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)
120 * HACK: FIXME: Bring back cpufreq_{get,put}_global_kobject()
121 * definition removed by upstream commit 8eec1020f0c0 "cpufreq:
122 * create cpu/cpufreq at boot time" to fix build failures.
124 static int cpufreq_global_kobject_usage;
126 int cpufreq_get_global_kobject(void)
128 if (!cpufreq_global_kobject_usage++)
129 return kobject_add(cpufreq_global_kobject,
130 &cpu_subsys.dev_root->kobj, "%s", "cpufreq");
135 void cpufreq_put_global_kobject(void)
137 if (!--cpufreq_global_kobject_usage)
138 kobject_del(cpufreq_global_kobject);
141 /* For cases where we have single governor instance for system */
142 static struct cpufreq_interactive_tunables *common_tunables;
144 static struct attribute_group *get_sysfs_attr(void);
146 static void cpufreq_interactive_timer_resched(
147 struct cpufreq_interactive_cpuinfo *pcpu)
149 struct cpufreq_interactive_tunables *tunables =
150 pcpu->policy->governor_data;
151 unsigned long expires;
154 spin_lock_irqsave(&pcpu->load_lock, flags);
156 get_cpu_idle_time(smp_processor_id(),
157 &pcpu->time_in_idle_timestamp,
158 tunables->io_is_busy);
159 pcpu->cputime_speedadj = 0;
160 pcpu->cputime_speedadj_timestamp = pcpu->time_in_idle_timestamp;
161 expires = jiffies + usecs_to_jiffies(tunables->timer_rate);
162 mod_timer_pinned(&pcpu->cpu_timer, expires);
164 if (tunables->timer_slack_val >= 0 &&
165 pcpu->target_freq > pcpu->policy->min) {
166 expires += usecs_to_jiffies(tunables->timer_slack_val);
167 mod_timer_pinned(&pcpu->cpu_slack_timer, expires);
170 spin_unlock_irqrestore(&pcpu->load_lock, flags);
173 /* The caller shall take enable_sem write semaphore to avoid any timer race.
174 * The cpu_timer and cpu_slack_timer must be deactivated when calling this
177 static void cpufreq_interactive_timer_start(
178 struct cpufreq_interactive_tunables *tunables, int cpu)
180 struct cpufreq_interactive_cpuinfo *pcpu = &per_cpu(cpuinfo, cpu);
181 unsigned long expires = jiffies +
182 usecs_to_jiffies(tunables->timer_rate);
185 pcpu->cpu_timer.expires = expires;
186 add_timer_on(&pcpu->cpu_timer, cpu);
187 if (tunables->timer_slack_val >= 0 &&
188 pcpu->target_freq > pcpu->policy->min) {
189 expires += usecs_to_jiffies(tunables->timer_slack_val);
190 pcpu->cpu_slack_timer.expires = expires;
191 add_timer_on(&pcpu->cpu_slack_timer, cpu);
194 spin_lock_irqsave(&pcpu->load_lock, flags);
196 get_cpu_idle_time(cpu, &pcpu->time_in_idle_timestamp,
197 tunables->io_is_busy);
198 pcpu->cputime_speedadj = 0;
199 pcpu->cputime_speedadj_timestamp = pcpu->time_in_idle_timestamp;
200 spin_unlock_irqrestore(&pcpu->load_lock, flags);
203 static unsigned int freq_to_above_hispeed_delay(
204 struct cpufreq_interactive_tunables *tunables,
211 spin_lock_irqsave(&tunables->above_hispeed_delay_lock, flags);
213 for (i = 0; i < tunables->nabove_hispeed_delay - 1 &&
214 freq >= tunables->above_hispeed_delay[i+1]; i += 2)
217 ret = tunables->above_hispeed_delay[i];
218 spin_unlock_irqrestore(&tunables->above_hispeed_delay_lock, flags);
222 static unsigned int freq_to_targetload(
223 struct cpufreq_interactive_tunables *tunables, unsigned int freq)
229 spin_lock_irqsave(&tunables->target_loads_lock, flags);
231 for (i = 0; i < tunables->ntarget_loads - 1 &&
232 freq >= tunables->target_loads[i+1]; i += 2)
235 ret = tunables->target_loads[i];
236 spin_unlock_irqrestore(&tunables->target_loads_lock, flags);
241 * If increasing frequencies never map to a lower target load then
242 * choose_freq() will find the minimum frequency that does not exceed its
243 * target load given the current load.
245 static unsigned int choose_freq(struct cpufreq_interactive_cpuinfo *pcpu,
246 unsigned int loadadjfreq)
248 unsigned int freq = pcpu->policy->cur;
249 unsigned int prevfreq, freqmin, freqmax;
258 tl = freq_to_targetload(pcpu->policy->governor_data, freq);
261 * Find the lowest frequency where the computed load is less
262 * than or equal to the target load.
265 if (cpufreq_frequency_table_target(
266 pcpu->policy, pcpu->freq_table, loadadjfreq / tl,
267 CPUFREQ_RELATION_L, &index))
269 freq = pcpu->freq_table[index].frequency;
271 if (freq > prevfreq) {
272 /* The previous frequency is too low. */
275 if (freq >= freqmax) {
277 * Find the highest frequency that is less
280 if (cpufreq_frequency_table_target(
281 pcpu->policy, pcpu->freq_table,
282 freqmax - 1, CPUFREQ_RELATION_H,
285 freq = pcpu->freq_table[index].frequency;
287 if (freq == freqmin) {
289 * The first frequency below freqmax
290 * has already been found to be too
291 * low. freqmax is the lowest speed
292 * we found that is fast enough.
298 } else if (freq < prevfreq) {
299 /* The previous frequency is high enough. */
302 if (freq <= freqmin) {
304 * Find the lowest frequency that is higher
307 if (cpufreq_frequency_table_target(
308 pcpu->policy, pcpu->freq_table,
309 freqmin + 1, CPUFREQ_RELATION_L,
312 freq = pcpu->freq_table[index].frequency;
315 * If freqmax is the first frequency above
316 * freqmin then we have already found that
317 * this speed is fast enough.
324 /* If same frequency chosen as previous then done. */
325 } while (freq != prevfreq);
330 static u64 update_load(int cpu)
332 struct cpufreq_interactive_cpuinfo *pcpu = &per_cpu(cpuinfo, cpu);
333 struct cpufreq_interactive_tunables *tunables =
334 pcpu->policy->governor_data;
337 unsigned int delta_idle;
338 unsigned int delta_time;
341 now_idle = get_cpu_idle_time(cpu, &now, tunables->io_is_busy);
342 delta_idle = (unsigned int)(now_idle - pcpu->time_in_idle);
343 delta_time = (unsigned int)(now - pcpu->time_in_idle_timestamp);
345 if (delta_time <= delta_idle)
348 active_time = delta_time - delta_idle;
350 pcpu->cputime_speedadj += active_time * pcpu->policy->cur;
352 pcpu->time_in_idle = now_idle;
353 pcpu->time_in_idle_timestamp = now;
357 static void cpufreq_interactive_timer(unsigned long data)
360 unsigned int delta_time;
361 u64 cputime_speedadj;
363 struct cpufreq_interactive_cpuinfo *pcpu =
364 &per_cpu(cpuinfo, data);
365 struct cpufreq_interactive_tunables *tunables =
366 pcpu->policy->governor_data;
367 unsigned int new_freq;
368 unsigned int loadadjfreq;
373 if (!down_read_trylock(&pcpu->enable_sem))
375 if (!pcpu->governor_enabled)
378 spin_lock_irqsave(&pcpu->load_lock, flags);
379 now = update_load(data);
380 delta_time = (unsigned int)(now - pcpu->cputime_speedadj_timestamp);
381 cputime_speedadj = pcpu->cputime_speedadj;
382 spin_unlock_irqrestore(&pcpu->load_lock, flags);
384 if (WARN_ON_ONCE(!delta_time))
387 spin_lock_irqsave(&pcpu->target_freq_lock, flags);
388 do_div(cputime_speedadj, delta_time);
389 loadadjfreq = (unsigned int)cputime_speedadj * 100;
390 cpu_load = loadadjfreq / pcpu->target_freq;
391 tunables->boosted = tunables->boost_val || now < tunables->boostpulse_endtime;
393 if (cpu_load >= tunables->go_hispeed_load || tunables->boosted) {
394 if (pcpu->policy->cur < tunables->hispeed_freq) {
395 new_freq = tunables->hispeed_freq;
397 new_freq = choose_freq(pcpu, loadadjfreq);
399 if (new_freq < tunables->hispeed_freq)
400 new_freq = tunables->hispeed_freq;
403 new_freq = choose_freq(pcpu, loadadjfreq);
404 if (new_freq > tunables->hispeed_freq &&
405 pcpu->policy->cur < tunables->hispeed_freq)
406 new_freq = tunables->hispeed_freq;
409 if (pcpu->policy->cur >= tunables->hispeed_freq &&
410 new_freq > pcpu->policy->cur &&
411 now - pcpu->pol_hispeed_val_time <
412 freq_to_above_hispeed_delay(tunables, pcpu->policy->cur)) {
413 trace_cpufreq_interactive_notyet(
414 data, cpu_load, pcpu->target_freq,
415 pcpu->policy->cur, new_freq);
416 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
420 pcpu->loc_hispeed_val_time = now;
422 if (cpufreq_frequency_table_target(pcpu->policy, pcpu->freq_table,
423 new_freq, CPUFREQ_RELATION_L,
425 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
429 new_freq = pcpu->freq_table[index].frequency;
432 * Do not scale below floor_freq unless we have been at or above the
433 * floor frequency for the minimum sample time since last validated.
435 max_fvtime = max(pcpu->pol_floor_val_time, pcpu->loc_floor_val_time);
436 if (new_freq < pcpu->floor_freq &&
437 pcpu->target_freq >= pcpu->policy->cur) {
438 if (now - max_fvtime < tunables->min_sample_time) {
439 trace_cpufreq_interactive_notyet(
440 data, cpu_load, pcpu->target_freq,
441 pcpu->policy->cur, new_freq);
442 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
448 * Update the timestamp for checking whether speed has been held at
449 * or above the selected frequency for a minimum of min_sample_time,
450 * if not boosted to hispeed_freq. If boosted to hispeed_freq then we
451 * allow the speed to drop as soon as the boostpulse duration expires
452 * (or the indefinite boost is turned off).
455 if (!tunables->boosted || new_freq > tunables->hispeed_freq) {
456 pcpu->floor_freq = new_freq;
457 if (pcpu->target_freq >= pcpu->policy->cur ||
458 new_freq >= pcpu->policy->cur)
459 pcpu->loc_floor_val_time = now;
462 if (pcpu->target_freq == new_freq &&
463 pcpu->target_freq <= pcpu->policy->cur) {
464 trace_cpufreq_interactive_already(
465 data, cpu_load, pcpu->target_freq,
466 pcpu->policy->cur, new_freq);
467 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
471 trace_cpufreq_interactive_target(data, cpu_load, pcpu->target_freq,
472 pcpu->policy->cur, new_freq);
474 pcpu->target_freq = new_freq;
475 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
476 spin_lock_irqsave(&speedchange_cpumask_lock, flags);
477 cpumask_set_cpu(data, &speedchange_cpumask);
478 spin_unlock_irqrestore(&speedchange_cpumask_lock, flags);
479 wake_up_process(speedchange_task);
482 if (!timer_pending(&pcpu->cpu_timer))
483 cpufreq_interactive_timer_resched(pcpu);
486 up_read(&pcpu->enable_sem);
490 static void cpufreq_interactive_idle_end(void)
492 struct cpufreq_interactive_cpuinfo *pcpu =
493 &per_cpu(cpuinfo, smp_processor_id());
495 if (!down_read_trylock(&pcpu->enable_sem))
497 if (!pcpu->governor_enabled) {
498 up_read(&pcpu->enable_sem);
502 /* Arm the timer for 1-2 ticks later if not already. */
503 if (!timer_pending(&pcpu->cpu_timer)) {
504 cpufreq_interactive_timer_resched(pcpu);
505 } else if (time_after_eq(jiffies, pcpu->cpu_timer.expires)) {
506 del_timer(&pcpu->cpu_timer);
507 del_timer(&pcpu->cpu_slack_timer);
508 cpufreq_interactive_timer(smp_processor_id());
511 up_read(&pcpu->enable_sem);
514 static int cpufreq_interactive_speedchange_task(void *data)
519 struct cpufreq_interactive_cpuinfo *pcpu;
522 set_current_state(TASK_INTERRUPTIBLE);
523 spin_lock_irqsave(&speedchange_cpumask_lock, flags);
525 if (cpumask_empty(&speedchange_cpumask)) {
526 spin_unlock_irqrestore(&speedchange_cpumask_lock,
530 if (kthread_should_stop())
533 spin_lock_irqsave(&speedchange_cpumask_lock, flags);
536 set_current_state(TASK_RUNNING);
537 tmp_mask = speedchange_cpumask;
538 cpumask_clear(&speedchange_cpumask);
539 spin_unlock_irqrestore(&speedchange_cpumask_lock, flags);
541 for_each_cpu(cpu, &tmp_mask) {
543 unsigned int max_freq = 0;
544 struct cpufreq_interactive_cpuinfo *pjcpu;
545 u64 hvt = ~0ULL, fvt = 0;
547 pcpu = &per_cpu(cpuinfo, cpu);
548 if (!down_read_trylock(&pcpu->enable_sem))
550 if (!pcpu->governor_enabled) {
551 up_read(&pcpu->enable_sem);
555 for_each_cpu(j, pcpu->policy->cpus) {
556 pjcpu = &per_cpu(cpuinfo, j);
558 fvt = max(fvt, pjcpu->loc_floor_val_time);
559 if (pjcpu->target_freq > max_freq) {
560 max_freq = pjcpu->target_freq;
561 hvt = pjcpu->loc_hispeed_val_time;
562 } else if (pjcpu->target_freq == max_freq) {
563 hvt = min(hvt, pjcpu->loc_hispeed_val_time);
566 for_each_cpu(j, pcpu->policy->cpus) {
567 pjcpu = &per_cpu(cpuinfo, j);
568 pjcpu->pol_floor_val_time = fvt;
571 if (max_freq != pcpu->policy->cur) {
572 __cpufreq_driver_target(pcpu->policy,
575 for_each_cpu(j, pcpu->policy->cpus) {
576 pjcpu = &per_cpu(cpuinfo, j);
577 pjcpu->pol_hispeed_val_time = hvt;
580 trace_cpufreq_interactive_setspeed(cpu,
584 up_read(&pcpu->enable_sem);
591 static void cpufreq_interactive_boost(struct cpufreq_interactive_tunables *tunables)
595 unsigned long flags[2];
596 struct cpufreq_interactive_cpuinfo *pcpu;
598 tunables->boosted = true;
600 spin_lock_irqsave(&speedchange_cpumask_lock, flags[0]);
602 for_each_online_cpu(i) {
603 pcpu = &per_cpu(cpuinfo, i);
604 if (tunables != pcpu->policy->governor_data)
607 spin_lock_irqsave(&pcpu->target_freq_lock, flags[1]);
608 if (pcpu->target_freq < tunables->hispeed_freq) {
609 pcpu->target_freq = tunables->hispeed_freq;
610 cpumask_set_cpu(i, &speedchange_cpumask);
611 pcpu->pol_hispeed_val_time =
612 ktime_to_us(ktime_get());
615 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags[1]);
618 spin_unlock_irqrestore(&speedchange_cpumask_lock, flags[0]);
621 wake_up_process(speedchange_task);
624 static int cpufreq_interactive_notifier(
625 struct notifier_block *nb, unsigned long val, void *data)
627 struct cpufreq_freqs *freq = data;
628 struct cpufreq_interactive_cpuinfo *pcpu;
632 if (val == CPUFREQ_POSTCHANGE) {
633 pcpu = &per_cpu(cpuinfo, freq->cpu);
634 if (!down_read_trylock(&pcpu->enable_sem))
636 if (!pcpu->governor_enabled) {
637 up_read(&pcpu->enable_sem);
641 for_each_cpu(cpu, pcpu->policy->cpus) {
642 struct cpufreq_interactive_cpuinfo *pjcpu =
643 &per_cpu(cpuinfo, cpu);
644 if (cpu != freq->cpu) {
645 if (!down_read_trylock(&pjcpu->enable_sem))
647 if (!pjcpu->governor_enabled) {
648 up_read(&pjcpu->enable_sem);
652 spin_lock_irqsave(&pjcpu->load_lock, flags);
654 spin_unlock_irqrestore(&pjcpu->load_lock, flags);
655 if (cpu != freq->cpu)
656 up_read(&pjcpu->enable_sem);
659 up_read(&pcpu->enable_sem);
664 static struct notifier_block cpufreq_notifier_block = {
665 .notifier_call = cpufreq_interactive_notifier,
668 static unsigned int *get_tokenized_data(const char *buf, int *num_tokens)
673 unsigned int *tokenized_data;
677 while ((cp = strpbrk(cp + 1, " :")))
680 if (!(ntokens & 0x1))
683 tokenized_data = kmalloc(ntokens * sizeof(unsigned int), GFP_KERNEL);
684 if (!tokenized_data) {
691 while (i < ntokens) {
692 if (sscanf(cp, "%u", &tokenized_data[i++]) != 1)
695 cp = strpbrk(cp, " :");
704 *num_tokens = ntokens;
705 return tokenized_data;
708 kfree(tokenized_data);
713 static ssize_t show_target_loads(
714 struct cpufreq_interactive_tunables *tunables,
721 spin_lock_irqsave(&tunables->target_loads_lock, flags);
723 for (i = 0; i < tunables->ntarget_loads; i++)
724 ret += sprintf(buf + ret, "%u%s", tunables->target_loads[i],
725 i & 0x1 ? ":" : " ");
727 sprintf(buf + ret - 1, "\n");
728 spin_unlock_irqrestore(&tunables->target_loads_lock, flags);
732 static ssize_t store_target_loads(
733 struct cpufreq_interactive_tunables *tunables,
734 const char *buf, size_t count)
737 unsigned int *new_target_loads = NULL;
740 new_target_loads = get_tokenized_data(buf, &ntokens);
741 if (IS_ERR(new_target_loads))
742 return PTR_RET(new_target_loads);
744 spin_lock_irqsave(&tunables->target_loads_lock, flags);
745 if (tunables->target_loads != default_target_loads)
746 kfree(tunables->target_loads);
747 tunables->target_loads = new_target_loads;
748 tunables->ntarget_loads = ntokens;
749 spin_unlock_irqrestore(&tunables->target_loads_lock, flags);
753 static ssize_t show_above_hispeed_delay(
754 struct cpufreq_interactive_tunables *tunables, char *buf)
760 spin_lock_irqsave(&tunables->above_hispeed_delay_lock, flags);
762 for (i = 0; i < tunables->nabove_hispeed_delay; i++)
763 ret += sprintf(buf + ret, "%u%s",
764 tunables->above_hispeed_delay[i],
765 i & 0x1 ? ":" : " ");
767 sprintf(buf + ret - 1, "\n");
768 spin_unlock_irqrestore(&tunables->above_hispeed_delay_lock, flags);
772 static ssize_t store_above_hispeed_delay(
773 struct cpufreq_interactive_tunables *tunables,
774 const char *buf, size_t count)
777 unsigned int *new_above_hispeed_delay = NULL;
780 new_above_hispeed_delay = get_tokenized_data(buf, &ntokens);
781 if (IS_ERR(new_above_hispeed_delay))
782 return PTR_RET(new_above_hispeed_delay);
784 spin_lock_irqsave(&tunables->above_hispeed_delay_lock, flags);
785 if (tunables->above_hispeed_delay != default_above_hispeed_delay)
786 kfree(tunables->above_hispeed_delay);
787 tunables->above_hispeed_delay = new_above_hispeed_delay;
788 tunables->nabove_hispeed_delay = ntokens;
789 spin_unlock_irqrestore(&tunables->above_hispeed_delay_lock, flags);
794 static ssize_t show_hispeed_freq(struct cpufreq_interactive_tunables *tunables,
797 return sprintf(buf, "%u\n", tunables->hispeed_freq);
800 static ssize_t store_hispeed_freq(struct cpufreq_interactive_tunables *tunables,
801 const char *buf, size_t count)
804 long unsigned int val;
806 ret = kstrtoul(buf, 0, &val);
809 tunables->hispeed_freq = val;
813 static ssize_t show_go_hispeed_load(struct cpufreq_interactive_tunables
814 *tunables, char *buf)
816 return sprintf(buf, "%lu\n", tunables->go_hispeed_load);
819 static ssize_t store_go_hispeed_load(struct cpufreq_interactive_tunables
820 *tunables, const char *buf, size_t count)
825 ret = kstrtoul(buf, 0, &val);
828 tunables->go_hispeed_load = val;
832 static ssize_t show_min_sample_time(struct cpufreq_interactive_tunables
833 *tunables, char *buf)
835 return sprintf(buf, "%lu\n", tunables->min_sample_time);
838 static ssize_t store_min_sample_time(struct cpufreq_interactive_tunables
839 *tunables, const char *buf, size_t count)
844 ret = kstrtoul(buf, 0, &val);
847 tunables->min_sample_time = val;
851 static ssize_t show_timer_rate(struct cpufreq_interactive_tunables *tunables,
854 return sprintf(buf, "%lu\n", tunables->timer_rate);
857 static ssize_t store_timer_rate(struct cpufreq_interactive_tunables *tunables,
858 const char *buf, size_t count)
861 unsigned long val, val_round;
863 ret = kstrtoul(buf, 0, &val);
867 val_round = jiffies_to_usecs(usecs_to_jiffies(val));
868 if (val != val_round)
869 pr_warn("timer_rate not aligned to jiffy. Rounded up to %lu\n",
872 tunables->timer_rate = val_round;
876 static ssize_t show_timer_slack(struct cpufreq_interactive_tunables *tunables,
879 return sprintf(buf, "%d\n", tunables->timer_slack_val);
882 static ssize_t store_timer_slack(struct cpufreq_interactive_tunables *tunables,
883 const char *buf, size_t count)
888 ret = kstrtol(buf, 10, &val);
892 tunables->timer_slack_val = val;
896 static ssize_t show_boost(struct cpufreq_interactive_tunables *tunables,
899 return sprintf(buf, "%d\n", tunables->boost_val);
902 static ssize_t store_boost(struct cpufreq_interactive_tunables *tunables,
903 const char *buf, size_t count)
908 ret = kstrtoul(buf, 0, &val);
912 tunables->boost_val = val;
914 if (tunables->boost_val) {
915 trace_cpufreq_interactive_boost("on");
916 if (!tunables->boosted)
917 cpufreq_interactive_boost(tunables);
919 tunables->boostpulse_endtime = ktime_to_us(ktime_get());
920 trace_cpufreq_interactive_unboost("off");
926 static ssize_t store_boostpulse(struct cpufreq_interactive_tunables *tunables,
927 const char *buf, size_t count)
932 ret = kstrtoul(buf, 0, &val);
936 tunables->boostpulse_endtime = ktime_to_us(ktime_get()) +
937 tunables->boostpulse_duration_val;
938 trace_cpufreq_interactive_boost("pulse");
939 if (!tunables->boosted)
940 cpufreq_interactive_boost(tunables);
944 static ssize_t show_boostpulse_duration(struct cpufreq_interactive_tunables
945 *tunables, char *buf)
947 return sprintf(buf, "%d\n", tunables->boostpulse_duration_val);
950 static ssize_t store_boostpulse_duration(struct cpufreq_interactive_tunables
951 *tunables, const char *buf, size_t count)
956 ret = kstrtoul(buf, 0, &val);
960 tunables->boostpulse_duration_val = val;
964 static ssize_t show_io_is_busy(struct cpufreq_interactive_tunables *tunables,
967 return sprintf(buf, "%u\n", tunables->io_is_busy);
970 static ssize_t store_io_is_busy(struct cpufreq_interactive_tunables *tunables,
971 const char *buf, size_t count)
976 ret = kstrtoul(buf, 0, &val);
979 tunables->io_is_busy = val;
984 * Create show/store routines
985 * - sys: One governor instance for complete SYSTEM
986 * - pol: One governor instance per struct cpufreq_policy
988 #define show_gov_pol_sys(file_name) \
989 static ssize_t show_##file_name##_gov_sys \
990 (struct kobject *kobj, struct attribute *attr, char *buf) \
992 return show_##file_name(common_tunables, buf); \
995 static ssize_t show_##file_name##_gov_pol \
996 (struct cpufreq_policy *policy, char *buf) \
998 return show_##file_name(policy->governor_data, buf); \
1001 #define store_gov_pol_sys(file_name) \
1002 static ssize_t store_##file_name##_gov_sys \
1003 (struct kobject *kobj, struct attribute *attr, const char *buf, \
1006 return store_##file_name(common_tunables, buf, count); \
1009 static ssize_t store_##file_name##_gov_pol \
1010 (struct cpufreq_policy *policy, const char *buf, size_t count) \
1012 return store_##file_name(policy->governor_data, buf, count); \
1015 #define show_store_gov_pol_sys(file_name) \
1016 show_gov_pol_sys(file_name); \
1017 store_gov_pol_sys(file_name)
1019 show_store_gov_pol_sys(target_loads);
1020 show_store_gov_pol_sys(above_hispeed_delay);
1021 show_store_gov_pol_sys(hispeed_freq);
1022 show_store_gov_pol_sys(go_hispeed_load);
1023 show_store_gov_pol_sys(min_sample_time);
1024 show_store_gov_pol_sys(timer_rate);
1025 show_store_gov_pol_sys(timer_slack);
1026 show_store_gov_pol_sys(boost);
1027 store_gov_pol_sys(boostpulse);
1028 show_store_gov_pol_sys(boostpulse_duration);
1029 show_store_gov_pol_sys(io_is_busy);
1031 #define gov_sys_attr_rw(_name) \
1032 static struct global_attr _name##_gov_sys = \
1033 __ATTR(_name, 0644, show_##_name##_gov_sys, store_##_name##_gov_sys)
1035 #define gov_pol_attr_rw(_name) \
1036 static struct freq_attr _name##_gov_pol = \
1037 __ATTR(_name, 0644, show_##_name##_gov_pol, store_##_name##_gov_pol)
1039 #define gov_sys_pol_attr_rw(_name) \
1040 gov_sys_attr_rw(_name); \
1041 gov_pol_attr_rw(_name)
1043 gov_sys_pol_attr_rw(target_loads);
1044 gov_sys_pol_attr_rw(above_hispeed_delay);
1045 gov_sys_pol_attr_rw(hispeed_freq);
1046 gov_sys_pol_attr_rw(go_hispeed_load);
1047 gov_sys_pol_attr_rw(min_sample_time);
1048 gov_sys_pol_attr_rw(timer_rate);
1049 gov_sys_pol_attr_rw(timer_slack);
1050 gov_sys_pol_attr_rw(boost);
1051 gov_sys_pol_attr_rw(boostpulse_duration);
1052 gov_sys_pol_attr_rw(io_is_busy);
1054 static struct global_attr boostpulse_gov_sys =
1055 __ATTR(boostpulse, 0200, NULL, store_boostpulse_gov_sys);
1057 static struct freq_attr boostpulse_gov_pol =
1058 __ATTR(boostpulse, 0200, NULL, store_boostpulse_gov_pol);
1060 /* One Governor instance for entire system */
1061 static struct attribute *interactive_attributes_gov_sys[] = {
1062 &target_loads_gov_sys.attr,
1063 &above_hispeed_delay_gov_sys.attr,
1064 &hispeed_freq_gov_sys.attr,
1065 &go_hispeed_load_gov_sys.attr,
1066 &min_sample_time_gov_sys.attr,
1067 &timer_rate_gov_sys.attr,
1068 &timer_slack_gov_sys.attr,
1069 &boost_gov_sys.attr,
1070 &boostpulse_gov_sys.attr,
1071 &boostpulse_duration_gov_sys.attr,
1072 &io_is_busy_gov_sys.attr,
1076 static struct attribute_group interactive_attr_group_gov_sys = {
1077 .attrs = interactive_attributes_gov_sys,
1078 .name = "interactive",
1081 /* Per policy governor instance */
1082 static struct attribute *interactive_attributes_gov_pol[] = {
1083 &target_loads_gov_pol.attr,
1084 &above_hispeed_delay_gov_pol.attr,
1085 &hispeed_freq_gov_pol.attr,
1086 &go_hispeed_load_gov_pol.attr,
1087 &min_sample_time_gov_pol.attr,
1088 &timer_rate_gov_pol.attr,
1089 &timer_slack_gov_pol.attr,
1090 &boost_gov_pol.attr,
1091 &boostpulse_gov_pol.attr,
1092 &boostpulse_duration_gov_pol.attr,
1093 &io_is_busy_gov_pol.attr,
1097 static struct attribute_group interactive_attr_group_gov_pol = {
1098 .attrs = interactive_attributes_gov_pol,
1099 .name = "interactive",
1102 static struct attribute_group *get_sysfs_attr(void)
1104 if (have_governor_per_policy())
1105 return &interactive_attr_group_gov_pol;
1107 return &interactive_attr_group_gov_sys;
1110 static int cpufreq_interactive_idle_notifier(struct notifier_block *nb,
1114 if (val == IDLE_END)
1115 cpufreq_interactive_idle_end();
1120 static struct notifier_block cpufreq_interactive_idle_nb = {
1121 .notifier_call = cpufreq_interactive_idle_notifier,
1124 static int cpufreq_governor_interactive(struct cpufreq_policy *policy,
1129 struct cpufreq_interactive_cpuinfo *pcpu;
1130 struct cpufreq_frequency_table *freq_table;
1131 struct cpufreq_interactive_tunables *tunables;
1132 unsigned long flags;
1134 if (have_governor_per_policy())
1135 tunables = policy->governor_data;
1137 tunables = common_tunables;
1139 WARN_ON(!tunables && (event != CPUFREQ_GOV_POLICY_INIT));
1142 case CPUFREQ_GOV_POLICY_INIT:
1143 if (have_governor_per_policy()) {
1145 } else if (tunables) {
1146 tunables->usage_count++;
1147 policy->governor_data = tunables;
1151 tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
1153 pr_err("%s: POLICY_INIT: kzalloc failed\n", __func__);
1157 tunables->usage_count = 1;
1158 tunables->above_hispeed_delay = default_above_hispeed_delay;
1159 tunables->nabove_hispeed_delay =
1160 ARRAY_SIZE(default_above_hispeed_delay);
1161 tunables->go_hispeed_load = DEFAULT_GO_HISPEED_LOAD;
1162 tunables->target_loads = default_target_loads;
1163 tunables->ntarget_loads = ARRAY_SIZE(default_target_loads);
1164 tunables->min_sample_time = DEFAULT_MIN_SAMPLE_TIME;
1165 tunables->timer_rate = DEFAULT_TIMER_RATE;
1166 tunables->boostpulse_duration_val = DEFAULT_MIN_SAMPLE_TIME;
1167 tunables->timer_slack_val = DEFAULT_TIMER_SLACK;
1169 spin_lock_init(&tunables->target_loads_lock);
1170 spin_lock_init(&tunables->above_hispeed_delay_lock);
1172 policy->governor_data = tunables;
1173 if (!have_governor_per_policy()) {
1174 common_tunables = tunables;
1175 WARN_ON(cpufreq_get_global_kobject());
1178 rc = sysfs_create_group(get_governor_parent_kobj(policy),
1182 policy->governor_data = NULL;
1183 if (!have_governor_per_policy()) {
1184 common_tunables = NULL;
1185 cpufreq_put_global_kobject();
1190 if (!policy->governor->initialized) {
1191 idle_notifier_register(&cpufreq_interactive_idle_nb);
1192 cpufreq_register_notifier(&cpufreq_notifier_block,
1193 CPUFREQ_TRANSITION_NOTIFIER);
1198 case CPUFREQ_GOV_POLICY_EXIT:
1199 if (!--tunables->usage_count) {
1200 if (policy->governor->initialized == 1) {
1201 cpufreq_unregister_notifier(&cpufreq_notifier_block,
1202 CPUFREQ_TRANSITION_NOTIFIER);
1203 idle_notifier_unregister(&cpufreq_interactive_idle_nb);
1206 sysfs_remove_group(get_governor_parent_kobj(policy),
1209 if (!have_governor_per_policy())
1210 cpufreq_put_global_kobject();
1213 common_tunables = NULL;
1216 policy->governor_data = NULL;
1219 case CPUFREQ_GOV_START:
1220 mutex_lock(&gov_lock);
1222 freq_table = cpufreq_frequency_get_table(policy->cpu);
1223 if (!tunables->hispeed_freq)
1224 tunables->hispeed_freq = policy->max;
1226 for_each_cpu(j, policy->cpus) {
1227 pcpu = &per_cpu(cpuinfo, j);
1228 pcpu->policy = policy;
1229 pcpu->target_freq = policy->cur;
1230 pcpu->freq_table = freq_table;
1231 pcpu->floor_freq = pcpu->target_freq;
1232 pcpu->pol_floor_val_time =
1233 ktime_to_us(ktime_get());
1234 pcpu->loc_floor_val_time = pcpu->pol_floor_val_time;
1235 pcpu->pol_hispeed_val_time = pcpu->pol_floor_val_time;
1236 pcpu->loc_hispeed_val_time = pcpu->pol_floor_val_time;
1237 down_write(&pcpu->enable_sem);
1238 del_timer_sync(&pcpu->cpu_timer);
1239 del_timer_sync(&pcpu->cpu_slack_timer);
1240 cpufreq_interactive_timer_start(tunables, j);
1241 pcpu->governor_enabled = 1;
1242 up_write(&pcpu->enable_sem);
1245 mutex_unlock(&gov_lock);
1248 case CPUFREQ_GOV_STOP:
1249 mutex_lock(&gov_lock);
1250 for_each_cpu(j, policy->cpus) {
1251 pcpu = &per_cpu(cpuinfo, j);
1252 down_write(&pcpu->enable_sem);
1253 pcpu->governor_enabled = 0;
1254 del_timer_sync(&pcpu->cpu_timer);
1255 del_timer_sync(&pcpu->cpu_slack_timer);
1256 up_write(&pcpu->enable_sem);
1259 mutex_unlock(&gov_lock);
1262 case CPUFREQ_GOV_LIMITS:
1263 if (policy->max < policy->cur)
1264 __cpufreq_driver_target(policy,
1265 policy->max, CPUFREQ_RELATION_H);
1266 else if (policy->min > policy->cur)
1267 __cpufreq_driver_target(policy,
1268 policy->min, CPUFREQ_RELATION_L);
1269 for_each_cpu(j, policy->cpus) {
1270 pcpu = &per_cpu(cpuinfo, j);
1272 down_read(&pcpu->enable_sem);
1273 if (pcpu->governor_enabled == 0) {
1274 up_read(&pcpu->enable_sem);
1278 spin_lock_irqsave(&pcpu->target_freq_lock, flags);
1279 if (policy->max < pcpu->target_freq)
1280 pcpu->target_freq = policy->max;
1281 else if (policy->min > pcpu->target_freq)
1282 pcpu->target_freq = policy->min;
1284 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
1285 up_read(&pcpu->enable_sem);
1292 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_INTERACTIVE
1295 struct cpufreq_governor cpufreq_gov_interactive = {
1296 .name = "interactive",
1297 .governor = cpufreq_governor_interactive,
1298 .max_transition_latency = 10000000,
1299 .owner = THIS_MODULE,
1302 static void cpufreq_interactive_nop_timer(unsigned long data)
1306 static int __init cpufreq_interactive_init(void)
1309 struct cpufreq_interactive_cpuinfo *pcpu;
1310 struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
1312 /* Initalize per-cpu timers */
1313 for_each_possible_cpu(i) {
1314 pcpu = &per_cpu(cpuinfo, i);
1315 init_timer_deferrable(&pcpu->cpu_timer);
1316 pcpu->cpu_timer.function = cpufreq_interactive_timer;
1317 pcpu->cpu_timer.data = i;
1318 init_timer(&pcpu->cpu_slack_timer);
1319 pcpu->cpu_slack_timer.function = cpufreq_interactive_nop_timer;
1320 spin_lock_init(&pcpu->load_lock);
1321 spin_lock_init(&pcpu->target_freq_lock);
1322 init_rwsem(&pcpu->enable_sem);
1325 spin_lock_init(&speedchange_cpumask_lock);
1326 mutex_init(&gov_lock);
1328 kthread_create(cpufreq_interactive_speedchange_task, NULL,
1330 if (IS_ERR(speedchange_task))
1331 return PTR_ERR(speedchange_task);
1333 sched_setscheduler_nocheck(speedchange_task, SCHED_FIFO, ¶m);
1334 get_task_struct(speedchange_task);
1336 /* NB: wake up so the thread does not look hung to the freezer */
1337 wake_up_process(speedchange_task);
1339 return cpufreq_register_governor(&cpufreq_gov_interactive);
1342 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_INTERACTIVE
1343 fs_initcall(cpufreq_interactive_init);
1345 module_init(cpufreq_interactive_init);
1348 static void __exit cpufreq_interactive_exit(void)
1350 cpufreq_unregister_governor(&cpufreq_gov_interactive);
1351 kthread_stop(speedchange_task);
1352 put_task_struct(speedchange_task);
1355 module_exit(cpufreq_interactive_exit);
1357 MODULE_AUTHOR("Mike Chan <mike@android.com>");
1358 MODULE_DESCRIPTION("'cpufreq_interactive' - A cpufreq governor for "
1359 "Latency sensitive workloads");
1360 MODULE_LICENSE("GPL");