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