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/mutex.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/mutex.h>
33 #include <linux/slab.h>
34 #include <asm/cputime.h>
36 #define CREATE_TRACE_POINTS
37 #include <trace/events/cpufreq_interactive.h>
39 static atomic_t active_count = ATOMIC_INIT(0);
41 struct cpufreq_interactive_cpuinfo {
42 struct timer_list cpu_timer;
47 u64 target_set_time_in_idle;
48 struct cpufreq_policy *policy;
49 struct cpufreq_frequency_table *freq_table;
50 unsigned int target_freq;
51 unsigned int floor_freq;
52 u64 floor_validate_time;
53 u64 hispeed_validate_time;
57 static DEFINE_PER_CPU(struct cpufreq_interactive_cpuinfo, cpuinfo);
59 /* realtime thread handles frequency scaling */
60 static struct task_struct *speedchange_task;
61 static cpumask_t speedchange_cpumask;
62 static spinlock_t speedchange_cpumask_lock;
64 /* Hi speed to bump to from lo speed when load burst (default max) */
65 static unsigned int hispeed_freq;
67 /* Go to hi speed when CPU load at or above this value. */
68 #define DEFAULT_GO_HISPEED_LOAD 85
69 static unsigned long go_hispeed_load;
72 * The minimum amount of time to spend at a frequency before we can ramp down.
74 #define DEFAULT_MIN_SAMPLE_TIME (80 * USEC_PER_MSEC)
75 static unsigned long min_sample_time;
78 * The sample rate of the timer used to increase frequency
80 #define DEFAULT_TIMER_RATE (20 * USEC_PER_MSEC)
81 static unsigned long timer_rate;
84 * Wait this long before raising speed above hispeed, by default a single
87 #define DEFAULT_ABOVE_HISPEED_DELAY DEFAULT_TIMER_RATE
88 static unsigned long above_hispeed_delay_val;
91 * Non-zero means longer-term speed boost active.
96 static bool governidle;
97 module_param(governidle, bool, S_IWUSR | S_IRUGO);
98 MODULE_PARM_DESC(governidle,
99 "Set to 1 to wake up CPUs from idle to reduce speed (default 0)");
101 static int cpufreq_governor_interactive(struct cpufreq_policy *policy,
104 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_INTERACTIVE
107 struct cpufreq_governor cpufreq_gov_interactive = {
108 .name = "interactive",
109 .governor = cpufreq_governor_interactive,
110 .max_transition_latency = 10000000,
111 .owner = THIS_MODULE,
114 static void cpufreq_interactive_timer(unsigned long data)
117 unsigned int delta_idle;
118 unsigned int delta_time;
120 int load_since_change;
123 struct cpufreq_interactive_cpuinfo *pcpu =
124 &per_cpu(cpuinfo, data);
126 unsigned int new_freq;
132 if (!pcpu->governor_enabled)
135 time_in_idle = pcpu->time_in_idle;
136 idle_exit_time = pcpu->idle_exit_time;
137 now_idle = get_cpu_idle_time_us(data, &now);
138 delta_idle = (unsigned int)(now_idle - time_in_idle);
139 delta_time = (unsigned int)(now - idle_exit_time);
142 * If timer ran less than 1ms after short-term sample started, retry.
144 if (delta_time < 1000)
147 if (delta_idle > delta_time)
150 cpu_load = 100 * (delta_time - delta_idle) / delta_time;
152 delta_idle = (unsigned int)(now_idle - pcpu->target_set_time_in_idle);
153 delta_time = (unsigned int)(now - pcpu->target_set_time);
155 if ((delta_time == 0) || (delta_idle > delta_time))
156 load_since_change = 0;
159 100 * (delta_time - delta_idle) / delta_time;
162 * Choose greater of short-term load (since last idle timer
163 * started or timer function re-armed itself) or long-term load
164 * (since last frequency change).
166 if (load_since_change > cpu_load)
167 cpu_load = load_since_change;
169 if (cpu_load >= go_hispeed_load || boost_val) {
170 if (pcpu->target_freq < hispeed_freq &&
171 hispeed_freq < pcpu->policy->max) {
172 new_freq = hispeed_freq;
174 new_freq = pcpu->policy->max * cpu_load / 100;
176 if (new_freq < hispeed_freq)
177 new_freq = hispeed_freq;
179 if (pcpu->target_freq == hispeed_freq &&
180 new_freq > hispeed_freq &&
181 now - pcpu->hispeed_validate_time
182 < above_hispeed_delay_val) {
183 trace_cpufreq_interactive_notyet(data, cpu_load,
190 new_freq = hispeed_freq * cpu_load / 100;
193 if (new_freq <= hispeed_freq)
194 pcpu->hispeed_validate_time = now;
196 if (cpufreq_frequency_table_target(pcpu->policy, pcpu->freq_table,
197 new_freq, CPUFREQ_RELATION_H,
199 pr_warn_once("timer %d: cpufreq_frequency_table_target error\n",
204 new_freq = pcpu->freq_table[index].frequency;
207 * Do not scale below floor_freq unless we have been at or above the
208 * floor frequency for the minimum sample time since last validated.
210 if (new_freq < pcpu->floor_freq) {
211 if (now - pcpu->floor_validate_time < min_sample_time) {
212 trace_cpufreq_interactive_notyet(data, cpu_load,
213 pcpu->target_freq, new_freq);
218 pcpu->floor_freq = new_freq;
219 pcpu->floor_validate_time = now;
221 if (pcpu->target_freq == new_freq) {
222 trace_cpufreq_interactive_already(data, cpu_load,
223 pcpu->target_freq, new_freq);
224 goto rearm_if_notmax;
227 trace_cpufreq_interactive_target(data, cpu_load, pcpu->target_freq,
229 pcpu->target_set_time_in_idle = now_idle;
230 pcpu->target_set_time = now;
232 pcpu->target_freq = new_freq;
233 spin_lock_irqsave(&speedchange_cpumask_lock, flags);
234 cpumask_set_cpu(data, &speedchange_cpumask);
235 spin_unlock_irqrestore(&speedchange_cpumask_lock, flags);
236 wake_up_process(speedchange_task);
240 * Already set max speed and don't see a need to change that,
241 * wait until next idle to re-evaluate, don't need timer.
243 if (pcpu->target_freq == pcpu->policy->max)
247 if (!timer_pending(&pcpu->cpu_timer)) {
249 * If governing speed in idle and already at min, cancel the
250 * timer if that CPU goes idle. We don't need to re-evaluate
251 * speed until the next idle exit.
253 if (governidle && pcpu->target_freq == pcpu->policy->min)
254 pcpu->timer_idlecancel = 1;
256 pcpu->time_in_idle = get_cpu_idle_time_us(
257 data, &pcpu->idle_exit_time);
258 mod_timer_pinned(&pcpu->cpu_timer,
259 jiffies + usecs_to_jiffies(timer_rate));
266 static void cpufreq_interactive_idle_start(void)
268 struct cpufreq_interactive_cpuinfo *pcpu =
269 &per_cpu(cpuinfo, smp_processor_id());
272 if (!pcpu->governor_enabled)
275 pending = timer_pending(&pcpu->cpu_timer);
277 if (pcpu->target_freq != pcpu->policy->min) {
279 * Entering idle while not at lowest speed. On some
280 * platforms this can hold the other CPU(s) at that speed
281 * even though the CPU is idle. Set a timer to re-evaluate
282 * speed so this idle CPU doesn't hold the other CPUs above
283 * min indefinitely. This should probably be a quirk of
284 * the CPUFreq driver.
287 pcpu->time_in_idle = get_cpu_idle_time_us(
288 smp_processor_id(), &pcpu->idle_exit_time);
289 pcpu->timer_idlecancel = 0;
292 jiffies + usecs_to_jiffies(timer_rate));
294 } else if (governidle) {
296 * If at min speed and entering idle after load has
297 * already been evaluated, and a timer has been set just in
298 * case the CPU suddenly goes busy, cancel that timer. The
299 * CPU didn't go busy; we'll recheck things upon idle exit.
301 if (pending && pcpu->timer_idlecancel) {
302 del_timer(&pcpu->cpu_timer);
303 pcpu->timer_idlecancel = 0;
309 static void cpufreq_interactive_idle_end(void)
311 struct cpufreq_interactive_cpuinfo *pcpu =
312 &per_cpu(cpuinfo, smp_processor_id());
314 if (!pcpu->governor_enabled)
317 /* Arm the timer for 1-2 ticks later if not already. */
318 if (!timer_pending(&pcpu->cpu_timer)) {
320 get_cpu_idle_time_us(smp_processor_id(),
321 &pcpu->idle_exit_time);
322 pcpu->timer_idlecancel = 0;
325 jiffies + usecs_to_jiffies(timer_rate));
330 static int cpufreq_interactive_speedchange_task(void *data)
335 struct cpufreq_interactive_cpuinfo *pcpu;
338 set_current_state(TASK_INTERRUPTIBLE);
339 spin_lock_irqsave(&speedchange_cpumask_lock, flags);
341 if (cpumask_empty(&speedchange_cpumask)) {
342 spin_unlock_irqrestore(&speedchange_cpumask_lock,
346 if (kthread_should_stop())
349 spin_lock_irqsave(&speedchange_cpumask_lock, flags);
352 set_current_state(TASK_RUNNING);
353 tmp_mask = speedchange_cpumask;
354 cpumask_clear(&speedchange_cpumask);
355 spin_unlock_irqrestore(&speedchange_cpumask_lock, flags);
357 for_each_cpu(cpu, &tmp_mask) {
359 unsigned int max_freq = 0;
361 pcpu = &per_cpu(cpuinfo, cpu);
364 if (!pcpu->governor_enabled)
367 for_each_cpu(j, pcpu->policy->cpus) {
368 struct cpufreq_interactive_cpuinfo *pjcpu =
369 &per_cpu(cpuinfo, j);
371 if (pjcpu->target_freq > max_freq)
372 max_freq = pjcpu->target_freq;
375 if (max_freq != pcpu->policy->cur)
376 __cpufreq_driver_target(pcpu->policy,
379 trace_cpufreq_interactive_setspeed(cpu,
388 static void cpufreq_interactive_boost(void)
393 struct cpufreq_interactive_cpuinfo *pcpu;
395 spin_lock_irqsave(&speedchange_cpumask_lock, flags);
397 for_each_online_cpu(i) {
398 pcpu = &per_cpu(cpuinfo, i);
400 if (pcpu->target_freq < hispeed_freq) {
401 pcpu->target_freq = hispeed_freq;
402 cpumask_set_cpu(i, &speedchange_cpumask);
403 pcpu->target_set_time_in_idle =
404 get_cpu_idle_time_us(i, &pcpu->target_set_time);
405 pcpu->hispeed_validate_time = pcpu->target_set_time;
410 * Set floor freq and (re)start timer for when last
414 pcpu->floor_freq = hispeed_freq;
415 pcpu->floor_validate_time = ktime_to_us(ktime_get());
418 spin_unlock_irqrestore(&speedchange_cpumask_lock, flags);
421 wake_up_process(speedchange_task);
424 static ssize_t show_hispeed_freq(struct kobject *kobj,
425 struct attribute *attr, char *buf)
427 return sprintf(buf, "%u\n", hispeed_freq);
430 static ssize_t store_hispeed_freq(struct kobject *kobj,
431 struct attribute *attr, const char *buf,
435 long unsigned int val;
437 ret = strict_strtoul(buf, 0, &val);
444 static struct global_attr hispeed_freq_attr = __ATTR(hispeed_freq, 0644,
445 show_hispeed_freq, store_hispeed_freq);
448 static ssize_t show_go_hispeed_load(struct kobject *kobj,
449 struct attribute *attr, char *buf)
451 return sprintf(buf, "%lu\n", go_hispeed_load);
454 static ssize_t store_go_hispeed_load(struct kobject *kobj,
455 struct attribute *attr, const char *buf, size_t count)
460 ret = strict_strtoul(buf, 0, &val);
463 go_hispeed_load = val;
467 static struct global_attr go_hispeed_load_attr = __ATTR(go_hispeed_load, 0644,
468 show_go_hispeed_load, store_go_hispeed_load);
470 static ssize_t show_min_sample_time(struct kobject *kobj,
471 struct attribute *attr, char *buf)
473 return sprintf(buf, "%lu\n", min_sample_time);
476 static ssize_t store_min_sample_time(struct kobject *kobj,
477 struct attribute *attr, const char *buf, size_t count)
482 ret = strict_strtoul(buf, 0, &val);
485 min_sample_time = val;
489 static struct global_attr min_sample_time_attr = __ATTR(min_sample_time, 0644,
490 show_min_sample_time, store_min_sample_time);
492 static ssize_t show_above_hispeed_delay(struct kobject *kobj,
493 struct attribute *attr, char *buf)
495 return sprintf(buf, "%lu\n", above_hispeed_delay_val);
498 static ssize_t store_above_hispeed_delay(struct kobject *kobj,
499 struct attribute *attr,
500 const char *buf, size_t count)
505 ret = strict_strtoul(buf, 0, &val);
508 above_hispeed_delay_val = val;
512 define_one_global_rw(above_hispeed_delay);
514 static ssize_t show_timer_rate(struct kobject *kobj,
515 struct attribute *attr, char *buf)
517 return sprintf(buf, "%lu\n", timer_rate);
520 static ssize_t store_timer_rate(struct kobject *kobj,
521 struct attribute *attr, const char *buf, size_t count)
526 ret = strict_strtoul(buf, 0, &val);
533 static struct global_attr timer_rate_attr = __ATTR(timer_rate, 0644,
534 show_timer_rate, store_timer_rate);
536 static ssize_t show_boost(struct kobject *kobj, struct attribute *attr,
539 return sprintf(buf, "%d\n", boost_val);
542 static ssize_t store_boost(struct kobject *kobj, struct attribute *attr,
543 const char *buf, size_t count)
548 ret = kstrtoul(buf, 0, &val);
555 trace_cpufreq_interactive_boost("on");
556 cpufreq_interactive_boost();
558 trace_cpufreq_interactive_unboost("off");
564 define_one_global_rw(boost);
566 static ssize_t store_boostpulse(struct kobject *kobj, struct attribute *attr,
567 const char *buf, size_t count)
572 ret = kstrtoul(buf, 0, &val);
576 trace_cpufreq_interactive_boost("pulse");
577 cpufreq_interactive_boost();
581 static struct global_attr boostpulse =
582 __ATTR(boostpulse, 0200, NULL, store_boostpulse);
584 static struct attribute *interactive_attributes[] = {
585 &hispeed_freq_attr.attr,
586 &go_hispeed_load_attr.attr,
587 &above_hispeed_delay.attr,
588 &min_sample_time_attr.attr,
589 &timer_rate_attr.attr,
595 static struct attribute_group interactive_attr_group = {
596 .attrs = interactive_attributes,
597 .name = "interactive",
600 static int cpufreq_interactive_idle_notifier(struct notifier_block *nb,
606 cpufreq_interactive_idle_start();
609 cpufreq_interactive_idle_end();
616 static struct notifier_block cpufreq_interactive_idle_nb = {
617 .notifier_call = cpufreq_interactive_idle_notifier,
620 static int cpufreq_governor_interactive(struct cpufreq_policy *policy,
625 struct cpufreq_interactive_cpuinfo *pcpu;
626 struct cpufreq_frequency_table *freq_table;
629 case CPUFREQ_GOV_START:
630 if (!cpu_online(policy->cpu))
634 cpufreq_frequency_get_table(policy->cpu);
636 hispeed_freq = policy->max;
638 for_each_cpu(j, policy->cpus) {
639 pcpu = &per_cpu(cpuinfo, j);
640 pcpu->policy = policy;
641 pcpu->target_freq = policy->cur;
642 pcpu->freq_table = freq_table;
643 pcpu->target_set_time_in_idle =
644 get_cpu_idle_time_us(j,
645 &pcpu->target_set_time);
646 pcpu->floor_freq = pcpu->target_freq;
647 pcpu->floor_validate_time =
648 pcpu->target_set_time;
649 pcpu->hispeed_validate_time =
650 pcpu->target_set_time;
651 pcpu->governor_enabled = 1;
653 pcpu->cpu_timer.expires =
654 jiffies + usecs_to_jiffies(timer_rate);
655 add_timer_on(&pcpu->cpu_timer, j);
659 * Do not register the idle hook and create sysfs
660 * entries if we have already done so.
662 if (atomic_inc_return(&active_count) > 1)
665 rc = sysfs_create_group(cpufreq_global_kobject,
666 &interactive_attr_group);
670 idle_notifier_register(&cpufreq_interactive_idle_nb);
673 case CPUFREQ_GOV_STOP:
674 for_each_cpu(j, policy->cpus) {
675 pcpu = &per_cpu(cpuinfo, j);
676 pcpu->governor_enabled = 0;
678 del_timer_sync(&pcpu->cpu_timer);
681 if (atomic_dec_return(&active_count) > 0)
684 idle_notifier_unregister(&cpufreq_interactive_idle_nb);
685 sysfs_remove_group(cpufreq_global_kobject,
686 &interactive_attr_group);
690 case CPUFREQ_GOV_LIMITS:
691 if (policy->max < policy->cur)
692 __cpufreq_driver_target(policy,
693 policy->max, CPUFREQ_RELATION_H);
694 else if (policy->min > policy->cur)
695 __cpufreq_driver_target(policy,
696 policy->min, CPUFREQ_RELATION_L);
702 static int __init cpufreq_interactive_init(void)
705 struct cpufreq_interactive_cpuinfo *pcpu;
706 struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
708 go_hispeed_load = DEFAULT_GO_HISPEED_LOAD;
709 min_sample_time = DEFAULT_MIN_SAMPLE_TIME;
710 above_hispeed_delay_val = DEFAULT_ABOVE_HISPEED_DELAY;
711 timer_rate = DEFAULT_TIMER_RATE;
713 /* Initalize per-cpu timers */
714 for_each_possible_cpu(i) {
715 pcpu = &per_cpu(cpuinfo, i);
717 init_timer(&pcpu->cpu_timer);
719 init_timer_deferrable(&pcpu->cpu_timer);
720 pcpu->cpu_timer.function = cpufreq_interactive_timer;
721 pcpu->cpu_timer.data = i;
724 spin_lock_init(&speedchange_cpumask_lock);
726 kthread_create(cpufreq_interactive_speedchange_task, NULL,
728 if (IS_ERR(speedchange_task))
729 return PTR_ERR(speedchange_task);
731 sched_setscheduler_nocheck(speedchange_task, SCHED_FIFO, ¶m);
732 get_task_struct(speedchange_task);
734 /* NB: wake up so the thread does not look hung to the freezer */
735 wake_up_process(speedchange_task);
737 return cpufreq_register_governor(&cpufreq_gov_interactive);
740 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_INTERACTIVE
741 fs_initcall(cpufreq_interactive_init);
743 module_init(cpufreq_interactive_init);
746 static void __exit cpufreq_interactive_exit(void)
748 cpufreq_unregister_governor(&cpufreq_gov_interactive);
749 kthread_stop(speedchange_task);
750 put_task_struct(speedchange_task);
753 module_exit(cpufreq_interactive_exit);
755 MODULE_AUTHOR("Mike Chan <mike@android.com>");
756 MODULE_DESCRIPTION("'cpufreq_interactive' - A cpufreq governor for "
757 "Latency sensitive workloads");
758 MODULE_LICENSE("GPL");