cpuidle.c revision 3af272ab75c7a0c7fa5ae5507724d961f7e7718b
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/kernel.h>
12#include <linux/mutex.h>
13#include <linux/sched.h>
14#include <linux/notifier.h>
15#include <linux/pm_qos.h>
16#include <linux/cpu.h>
17#include <linux/cpuidle.h>
18#include <linux/ktime.h>
19#include <linux/hrtimer.h>
20#include <linux/module.h>
21#include <trace/events/power.h>
22
23#include "cpuidle.h"
24
25DEFINE_PER_CPU(struct cpuidle_device *, cpuidle_devices);
26
27DEFINE_MUTEX(cpuidle_lock);
28LIST_HEAD(cpuidle_detected_devices);
29
30static int enabled_devices;
31static int off __read_mostly;
32static int initialized __read_mostly;
33
34int cpuidle_disabled(void)
35{
36	return off;
37}
38void disable_cpuidle(void)
39{
40	off = 1;
41}
42
43#if defined(CONFIG_ARCH_HAS_CPU_IDLE_WAIT)
44static void cpuidle_kick_cpus(void)
45{
46	cpu_idle_wait();
47}
48#elif defined(CONFIG_SMP)
49# error "Arch needs cpu_idle_wait() equivalent here"
50#else /* !CONFIG_ARCH_HAS_CPU_IDLE_WAIT && !CONFIG_SMP */
51static void cpuidle_kick_cpus(void) {}
52#endif
53
54static int __cpuidle_register_device(struct cpuidle_device *dev);
55
56static inline int cpuidle_enter(struct cpuidle_device *dev,
57				struct cpuidle_driver *drv, int index)
58{
59	struct cpuidle_state *target_state = &drv->states[index];
60	return target_state->enter(dev, drv, index);
61}
62
63static inline int cpuidle_enter_tk(struct cpuidle_device *dev,
64			       struct cpuidle_driver *drv, int index)
65{
66	return cpuidle_wrap_enter(dev, drv, index, cpuidle_enter);
67}
68
69typedef int (*cpuidle_enter_t)(struct cpuidle_device *dev,
70			       struct cpuidle_driver *drv, int index);
71
72static cpuidle_enter_t cpuidle_enter_ops;
73
74/**
75 * cpuidle_play_dead - cpu off-lining
76 *
77 * Returns in case of an error or no driver
78 */
79int cpuidle_play_dead(void)
80{
81	struct cpuidle_device *dev = __this_cpu_read(cpuidle_devices);
82	struct cpuidle_driver *drv = cpuidle_get_driver();
83	int i, dead_state = -1;
84	int power_usage = -1;
85
86	if (!drv)
87		return -ENODEV;
88
89	/* Find lowest-power state that supports long-term idle */
90	for (i = CPUIDLE_DRIVER_STATE_START; i < drv->state_count; i++) {
91		struct cpuidle_state *s = &drv->states[i];
92
93		if (s->power_usage < power_usage && s->enter_dead) {
94			power_usage = s->power_usage;
95			dead_state = i;
96		}
97	}
98
99	if (dead_state != -1)
100		return drv->states[dead_state].enter_dead(dev, dead_state);
101
102	return -ENODEV;
103}
104
105/**
106 * cpuidle_enter_state - enter the state and update stats
107 * @dev: cpuidle device for this cpu
108 * @drv: cpuidle driver for this cpu
109 * @next_state: index into drv->states of the state to enter
110 */
111int cpuidle_enter_state(struct cpuidle_device *dev, struct cpuidle_driver *drv,
112		int next_state)
113{
114	int entered_state;
115
116	entered_state = cpuidle_enter_ops(dev, drv, next_state);
117
118	if (entered_state >= 0) {
119		/* Update cpuidle counters */
120		/* This can be moved to within driver enter routine
121		 * but that results in multiple copies of same code.
122		 */
123		dev->states_usage[entered_state].time +=
124				(unsigned long long)dev->last_residency;
125		dev->states_usage[entered_state].usage++;
126	} else {
127		dev->last_residency = 0;
128	}
129
130	return entered_state;
131}
132
133/**
134 * cpuidle_idle_call - the main idle loop
135 *
136 * NOTE: no locks or semaphores should be used here
137 * return non-zero on failure
138 */
139int cpuidle_idle_call(void)
140{
141	struct cpuidle_device *dev = __this_cpu_read(cpuidle_devices);
142	struct cpuidle_driver *drv = cpuidle_get_driver();
143	int next_state, entered_state;
144
145	if (off)
146		return -ENODEV;
147
148	if (!initialized)
149		return -ENODEV;
150
151	/* check if the device is ready */
152	if (!dev || !dev->enabled)
153		return -EBUSY;
154
155#if 0
156	/* shows regressions, re-enable for 2.6.29 */
157	/*
158	 * run any timers that can be run now, at this point
159	 * before calculating the idle duration etc.
160	 */
161	hrtimer_peek_ahead_timers();
162#endif
163
164	/* ask the governor for the next state */
165	next_state = cpuidle_curr_governor->select(drv, dev);
166	if (need_resched()) {
167		local_irq_enable();
168		return 0;
169	}
170
171	trace_power_start_rcuidle(POWER_CSTATE, next_state, dev->cpu);
172	trace_cpu_idle_rcuidle(next_state, dev->cpu);
173
174	entered_state = cpuidle_enter_state(dev, drv, next_state);
175
176	trace_power_end_rcuidle(dev->cpu);
177	trace_cpu_idle_rcuidle(PWR_EVENT_EXIT, dev->cpu);
178
179	/* give the governor an opportunity to reflect on the outcome */
180	if (cpuidle_curr_governor->reflect)
181		cpuidle_curr_governor->reflect(dev, entered_state);
182
183	return 0;
184}
185
186/**
187 * cpuidle_install_idle_handler - installs the cpuidle idle loop handler
188 */
189void cpuidle_install_idle_handler(void)
190{
191	if (enabled_devices) {
192		/* Make sure all changes finished before we switch to new idle */
193		smp_wmb();
194		initialized = 1;
195	}
196}
197
198/**
199 * cpuidle_uninstall_idle_handler - uninstalls the cpuidle idle loop handler
200 */
201void cpuidle_uninstall_idle_handler(void)
202{
203	if (enabled_devices) {
204		initialized = 0;
205		cpuidle_kick_cpus();
206	}
207}
208
209/**
210 * cpuidle_pause_and_lock - temporarily disables CPUIDLE
211 */
212void cpuidle_pause_and_lock(void)
213{
214	mutex_lock(&cpuidle_lock);
215	cpuidle_uninstall_idle_handler();
216}
217
218EXPORT_SYMBOL_GPL(cpuidle_pause_and_lock);
219
220/**
221 * cpuidle_resume_and_unlock - resumes CPUIDLE operation
222 */
223void cpuidle_resume_and_unlock(void)
224{
225	cpuidle_install_idle_handler();
226	mutex_unlock(&cpuidle_lock);
227}
228
229EXPORT_SYMBOL_GPL(cpuidle_resume_and_unlock);
230
231/**
232 * cpuidle_wrap_enter - performs timekeeping and irqen around enter function
233 * @dev: pointer to a valid cpuidle_device object
234 * @drv: pointer to a valid cpuidle_driver object
235 * @index: index of the target cpuidle state.
236 */
237int cpuidle_wrap_enter(struct cpuidle_device *dev,
238				struct cpuidle_driver *drv, int index,
239				int (*enter)(struct cpuidle_device *dev,
240					struct cpuidle_driver *drv, int index))
241{
242	ktime_t time_start, time_end;
243	s64 diff;
244
245	time_start = ktime_get();
246
247	index = enter(dev, drv, index);
248
249	time_end = ktime_get();
250
251	local_irq_enable();
252
253	diff = ktime_to_us(ktime_sub(time_end, time_start));
254	if (diff > INT_MAX)
255		diff = INT_MAX;
256
257	dev->last_residency = (int) diff;
258
259	return index;
260}
261
262#ifdef CONFIG_ARCH_HAS_CPU_RELAX
263static int poll_idle(struct cpuidle_device *dev,
264		struct cpuidle_driver *drv, int index)
265{
266	ktime_t	t1, t2;
267	s64 diff;
268
269	t1 = ktime_get();
270	local_irq_enable();
271	while (!need_resched())
272		cpu_relax();
273
274	t2 = ktime_get();
275	diff = ktime_to_us(ktime_sub(t2, t1));
276	if (diff > INT_MAX)
277		diff = INT_MAX;
278
279	dev->last_residency = (int) diff;
280
281	return index;
282}
283
284static void poll_idle_init(struct cpuidle_driver *drv)
285{
286	struct cpuidle_state *state = &drv->states[0];
287
288	snprintf(state->name, CPUIDLE_NAME_LEN, "POLL");
289	snprintf(state->desc, CPUIDLE_DESC_LEN, "CPUIDLE CORE POLL IDLE");
290	state->exit_latency = 0;
291	state->target_residency = 0;
292	state->power_usage = -1;
293	state->flags = 0;
294	state->enter = poll_idle;
295	state->disable = 0;
296}
297#else
298static void poll_idle_init(struct cpuidle_driver *drv) {}
299#endif /* CONFIG_ARCH_HAS_CPU_RELAX */
300
301/**
302 * cpuidle_enable_device - enables idle PM for a CPU
303 * @dev: the CPU
304 *
305 * This function must be called between cpuidle_pause_and_lock and
306 * cpuidle_resume_and_unlock when used externally.
307 */
308int cpuidle_enable_device(struct cpuidle_device *dev)
309{
310	int ret, i;
311	struct cpuidle_driver *drv = cpuidle_get_driver();
312
313	if (dev->enabled)
314		return 0;
315	if (!drv || !cpuidle_curr_governor)
316		return -EIO;
317	if (!dev->state_count)
318		dev->state_count = drv->state_count;
319
320	if (dev->registered == 0) {
321		ret = __cpuidle_register_device(dev);
322		if (ret)
323			return ret;
324	}
325
326	cpuidle_enter_ops = drv->en_core_tk_irqen ?
327		cpuidle_enter_tk : cpuidle_enter;
328
329	poll_idle_init(drv);
330
331	if ((ret = cpuidle_add_state_sysfs(dev)))
332		return ret;
333
334	if (cpuidle_curr_governor->enable &&
335	    (ret = cpuidle_curr_governor->enable(drv, dev)))
336		goto fail_sysfs;
337
338	for (i = 0; i < dev->state_count; i++) {
339		dev->states_usage[i].usage = 0;
340		dev->states_usage[i].time = 0;
341	}
342	dev->last_residency = 0;
343
344	smp_wmb();
345
346	dev->enabled = 1;
347
348	enabled_devices++;
349	return 0;
350
351fail_sysfs:
352	cpuidle_remove_state_sysfs(dev);
353
354	return ret;
355}
356
357EXPORT_SYMBOL_GPL(cpuidle_enable_device);
358
359/**
360 * cpuidle_disable_device - disables idle PM for a CPU
361 * @dev: the CPU
362 *
363 * This function must be called between cpuidle_pause_and_lock and
364 * cpuidle_resume_and_unlock when used externally.
365 */
366void cpuidle_disable_device(struct cpuidle_device *dev)
367{
368	if (!dev->enabled)
369		return;
370	if (!cpuidle_get_driver() || !cpuidle_curr_governor)
371		return;
372
373	dev->enabled = 0;
374
375	if (cpuidle_curr_governor->disable)
376		cpuidle_curr_governor->disable(cpuidle_get_driver(), dev);
377
378	cpuidle_remove_state_sysfs(dev);
379	enabled_devices--;
380}
381
382EXPORT_SYMBOL_GPL(cpuidle_disable_device);
383
384/**
385 * __cpuidle_register_device - internal register function called before register
386 * and enable routines
387 * @dev: the cpu
388 *
389 * cpuidle_lock mutex must be held before this is called
390 */
391static int __cpuidle_register_device(struct cpuidle_device *dev)
392{
393	int ret;
394	struct device *cpu_dev = get_cpu_device((unsigned long)dev->cpu);
395	struct cpuidle_driver *cpuidle_driver = cpuidle_get_driver();
396
397	if (!dev)
398		return -EINVAL;
399	if (!try_module_get(cpuidle_driver->owner))
400		return -EINVAL;
401
402	init_completion(&dev->kobj_unregister);
403
404	per_cpu(cpuidle_devices, dev->cpu) = dev;
405	list_add(&dev->device_list, &cpuidle_detected_devices);
406	ret = cpuidle_add_sysfs(cpu_dev);
407	if (ret)
408		goto err_sysfs;
409
410	dev->registered = 1;
411	return 0;
412
413err_sysfs:
414	list_del(&dev->device_list);
415	per_cpu(cpuidle_devices, dev->cpu) = NULL;
416	module_put(cpuidle_driver->owner);
417	return ret;
418}
419
420/**
421 * cpuidle_register_device - registers a CPU's idle PM feature
422 * @dev: the cpu
423 */
424int cpuidle_register_device(struct cpuidle_device *dev)
425{
426	int ret;
427
428	mutex_lock(&cpuidle_lock);
429
430	if ((ret = __cpuidle_register_device(dev))) {
431		mutex_unlock(&cpuidle_lock);
432		return ret;
433	}
434
435	cpuidle_enable_device(dev);
436	cpuidle_install_idle_handler();
437
438	mutex_unlock(&cpuidle_lock);
439
440	return 0;
441
442}
443
444EXPORT_SYMBOL_GPL(cpuidle_register_device);
445
446/**
447 * cpuidle_unregister_device - unregisters a CPU's idle PM feature
448 * @dev: the cpu
449 */
450void cpuidle_unregister_device(struct cpuidle_device *dev)
451{
452	struct device *cpu_dev = get_cpu_device((unsigned long)dev->cpu);
453	struct cpuidle_driver *cpuidle_driver = cpuidle_get_driver();
454
455	if (dev->registered == 0)
456		return;
457
458	cpuidle_pause_and_lock();
459
460	cpuidle_disable_device(dev);
461
462	cpuidle_remove_sysfs(cpu_dev);
463	list_del(&dev->device_list);
464	wait_for_completion(&dev->kobj_unregister);
465	per_cpu(cpuidle_devices, dev->cpu) = NULL;
466
467	cpuidle_resume_and_unlock();
468
469	module_put(cpuidle_driver->owner);
470}
471
472EXPORT_SYMBOL_GPL(cpuidle_unregister_device);
473
474#ifdef CONFIG_SMP
475
476static void smp_callback(void *v)
477{
478	/* we already woke the CPU up, nothing more to do */
479}
480
481/*
482 * This function gets called when a part of the kernel has a new latency
483 * requirement.  This means we need to get all processors out of their C-state,
484 * and then recalculate a new suitable C-state. Just do a cross-cpu IPI; that
485 * wakes them all right up.
486 */
487static int cpuidle_latency_notify(struct notifier_block *b,
488		unsigned long l, void *v)
489{
490	smp_call_function(smp_callback, NULL, 1);
491	return NOTIFY_OK;
492}
493
494static struct notifier_block cpuidle_latency_notifier = {
495	.notifier_call = cpuidle_latency_notify,
496};
497
498static inline void latency_notifier_init(struct notifier_block *n)
499{
500	pm_qos_add_notifier(PM_QOS_CPU_DMA_LATENCY, n);
501}
502
503#else /* CONFIG_SMP */
504
505#define latency_notifier_init(x) do { } while (0)
506
507#endif /* CONFIG_SMP */
508
509/**
510 * cpuidle_init - core initializer
511 */
512static int __init cpuidle_init(void)
513{
514	int ret;
515
516	if (cpuidle_disabled())
517		return -ENODEV;
518
519	ret = cpuidle_add_interface(cpu_subsys.dev_root);
520	if (ret)
521		return ret;
522
523	latency_notifier_init(&cpuidle_latency_notifier);
524
525	return 0;
526}
527
528module_param(off, int, 0444);
529core_initcall(cpuidle_init);
530