2 * cpuidle.c - core cpuidle infrastructure
4 * (C) 2006-2007 Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
5 * Shaohua Li <shaohua.li@intel.com>
6 * Adam Belay <abelay@novell.com>
8 * This code is licenced under the GPL.
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>
33 DEFINE_PER_CPU(struct cpuidle_device *, cpuidle_devices);
34 DEFINE_PER_CPU(struct cpuidle_device, cpuidle_dev);
36 DEFINE_MUTEX(cpuidle_lock);
37 LIST_HEAD(cpuidle_detected_devices);
39 static int enabled_devices;
40 static int off __read_mostly;
41 static int initialized __read_mostly;
43 int cpuidle_disabled(void)
47 void disable_cpuidle(void)
52 bool cpuidle_not_available(struct cpuidle_driver *drv,
53 struct cpuidle_device *dev)
55 return off || !initialized || !drv || !dev || !dev->enabled;
59 * cpuidle_play_dead - cpu off-lining
61 * Returns in case of an error or no driver
63 int cpuidle_play_dead(void)
65 struct cpuidle_device *dev = __this_cpu_read(cpuidle_devices);
66 struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);
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);
80 static int find_deepest_state(struct cpuidle_driver *drv,
81 struct cpuidle_device *dev,
83 unsigned int forbidden_flags,
89 for (i = 1; i < drv->state_count; i++) {
90 struct cpuidle_state *s = &drv->states[i];
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))
99 latency_req = s->exit_latency_ns;
106 * cpuidle_use_deepest_state - Set/unset governor override mode.
107 * @latency_limit_ns: Idle state exit latency limit (or no override if 0).
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.
113 void cpuidle_use_deepest_state(u64 latency_limit_ns)
115 struct cpuidle_device *dev;
118 dev = cpuidle_get_device();
120 dev->forced_idle_latency_limit_ns = latency_limit_ns;
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
130 * Return: the index of the deepest available idle state.
132 int cpuidle_find_deepest_state(struct cpuidle_driver *drv,
133 struct cpuidle_device *dev,
134 u64 latency_limit_ns)
136 return find_deepest_state(drv, dev, latency_limit_ns, 0, false);
139 #ifdef CONFIG_SUSPEND
140 static noinstr void enter_s2idle_proper(struct cpuidle_driver *drv,
141 struct cpuidle_device *dev, int index)
143 struct cpuidle_state *target_state = &drv->states[index];
144 ktime_t time_start, time_end;
146 instrumentation_begin();
148 time_start = ns_to_ktime(local_clock_noinstr());
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.
156 stop_critical_timings();
157 if (!(target_state->flags & CPUIDLE_FLAG_RCU_IDLE)) {
159 /* Annotate away the indirect call */
160 instrumentation_begin();
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();
170 start_critical_timings();
172 time_end = ns_to_ktime(local_clock_noinstr());
174 dev->states_usage[index].s2idle_time += ktime_us_delta(time_end, time_start);
175 dev->states_usage[index].s2idle_usage++;
176 instrumentation_end();
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.
184 * If there are states with the ->enter_s2idle callback, find the deepest of
185 * them and enter it with frozen tick.
187 int cpuidle_enter_s2idle(struct cpuidle_driver *drv, struct cpuidle_device *dev)
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
196 index = find_deepest_state(drv, dev, U64_MAX, 0, true);
198 enter_s2idle_proper(drv, dev, index);
203 #endif /* CONFIG_SUSPEND */
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
211 noinstr int cpuidle_enter_state(struct cpuidle_device *dev,
212 struct cpuidle_driver *drv,
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;
221 instrumentation_begin();
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.
228 if (broadcast && tick_broadcast_enter()) {
229 index = find_deepest_state(drv, dev, target_state->exit_latency_ns,
230 CPUIDLE_FLAG_TIMER_STOP, false);
235 target_state = &drv->states[index];
239 if (target_state->flags & CPUIDLE_FLAG_TLB_FLUSHED)
242 /* Take note of the planned idle state. */
243 sched_idle_set_state(target_state);
245 trace_cpu_idle(index, dev->cpu);
246 time_start = ns_to_ktime(local_clock_noinstr());
248 stop_critical_timings();
249 if (!(target_state->flags & CPUIDLE_FLAG_RCU_IDLE)) {
251 /* Annotate away the indirect call */
252 instrumentation_begin();
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.
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.
267 entered_state = target_state->enter(dev, drv, index);
269 if (WARN_ONCE(!irqs_disabled(), "%ps leaked IRQ state", target_state->enter))
270 raw_local_irq_disable();
272 if (!(target_state->flags & CPUIDLE_FLAG_RCU_IDLE)) {
273 instrumentation_end();
276 start_critical_timings();
278 sched_clock_idle_wakeup_event();
279 time_end = ns_to_ktime(local_clock_noinstr());
280 trace_cpu_idle(PWR_EVENT_EXIT, dev->cpu);
282 /* The cpu is no longer idle or about to enter idle. */
283 sched_idle_set_state(NULL);
286 tick_broadcast_exit();
288 if (!cpuidle_state_is_coupled(drv, index))
291 if (entered_state >= 0) {
292 s64 diff, delay = drv->states[entered_state].exit_latency_ns;
296 * Update cpuidle counters
297 * This can be moved to within driver enter routine,
298 * but that results in multiple copies of same code.
300 diff = ktime_sub(time_end, time_start);
302 dev->last_residency_ns = diff;
303 dev->states_usage[entered_state].time_ns += diff;
304 dev->states_usage[entered_state].usage++;
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)
311 /* Shallower states are enabled, so update. */
312 dev->states_usage[entered_state].above++;
313 trace_cpu_idle_miss(dev->cpu, entered_state, false);
316 } else if (diff > delay) {
317 for (i = entered_state + 1; i < drv->state_count; i++) {
318 if (dev->states_usage[i].disable)
322 * Update if a deeper state would have been a
323 * better match for the observed idle duration.
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);
334 dev->last_residency_ns = 0;
335 dev->states_usage[index].rejected++;
338 instrumentation_end();
340 return entered_state;
344 * cpuidle_select - ask the cpuidle framework to choose an idle state
346 * @drv: the cpuidle driver
347 * @dev: the cpuidle device
348 * @stop_tick: indication on whether or not to stop the tick
350 * Returns the index of the idle state. The return value must not be negative.
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.
356 int cpuidle_select(struct cpuidle_driver *drv, struct cpuidle_device *dev,
359 return cpuidle_curr_governor->select(drv, dev, stop_tick);
363 * cpuidle_enter - enter into the specified idle state
365 * @drv: the cpuidle driver tied with the cpu
366 * @dev: the cpuidle device
367 * @index: the index in the idle state table
369 * Returns the index in the idle state, < 0 in case of error.
370 * The error code depends on the backend driver
372 int cpuidle_enter(struct cpuidle_driver *drv, struct cpuidle_device *dev,
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.
383 WRITE_ONCE(dev->next_hrtimer, tick_nohz_get_next_hrtimer());
385 if (cpuidle_state_is_coupled(drv, index))
386 ret = cpuidle_enter_state_coupled(dev, drv, index);
388 ret = cpuidle_enter_state(dev, drv, index);
390 WRITE_ONCE(dev->next_hrtimer, 0);
395 * cpuidle_reflect - tell the underlying governor what was the state
398 * @dev : the cpuidle device
399 * @index: the index in the idle state table
402 void cpuidle_reflect(struct cpuidle_device *dev, int index)
404 if (cpuidle_curr_governor->reflect && index >= 0)
405 cpuidle_curr_governor->reflect(dev, index);
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
414 * perf bench sched pipe -l 10000
416 * Run multiple times to avoid cpufreq effects.
418 #define CPUIDLE_POLL_MIN 10000
419 #define CPUIDLE_POLL_MAX (TICK_NSEC / 16)
422 * cpuidle_poll_time - return amount of time to poll for,
423 * governors can override dev->poll_limit_ns if necessary
425 * @drv: the cpuidle driver tied with the cpu
426 * @dev: the cpuidle device
429 __cpuidle u64 cpuidle_poll_time(struct cpuidle_driver *drv,
430 struct cpuidle_device *dev)
435 BUILD_BUG_ON(CPUIDLE_POLL_MIN > CPUIDLE_POLL_MAX);
437 if (dev->poll_limit_ns)
438 return dev->poll_limit_ns;
440 limit_ns = CPUIDLE_POLL_MAX;
441 for (i = 1; i < drv->state_count; i++) {
444 if (dev->states_usage[i].disable)
447 state_limit = drv->states[i].target_residency_ns;
448 if (state_limit < CPUIDLE_POLL_MIN)
451 limit_ns = min_t(u64, state_limit, CPUIDLE_POLL_MAX);
455 dev->poll_limit_ns = limit_ns;
457 return dev->poll_limit_ns;
461 * cpuidle_install_idle_handler - installs the cpuidle idle loop handler
463 void cpuidle_install_idle_handler(void)
465 if (enabled_devices) {
466 /* Make sure all changes finished before we switch to new idle */
473 * cpuidle_uninstall_idle_handler - uninstalls the cpuidle idle loop handler
475 void cpuidle_uninstall_idle_handler(void)
477 if (enabled_devices) {
479 wake_up_all_idle_cpus();
483 * Make sure external observers (such as the scheduler)
484 * are done looking at pointed idle states.
490 * cpuidle_pause_and_lock - temporarily disables CPUIDLE
492 void cpuidle_pause_and_lock(void)
494 mutex_lock(&cpuidle_lock);
495 cpuidle_uninstall_idle_handler();
498 EXPORT_SYMBOL_GPL(cpuidle_pause_and_lock);
501 * cpuidle_resume_and_unlock - resumes CPUIDLE operation
503 void cpuidle_resume_and_unlock(void)
505 cpuidle_install_idle_handler();
506 mutex_unlock(&cpuidle_lock);
509 EXPORT_SYMBOL_GPL(cpuidle_resume_and_unlock);
511 /* Currently used in suspend/resume path to suspend cpuidle */
512 void cpuidle_pause(void)
514 mutex_lock(&cpuidle_lock);
515 cpuidle_uninstall_idle_handler();
516 mutex_unlock(&cpuidle_lock);
519 /* Currently used in suspend/resume path to resume cpuidle */
520 void cpuidle_resume(void)
522 mutex_lock(&cpuidle_lock);
523 cpuidle_install_idle_handler();
524 mutex_unlock(&cpuidle_lock);
528 * cpuidle_enable_device - enables idle PM for a CPU
531 * This function must be called between cpuidle_pause_and_lock and
532 * cpuidle_resume_and_unlock when used externally.
534 int cpuidle_enable_device(struct cpuidle_device *dev)
537 struct cpuidle_driver *drv;
545 if (!cpuidle_curr_governor)
548 drv = cpuidle_get_cpu_driver(dev);
553 if (!dev->registered)
556 ret = cpuidle_add_device_sysfs(dev);
560 if (cpuidle_curr_governor->enable) {
561 ret = cpuidle_curr_governor->enable(drv, dev);
574 cpuidle_remove_device_sysfs(dev);
579 EXPORT_SYMBOL_GPL(cpuidle_enable_device);
582 * cpuidle_disable_device - disables idle PM for a CPU
585 * This function must be called between cpuidle_pause_and_lock and
586 * cpuidle_resume_and_unlock when used externally.
588 void cpuidle_disable_device(struct cpuidle_device *dev)
590 struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);
592 if (!dev || !dev->enabled)
595 if (!drv || !cpuidle_curr_governor)
600 if (cpuidle_curr_governor->disable)
601 cpuidle_curr_governor->disable(drv, dev);
603 cpuidle_remove_device_sysfs(dev);
607 EXPORT_SYMBOL_GPL(cpuidle_disable_device);
609 static void __cpuidle_unregister_device(struct cpuidle_device *dev)
611 struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);
613 list_del(&dev->device_list);
614 per_cpu(cpuidle_devices, dev->cpu) = NULL;
615 module_put(drv->owner);
620 static void __cpuidle_device_init(struct cpuidle_device *dev)
622 memset(dev->states_usage, 0, sizeof(dev->states_usage));
623 dev->last_residency_ns = 0;
624 dev->next_hrtimer = 0;
628 * __cpuidle_register_device - internal register function called before register
629 * and enable routines
632 * cpuidle_lock mutex must be held before this is called
634 static int __cpuidle_register_device(struct cpuidle_device *dev)
636 struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);
639 if (!try_module_get(drv->owner))
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;
646 if (drv->states[i].flags & CPUIDLE_FLAG_OFF)
647 dev->states_usage[i].disable |= CPUIDLE_STATE_DISABLED_BY_USER;
650 per_cpu(cpuidle_devices, dev->cpu) = dev;
651 list_add(&dev->device_list, &cpuidle_detected_devices);
653 ret = cpuidle_coupled_register_device(dev);
655 __cpuidle_unregister_device(dev);
663 * cpuidle_register_device - registers a CPU's idle PM feature
666 int cpuidle_register_device(struct cpuidle_device *dev)
673 mutex_lock(&cpuidle_lock);
678 __cpuidle_device_init(dev);
680 ret = __cpuidle_register_device(dev);
684 ret = cpuidle_add_sysfs(dev);
688 ret = cpuidle_enable_device(dev);
692 cpuidle_install_idle_handler();
695 mutex_unlock(&cpuidle_lock);
700 cpuidle_remove_sysfs(dev);
702 __cpuidle_unregister_device(dev);
706 EXPORT_SYMBOL_GPL(cpuidle_register_device);
709 * cpuidle_unregister_device - unregisters a CPU's idle PM feature
712 void cpuidle_unregister_device(struct cpuidle_device *dev)
714 if (!dev || dev->registered == 0)
717 cpuidle_pause_and_lock();
719 cpuidle_disable_device(dev);
721 cpuidle_remove_sysfs(dev);
723 __cpuidle_unregister_device(dev);
725 cpuidle_coupled_unregister_device(dev);
727 cpuidle_resume_and_unlock();
730 EXPORT_SYMBOL_GPL(cpuidle_unregister_device);
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.
737 * @drv: a valid pointer to a struct cpuidle_driver
739 void cpuidle_unregister(struct cpuidle_driver *drv)
742 struct cpuidle_device *device;
744 for_each_cpu(cpu, drv->cpumask) {
745 device = &per_cpu(cpuidle_dev, cpu);
746 cpuidle_unregister_device(device);
749 cpuidle_unregister_driver(drv);
751 EXPORT_SYMBOL_GPL(cpuidle_unregister);
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.
759 * @drv : a valid pointer to a struct cpuidle_driver
760 * @coupled_cpus: a cpumask for the coupled states
762 * Returns 0 on success, < 0 otherwise
764 int cpuidle_register(struct cpuidle_driver *drv,
765 const struct cpumask *const coupled_cpus)
768 struct cpuidle_device *device;
770 ret = cpuidle_register_driver(drv);
772 pr_err("failed to register cpuidle driver\n");
776 for_each_cpu(cpu, drv->cpumask) {
777 device = &per_cpu(cpuidle_dev, cpu);
780 #ifdef CONFIG_ARCH_NEEDS_CPU_IDLE_COUPLED
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.
787 device->coupled_cpus = *coupled_cpus;
789 ret = cpuidle_register_device(device);
793 pr_err("Failed to register cpuidle device for cpu%d\n", cpu);
795 cpuidle_unregister(drv);
801 EXPORT_SYMBOL_GPL(cpuidle_register);
804 * cpuidle_init - core initializer
806 static int __init cpuidle_init(void)
808 if (cpuidle_disabled())
811 return cpuidle_add_interface();
814 module_param(off, int, 0444);
815 module_param_string(governor, param_governor, CPUIDLE_NAME_LEN, 0444);
816 core_initcall(cpuidle_init);