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;
45 u64 time_in_idle_timestamp;
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_resched(
115 struct cpufreq_interactive_cpuinfo *pcpu)
117 mod_timer_pinned(&pcpu->cpu_timer,
118 jiffies + usecs_to_jiffies(timer_rate));
120 get_cpu_idle_time_us(smp_processor_id(),
121 &pcpu->time_in_idle_timestamp);
124 static void cpufreq_interactive_timer(unsigned long data)
127 unsigned int delta_idle;
128 unsigned int delta_time;
130 int load_since_change;
131 struct cpufreq_interactive_cpuinfo *pcpu =
132 &per_cpu(cpuinfo, data);
134 unsigned int new_freq;
140 if (!pcpu->governor_enabled)
143 now_idle = get_cpu_idle_time_us(data, &now);
144 delta_idle = (unsigned int)(now_idle - pcpu->time_in_idle);
145 delta_time = (unsigned int)(now - pcpu->time_in_idle_timestamp);
148 * If timer ran less than 1ms after short-term sample started, retry.
150 if (delta_time < 1000)
153 if (delta_idle > delta_time)
156 cpu_load = 100 * (delta_time - delta_idle) / delta_time;
158 delta_idle = (unsigned int)(now_idle - pcpu->target_set_time_in_idle);
159 delta_time = (unsigned int)(now - pcpu->target_set_time);
161 if ((delta_time == 0) || (delta_idle > delta_time))
162 load_since_change = 0;
165 100 * (delta_time - delta_idle) / delta_time;
168 * Choose greater of short-term load (since last idle timer
169 * started or timer function re-armed itself) or long-term load
170 * (since last frequency change).
172 if (load_since_change > cpu_load)
173 cpu_load = load_since_change;
175 if (cpu_load >= go_hispeed_load || boost_val) {
176 if (pcpu->target_freq < hispeed_freq &&
177 hispeed_freq < pcpu->policy->max) {
178 new_freq = hispeed_freq;
180 new_freq = pcpu->policy->max * cpu_load / 100;
182 if (new_freq < hispeed_freq)
183 new_freq = hispeed_freq;
185 if (pcpu->target_freq == hispeed_freq &&
186 new_freq > hispeed_freq &&
187 now - pcpu->hispeed_validate_time
188 < above_hispeed_delay_val) {
189 trace_cpufreq_interactive_notyet(
190 data, cpu_load, pcpu->target_freq,
191 pcpu->policy->cur, new_freq);
196 new_freq = hispeed_freq * cpu_load / 100;
199 if (new_freq <= hispeed_freq)
200 pcpu->hispeed_validate_time = now;
202 if (cpufreq_frequency_table_target(pcpu->policy, pcpu->freq_table,
203 new_freq, CPUFREQ_RELATION_H,
205 pr_warn_once("timer %d: cpufreq_frequency_table_target error\n",
210 new_freq = pcpu->freq_table[index].frequency;
213 * Do not scale below floor_freq unless we have been at or above the
214 * floor frequency for the minimum sample time since last validated.
216 if (new_freq < pcpu->floor_freq) {
217 if (now - pcpu->floor_validate_time < min_sample_time) {
218 trace_cpufreq_interactive_notyet(
219 data, cpu_load, pcpu->target_freq,
220 pcpu->policy->cur, new_freq);
225 pcpu->floor_freq = new_freq;
226 pcpu->floor_validate_time = now;
228 if (pcpu->target_freq == new_freq) {
229 trace_cpufreq_interactive_already(
230 data, cpu_load, pcpu->target_freq,
231 pcpu->policy->cur, new_freq);
232 goto rearm_if_notmax;
235 trace_cpufreq_interactive_target(data, cpu_load, pcpu->target_freq,
236 pcpu->policy->cur, new_freq);
237 pcpu->target_set_time_in_idle = now_idle;
238 pcpu->target_set_time = now;
240 pcpu->target_freq = new_freq;
241 spin_lock_irqsave(&speedchange_cpumask_lock, flags);
242 cpumask_set_cpu(data, &speedchange_cpumask);
243 spin_unlock_irqrestore(&speedchange_cpumask_lock, flags);
244 wake_up_process(speedchange_task);
248 * Already set max speed and don't see a need to change that,
249 * wait until next idle to re-evaluate, don't need timer.
251 if (pcpu->target_freq == pcpu->policy->max)
255 if (!timer_pending(&pcpu->cpu_timer)) {
257 * If governing speed in idle and already at min, cancel the
258 * timer if that CPU goes idle. We don't need to re-evaluate
259 * speed until the next idle exit.
261 if (governidle && pcpu->target_freq == pcpu->policy->min)
262 pcpu->timer_idlecancel = 1;
264 cpufreq_interactive_timer_resched(pcpu);
271 static void cpufreq_interactive_idle_start(void)
273 struct cpufreq_interactive_cpuinfo *pcpu =
274 &per_cpu(cpuinfo, smp_processor_id());
277 if (!pcpu->governor_enabled)
280 pending = timer_pending(&pcpu->cpu_timer);
282 if (pcpu->target_freq != pcpu->policy->min) {
284 * Entering idle while not at lowest speed. On some
285 * platforms this can hold the other CPU(s) at that speed
286 * even though the CPU is idle. Set a timer to re-evaluate
287 * speed so this idle CPU doesn't hold the other CPUs above
288 * min indefinitely. This should probably be a quirk of
289 * the CPUFreq driver.
292 pcpu->timer_idlecancel = 0;
293 cpufreq_interactive_timer_resched(pcpu);
295 } else if (governidle) {
297 * If at min speed and entering idle after load has
298 * already been evaluated, and a timer has been set just in
299 * case the CPU suddenly goes busy, cancel that timer. The
300 * CPU didn't go busy; we'll recheck things upon idle exit.
302 if (pending && pcpu->timer_idlecancel) {
303 del_timer(&pcpu->cpu_timer);
304 pcpu->timer_idlecancel = 0;
310 static void cpufreq_interactive_idle_end(void)
312 struct cpufreq_interactive_cpuinfo *pcpu =
313 &per_cpu(cpuinfo, smp_processor_id());
315 if (!pcpu->governor_enabled)
318 /* Arm the timer for 1-2 ticks later if not already. */
319 if (!timer_pending(&pcpu->cpu_timer)) {
320 pcpu->timer_idlecancel = 0;
321 cpufreq_interactive_timer_resched(pcpu);
322 } else if (!governidle &&
323 time_after_eq(jiffies, pcpu->cpu_timer.expires)) {
324 del_timer(&pcpu->cpu_timer);
325 cpufreq_interactive_timer(smp_processor_id());
329 static int cpufreq_interactive_speedchange_task(void *data)
334 struct cpufreq_interactive_cpuinfo *pcpu;
337 set_current_state(TASK_INTERRUPTIBLE);
338 spin_lock_irqsave(&speedchange_cpumask_lock, flags);
340 if (cpumask_empty(&speedchange_cpumask)) {
341 spin_unlock_irqrestore(&speedchange_cpumask_lock,
345 if (kthread_should_stop())
348 spin_lock_irqsave(&speedchange_cpumask_lock, flags);
351 set_current_state(TASK_RUNNING);
352 tmp_mask = speedchange_cpumask;
353 cpumask_clear(&speedchange_cpumask);
354 spin_unlock_irqrestore(&speedchange_cpumask_lock, flags);
356 for_each_cpu(cpu, &tmp_mask) {
358 unsigned int max_freq = 0;
360 pcpu = &per_cpu(cpuinfo, cpu);
363 if (!pcpu->governor_enabled)
366 for_each_cpu(j, pcpu->policy->cpus) {
367 struct cpufreq_interactive_cpuinfo *pjcpu =
368 &per_cpu(cpuinfo, j);
370 if (pjcpu->target_freq > max_freq)
371 max_freq = pjcpu->target_freq;
374 if (max_freq != pcpu->policy->cur)
375 __cpufreq_driver_target(pcpu->policy,
378 trace_cpufreq_interactive_setspeed(cpu,
387 static void cpufreq_interactive_boost(void)
392 struct cpufreq_interactive_cpuinfo *pcpu;
394 spin_lock_irqsave(&speedchange_cpumask_lock, flags);
396 for_each_online_cpu(i) {
397 pcpu = &per_cpu(cpuinfo, i);
399 if (pcpu->target_freq < hispeed_freq) {
400 pcpu->target_freq = hispeed_freq;
401 cpumask_set_cpu(i, &speedchange_cpumask);
402 pcpu->target_set_time_in_idle =
403 get_cpu_idle_time_us(i, &pcpu->target_set_time);
404 pcpu->hispeed_validate_time = pcpu->target_set_time;
409 * Set floor freq and (re)start timer for when last
413 pcpu->floor_freq = hispeed_freq;
414 pcpu->floor_validate_time = ktime_to_us(ktime_get());
417 spin_unlock_irqrestore(&speedchange_cpumask_lock, flags);
420 wake_up_process(speedchange_task);
423 static ssize_t show_hispeed_freq(struct kobject *kobj,
424 struct attribute *attr, char *buf)
426 return sprintf(buf, "%u\n", hispeed_freq);
429 static ssize_t store_hispeed_freq(struct kobject *kobj,
430 struct attribute *attr, const char *buf,
434 long unsigned int val;
436 ret = strict_strtoul(buf, 0, &val);
443 static struct global_attr hispeed_freq_attr = __ATTR(hispeed_freq, 0644,
444 show_hispeed_freq, store_hispeed_freq);
447 static ssize_t show_go_hispeed_load(struct kobject *kobj,
448 struct attribute *attr, char *buf)
450 return sprintf(buf, "%lu\n", go_hispeed_load);
453 static ssize_t store_go_hispeed_load(struct kobject *kobj,
454 struct attribute *attr, const char *buf, size_t count)
459 ret = strict_strtoul(buf, 0, &val);
462 go_hispeed_load = val;
466 static struct global_attr go_hispeed_load_attr = __ATTR(go_hispeed_load, 0644,
467 show_go_hispeed_load, store_go_hispeed_load);
469 static ssize_t show_min_sample_time(struct kobject *kobj,
470 struct attribute *attr, char *buf)
472 return sprintf(buf, "%lu\n", min_sample_time);
475 static ssize_t store_min_sample_time(struct kobject *kobj,
476 struct attribute *attr, const char *buf, size_t count)
481 ret = strict_strtoul(buf, 0, &val);
484 min_sample_time = val;
488 static struct global_attr min_sample_time_attr = __ATTR(min_sample_time, 0644,
489 show_min_sample_time, store_min_sample_time);
491 static ssize_t show_above_hispeed_delay(struct kobject *kobj,
492 struct attribute *attr, char *buf)
494 return sprintf(buf, "%lu\n", above_hispeed_delay_val);
497 static ssize_t store_above_hispeed_delay(struct kobject *kobj,
498 struct attribute *attr,
499 const char *buf, size_t count)
504 ret = strict_strtoul(buf, 0, &val);
507 above_hispeed_delay_val = val;
511 define_one_global_rw(above_hispeed_delay);
513 static ssize_t show_timer_rate(struct kobject *kobj,
514 struct attribute *attr, char *buf)
516 return sprintf(buf, "%lu\n", timer_rate);
519 static ssize_t store_timer_rate(struct kobject *kobj,
520 struct attribute *attr, const char *buf, size_t count)
525 ret = strict_strtoul(buf, 0, &val);
532 static struct global_attr timer_rate_attr = __ATTR(timer_rate, 0644,
533 show_timer_rate, store_timer_rate);
535 static ssize_t show_boost(struct kobject *kobj, struct attribute *attr,
538 return sprintf(buf, "%d\n", boost_val);
541 static ssize_t store_boost(struct kobject *kobj, struct attribute *attr,
542 const char *buf, size_t count)
547 ret = kstrtoul(buf, 0, &val);
554 trace_cpufreq_interactive_boost("on");
555 cpufreq_interactive_boost();
557 trace_cpufreq_interactive_unboost("off");
563 define_one_global_rw(boost);
565 static ssize_t store_boostpulse(struct kobject *kobj, struct attribute *attr,
566 const char *buf, size_t count)
571 ret = kstrtoul(buf, 0, &val);
575 trace_cpufreq_interactive_boost("pulse");
576 cpufreq_interactive_boost();
580 static struct global_attr boostpulse =
581 __ATTR(boostpulse, 0200, NULL, store_boostpulse);
583 static struct attribute *interactive_attributes[] = {
584 &hispeed_freq_attr.attr,
585 &go_hispeed_load_attr.attr,
586 &above_hispeed_delay.attr,
587 &min_sample_time_attr.attr,
588 &timer_rate_attr.attr,
594 static struct attribute_group interactive_attr_group = {
595 .attrs = interactive_attributes,
596 .name = "interactive",
599 static int cpufreq_interactive_idle_notifier(struct notifier_block *nb,
605 cpufreq_interactive_idle_start();
608 cpufreq_interactive_idle_end();
615 static struct notifier_block cpufreq_interactive_idle_nb = {
616 .notifier_call = cpufreq_interactive_idle_notifier,
619 static int cpufreq_governor_interactive(struct cpufreq_policy *policy,
624 struct cpufreq_interactive_cpuinfo *pcpu;
625 struct cpufreq_frequency_table *freq_table;
628 case CPUFREQ_GOV_START:
629 if (!cpu_online(policy->cpu))
633 cpufreq_frequency_get_table(policy->cpu);
635 hispeed_freq = policy->max;
637 for_each_cpu(j, policy->cpus) {
638 pcpu = &per_cpu(cpuinfo, j);
639 pcpu->policy = policy;
640 pcpu->target_freq = policy->cur;
641 pcpu->freq_table = freq_table;
642 pcpu->target_set_time_in_idle =
643 get_cpu_idle_time_us(j,
644 &pcpu->target_set_time);
645 pcpu->floor_freq = pcpu->target_freq;
646 pcpu->floor_validate_time =
647 pcpu->target_set_time;
648 pcpu->hispeed_validate_time =
649 pcpu->target_set_time;
650 pcpu->governor_enabled = 1;
652 pcpu->cpu_timer.expires =
653 jiffies + usecs_to_jiffies(timer_rate);
654 add_timer_on(&pcpu->cpu_timer, j);
658 * Do not register the idle hook and create sysfs
659 * entries if we have already done so.
661 if (atomic_inc_return(&active_count) > 1)
664 rc = sysfs_create_group(cpufreq_global_kobject,
665 &interactive_attr_group);
669 idle_notifier_register(&cpufreq_interactive_idle_nb);
672 case CPUFREQ_GOV_STOP:
673 for_each_cpu(j, policy->cpus) {
674 pcpu = &per_cpu(cpuinfo, j);
675 pcpu->governor_enabled = 0;
677 del_timer_sync(&pcpu->cpu_timer);
680 if (atomic_dec_return(&active_count) > 0)
683 idle_notifier_unregister(&cpufreq_interactive_idle_nb);
684 sysfs_remove_group(cpufreq_global_kobject,
685 &interactive_attr_group);
689 case CPUFREQ_GOV_LIMITS:
690 if (policy->max < policy->cur)
691 __cpufreq_driver_target(policy,
692 policy->max, CPUFREQ_RELATION_H);
693 else if (policy->min > policy->cur)
694 __cpufreq_driver_target(policy,
695 policy->min, CPUFREQ_RELATION_L);
701 static int __init cpufreq_interactive_init(void)
704 struct cpufreq_interactive_cpuinfo *pcpu;
705 struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
707 go_hispeed_load = DEFAULT_GO_HISPEED_LOAD;
708 min_sample_time = DEFAULT_MIN_SAMPLE_TIME;
709 above_hispeed_delay_val = DEFAULT_ABOVE_HISPEED_DELAY;
710 timer_rate = DEFAULT_TIMER_RATE;
712 /* Initalize per-cpu timers */
713 for_each_possible_cpu(i) {
714 pcpu = &per_cpu(cpuinfo, i);
716 init_timer(&pcpu->cpu_timer);
718 init_timer_deferrable(&pcpu->cpu_timer);
719 pcpu->cpu_timer.function = cpufreq_interactive_timer;
720 pcpu->cpu_timer.data = i;
723 spin_lock_init(&speedchange_cpumask_lock);
725 kthread_create(cpufreq_interactive_speedchange_task, NULL,
727 if (IS_ERR(speedchange_task))
728 return PTR_ERR(speedchange_task);
730 sched_setscheduler_nocheck(speedchange_task, SCHED_FIFO, ¶m);
731 get_task_struct(speedchange_task);
733 /* NB: wake up so the thread does not look hung to the freezer */
734 wake_up_process(speedchange_task);
736 return cpufreq_register_governor(&cpufreq_gov_interactive);
739 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_INTERACTIVE
740 fs_initcall(cpufreq_interactive_init);
742 module_init(cpufreq_interactive_init);
745 static void __exit cpufreq_interactive_exit(void)
747 cpufreq_unregister_governor(&cpufreq_gov_interactive);
748 kthread_stop(speedchange_task);
749 put_task_struct(speedchange_task);
752 module_exit(cpufreq_interactive_exit);
754 MODULE_AUTHOR("Mike Chan <mike@android.com>");
755 MODULE_DESCRIPTION("'cpufreq_interactive' - A cpufreq governor for "
756 "Latency sensitive workloads");
757 MODULE_LICENSE("GPL");