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