Loading...
1/*
2 * cpuidle.c - core cpuidle infrastructure
3 *
4 * (C) 2006-2007 Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
5 * Shaohua Li <shaohua.li@intel.com>
6 * Adam Belay <abelay@novell.com>
7 *
8 * This code is licenced under the GPL.
9 */
10
11#include "linux/percpu-defs.h"
12#include <linux/clockchips.h>
13#include <linux/kernel.h>
14#include <linux/mutex.h>
15#include <linux/sched.h>
16#include <linux/sched/clock.h>
17#include <linux/sched/idle.h>
18#include <linux/notifier.h>
19#include <linux/pm_qos.h>
20#include <linux/cpu.h>
21#include <linux/cpuidle.h>
22#include <linux/ktime.h>
23#include <linux/hrtimer.h>
24#include <linux/module.h>
25#include <linux/suspend.h>
26#include <linux/tick.h>
27#include <linux/mmu_context.h>
28#include <linux/context_tracking.h>
29#include <trace/events/power.h>
30
31#include "cpuidle.h"
32
33DEFINE_PER_CPU(struct cpuidle_device *, cpuidle_devices);
34DEFINE_PER_CPU(struct cpuidle_device, cpuidle_dev);
35
36DEFINE_MUTEX(cpuidle_lock);
37LIST_HEAD(cpuidle_detected_devices);
38
39static int enabled_devices;
40static int off __read_mostly;
41static int initialized __read_mostly;
42
43int cpuidle_disabled(void)
44{
45 return off;
46}
47void disable_cpuidle(void)
48{
49 off = 1;
50}
51
52bool cpuidle_not_available(struct cpuidle_driver *drv,
53 struct cpuidle_device *dev)
54{
55 return off || !initialized || !drv || !dev || !dev->enabled;
56}
57
58/**
59 * cpuidle_play_dead - cpu off-lining
60 *
61 * Returns in case of an error or no driver
62 */
63int cpuidle_play_dead(void)
64{
65 struct cpuidle_device *dev = __this_cpu_read(cpuidle_devices);
66 struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);
67 int i;
68
69 if (!drv)
70 return -ENODEV;
71
72 for (i = drv->state_count - 1; i >= 0; i--) {
73 if (drv->states[i].enter_dead)
74 drv->states[i].enter_dead(dev, i);
75 }
76
77 /*
78 * If :enter_dead() is successful, it will never return, so reaching
79 * here means that all of them failed above or were not present.
80 */
81 return -ENODEV;
82}
83
84static int find_deepest_state(struct cpuidle_driver *drv,
85 struct cpuidle_device *dev,
86 u64 max_latency_ns,
87 unsigned int forbidden_flags,
88 bool s2idle)
89{
90 u64 latency_req = 0;
91 int i, ret = 0;
92
93 for (i = 1; i < drv->state_count; i++) {
94 struct cpuidle_state *s = &drv->states[i];
95
96 if (dev->states_usage[i].disable ||
97 s->exit_latency_ns <= latency_req ||
98 s->exit_latency_ns > max_latency_ns ||
99 (s->flags & forbidden_flags) ||
100 (s2idle && !s->enter_s2idle))
101 continue;
102
103 latency_req = s->exit_latency_ns;
104 ret = i;
105 }
106 return ret;
107}
108
109/**
110 * cpuidle_use_deepest_state - Set/unset governor override mode.
111 * @latency_limit_ns: Idle state exit latency limit (or no override if 0).
112 *
113 * If @latency_limit_ns is nonzero, set the current CPU to use the deepest idle
114 * state with exit latency within @latency_limit_ns (override governors going
115 * forward), or do not override governors if it is zero.
116 */
117void cpuidle_use_deepest_state(u64 latency_limit_ns)
118{
119 struct cpuidle_device *dev;
120
121 preempt_disable();
122 dev = cpuidle_get_device();
123 if (dev)
124 dev->forced_idle_latency_limit_ns = latency_limit_ns;
125 preempt_enable();
126}
127
128/**
129 * cpuidle_find_deepest_state - Find the deepest available idle state.
130 * @drv: cpuidle driver for the given CPU.
131 * @dev: cpuidle device for the given CPU.
132 * @latency_limit_ns: Idle state exit latency limit
133 *
134 * Return: the index of the deepest available idle state.
135 */
136int cpuidle_find_deepest_state(struct cpuidle_driver *drv,
137 struct cpuidle_device *dev,
138 u64 latency_limit_ns)
139{
140 return find_deepest_state(drv, dev, latency_limit_ns, 0, false);
141}
142
143#ifdef CONFIG_SUSPEND
144static noinstr void enter_s2idle_proper(struct cpuidle_driver *drv,
145 struct cpuidle_device *dev, int index)
146{
147 struct cpuidle_state *target_state = &drv->states[index];
148 ktime_t time_start, time_end;
149
150 instrumentation_begin();
151
152 time_start = ns_to_ktime(local_clock_noinstr());
153
154 tick_freeze();
155 /*
156 * The state used here cannot be a "coupled" one, because the "coupled"
157 * cpuidle mechanism enables interrupts and doing that with timekeeping
158 * suspended is generally unsafe.
159 */
160 stop_critical_timings();
161 if (!(target_state->flags & CPUIDLE_FLAG_RCU_IDLE)) {
162 ct_cpuidle_enter();
163 /* Annotate away the indirect call */
164 instrumentation_begin();
165 }
166 target_state->enter_s2idle(dev, drv, index);
167 if (WARN_ON_ONCE(!irqs_disabled()))
168 raw_local_irq_disable();
169 if (!(target_state->flags & CPUIDLE_FLAG_RCU_IDLE)) {
170 instrumentation_end();
171 ct_cpuidle_exit();
172 }
173 tick_unfreeze();
174 start_critical_timings();
175
176 time_end = ns_to_ktime(local_clock_noinstr());
177
178 dev->states_usage[index].s2idle_time += ktime_us_delta(time_end, time_start);
179 dev->states_usage[index].s2idle_usage++;
180 instrumentation_end();
181}
182
183/**
184 * cpuidle_enter_s2idle - Enter an idle state suitable for suspend-to-idle.
185 * @drv: cpuidle driver for the given CPU.
186 * @dev: cpuidle device for the given CPU.
187 *
188 * If there are states with the ->enter_s2idle callback, find the deepest of
189 * them and enter it with frozen tick.
190 */
191int cpuidle_enter_s2idle(struct cpuidle_driver *drv, struct cpuidle_device *dev)
192{
193 int index;
194
195 /*
196 * Find the deepest state with ->enter_s2idle present, which guarantees
197 * that interrupts won't be enabled when it exits and allows the tick to
198 * be frozen safely.
199 */
200 index = find_deepest_state(drv, dev, U64_MAX, 0, true);
201 if (index > 0) {
202 enter_s2idle_proper(drv, dev, index);
203 local_irq_enable();
204 }
205 return index;
206}
207#endif /* CONFIG_SUSPEND */
208
209/**
210 * cpuidle_enter_state - enter the state and update stats
211 * @dev: cpuidle device for this cpu
212 * @drv: cpuidle driver for this cpu
213 * @index: index into the states table in @drv of the state to enter
214 */
215noinstr int cpuidle_enter_state(struct cpuidle_device *dev,
216 struct cpuidle_driver *drv,
217 int index)
218{
219 int entered_state;
220
221 struct cpuidle_state *target_state = &drv->states[index];
222 bool broadcast = !!(target_state->flags & CPUIDLE_FLAG_TIMER_STOP);
223 ktime_t time_start, time_end;
224
225 instrumentation_begin();
226
227 /*
228 * Tell the time framework to switch to a broadcast timer because our
229 * local timer will be shut down. If a local timer is used from another
230 * CPU as a broadcast timer, this call may fail if it is not available.
231 */
232 if (broadcast && tick_broadcast_enter()) {
233 index = find_deepest_state(drv, dev, target_state->exit_latency_ns,
234 CPUIDLE_FLAG_TIMER_STOP, false);
235
236 target_state = &drv->states[index];
237 broadcast = false;
238 }
239
240 if (target_state->flags & CPUIDLE_FLAG_TLB_FLUSHED)
241 leave_mm();
242
243 /* Take note of the planned idle state. */
244 sched_idle_set_state(target_state);
245
246 trace_cpu_idle(index, dev->cpu);
247 time_start = ns_to_ktime(local_clock_noinstr());
248
249 stop_critical_timings();
250 if (!(target_state->flags & CPUIDLE_FLAG_RCU_IDLE)) {
251 ct_cpuidle_enter();
252 /* Annotate away the indirect call */
253 instrumentation_begin();
254 }
255
256 /*
257 * NOTE!!
258 *
259 * For cpuidle_state::enter() methods that do *NOT* set
260 * CPUIDLE_FLAG_RCU_IDLE RCU will be disabled here and these functions
261 * must be marked either noinstr or __cpuidle.
262 *
263 * For cpuidle_state::enter() methods that *DO* set
264 * CPUIDLE_FLAG_RCU_IDLE this isn't required, but they must mark the
265 * function calling ct_cpuidle_enter() as noinstr/__cpuidle and all
266 * functions called within the RCU-idle region.
267 */
268 entered_state = target_state->enter(dev, drv, index);
269
270 if (WARN_ONCE(!irqs_disabled(), "%ps leaked IRQ state", target_state->enter))
271 raw_local_irq_disable();
272
273 if (!(target_state->flags & CPUIDLE_FLAG_RCU_IDLE)) {
274 instrumentation_end();
275 ct_cpuidle_exit();
276 }
277 start_critical_timings();
278
279 sched_clock_idle_wakeup_event();
280 time_end = ns_to_ktime(local_clock_noinstr());
281 trace_cpu_idle(PWR_EVENT_EXIT, dev->cpu);
282
283 /* The cpu is no longer idle or about to enter idle. */
284 sched_idle_set_state(NULL);
285
286 if (broadcast)
287 tick_broadcast_exit();
288
289 if (!cpuidle_state_is_coupled(drv, index))
290 local_irq_enable();
291
292 if (entered_state >= 0) {
293 s64 diff, delay = drv->states[entered_state].exit_latency_ns;
294 int i;
295
296 /*
297 * Update cpuidle counters
298 * This can be moved to within driver enter routine,
299 * but that results in multiple copies of same code.
300 */
301 diff = ktime_sub(time_end, time_start);
302
303 dev->last_residency_ns = diff;
304 dev->states_usage[entered_state].time_ns += diff;
305 dev->states_usage[entered_state].usage++;
306
307 if (diff < drv->states[entered_state].target_residency_ns) {
308 for (i = entered_state - 1; i >= 0; i--) {
309 if (dev->states_usage[i].disable)
310 continue;
311
312 /* Shallower states are enabled, so update. */
313 dev->states_usage[entered_state].above++;
314 trace_cpu_idle_miss(dev->cpu, entered_state, false);
315 break;
316 }
317 } else if (diff > delay) {
318 for (i = entered_state + 1; i < drv->state_count; i++) {
319 if (dev->states_usage[i].disable)
320 continue;
321
322 /*
323 * Update if a deeper state would have been a
324 * better match for the observed idle duration.
325 */
326 if (diff - delay >= drv->states[i].target_residency_ns) {
327 dev->states_usage[entered_state].below++;
328 trace_cpu_idle_miss(dev->cpu, entered_state, true);
329 }
330
331 break;
332 }
333 }
334 } else {
335 dev->last_residency_ns = 0;
336 dev->states_usage[index].rejected++;
337 }
338
339 instrumentation_end();
340
341 return entered_state;
342}
343
344/**
345 * cpuidle_select - ask the cpuidle framework to choose an idle state
346 *
347 * @drv: the cpuidle driver
348 * @dev: the cpuidle device
349 * @stop_tick: indication on whether or not to stop the tick
350 *
351 * Returns the index of the idle state. The return value must not be negative.
352 *
353 * The memory location pointed to by @stop_tick is expected to be written the
354 * 'false' boolean value if the scheduler tick should not be stopped before
355 * entering the returned state.
356 */
357int cpuidle_select(struct cpuidle_driver *drv, struct cpuidle_device *dev,
358 bool *stop_tick)
359{
360 return cpuidle_curr_governor->select(drv, dev, stop_tick);
361}
362
363/**
364 * cpuidle_enter - enter into the specified idle state
365 *
366 * @drv: the cpuidle driver tied with the cpu
367 * @dev: the cpuidle device
368 * @index: the index in the idle state table
369 *
370 * Returns the index in the idle state, < 0 in case of error.
371 * The error code depends on the backend driver
372 */
373int cpuidle_enter(struct cpuidle_driver *drv, struct cpuidle_device *dev,
374 int index)
375{
376 int ret = 0;
377
378 /*
379 * Store the next hrtimer, which becomes either next tick or the next
380 * timer event, whatever expires first. Additionally, to make this data
381 * useful for consumers outside cpuidle, we rely on that the governor's
382 * ->select() callback have decided, whether to stop the tick or not.
383 */
384 WRITE_ONCE(dev->next_hrtimer, tick_nohz_get_next_hrtimer());
385
386 if (cpuidle_state_is_coupled(drv, index))
387 ret = cpuidle_enter_state_coupled(dev, drv, index);
388 else
389 ret = cpuidle_enter_state(dev, drv, index);
390
391 WRITE_ONCE(dev->next_hrtimer, 0);
392 return ret;
393}
394
395/**
396 * cpuidle_reflect - tell the underlying governor what was the state
397 * we were in
398 *
399 * @dev : the cpuidle device
400 * @index: the index in the idle state table
401 *
402 */
403void cpuidle_reflect(struct cpuidle_device *dev, int index)
404{
405 if (cpuidle_curr_governor->reflect && index >= 0)
406 cpuidle_curr_governor->reflect(dev, index);
407}
408
409/*
410 * Min polling interval of 10usec is a guess. It is assuming that
411 * for most users, the time for a single ping-pong workload like
412 * perf bench pipe would generally complete within 10usec but
413 * this is hardware dependent. Actual time can be estimated with
414 *
415 * perf bench sched pipe -l 10000
416 *
417 * Run multiple times to avoid cpufreq effects.
418 */
419#define CPUIDLE_POLL_MIN 10000
420#define CPUIDLE_POLL_MAX (TICK_NSEC / 16)
421
422/**
423 * cpuidle_poll_time - return amount of time to poll for,
424 * governors can override dev->poll_limit_ns if necessary
425 *
426 * @drv: the cpuidle driver tied with the cpu
427 * @dev: the cpuidle device
428 *
429 */
430__cpuidle u64 cpuidle_poll_time(struct cpuidle_driver *drv,
431 struct cpuidle_device *dev)
432{
433 int i;
434 u64 limit_ns;
435
436 BUILD_BUG_ON(CPUIDLE_POLL_MIN > CPUIDLE_POLL_MAX);
437
438 if (dev->poll_limit_ns)
439 return dev->poll_limit_ns;
440
441 limit_ns = CPUIDLE_POLL_MAX;
442 for (i = 1; i < drv->state_count; i++) {
443 u64 state_limit;
444
445 if (dev->states_usage[i].disable)
446 continue;
447
448 state_limit = drv->states[i].target_residency_ns;
449 if (state_limit < CPUIDLE_POLL_MIN)
450 continue;
451
452 limit_ns = min_t(u64, state_limit, CPUIDLE_POLL_MAX);
453 break;
454 }
455
456 dev->poll_limit_ns = limit_ns;
457
458 return dev->poll_limit_ns;
459}
460
461/**
462 * cpuidle_install_idle_handler - installs the cpuidle idle loop handler
463 */
464void cpuidle_install_idle_handler(void)
465{
466 if (enabled_devices) {
467 /* Make sure all changes finished before we switch to new idle */
468 smp_wmb();
469 initialized = 1;
470 }
471}
472
473/**
474 * cpuidle_uninstall_idle_handler - uninstalls the cpuidle idle loop handler
475 */
476void cpuidle_uninstall_idle_handler(void)
477{
478 if (enabled_devices) {
479 initialized = 0;
480 wake_up_all_idle_cpus();
481 }
482
483 /*
484 * Make sure external observers (such as the scheduler)
485 * are done looking at pointed idle states.
486 */
487 synchronize_rcu();
488}
489
490/**
491 * cpuidle_pause_and_lock - temporarily disables CPUIDLE
492 */
493void cpuidle_pause_and_lock(void)
494{
495 mutex_lock(&cpuidle_lock);
496 cpuidle_uninstall_idle_handler();
497}
498
499EXPORT_SYMBOL_GPL(cpuidle_pause_and_lock);
500
501/**
502 * cpuidle_resume_and_unlock - resumes CPUIDLE operation
503 */
504void cpuidle_resume_and_unlock(void)
505{
506 cpuidle_install_idle_handler();
507 mutex_unlock(&cpuidle_lock);
508}
509
510EXPORT_SYMBOL_GPL(cpuidle_resume_and_unlock);
511
512/* Currently used in suspend/resume path to suspend cpuidle */
513void cpuidle_pause(void)
514{
515 mutex_lock(&cpuidle_lock);
516 cpuidle_uninstall_idle_handler();
517 mutex_unlock(&cpuidle_lock);
518}
519
520/* Currently used in suspend/resume path to resume cpuidle */
521void cpuidle_resume(void)
522{
523 mutex_lock(&cpuidle_lock);
524 cpuidle_install_idle_handler();
525 mutex_unlock(&cpuidle_lock);
526}
527
528/**
529 * cpuidle_enable_device - enables idle PM for a CPU
530 * @dev: the CPU
531 *
532 * This function must be called between cpuidle_pause_and_lock and
533 * cpuidle_resume_and_unlock when used externally.
534 */
535int cpuidle_enable_device(struct cpuidle_device *dev)
536{
537 int ret;
538 struct cpuidle_driver *drv;
539
540 if (!dev)
541 return -EINVAL;
542
543 if (dev->enabled)
544 return 0;
545
546 if (!cpuidle_curr_governor)
547 return -EIO;
548
549 drv = cpuidle_get_cpu_driver(dev);
550
551 if (!drv)
552 return -EIO;
553
554 if (!dev->registered)
555 return -EINVAL;
556
557 ret = cpuidle_add_device_sysfs(dev);
558 if (ret)
559 return ret;
560
561 if (cpuidle_curr_governor->enable) {
562 ret = cpuidle_curr_governor->enable(drv, dev);
563 if (ret)
564 goto fail_sysfs;
565 }
566
567 smp_wmb();
568
569 dev->enabled = 1;
570
571 enabled_devices++;
572 return 0;
573
574fail_sysfs:
575 cpuidle_remove_device_sysfs(dev);
576
577 return ret;
578}
579
580EXPORT_SYMBOL_GPL(cpuidle_enable_device);
581
582/**
583 * cpuidle_disable_device - disables idle PM for a CPU
584 * @dev: the CPU
585 *
586 * This function must be called between cpuidle_pause_and_lock and
587 * cpuidle_resume_and_unlock when used externally.
588 */
589void cpuidle_disable_device(struct cpuidle_device *dev)
590{
591 struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);
592
593 if (!dev || !dev->enabled)
594 return;
595
596 if (!drv || !cpuidle_curr_governor)
597 return;
598
599 dev->enabled = 0;
600
601 if (cpuidle_curr_governor->disable)
602 cpuidle_curr_governor->disable(drv, dev);
603
604 cpuidle_remove_device_sysfs(dev);
605 enabled_devices--;
606}
607
608EXPORT_SYMBOL_GPL(cpuidle_disable_device);
609
610static void __cpuidle_unregister_device(struct cpuidle_device *dev)
611{
612 struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);
613
614 list_del(&dev->device_list);
615 per_cpu(cpuidle_devices, dev->cpu) = NULL;
616 module_put(drv->owner);
617
618 dev->registered = 0;
619}
620
621static void __cpuidle_device_init(struct cpuidle_device *dev)
622{
623 memset(dev->states_usage, 0, sizeof(dev->states_usage));
624 dev->last_residency_ns = 0;
625 dev->next_hrtimer = 0;
626}
627
628/**
629 * __cpuidle_register_device - internal register function called before register
630 * and enable routines
631 * @dev: the cpu
632 *
633 * cpuidle_lock mutex must be held before this is called
634 */
635static int __cpuidle_register_device(struct cpuidle_device *dev)
636{
637 struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);
638 int i, ret;
639
640 if (!try_module_get(drv->owner))
641 return -EINVAL;
642
643 for (i = 0; i < drv->state_count; i++) {
644 if (drv->states[i].flags & CPUIDLE_FLAG_UNUSABLE)
645 dev->states_usage[i].disable |= CPUIDLE_STATE_DISABLED_BY_DRIVER;
646
647 if (drv->states[i].flags & CPUIDLE_FLAG_OFF)
648 dev->states_usage[i].disable |= CPUIDLE_STATE_DISABLED_BY_USER;
649 }
650
651 per_cpu(cpuidle_devices, dev->cpu) = dev;
652 list_add(&dev->device_list, &cpuidle_detected_devices);
653
654 ret = cpuidle_coupled_register_device(dev);
655 if (ret)
656 __cpuidle_unregister_device(dev);
657 else
658 dev->registered = 1;
659
660 return ret;
661}
662
663/**
664 * cpuidle_register_device - registers a CPU's idle PM feature
665 * @dev: the cpu
666 */
667int cpuidle_register_device(struct cpuidle_device *dev)
668{
669 int ret = -EBUSY;
670
671 if (!dev)
672 return -EINVAL;
673
674 mutex_lock(&cpuidle_lock);
675
676 if (dev->registered)
677 goto out_unlock;
678
679 __cpuidle_device_init(dev);
680
681 ret = __cpuidle_register_device(dev);
682 if (ret)
683 goto out_unlock;
684
685 ret = cpuidle_add_sysfs(dev);
686 if (ret)
687 goto out_unregister;
688
689 ret = cpuidle_enable_device(dev);
690 if (ret)
691 goto out_sysfs;
692
693 cpuidle_install_idle_handler();
694
695out_unlock:
696 mutex_unlock(&cpuidle_lock);
697
698 return ret;
699
700out_sysfs:
701 cpuidle_remove_sysfs(dev);
702out_unregister:
703 __cpuidle_unregister_device(dev);
704 goto out_unlock;
705}
706
707EXPORT_SYMBOL_GPL(cpuidle_register_device);
708
709/**
710 * cpuidle_unregister_device - unregisters a CPU's idle PM feature
711 * @dev: the cpu
712 */
713void cpuidle_unregister_device(struct cpuidle_device *dev)
714{
715 if (!dev || dev->registered == 0)
716 return;
717
718 cpuidle_pause_and_lock();
719
720 cpuidle_disable_device(dev);
721
722 cpuidle_remove_sysfs(dev);
723
724 __cpuidle_unregister_device(dev);
725
726 cpuidle_coupled_unregister_device(dev);
727
728 cpuidle_resume_and_unlock();
729}
730
731EXPORT_SYMBOL_GPL(cpuidle_unregister_device);
732
733/**
734 * cpuidle_unregister: unregister a driver and the devices. This function
735 * can be used only if the driver has been previously registered through
736 * the cpuidle_register function.
737 *
738 * @drv: a valid pointer to a struct cpuidle_driver
739 */
740void cpuidle_unregister(struct cpuidle_driver *drv)
741{
742 int cpu;
743 struct cpuidle_device *device;
744
745 for_each_cpu(cpu, drv->cpumask) {
746 device = &per_cpu(cpuidle_dev, cpu);
747 cpuidle_unregister_device(device);
748 }
749
750 cpuidle_unregister_driver(drv);
751}
752EXPORT_SYMBOL_GPL(cpuidle_unregister);
753
754/**
755 * cpuidle_register: registers the driver and the cpu devices with the
756 * coupled_cpus passed as parameter. This function is used for all common
757 * initialization pattern there are in the arch specific drivers. The
758 * devices is globally defined in this file.
759 *
760 * @drv : a valid pointer to a struct cpuidle_driver
761 * @coupled_cpus: a cpumask for the coupled states
762 *
763 * Returns 0 on success, < 0 otherwise
764 */
765int cpuidle_register(struct cpuidle_driver *drv,
766 const struct cpumask *const coupled_cpus)
767{
768 int ret, cpu;
769 struct cpuidle_device *device;
770
771 ret = cpuidle_register_driver(drv);
772 if (ret) {
773 pr_err("failed to register cpuidle driver\n");
774 return ret;
775 }
776
777 for_each_cpu(cpu, drv->cpumask) {
778 device = &per_cpu(cpuidle_dev, cpu);
779 device->cpu = cpu;
780
781#ifdef CONFIG_ARCH_NEEDS_CPU_IDLE_COUPLED
782 /*
783 * On multiplatform for ARM, the coupled idle states could be
784 * enabled in the kernel even if the cpuidle driver does not
785 * use it. Note, coupled_cpus is a struct copy.
786 */
787 if (coupled_cpus)
788 device->coupled_cpus = *coupled_cpus;
789#endif
790 ret = cpuidle_register_device(device);
791 if (!ret)
792 continue;
793
794 pr_err("Failed to register cpuidle device for cpu%d\n", cpu);
795
796 cpuidle_unregister(drv);
797 break;
798 }
799
800 return ret;
801}
802EXPORT_SYMBOL_GPL(cpuidle_register);
803
804/**
805 * cpuidle_init - core initializer
806 */
807static int __init cpuidle_init(void)
808{
809 if (cpuidle_disabled())
810 return -ENODEV;
811
812 return cpuidle_add_interface();
813}
814
815module_param(off, int, 0444);
816module_param_string(governor, param_governor, CPUIDLE_NAME_LEN, 0444);
817core_initcall(cpuidle_init);
1/*
2 * cpuidle.c - core cpuidle infrastructure
3 *
4 * (C) 2006-2007 Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
5 * Shaohua Li <shaohua.li@intel.com>
6 * Adam Belay <abelay@novell.com>
7 *
8 * This code is licenced under the GPL.
9 */
10
11#include <linux/clockchips.h>
12#include <linux/kernel.h>
13#include <linux/mutex.h>
14#include <linux/sched.h>
15#include <linux/sched/clock.h>
16#include <linux/notifier.h>
17#include <linux/pm_qos.h>
18#include <linux/cpu.h>
19#include <linux/cpuidle.h>
20#include <linux/ktime.h>
21#include <linux/hrtimer.h>
22#include <linux/module.h>
23#include <linux/suspend.h>
24#include <linux/tick.h>
25#include <trace/events/power.h>
26
27#include "cpuidle.h"
28
29DEFINE_PER_CPU(struct cpuidle_device *, cpuidle_devices);
30DEFINE_PER_CPU(struct cpuidle_device, cpuidle_dev);
31
32DEFINE_MUTEX(cpuidle_lock);
33LIST_HEAD(cpuidle_detected_devices);
34
35static int enabled_devices;
36static int off __read_mostly;
37static int initialized __read_mostly;
38
39int cpuidle_disabled(void)
40{
41 return off;
42}
43void disable_cpuidle(void)
44{
45 off = 1;
46}
47
48bool cpuidle_not_available(struct cpuidle_driver *drv,
49 struct cpuidle_device *dev)
50{
51 return off || !initialized || !drv || !dev || !dev->enabled;
52}
53
54/**
55 * cpuidle_play_dead - cpu off-lining
56 *
57 * Returns in case of an error or no driver
58 */
59int cpuidle_play_dead(void)
60{
61 struct cpuidle_device *dev = __this_cpu_read(cpuidle_devices);
62 struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);
63 int i;
64
65 if (!drv)
66 return -ENODEV;
67
68 /* Find lowest-power state that supports long-term idle */
69 for (i = drv->state_count - 1; i >= 0; i--)
70 if (drv->states[i].enter_dead)
71 return drv->states[i].enter_dead(dev, i);
72
73 return -ENODEV;
74}
75
76static int find_deepest_state(struct cpuidle_driver *drv,
77 struct cpuidle_device *dev,
78 unsigned int max_latency,
79 unsigned int forbidden_flags,
80 bool s2idle)
81{
82 unsigned int latency_req = 0;
83 int i, ret = 0;
84
85 for (i = 1; i < drv->state_count; i++) {
86 struct cpuidle_state *s = &drv->states[i];
87 struct cpuidle_state_usage *su = &dev->states_usage[i];
88
89 if (s->disabled || su->disable || s->exit_latency <= latency_req
90 || s->exit_latency > max_latency
91 || (s->flags & forbidden_flags)
92 || (s2idle && !s->enter_s2idle))
93 continue;
94
95 latency_req = s->exit_latency;
96 ret = i;
97 }
98 return ret;
99}
100
101/**
102 * cpuidle_use_deepest_state - Set/clear governor override flag.
103 * @enable: New value of the flag.
104 *
105 * Set/unset the current CPU to use the deepest idle state (override governors
106 * going forward if set).
107 */
108void cpuidle_use_deepest_state(bool enable)
109{
110 struct cpuidle_device *dev;
111
112 preempt_disable();
113 dev = cpuidle_get_device();
114 if (dev)
115 dev->use_deepest_state = enable;
116 preempt_enable();
117}
118
119/**
120 * cpuidle_find_deepest_state - Find the deepest available idle state.
121 * @drv: cpuidle driver for the given CPU.
122 * @dev: cpuidle device for the given CPU.
123 */
124int cpuidle_find_deepest_state(struct cpuidle_driver *drv,
125 struct cpuidle_device *dev)
126{
127 return find_deepest_state(drv, dev, UINT_MAX, 0, false);
128}
129
130#ifdef CONFIG_SUSPEND
131static void enter_s2idle_proper(struct cpuidle_driver *drv,
132 struct cpuidle_device *dev, int index)
133{
134 ktime_t time_start, time_end;
135
136 time_start = ns_to_ktime(local_clock());
137
138 /*
139 * trace_suspend_resume() called by tick_freeze() for the last CPU
140 * executing it contains RCU usage regarded as invalid in the idle
141 * context, so tell RCU about that.
142 */
143 RCU_NONIDLE(tick_freeze());
144 /*
145 * The state used here cannot be a "coupled" one, because the "coupled"
146 * cpuidle mechanism enables interrupts and doing that with timekeeping
147 * suspended is generally unsafe.
148 */
149 stop_critical_timings();
150 drv->states[index].enter_s2idle(dev, drv, index);
151 WARN_ON(!irqs_disabled());
152 /*
153 * timekeeping_resume() that will be called by tick_unfreeze() for the
154 * first CPU executing it calls functions containing RCU read-side
155 * critical sections, so tell RCU about that.
156 */
157 RCU_NONIDLE(tick_unfreeze());
158 start_critical_timings();
159
160 time_end = ns_to_ktime(local_clock());
161
162 dev->states_usage[index].s2idle_time += ktime_us_delta(time_end, time_start);
163 dev->states_usage[index].s2idle_usage++;
164}
165
166/**
167 * cpuidle_enter_s2idle - Enter an idle state suitable for suspend-to-idle.
168 * @drv: cpuidle driver for the given CPU.
169 * @dev: cpuidle device for the given CPU.
170 *
171 * If there are states with the ->enter_s2idle callback, find the deepest of
172 * them and enter it with frozen tick.
173 */
174int cpuidle_enter_s2idle(struct cpuidle_driver *drv, struct cpuidle_device *dev)
175{
176 int index;
177
178 /*
179 * Find the deepest state with ->enter_s2idle present, which guarantees
180 * that interrupts won't be enabled when it exits and allows the tick to
181 * be frozen safely.
182 */
183 index = find_deepest_state(drv, dev, UINT_MAX, 0, true);
184 if (index > 0)
185 enter_s2idle_proper(drv, dev, index);
186
187 return index;
188}
189#endif /* CONFIG_SUSPEND */
190
191/**
192 * cpuidle_enter_state - enter the state and update stats
193 * @dev: cpuidle device for this cpu
194 * @drv: cpuidle driver for this cpu
195 * @index: index into the states table in @drv of the state to enter
196 */
197int cpuidle_enter_state(struct cpuidle_device *dev, struct cpuidle_driver *drv,
198 int index)
199{
200 int entered_state;
201
202 struct cpuidle_state *target_state = &drv->states[index];
203 bool broadcast = !!(target_state->flags & CPUIDLE_FLAG_TIMER_STOP);
204 ktime_t time_start, time_end;
205
206 /*
207 * Tell the time framework to switch to a broadcast timer because our
208 * local timer will be shut down. If a local timer is used from another
209 * CPU as a broadcast timer, this call may fail if it is not available.
210 */
211 if (broadcast && tick_broadcast_enter()) {
212 index = find_deepest_state(drv, dev, target_state->exit_latency,
213 CPUIDLE_FLAG_TIMER_STOP, false);
214 if (index < 0) {
215 default_idle_call();
216 return -EBUSY;
217 }
218 target_state = &drv->states[index];
219 broadcast = false;
220 }
221
222 /* Take note of the planned idle state. */
223 sched_idle_set_state(target_state);
224
225 trace_cpu_idle_rcuidle(index, dev->cpu);
226 time_start = ns_to_ktime(local_clock());
227
228 stop_critical_timings();
229 entered_state = target_state->enter(dev, drv, index);
230 start_critical_timings();
231
232 sched_clock_idle_wakeup_event();
233 time_end = ns_to_ktime(local_clock());
234 trace_cpu_idle_rcuidle(PWR_EVENT_EXIT, dev->cpu);
235
236 /* The cpu is no longer idle or about to enter idle. */
237 sched_idle_set_state(NULL);
238
239 if (broadcast) {
240 if (WARN_ON_ONCE(!irqs_disabled()))
241 local_irq_disable();
242
243 tick_broadcast_exit();
244 }
245
246 if (!cpuidle_state_is_coupled(drv, index))
247 local_irq_enable();
248
249 if (entered_state >= 0) {
250 s64 diff, delay = drv->states[entered_state].exit_latency;
251 int i;
252
253 /*
254 * Update cpuidle counters
255 * This can be moved to within driver enter routine,
256 * but that results in multiple copies of same code.
257 */
258 diff = ktime_us_delta(time_end, time_start);
259 if (diff > INT_MAX)
260 diff = INT_MAX;
261
262 dev->last_residency = (int)diff;
263 dev->states_usage[entered_state].time += dev->last_residency;
264 dev->states_usage[entered_state].usage++;
265
266 if (diff < drv->states[entered_state].target_residency) {
267 for (i = entered_state - 1; i >= 0; i--) {
268 if (drv->states[i].disabled ||
269 dev->states_usage[i].disable)
270 continue;
271
272 /* Shallower states are enabled, so update. */
273 dev->states_usage[entered_state].above++;
274 break;
275 }
276 } else if (diff > delay) {
277 for (i = entered_state + 1; i < drv->state_count; i++) {
278 if (drv->states[i].disabled ||
279 dev->states_usage[i].disable)
280 continue;
281
282 /*
283 * Update if a deeper state would have been a
284 * better match for the observed idle duration.
285 */
286 if (diff - delay >= drv->states[i].target_residency)
287 dev->states_usage[entered_state].below++;
288
289 break;
290 }
291 }
292 } else {
293 dev->last_residency = 0;
294 }
295
296 return entered_state;
297}
298
299/**
300 * cpuidle_select - ask the cpuidle framework to choose an idle state
301 *
302 * @drv: the cpuidle driver
303 * @dev: the cpuidle device
304 * @stop_tick: indication on whether or not to stop the tick
305 *
306 * Returns the index of the idle state. The return value must not be negative.
307 *
308 * The memory location pointed to by @stop_tick is expected to be written the
309 * 'false' boolean value if the scheduler tick should not be stopped before
310 * entering the returned state.
311 */
312int cpuidle_select(struct cpuidle_driver *drv, struct cpuidle_device *dev,
313 bool *stop_tick)
314{
315 return cpuidle_curr_governor->select(drv, dev, stop_tick);
316}
317
318/**
319 * cpuidle_enter - enter into the specified idle state
320 *
321 * @drv: the cpuidle driver tied with the cpu
322 * @dev: the cpuidle device
323 * @index: the index in the idle state table
324 *
325 * Returns the index in the idle state, < 0 in case of error.
326 * The error code depends on the backend driver
327 */
328int cpuidle_enter(struct cpuidle_driver *drv, struct cpuidle_device *dev,
329 int index)
330{
331 int ret = 0;
332
333 /*
334 * Store the next hrtimer, which becomes either next tick or the next
335 * timer event, whatever expires first. Additionally, to make this data
336 * useful for consumers outside cpuidle, we rely on that the governor's
337 * ->select() callback have decided, whether to stop the tick or not.
338 */
339 WRITE_ONCE(dev->next_hrtimer, tick_nohz_get_next_hrtimer());
340
341 if (cpuidle_state_is_coupled(drv, index))
342 ret = cpuidle_enter_state_coupled(dev, drv, index);
343 else
344 ret = cpuidle_enter_state(dev, drv, index);
345
346 WRITE_ONCE(dev->next_hrtimer, 0);
347 return ret;
348}
349
350/**
351 * cpuidle_reflect - tell the underlying governor what was the state
352 * we were in
353 *
354 * @dev : the cpuidle device
355 * @index: the index in the idle state table
356 *
357 */
358void cpuidle_reflect(struct cpuidle_device *dev, int index)
359{
360 if (cpuidle_curr_governor->reflect && index >= 0)
361 cpuidle_curr_governor->reflect(dev, index);
362}
363
364/**
365 * cpuidle_poll_time - return amount of time to poll for,
366 * governors can override dev->poll_limit_ns if necessary
367 *
368 * @drv: the cpuidle driver tied with the cpu
369 * @dev: the cpuidle device
370 *
371 */
372u64 cpuidle_poll_time(struct cpuidle_driver *drv,
373 struct cpuidle_device *dev)
374{
375 int i;
376 u64 limit_ns;
377
378 if (dev->poll_limit_ns)
379 return dev->poll_limit_ns;
380
381 limit_ns = TICK_NSEC;
382 for (i = 1; i < drv->state_count; i++) {
383 if (drv->states[i].disabled || dev->states_usage[i].disable)
384 continue;
385
386 limit_ns = (u64)drv->states[i].target_residency * NSEC_PER_USEC;
387 }
388
389 dev->poll_limit_ns = limit_ns;
390
391 return dev->poll_limit_ns;
392}
393
394/**
395 * cpuidle_install_idle_handler - installs the cpuidle idle loop handler
396 */
397void cpuidle_install_idle_handler(void)
398{
399 if (enabled_devices) {
400 /* Make sure all changes finished before we switch to new idle */
401 smp_wmb();
402 initialized = 1;
403 }
404}
405
406/**
407 * cpuidle_uninstall_idle_handler - uninstalls the cpuidle idle loop handler
408 */
409void cpuidle_uninstall_idle_handler(void)
410{
411 if (enabled_devices) {
412 initialized = 0;
413 wake_up_all_idle_cpus();
414 }
415
416 /*
417 * Make sure external observers (such as the scheduler)
418 * are done looking at pointed idle states.
419 */
420 synchronize_rcu();
421}
422
423/**
424 * cpuidle_pause_and_lock - temporarily disables CPUIDLE
425 */
426void cpuidle_pause_and_lock(void)
427{
428 mutex_lock(&cpuidle_lock);
429 cpuidle_uninstall_idle_handler();
430}
431
432EXPORT_SYMBOL_GPL(cpuidle_pause_and_lock);
433
434/**
435 * cpuidle_resume_and_unlock - resumes CPUIDLE operation
436 */
437void cpuidle_resume_and_unlock(void)
438{
439 cpuidle_install_idle_handler();
440 mutex_unlock(&cpuidle_lock);
441}
442
443EXPORT_SYMBOL_GPL(cpuidle_resume_and_unlock);
444
445/* Currently used in suspend/resume path to suspend cpuidle */
446void cpuidle_pause(void)
447{
448 mutex_lock(&cpuidle_lock);
449 cpuidle_uninstall_idle_handler();
450 mutex_unlock(&cpuidle_lock);
451}
452
453/* Currently used in suspend/resume path to resume cpuidle */
454void cpuidle_resume(void)
455{
456 mutex_lock(&cpuidle_lock);
457 cpuidle_install_idle_handler();
458 mutex_unlock(&cpuidle_lock);
459}
460
461/**
462 * cpuidle_enable_device - enables idle PM for a CPU
463 * @dev: the CPU
464 *
465 * This function must be called between cpuidle_pause_and_lock and
466 * cpuidle_resume_and_unlock when used externally.
467 */
468int cpuidle_enable_device(struct cpuidle_device *dev)
469{
470 int ret;
471 struct cpuidle_driver *drv;
472
473 if (!dev)
474 return -EINVAL;
475
476 if (dev->enabled)
477 return 0;
478
479 if (!cpuidle_curr_governor)
480 return -EIO;
481
482 drv = cpuidle_get_cpu_driver(dev);
483
484 if (!drv)
485 return -EIO;
486
487 if (!dev->registered)
488 return -EINVAL;
489
490 ret = cpuidle_add_device_sysfs(dev);
491 if (ret)
492 return ret;
493
494 if (cpuidle_curr_governor->enable) {
495 ret = cpuidle_curr_governor->enable(drv, dev);
496 if (ret)
497 goto fail_sysfs;
498 }
499
500 smp_wmb();
501
502 dev->enabled = 1;
503
504 enabled_devices++;
505 return 0;
506
507fail_sysfs:
508 cpuidle_remove_device_sysfs(dev);
509
510 return ret;
511}
512
513EXPORT_SYMBOL_GPL(cpuidle_enable_device);
514
515/**
516 * cpuidle_disable_device - disables idle PM for a CPU
517 * @dev: the CPU
518 *
519 * This function must be called between cpuidle_pause_and_lock and
520 * cpuidle_resume_and_unlock when used externally.
521 */
522void cpuidle_disable_device(struct cpuidle_device *dev)
523{
524 struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);
525
526 if (!dev || !dev->enabled)
527 return;
528
529 if (!drv || !cpuidle_curr_governor)
530 return;
531
532 dev->enabled = 0;
533
534 if (cpuidle_curr_governor->disable)
535 cpuidle_curr_governor->disable(drv, dev);
536
537 cpuidle_remove_device_sysfs(dev);
538 enabled_devices--;
539}
540
541EXPORT_SYMBOL_GPL(cpuidle_disable_device);
542
543static void __cpuidle_unregister_device(struct cpuidle_device *dev)
544{
545 struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);
546
547 list_del(&dev->device_list);
548 per_cpu(cpuidle_devices, dev->cpu) = NULL;
549 module_put(drv->owner);
550
551 dev->registered = 0;
552}
553
554static void __cpuidle_device_init(struct cpuidle_device *dev)
555{
556 memset(dev->states_usage, 0, sizeof(dev->states_usage));
557 dev->last_residency = 0;
558 dev->next_hrtimer = 0;
559}
560
561/**
562 * __cpuidle_register_device - internal register function called before register
563 * and enable routines
564 * @dev: the cpu
565 *
566 * cpuidle_lock mutex must be held before this is called
567 */
568static int __cpuidle_register_device(struct cpuidle_device *dev)
569{
570 int ret;
571 struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);
572
573 if (!try_module_get(drv->owner))
574 return -EINVAL;
575
576 per_cpu(cpuidle_devices, dev->cpu) = dev;
577 list_add(&dev->device_list, &cpuidle_detected_devices);
578
579 ret = cpuidle_coupled_register_device(dev);
580 if (ret)
581 __cpuidle_unregister_device(dev);
582 else
583 dev->registered = 1;
584
585 return ret;
586}
587
588/**
589 * cpuidle_register_device - registers a CPU's idle PM feature
590 * @dev: the cpu
591 */
592int cpuidle_register_device(struct cpuidle_device *dev)
593{
594 int ret = -EBUSY;
595
596 if (!dev)
597 return -EINVAL;
598
599 mutex_lock(&cpuidle_lock);
600
601 if (dev->registered)
602 goto out_unlock;
603
604 __cpuidle_device_init(dev);
605
606 ret = __cpuidle_register_device(dev);
607 if (ret)
608 goto out_unlock;
609
610 ret = cpuidle_add_sysfs(dev);
611 if (ret)
612 goto out_unregister;
613
614 ret = cpuidle_enable_device(dev);
615 if (ret)
616 goto out_sysfs;
617
618 cpuidle_install_idle_handler();
619
620out_unlock:
621 mutex_unlock(&cpuidle_lock);
622
623 return ret;
624
625out_sysfs:
626 cpuidle_remove_sysfs(dev);
627out_unregister:
628 __cpuidle_unregister_device(dev);
629 goto out_unlock;
630}
631
632EXPORT_SYMBOL_GPL(cpuidle_register_device);
633
634/**
635 * cpuidle_unregister_device - unregisters a CPU's idle PM feature
636 * @dev: the cpu
637 */
638void cpuidle_unregister_device(struct cpuidle_device *dev)
639{
640 if (!dev || dev->registered == 0)
641 return;
642
643 cpuidle_pause_and_lock();
644
645 cpuidle_disable_device(dev);
646
647 cpuidle_remove_sysfs(dev);
648
649 __cpuidle_unregister_device(dev);
650
651 cpuidle_coupled_unregister_device(dev);
652
653 cpuidle_resume_and_unlock();
654}
655
656EXPORT_SYMBOL_GPL(cpuidle_unregister_device);
657
658/**
659 * cpuidle_unregister: unregister a driver and the devices. This function
660 * can be used only if the driver has been previously registered through
661 * the cpuidle_register function.
662 *
663 * @drv: a valid pointer to a struct cpuidle_driver
664 */
665void cpuidle_unregister(struct cpuidle_driver *drv)
666{
667 int cpu;
668 struct cpuidle_device *device;
669
670 for_each_cpu(cpu, drv->cpumask) {
671 device = &per_cpu(cpuidle_dev, cpu);
672 cpuidle_unregister_device(device);
673 }
674
675 cpuidle_unregister_driver(drv);
676}
677EXPORT_SYMBOL_GPL(cpuidle_unregister);
678
679/**
680 * cpuidle_register: registers the driver and the cpu devices with the
681 * coupled_cpus passed as parameter. This function is used for all common
682 * initialization pattern there are in the arch specific drivers. The
683 * devices is globally defined in this file.
684 *
685 * @drv : a valid pointer to a struct cpuidle_driver
686 * @coupled_cpus: a cpumask for the coupled states
687 *
688 * Returns 0 on success, < 0 otherwise
689 */
690int cpuidle_register(struct cpuidle_driver *drv,
691 const struct cpumask *const coupled_cpus)
692{
693 int ret, cpu;
694 struct cpuidle_device *device;
695
696 ret = cpuidle_register_driver(drv);
697 if (ret) {
698 pr_err("failed to register cpuidle driver\n");
699 return ret;
700 }
701
702 for_each_cpu(cpu, drv->cpumask) {
703 device = &per_cpu(cpuidle_dev, cpu);
704 device->cpu = cpu;
705
706#ifdef CONFIG_ARCH_NEEDS_CPU_IDLE_COUPLED
707 /*
708 * On multiplatform for ARM, the coupled idle states could be
709 * enabled in the kernel even if the cpuidle driver does not
710 * use it. Note, coupled_cpus is a struct copy.
711 */
712 if (coupled_cpus)
713 device->coupled_cpus = *coupled_cpus;
714#endif
715 ret = cpuidle_register_device(device);
716 if (!ret)
717 continue;
718
719 pr_err("Failed to register cpuidle device for cpu%d\n", cpu);
720
721 cpuidle_unregister(drv);
722 break;
723 }
724
725 return ret;
726}
727EXPORT_SYMBOL_GPL(cpuidle_register);
728
729#ifdef CONFIG_SMP
730
731/*
732 * This function gets called when a part of the kernel has a new latency
733 * requirement. This means we need to get all processors out of their C-state,
734 * and then recalculate a new suitable C-state. Just do a cross-cpu IPI; that
735 * wakes them all right up.
736 */
737static int cpuidle_latency_notify(struct notifier_block *b,
738 unsigned long l, void *v)
739{
740 wake_up_all_idle_cpus();
741 return NOTIFY_OK;
742}
743
744static struct notifier_block cpuidle_latency_notifier = {
745 .notifier_call = cpuidle_latency_notify,
746};
747
748static inline void latency_notifier_init(struct notifier_block *n)
749{
750 pm_qos_add_notifier(PM_QOS_CPU_DMA_LATENCY, n);
751}
752
753#else /* CONFIG_SMP */
754
755#define latency_notifier_init(x) do { } while (0)
756
757#endif /* CONFIG_SMP */
758
759/**
760 * cpuidle_init - core initializer
761 */
762static int __init cpuidle_init(void)
763{
764 int ret;
765
766 if (cpuidle_disabled())
767 return -ENODEV;
768
769 ret = cpuidle_add_interface(cpu_subsys.dev_root);
770 if (ret)
771 return ret;
772
773 latency_notifier_init(&cpuidle_latency_notifier);
774
775 return 0;
776}
777
778module_param(off, int, 0444);
779module_param_string(governor, param_governor, CPUIDLE_NAME_LEN, 0444);
780core_initcall(cpuidle_init);