2 * drivers/cpufreq/cpufreq_interactive.c
4 * Copyright (C) 2010 Google, Inc.
6 * This software is licensed under the terms of the GNU General Public
7 * License version 2, as published by the Free Software Foundation, and
8 * may be copied, distributed, and modified under those terms.
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
15 * Author: Mike Chan (mike@android.com)
19 #include <linux/cpu.h>
20 #include <linux/cpumask.h>
21 #include <linux/cpufreq.h>
22 #include <linux/module.h>
23 #include <linux/moduleparam.h>
24 #include <linux/rwsem.h>
25 #include <linux/sched.h>
26 #include <linux/sched/rt.h>
27 #include <linux/tick.h>
28 #include <linux/time.h>
29 #include <linux/timer.h>
30 #include <linux/workqueue.h>
31 #include <linux/kthread.h>
32 #include <linux/slab.h>
33 #include "cpufreq_governor.h"
35 #define CREATE_TRACE_POINTS
36 #include <trace/events/cpufreq_interactive.h>
38 struct cpufreq_interactive_cpuinfo {
39 struct timer_list cpu_timer;
40 struct timer_list cpu_slack_timer;
41 spinlock_t load_lock; /* protects the next 4 fields */
43 u64 time_in_idle_timestamp;
45 u64 cputime_speedadj_timestamp;
46 struct cpufreq_policy *policy;
47 struct cpufreq_frequency_table *freq_table;
48 spinlock_t target_freq_lock; /*protects target freq */
49 unsigned int target_freq;
50 unsigned int floor_freq;
51 unsigned int max_freq;
52 u64 floor_validate_time;
53 u64 hispeed_validate_time;
54 struct rw_semaphore enable_sem;
58 static DEFINE_PER_CPU(struct cpufreq_interactive_cpuinfo, cpuinfo);
60 /* realtime thread handles frequency scaling */
61 static struct task_struct *speedchange_task;
62 static cpumask_t speedchange_cpumask;
63 static spinlock_t speedchange_cpumask_lock;
64 static struct mutex gov_lock;
66 /* Target load. Lower values result in higher CPU speeds. */
67 #define DEFAULT_TARGET_LOAD 90
68 static unsigned int default_target_loads[] = {DEFAULT_TARGET_LOAD};
70 #define DEFAULT_TIMER_RATE (20 * USEC_PER_MSEC)
71 #define DEFAULT_ABOVE_HISPEED_DELAY DEFAULT_TIMER_RATE
72 static unsigned int default_above_hispeed_delay[] = {
73 DEFAULT_ABOVE_HISPEED_DELAY };
75 struct cpufreq_interactive_tunables {
77 /* Hi speed to bump to from lo speed when load burst (default max) */
78 unsigned int hispeed_freq;
79 /* Go to hi speed when CPU load at or above this value. */
80 #define DEFAULT_GO_HISPEED_LOAD 99
81 unsigned long go_hispeed_load;
82 /* Target load. Lower values result in higher CPU speeds. */
83 spinlock_t target_loads_lock;
84 unsigned int *target_loads;
87 * The minimum amount of time to spend at a frequency before we can ramp
90 #define DEFAULT_MIN_SAMPLE_TIME (80 * USEC_PER_MSEC)
91 unsigned long min_sample_time;
93 * The sample rate of the timer used to increase frequency
95 unsigned long timer_rate;
97 * Wait this long before raising speed above hispeed, by default a
98 * single timer interval.
100 spinlock_t above_hispeed_delay_lock;
101 unsigned int *above_hispeed_delay;
102 int nabove_hispeed_delay;
103 /* Non-zero means indefinite speed boost active */
105 /* Duration of a boot pulse in usecs */
106 int boostpulse_duration_val;
107 /* End time of boost pulse in ktime converted to usecs */
108 u64 boostpulse_endtime;
110 * Max additional time to wait in idle, beyond timer_rate, at speeds
111 * above minimum before wakeup to reduce speed, or -1 if unnecessary.
113 #define DEFAULT_TIMER_SLACK (4 * DEFAULT_TIMER_RATE)
118 /* For cases where we have single governor instance for system */
119 struct cpufreq_interactive_tunables *common_tunables;
121 static struct attribute_group *get_sysfs_attr(void);
123 static void cpufreq_interactive_timer_resched(
124 struct cpufreq_interactive_cpuinfo *pcpu)
126 struct cpufreq_interactive_tunables *tunables =
127 pcpu->policy->governor_data;
128 unsigned long expires;
131 spin_lock_irqsave(&pcpu->load_lock, flags);
133 get_cpu_idle_time(smp_processor_id(),
134 &pcpu->time_in_idle_timestamp,
135 tunables->io_is_busy);
136 pcpu->cputime_speedadj = 0;
137 pcpu->cputime_speedadj_timestamp = pcpu->time_in_idle_timestamp;
138 expires = jiffies + usecs_to_jiffies(tunables->timer_rate);
139 mod_timer_pinned(&pcpu->cpu_timer, expires);
141 if (tunables->timer_slack_val >= 0 &&
142 pcpu->target_freq > pcpu->policy->min) {
143 expires += usecs_to_jiffies(tunables->timer_slack_val);
144 mod_timer_pinned(&pcpu->cpu_slack_timer, expires);
147 spin_unlock_irqrestore(&pcpu->load_lock, flags);
150 /* The caller shall take enable_sem write semaphore to avoid any timer race.
151 * The cpu_timer and cpu_slack_timer must be deactivated when calling this
154 static void cpufreq_interactive_timer_start(
155 struct cpufreq_interactive_tunables *tunables, int cpu)
157 struct cpufreq_interactive_cpuinfo *pcpu = &per_cpu(cpuinfo, cpu);
158 unsigned long expires = jiffies +
159 usecs_to_jiffies(tunables->timer_rate);
162 pcpu->cpu_timer.expires = expires;
163 add_timer_on(&pcpu->cpu_timer, cpu);
164 if (tunables->timer_slack_val >= 0 &&
165 pcpu->target_freq > pcpu->policy->min) {
166 expires += usecs_to_jiffies(tunables->timer_slack_val);
167 pcpu->cpu_slack_timer.expires = expires;
168 add_timer_on(&pcpu->cpu_slack_timer, cpu);
171 spin_lock_irqsave(&pcpu->load_lock, flags);
173 get_cpu_idle_time(cpu, &pcpu->time_in_idle_timestamp,
174 tunables->io_is_busy);
175 pcpu->cputime_speedadj = 0;
176 pcpu->cputime_speedadj_timestamp = pcpu->time_in_idle_timestamp;
177 spin_unlock_irqrestore(&pcpu->load_lock, flags);
180 static unsigned int freq_to_above_hispeed_delay(
181 struct cpufreq_interactive_tunables *tunables,
188 spin_lock_irqsave(&tunables->above_hispeed_delay_lock, flags);
190 for (i = 0; i < tunables->nabove_hispeed_delay - 1 &&
191 freq >= tunables->above_hispeed_delay[i+1]; i += 2)
194 ret = tunables->above_hispeed_delay[i];
195 spin_unlock_irqrestore(&tunables->above_hispeed_delay_lock, flags);
199 static unsigned int freq_to_targetload(
200 struct cpufreq_interactive_tunables *tunables, unsigned int freq)
206 spin_lock_irqsave(&tunables->target_loads_lock, flags);
208 for (i = 0; i < tunables->ntarget_loads - 1 &&
209 freq >= tunables->target_loads[i+1]; i += 2)
212 ret = tunables->target_loads[i];
213 spin_unlock_irqrestore(&tunables->target_loads_lock, flags);
218 * If increasing frequencies never map to a lower target load then
219 * choose_freq() will find the minimum frequency that does not exceed its
220 * target load given the current load.
222 static unsigned int choose_freq(struct cpufreq_interactive_cpuinfo *pcpu,
223 unsigned int loadadjfreq)
225 unsigned int freq = pcpu->policy->cur;
226 unsigned int prevfreq, freqmin, freqmax;
235 tl = freq_to_targetload(pcpu->policy->governor_data, freq);
238 * Find the lowest frequency where the computed load is less
239 * than or equal to the target load.
242 if (cpufreq_frequency_table_target(
243 pcpu->policy, pcpu->freq_table, loadadjfreq / tl,
244 CPUFREQ_RELATION_L, &index))
246 freq = pcpu->freq_table[index].frequency;
248 if (freq > prevfreq) {
249 /* The previous frequency is too low. */
252 if (freq >= freqmax) {
254 * Find the highest frequency that is less
257 if (cpufreq_frequency_table_target(
258 pcpu->policy, pcpu->freq_table,
259 freqmax - 1, CPUFREQ_RELATION_H,
262 freq = pcpu->freq_table[index].frequency;
264 if (freq == freqmin) {
266 * The first frequency below freqmax
267 * has already been found to be too
268 * low. freqmax is the lowest speed
269 * we found that is fast enough.
275 } else if (freq < prevfreq) {
276 /* The previous frequency is high enough. */
279 if (freq <= freqmin) {
281 * Find the lowest frequency that is higher
284 if (cpufreq_frequency_table_target(
285 pcpu->policy, pcpu->freq_table,
286 freqmin + 1, CPUFREQ_RELATION_L,
289 freq = pcpu->freq_table[index].frequency;
292 * If freqmax is the first frequency above
293 * freqmin then we have already found that
294 * this speed is fast enough.
301 /* If same frequency chosen as previous then done. */
302 } while (freq != prevfreq);
307 static u64 update_load(int cpu)
309 struct cpufreq_interactive_cpuinfo *pcpu = &per_cpu(cpuinfo, cpu);
310 struct cpufreq_interactive_tunables *tunables =
311 pcpu->policy->governor_data;
318 now_idle = get_cpu_idle_time(cpu, &now, tunables->io_is_busy);
319 delta_idle = (now_idle - pcpu->time_in_idle);
320 delta_time = (now - pcpu->time_in_idle_timestamp);
322 if (delta_time <= delta_idle)
325 active_time = delta_time - delta_idle;
327 pcpu->cputime_speedadj += active_time * pcpu->policy->cur;
329 pcpu->time_in_idle = now_idle;
330 pcpu->time_in_idle_timestamp = now;
334 static void cpufreq_interactive_timer(unsigned long data)
337 unsigned int delta_time;
338 u64 cputime_speedadj;
340 struct cpufreq_interactive_cpuinfo *pcpu =
341 &per_cpu(cpuinfo, data);
342 struct cpufreq_interactive_tunables *tunables =
343 pcpu->policy->governor_data;
344 unsigned int new_freq;
345 unsigned int loadadjfreq;
350 if (!down_read_trylock(&pcpu->enable_sem))
352 if (!pcpu->governor_enabled)
355 spin_lock_irqsave(&pcpu->load_lock, flags);
356 now = update_load(data);
357 delta_time = (unsigned int)(now - pcpu->cputime_speedadj_timestamp);
358 cputime_speedadj = pcpu->cputime_speedadj;
359 spin_unlock_irqrestore(&pcpu->load_lock, flags);
361 if (WARN_ON_ONCE(!delta_time))
364 spin_lock_irqsave(&pcpu->target_freq_lock, flags);
365 do_div(cputime_speedadj, delta_time);
366 loadadjfreq = (unsigned int)cputime_speedadj * 100;
367 cpu_load = loadadjfreq / pcpu->target_freq;
368 boosted = tunables->boost_val || now < tunables->boostpulse_endtime;
370 if (cpu_load >= tunables->go_hispeed_load || boosted) {
371 if (pcpu->target_freq < tunables->hispeed_freq) {
372 new_freq = tunables->hispeed_freq;
374 new_freq = choose_freq(pcpu, loadadjfreq);
376 if (new_freq < tunables->hispeed_freq)
377 new_freq = tunables->hispeed_freq;
380 new_freq = choose_freq(pcpu, loadadjfreq);
383 if (pcpu->target_freq >= tunables->hispeed_freq &&
384 new_freq > pcpu->target_freq &&
385 now - pcpu->hispeed_validate_time <
386 freq_to_above_hispeed_delay(tunables, pcpu->target_freq)) {
387 trace_cpufreq_interactive_notyet(
388 data, cpu_load, pcpu->target_freq,
389 pcpu->policy->cur, new_freq);
390 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
394 pcpu->hispeed_validate_time = now;
396 if (cpufreq_frequency_table_target(pcpu->policy, pcpu->freq_table,
397 new_freq, CPUFREQ_RELATION_L,
399 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
403 new_freq = pcpu->freq_table[index].frequency;
406 * Do not scale below floor_freq unless we have been at or above the
407 * floor frequency for the minimum sample time since last validated.
409 if (new_freq < pcpu->floor_freq) {
410 if (now - pcpu->floor_validate_time <
411 tunables->min_sample_time) {
412 trace_cpufreq_interactive_notyet(
413 data, cpu_load, pcpu->target_freq,
414 pcpu->policy->cur, new_freq);
415 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
421 * Update the timestamp for checking whether speed has been held at
422 * or above the selected frequency for a minimum of min_sample_time,
423 * if not boosted to hispeed_freq. If boosted to hispeed_freq then we
424 * allow the speed to drop as soon as the boostpulse duration expires
425 * (or the indefinite boost is turned off).
428 if (!boosted || new_freq > tunables->hispeed_freq) {
429 pcpu->floor_freq = new_freq;
430 pcpu->floor_validate_time = now;
433 if (pcpu->target_freq == new_freq) {
434 trace_cpufreq_interactive_already(
435 data, cpu_load, pcpu->target_freq,
436 pcpu->policy->cur, new_freq);
437 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
438 goto rearm_if_notmax;
441 trace_cpufreq_interactive_target(data, cpu_load, pcpu->target_freq,
442 pcpu->policy->cur, new_freq);
444 pcpu->target_freq = new_freq;
445 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
446 spin_lock_irqsave(&speedchange_cpumask_lock, flags);
447 cpumask_set_cpu(data, &speedchange_cpumask);
448 spin_unlock_irqrestore(&speedchange_cpumask_lock, flags);
449 wake_up_process(speedchange_task);
453 * Already set max speed and don't see a need to change that,
454 * wait until next idle to re-evaluate, don't need timer.
456 if (pcpu->target_freq == pcpu->policy->max)
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_start(void)
470 struct cpufreq_interactive_cpuinfo *pcpu =
471 &per_cpu(cpuinfo, smp_processor_id());
474 if (!down_read_trylock(&pcpu->enable_sem))
476 if (!pcpu->governor_enabled) {
477 up_read(&pcpu->enable_sem);
481 pending = timer_pending(&pcpu->cpu_timer);
483 if (pcpu->target_freq != pcpu->policy->min) {
485 * Entering idle while not at lowest speed. On some
486 * platforms this can hold the other CPU(s) at that speed
487 * even though the CPU is idle. Set a timer to re-evaluate
488 * speed so this idle CPU doesn't hold the other CPUs above
489 * min indefinitely. This should probably be a quirk of
490 * the CPUFreq driver.
493 cpufreq_interactive_timer_resched(pcpu);
496 up_read(&pcpu->enable_sem);
499 static void cpufreq_interactive_idle_end(void)
501 struct cpufreq_interactive_cpuinfo *pcpu =
502 &per_cpu(cpuinfo, smp_processor_id());
504 if (!down_read_trylock(&pcpu->enable_sem))
506 if (!pcpu->governor_enabled) {
507 up_read(&pcpu->enable_sem);
511 /* Arm the timer for 1-2 ticks later if not already. */
512 if (!timer_pending(&pcpu->cpu_timer)) {
513 cpufreq_interactive_timer_resched(pcpu);
514 } else if (time_after_eq(jiffies, pcpu->cpu_timer.expires)) {
515 del_timer(&pcpu->cpu_timer);
516 del_timer(&pcpu->cpu_slack_timer);
517 cpufreq_interactive_timer(smp_processor_id());
520 up_read(&pcpu->enable_sem);
523 static int cpufreq_interactive_speedchange_task(void *data)
528 struct cpufreq_interactive_cpuinfo *pcpu;
531 set_current_state(TASK_INTERRUPTIBLE);
532 spin_lock_irqsave(&speedchange_cpumask_lock, flags);
534 if (cpumask_empty(&speedchange_cpumask)) {
535 spin_unlock_irqrestore(&speedchange_cpumask_lock,
539 if (kthread_should_stop())
542 spin_lock_irqsave(&speedchange_cpumask_lock, flags);
545 set_current_state(TASK_RUNNING);
546 tmp_mask = speedchange_cpumask;
547 cpumask_clear(&speedchange_cpumask);
548 spin_unlock_irqrestore(&speedchange_cpumask_lock, flags);
550 for_each_cpu(cpu, &tmp_mask) {
552 unsigned int max_freq = 0;
554 pcpu = &per_cpu(cpuinfo, cpu);
555 if (!down_read_trylock(&pcpu->enable_sem))
557 if (!pcpu->governor_enabled) {
558 up_read(&pcpu->enable_sem);
562 for_each_cpu(j, pcpu->policy->cpus) {
563 struct cpufreq_interactive_cpuinfo *pjcpu =
564 &per_cpu(cpuinfo, j);
566 if (pjcpu->target_freq > max_freq)
567 max_freq = pjcpu->target_freq;
570 if (max_freq != pcpu->policy->cur)
571 __cpufreq_driver_target(pcpu->policy,
574 trace_cpufreq_interactive_setspeed(cpu,
578 up_read(&pcpu->enable_sem);
585 static void cpufreq_interactive_boost(void)
589 unsigned long flags[2];
590 struct cpufreq_interactive_cpuinfo *pcpu;
591 struct cpufreq_interactive_tunables *tunables;
593 spin_lock_irqsave(&speedchange_cpumask_lock, flags[0]);
595 for_each_online_cpu(i) {
596 pcpu = &per_cpu(cpuinfo, i);
597 tunables = pcpu->policy->governor_data;
599 spin_lock_irqsave(&pcpu->target_freq_lock, flags[1]);
600 if (pcpu->target_freq < tunables->hispeed_freq) {
601 pcpu->target_freq = tunables->hispeed_freq;
602 cpumask_set_cpu(i, &speedchange_cpumask);
603 pcpu->hispeed_validate_time =
604 ktime_to_us(ktime_get());
609 * Set floor freq and (re)start timer for when last
613 pcpu->floor_freq = tunables->hispeed_freq;
614 pcpu->floor_validate_time = 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 = strict_strtoul(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 = strict_strtoul(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 = strict_strtoul(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)
863 ret = strict_strtoul(buf, 0, &val);
866 tunables->timer_rate = val;
870 static ssize_t show_timer_slack(struct cpufreq_interactive_tunables *tunables,
873 return sprintf(buf, "%d\n", tunables->timer_slack_val);
876 static ssize_t store_timer_slack(struct cpufreq_interactive_tunables *tunables,
877 const char *buf, size_t count)
882 ret = kstrtol(buf, 10, &val);
886 tunables->timer_slack_val = val;
890 static ssize_t show_boost(struct cpufreq_interactive_tunables *tunables,
893 return sprintf(buf, "%d\n", tunables->boost_val);
896 static ssize_t store_boost(struct cpufreq_interactive_tunables *tunables,
897 const char *buf, size_t count)
902 ret = kstrtoul(buf, 0, &val);
906 tunables->boost_val = val;
908 if (tunables->boost_val) {
909 trace_cpufreq_interactive_boost("on");
910 cpufreq_interactive_boost();
912 tunables->boostpulse_endtime = ktime_to_us(ktime_get());
913 trace_cpufreq_interactive_unboost("off");
919 static ssize_t store_boostpulse(struct cpufreq_interactive_tunables *tunables,
920 const char *buf, size_t count)
925 ret = kstrtoul(buf, 0, &val);
929 tunables->boostpulse_endtime = ktime_to_us(ktime_get()) +
930 tunables->boostpulse_duration_val;
931 trace_cpufreq_interactive_boost("pulse");
932 cpufreq_interactive_boost();
936 static ssize_t show_boostpulse_duration(struct cpufreq_interactive_tunables
937 *tunables, char *buf)
939 return sprintf(buf, "%d\n", tunables->boostpulse_duration_val);
942 static ssize_t store_boostpulse_duration(struct cpufreq_interactive_tunables
943 *tunables, const char *buf, size_t count)
948 ret = kstrtoul(buf, 0, &val);
952 tunables->boostpulse_duration_val = val;
956 static ssize_t show_io_is_busy(struct cpufreq_interactive_tunables *tunables,
959 return sprintf(buf, "%u\n", tunables->io_is_busy);
962 static ssize_t store_io_is_busy(struct cpufreq_interactive_tunables *tunables,
963 const char *buf, size_t count)
968 ret = kstrtoul(buf, 0, &val);
971 tunables->io_is_busy = val;
976 * Create show/store routines
977 * - sys: One governor instance for complete SYSTEM
978 * - pol: One governor instance per struct cpufreq_policy
980 #define show_gov_pol_sys(file_name) \
981 static ssize_t show_##file_name##_gov_sys \
982 (struct kobject *kobj, struct attribute *attr, char *buf) \
984 return show_##file_name(common_tunables, buf); \
987 static ssize_t show_##file_name##_gov_pol \
988 (struct cpufreq_policy *policy, char *buf) \
990 return show_##file_name(policy->governor_data, buf); \
993 #define store_gov_pol_sys(file_name) \
994 static ssize_t store_##file_name##_gov_sys \
995 (struct kobject *kobj, struct attribute *attr, const char *buf, \
998 return store_##file_name(common_tunables, buf, count); \
1001 static ssize_t store_##file_name##_gov_pol \
1002 (struct cpufreq_policy *policy, const char *buf, size_t count) \
1004 return store_##file_name(policy->governor_data, buf, count); \
1007 #define show_store_gov_pol_sys(file_name) \
1008 show_gov_pol_sys(file_name); \
1009 store_gov_pol_sys(file_name)
1011 show_store_gov_pol_sys(target_loads);
1012 show_store_gov_pol_sys(above_hispeed_delay);
1013 show_store_gov_pol_sys(hispeed_freq);
1014 show_store_gov_pol_sys(go_hispeed_load);
1015 show_store_gov_pol_sys(min_sample_time);
1016 show_store_gov_pol_sys(timer_rate);
1017 show_store_gov_pol_sys(timer_slack);
1018 show_store_gov_pol_sys(boost);
1019 store_gov_pol_sys(boostpulse);
1020 show_store_gov_pol_sys(boostpulse_duration);
1021 show_store_gov_pol_sys(io_is_busy);
1023 #define gov_sys_attr_rw(_name) \
1024 static struct global_attr _name##_gov_sys = \
1025 __ATTR(_name, 0644, show_##_name##_gov_sys, store_##_name##_gov_sys)
1027 #define gov_pol_attr_rw(_name) \
1028 static struct freq_attr _name##_gov_pol = \
1029 __ATTR(_name, 0644, show_##_name##_gov_pol, store_##_name##_gov_pol)
1031 #define gov_sys_pol_attr_rw(_name) \
1032 gov_sys_attr_rw(_name); \
1033 gov_pol_attr_rw(_name)
1035 gov_sys_pol_attr_rw(target_loads);
1036 gov_sys_pol_attr_rw(above_hispeed_delay);
1037 gov_sys_pol_attr_rw(hispeed_freq);
1038 gov_sys_pol_attr_rw(go_hispeed_load);
1039 gov_sys_pol_attr_rw(min_sample_time);
1040 gov_sys_pol_attr_rw(timer_rate);
1041 gov_sys_pol_attr_rw(timer_slack);
1042 gov_sys_pol_attr_rw(boost);
1043 gov_sys_pol_attr_rw(boostpulse_duration);
1044 gov_sys_pol_attr_rw(io_is_busy);
1046 static struct global_attr boostpulse_gov_sys =
1047 __ATTR(boostpulse, 0200, NULL, store_boostpulse_gov_sys);
1049 static struct freq_attr boostpulse_gov_pol =
1050 __ATTR(boostpulse, 0200, NULL, store_boostpulse_gov_pol);
1052 /* One Governor instance for entire system */
1053 static struct attribute *interactive_attributes_gov_sys[] = {
1054 &target_loads_gov_sys.attr,
1055 &above_hispeed_delay_gov_sys.attr,
1056 &hispeed_freq_gov_sys.attr,
1057 &go_hispeed_load_gov_sys.attr,
1058 &min_sample_time_gov_sys.attr,
1059 &timer_rate_gov_sys.attr,
1060 &timer_slack_gov_sys.attr,
1061 &boost_gov_sys.attr,
1062 &boostpulse_gov_sys.attr,
1063 &boostpulse_duration_gov_sys.attr,
1064 &io_is_busy_gov_sys.attr,
1068 static struct attribute_group interactive_attr_group_gov_sys = {
1069 .attrs = interactive_attributes_gov_sys,
1070 .name = "interactive",
1073 /* Per policy governor instance */
1074 static struct attribute *interactive_attributes_gov_pol[] = {
1075 &target_loads_gov_pol.attr,
1076 &above_hispeed_delay_gov_pol.attr,
1077 &hispeed_freq_gov_pol.attr,
1078 &go_hispeed_load_gov_pol.attr,
1079 &min_sample_time_gov_pol.attr,
1080 &timer_rate_gov_pol.attr,
1081 &timer_slack_gov_pol.attr,
1082 &boost_gov_pol.attr,
1083 &boostpulse_gov_pol.attr,
1084 &boostpulse_duration_gov_pol.attr,
1085 &io_is_busy_gov_pol.attr,
1089 static struct attribute_group interactive_attr_group_gov_pol = {
1090 .attrs = interactive_attributes_gov_pol,
1091 .name = "interactive",
1094 static struct attribute_group *get_sysfs_attr(void)
1096 if (have_governor_per_policy())
1097 return &interactive_attr_group_gov_pol;
1099 return &interactive_attr_group_gov_sys;
1102 static int cpufreq_interactive_idle_notifier(struct notifier_block *nb,
1108 cpufreq_interactive_idle_start();
1111 cpufreq_interactive_idle_end();
1118 static struct notifier_block cpufreq_interactive_idle_nb = {
1119 .notifier_call = cpufreq_interactive_idle_notifier,
1122 static int cpufreq_governor_interactive(struct cpufreq_policy *policy,
1127 struct cpufreq_interactive_cpuinfo *pcpu;
1128 struct cpufreq_frequency_table *freq_table;
1129 struct cpufreq_interactive_tunables *tunables;
1130 unsigned long flags;
1132 if (have_governor_per_policy())
1133 tunables = policy->governor_data;
1135 tunables = common_tunables;
1137 WARN_ON(!tunables && (event != CPUFREQ_GOV_POLICY_INIT));
1140 case CPUFREQ_GOV_POLICY_INIT:
1141 if (have_governor_per_policy()) {
1143 } else if (tunables) {
1144 tunables->usage_count++;
1145 policy->governor_data = tunables;
1149 tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
1151 pr_err("%s: POLICY_INIT: kzalloc failed\n", __func__);
1155 tunables->usage_count = 1;
1156 tunables->above_hispeed_delay = default_above_hispeed_delay;
1157 tunables->nabove_hispeed_delay =
1158 ARRAY_SIZE(default_above_hispeed_delay);
1159 tunables->go_hispeed_load = DEFAULT_GO_HISPEED_LOAD;
1160 tunables->target_loads = default_target_loads;
1161 tunables->ntarget_loads = ARRAY_SIZE(default_target_loads);
1162 tunables->min_sample_time = DEFAULT_MIN_SAMPLE_TIME;
1163 tunables->timer_rate = DEFAULT_TIMER_RATE;
1164 tunables->boostpulse_duration_val = DEFAULT_MIN_SAMPLE_TIME;
1165 tunables->timer_slack_val = DEFAULT_TIMER_SLACK;
1167 spin_lock_init(&tunables->target_loads_lock);
1168 spin_lock_init(&tunables->above_hispeed_delay_lock);
1170 policy->governor_data = tunables;
1171 if (!have_governor_per_policy())
1172 common_tunables = tunables;
1174 rc = sysfs_create_group(get_governor_parent_kobj(policy),
1178 policy->governor_data = NULL;
1179 if (!have_governor_per_policy())
1180 common_tunables = NULL;
1184 if (!policy->governor->initialized) {
1185 idle_notifier_register(&cpufreq_interactive_idle_nb);
1186 cpufreq_register_notifier(&cpufreq_notifier_block,
1187 CPUFREQ_TRANSITION_NOTIFIER);
1192 case CPUFREQ_GOV_POLICY_EXIT:
1193 if (!--tunables->usage_count) {
1194 if (policy->governor->initialized == 1) {
1195 cpufreq_unregister_notifier(&cpufreq_notifier_block,
1196 CPUFREQ_TRANSITION_NOTIFIER);
1197 idle_notifier_unregister(&cpufreq_interactive_idle_nb);
1200 sysfs_remove_group(get_governor_parent_kobj(policy),
1203 common_tunables = NULL;
1206 policy->governor_data = NULL;
1209 case CPUFREQ_GOV_START:
1210 mutex_lock(&gov_lock);
1212 freq_table = cpufreq_frequency_get_table(policy->cpu);
1213 if (!tunables->hispeed_freq)
1214 tunables->hispeed_freq = policy->max;
1216 for_each_cpu(j, policy->cpus) {
1217 pcpu = &per_cpu(cpuinfo, j);
1218 pcpu->policy = policy;
1219 pcpu->target_freq = policy->cur;
1220 pcpu->freq_table = freq_table;
1221 pcpu->floor_freq = pcpu->target_freq;
1222 pcpu->floor_validate_time =
1223 ktime_to_us(ktime_get());
1224 pcpu->hispeed_validate_time =
1225 pcpu->floor_validate_time;
1226 pcpu->max_freq = policy->max;
1227 down_write(&pcpu->enable_sem);
1228 del_timer_sync(&pcpu->cpu_timer);
1229 del_timer_sync(&pcpu->cpu_slack_timer);
1230 cpufreq_interactive_timer_start(tunables, j);
1231 pcpu->governor_enabled = 1;
1232 up_write(&pcpu->enable_sem);
1235 mutex_unlock(&gov_lock);
1238 case CPUFREQ_GOV_STOP:
1239 mutex_lock(&gov_lock);
1240 for_each_cpu(j, policy->cpus) {
1241 pcpu = &per_cpu(cpuinfo, j);
1242 down_write(&pcpu->enable_sem);
1243 pcpu->governor_enabled = 0;
1244 del_timer_sync(&pcpu->cpu_timer);
1245 del_timer_sync(&pcpu->cpu_slack_timer);
1246 up_write(&pcpu->enable_sem);
1249 mutex_unlock(&gov_lock);
1252 case CPUFREQ_GOV_LIMITS:
1253 if (policy->max < policy->cur)
1254 __cpufreq_driver_target(policy,
1255 policy->max, CPUFREQ_RELATION_H);
1256 else if (policy->min > policy->cur)
1257 __cpufreq_driver_target(policy,
1258 policy->min, CPUFREQ_RELATION_L);
1259 for_each_cpu(j, policy->cpus) {
1260 pcpu = &per_cpu(cpuinfo, j);
1262 down_read(&pcpu->enable_sem);
1263 if (pcpu->governor_enabled == 0) {
1264 up_read(&pcpu->enable_sem);
1268 spin_lock_irqsave(&pcpu->target_freq_lock, flags);
1269 if (policy->max < pcpu->target_freq)
1270 pcpu->target_freq = policy->max;
1271 else if (policy->min > pcpu->target_freq)
1272 pcpu->target_freq = policy->min;
1274 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
1275 up_read(&pcpu->enable_sem);
1277 /* Reschedule timer only if policy->max is raised.
1278 * Delete the timers, else the timer callback may
1279 * return without re-arm the timer when failed
1280 * acquire the semaphore. This race may cause timer
1281 * stopped unexpectedly.
1284 if (policy->max > pcpu->max_freq) {
1285 down_write(&pcpu->enable_sem);
1286 del_timer_sync(&pcpu->cpu_timer);
1287 del_timer_sync(&pcpu->cpu_slack_timer);
1288 cpufreq_interactive_timer_start(tunables, j);
1289 up_write(&pcpu->enable_sem);
1292 pcpu->max_freq = policy->max;
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");