rate);
1037 show_store_gov_pol_sys(timer_slack);
1038 show_store_gov_pol_sys(boost);
1039 store_gov_pol_sys(boostpulse);
1040 show_store_gov_pol_sys(boostpulse_duration);
1041 show_store_gov_pol_sys(io_is_busy);
1043 #define gov_sys_attr_rw(_name) \
1044 static struct global_attr _name##_gov_sys = \
1045 __ATTR(_name, 0644, show_##_name##_gov_sys, store_##_name##_gov_sys)
1047 #define gov_pol_attr_rw(_name) \
1048 static struct freq_attr _name##_gov_pol = \
1049 __ATTR(_name, 0644, show_##_name##_gov_pol, store_##_name##_gov_pol)
1051 #define gov_sys_pol_attr_rw(_name) \
1052 gov_sys_attr_rw(_name); \
1053 gov_pol_attr_rw(_name)
1055 gov_sys_pol_attr_rw(target_loads);
1056 gov_sys_pol_attr_rw(above_hispeed_delay);
1057 gov_sys_pol_attr_rw(hispeed_freq);
1058 gov_sys_pol_attr_rw(go_hispeed_load);
1059 gov_sys_pol_attr_rw(min_sample_time);
1060 gov_sys_pol_attr_rw(timer_rate);
1061 gov_sys_pol_attr_rw(timer_slack);
1062 gov_sys_pol_attr_rw(boost);
1063 gov_sys_pol_attr_rw(boostpulse_duration);
1064 gov_sys_pol_attr_rw(io_is_busy);
1066 static struct global_attr boostpulse_gov_sys =
1067 __ATTR(boostpulse, 0200, NULL, store_boostpulse_gov_sys);
1069 static struct freq_attr boostpulse_gov_pol =
1070 __ATTR(boostpulse, 0200, NULL, store_boostpulse_gov_pol);
1072 /* One Governor instance for entire system */
1073 static struct attribute *interactive_attributes_gov_sys[] = {
1074 &target_loads_gov_sys.attr,
1075 &above_hispeed_delay_gov_sys.attr,
1076 &hispeed_freq_gov_sys.attr,
1077 &go_hispeed_load_gov_sys.attr,
1078 &min_sample_time_gov_sys.attr,
1079 &timer_rate_gov_sys.attr,
1080 &timer_slack_gov_sys.attr,
1081 &boost_gov_sys.attr,
1082 &boostpulse_gov_sys.attr,
1083 &boostpulse_duration_gov_sys.attr,
1084 &io_is_busy_gov_sys.attr,
1088 static struct attribute_group interactive_attr_group_gov_sys = {
1089 .attrs = interactive_attributes_gov_sys,
1090 .name = "interactive",
1093 /* Per policy governor instance */
1094 static struct attribute *interactive_attributes_gov_pol[] = {
1095 &target_loads_gov_pol.attr,
1096 &above_hispeed_delay_gov_pol.attr,
1097 &hispeed_freq_gov_pol.attr,
1098 &go_hispeed_load_gov_pol.attr,
1099 &min_sample_time_gov_pol.attr,
1100 &timer_rate_gov_pol.attr,
1101 &timer_slack_gov_pol.attr,
1102 &boost_gov_pol.attr,
1103 &boostpulse_gov_pol.attr,
1104 &boostpulse_duration_gov_pol.attr,
1105 &io_is_busy_gov_pol.attr,
1109 static struct attribute_group interactive_attr_group_gov_pol = {
1110 .attrs = interactive_attributes_gov_pol,
1111 .name = "interactive",
1114 static struct attribute_group *get_sysfs_attr(void)
1116 if (have_governor_per_policy())
1117 return &interactive_attr_group_gov_pol;
1119 return &interactive_attr_group_gov_sys;
1122 static int cpufreq_interactive_idle_notifier(struct notifier_block *nb,
1126 if (val == IDLE_END)
1127 cpufreq_interactive_idle_end();
1132 static struct notifier_block cpufreq_interactive_idle_nb = {
1133 .notifier_call = cpufreq_interactive_idle_notifier,
1136 static int cpufreq_governor_interactive(struct cpufreq_policy *policy,
1141 struct cpufreq_interactive_cpuinfo *pcpu;
1142 struct cpufreq_frequency_table *freq_table;
1143 struct cpufreq_interactive_tunables *tunables;
1144 unsigned long flags;
1146 if (have_governor_per_policy())
1147 tunables = policy->governor_data;
1149 tunables = common_tunables;
1151 WARN_ON(!tunables && (event != CPUFREQ_GOV_POLICY_INIT));
1154 case CPUFREQ_GOV_POLICY_INIT:
1155 if (have_governor_per_policy()) {
1157 } else if (tunables) {
1158 tunables->usage_count++;
1159 policy->governor_data = tunables;
1163 tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
1165 pr_err("%s: POLICY_INIT: kzalloc failed\n", __func__);
1169 tunables->usage_count = 1;
1170 tunables->above_hispeed_delay = default_above_hispeed_delay;
1171 tunables->nabove_hispeed_delay =
1172 ARRAY_SIZE(default_above_hispeed_delay);
1173 tunables->go_hispeed_load = DEFAULT_GO_HISPEED_LOAD;
1174 tunables->target_loads = default_target_loads;
1175 tunables->ntarget_loads = ARRAY_SIZE(default_target_loads);
1176 tunables->min_sample_time = DEFAULT_MIN_SAMPLE_TIME;
1177 tunables->timer_rate = DEFAULT_TIMER_RATE;
1178 tunables->boostpulse_duration_val = DEFAULT_MIN_SAMPLE_TIME;
1179 tunables->timer_slack_val = DEFAULT_TIMER_SLACK;
1181 spin_lock_init(&tunables->target_loads_lock);
1182 spin_lock_init(&tunables->above_hispeed_delay_lock);
1184 policy->governor_data = tunables;
1185 if (!have_governor_per_policy()) {
1186 common_tunables = tunables;
1189 rc = sysfs_create_group(get_governor_parent_kobj(policy),
1193 policy->governor_data = NULL;
1194 if (!have_governor_per_policy()) {
1195 common_tunables = NULL;
1200 if (!policy->governor->initialized) {
1201 idle_notifier_register(&cpufreq_interactive_idle_nb);
1202 cpufreq_register_notifier(&cpufreq_notifier_block,
1203 CPUFREQ_TRANSITION_NOTIFIER);
1208 case CPUFREQ_GOV_POLICY_EXIT:
1209 if (!--tunables->usage_count) {
1210 if (policy->governor->initialized == 1) {
1211 cpufreq_unregister_notifier(&cpufreq_notifier_block,
1212 CPUFREQ_TRANSITION_NOTIFIER);
1213 idle_notifier_unregister(&cpufreq_interactive_idle_nb);
1216 sysfs_remove_group(get_governor_parent_kobj(policy),
1220 common_tunables = NULL;
1223 policy->governor_data = NULL;
1226 case CPUFREQ_GOV_START:
1227 mutex_lock(&gov_lock);
1229 freq_table = cpufreq_frequency_get_table(policy->cpu);
1230 if (!tunables->hispeed_freq)
1231 tunables->hispeed_freq = policy->max;
1233 for_each_cpu(j, policy->cpus) {
1234 pcpu = &per_cpu(cpuinfo, j);
1235 pcpu->policy = policy;
1236 pcpu->target_freq = policy->cur;
1237 pcpu->freq_table = freq_table;
1238 pcpu->floor_freq = pcpu->target_freq;
1239 pcpu->pol_floor_val_time =
1240 ktime_to_us(ktime_get());
1241 pcpu->loc_floor_val_time = pcpu->pol_floor_val_time;
1242 pcpu->pol_hispeed_val_time = pcpu->pol_floor_val_time;
1243 pcpu->loc_hispeed_val_time = pcpu->pol_floor_val_time;
1244 down_write(&pcpu->enable_sem);
1245 del_timer_sync(&pcpu->cpu_timer);
1246 del_timer_sync(&pcpu->cpu_slack_timer);
1247 cpufreq_interactive_timer_start(tunables, j);
1248 pcpu->governor_enabled = 1;
1249 up_write(&pcpu->enable_sem);
1252 mutex_unlock(&gov_lock);
1255 case CPUFREQ_GOV_STOP:
1256 mutex_lock(&gov_lock);
1257 for_each_cpu(j, policy->cpus) {
1258 pcpu = &per_cpu(cpuinfo, j);
1259 down_write(&pcpu->enable_sem);
1260 pcpu->governor_enabled = 0;
1261 del_timer_sync(&pcpu->cpu_timer);
1262 del_timer_sync(&pcpu->cpu_slack_timer);
1263 up_write(&pcpu->enable_sem);
1266 mutex_unlock(&gov_lock);
1269 case CPUFREQ_GOV_LIMITS:
1270 if (policy->max < policy->cur)
1271 __cpufreq_driver_target(policy,
1272 policy->max, CPUFREQ_RELATION_H);
1273 else if (policy->min > policy->cur)
1274 __cpufreq_driver_target(policy,
1275 policy->min, CPUFREQ_RELATION_L);
1276 for_each_cpu(j, policy->cpus) {
1277 pcpu = &per_cpu(cpuinfo, j);
1279 down_read(&pcpu->enable_sem);
1280 if (pcpu->governor_enabled == 0) {
1281 up_read(&pcpu->enable_sem);
1285 spin_lock_irqsave(&pcpu->target_freq_lock, flags);
1286 if (policy->max < pcpu->target_freq)
1287 pcpu->target_freq = policy->max;
1288 else if (policy->min > pcpu->target_freq)
1289 pcpu->target_freq = policy->min;
1291 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
1292 up_read(&pcpu->enable_sem);
1299 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_INTERACTIVE
1302 struct cpufreq_governor cpufreq_gov_interactive = {
1303 .name = "interactive",
1304 .governor = cpufreq_governor_interactive,
1305 .max_transition_latency = 10000000,
1306 .owner = THIS_MODULE,
1309 static void cpufreq_interactive_nop_timer(unsigned long data)
1313 static int __init cpufreq_interactive_init(void)
1316 struct cpufreq_interactive_cpuinfo *pcpu;
1317 struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
1319 /* Initalize per-cpu timers */
1320 for_each_possible_cpu(i) {
1321 pcpu = &per_cpu(cpuinfo, i);
1322 init_timer_deferrable(&pcpu->cpu_timer);
1323 pcpu->cpu_timer.function = cpufreq_interactive_timer;
1324 pcpu->cpu_timer.data = i;
1325 init_timer(&pcpu->cpu_slack_timer);
1326 pcpu->cpu_slack_timer.function = cpufreq_interactive_nop_timer;
1327 spin_lock_init(&pcpu->load_lock);
1328 spin_lock_init(&pcpu->target_freq_lock);
1329 init_rwsem(&pcpu->enable_sem);
1332 spin_lock_init(&speedchange_cpumask_lock);
1333 mutex_init(&gov_lock);
1335 kthread_create(cpufreq_interactive_speedchange_task, NULL,
1337 if (IS_ERR(speedchange_task))
1338 return PTR_ERR(speedchange_task);
1340 sched_setscheduler_nocheck(speedchange_task, SCHED_FIFO, ¶m);
1341 get_task_struct(speedchange_task);
1343 /* NB: wake up so the thread does not look hung to the freezer */
1344 wake_up_process(speedchange_task);
1346 return cpufreq_register_governor(&cpufreq_gov_interactive);
1349 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_INTERACTIVE
1350 fs_initcall(cpufreq_interactive_init);
1352 module_init(cpufreq_interactive_init);
1355 static void __exit cpufreq_interactive_exit(void)
1357 cpufreq_unregister_governor(&cpufreq_gov_interactive);
1358 kthread_stop(speedchange_task);
1359 put_task_struct(speedchange_task);
1362 module_exit(cpufreq_interactive_exit);
1364 MODULE_AUTHOR("Mike Chan <mike@android.com>");
1365 MODULE_DESCRIPTION("'cpufreq_interactive' - A cpufreq governor for "
1366 "Latency sensitive workloads");
1367 MODULE_LICENSE("GPL");