cpuidle.c revision b25edc42bfb9602f0503474b2c94701d5536ce60 
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_params.h>
16#include <linux/cpu.h>
17#include <linux/cpuidle.h>
18#include <linux/ktime.h>
19#include <linux/hrtimer.h>
20#include <trace/events/power.h>
21
22#include "cpuidle.h"
23
24DEFINE_PER_CPU(struct cpuidle_device *, cpuidle_devices);
25
26DEFINE_MUTEX(cpuidle_lock);
27LIST_HEAD(cpuidle_detected_devices);
28
29static int enabled_devices;
30static int off __read_mostly;
31static int initialized __read_mostly;
32
33int cpuidle_disabled(void)
34{
35	return off;
36}
37void disable_cpuidle(void)
38{
39	off = 1;
40}
41
42#if defined(CONFIG_ARCH_HAS_CPU_IDLE_WAIT)
43static void cpuidle_kick_cpus(void)
44{
45	cpu_idle_wait();
46}
47#elif defined(CONFIG_SMP)
48# error "Arch needs cpu_idle_wait() equivalent here"
49#else /* !CONFIG_ARCH_HAS_CPU_IDLE_WAIT && !CONFIG_SMP */
50static void cpuidle_kick_cpus(void) {}
51#endif
52
53static int __cpuidle_register_device(struct cpuidle_device *dev);
54
55/**
56 * cpuidle_idle_call - the main idle loop
57 *
58 * NOTE: no locks or semaphores should be used here
59 * return non-zero on failure
60 */
61int cpuidle_idle_call(void)
62{
63	struct cpuidle_device *dev = __this_cpu_read(cpuidle_devices);
64	struct cpuidle_state *target_state;
65	int next_state, entered_state;
66
67	if (off)
68		return -ENODEV;
69
70	if (!initialized)
71		return -ENODEV;
72
73	/* check if the device is ready */
74	if (!dev || !dev->enabled)
75		return -EBUSY;
76
77#if 0
78	/* shows regressions, re-enable for 2.6.29 */
79	/*
80	 * run any timers that can be run now, at this point
81	 * before calculating the idle duration etc.
82	 */
83	hrtimer_peek_ahead_timers();
84#endif
85
86	/* ask the governor for the next state */
87	next_state = cpuidle_curr_governor->select(dev);
88	if (need_resched()) {
89		local_irq_enable();
90		return 0;
91	}
92
93	target_state = &dev->states[next_state];
94
95	trace_power_start(POWER_CSTATE, next_state, dev->cpu);
96	trace_cpu_idle(next_state, dev->cpu);
97
98	entered_state = target_state->enter(dev, next_state);
99
100	trace_power_end(dev->cpu);
101	trace_cpu_idle(PWR_EVENT_EXIT, dev->cpu);
102
103	if (entered_state >= 0) {
104		/* Update cpuidle counters */
105		/* This can be moved to within driver enter routine
106		 * but that results in multiple copies of same code.
107		 */
108		dev->states[entered_state].time +=
109				(unsigned long long)dev->last_residency;
110		dev->states[entered_state].usage++;
111	}
112
113	/* give the governor an opportunity to reflect on the outcome */
114	if (cpuidle_curr_governor->reflect)
115		cpuidle_curr_governor->reflect(dev, entered_state);
116
117	return 0;
118}
119
120/**
121 * cpuidle_install_idle_handler - installs the cpuidle idle loop handler
122 */
123void cpuidle_install_idle_handler(void)
124{
125	if (enabled_devices) {
126		/* Make sure all changes finished before we switch to new idle */
127		smp_wmb();
128		initialized = 1;
129	}
130}
131
132/**
133 * cpuidle_uninstall_idle_handler - uninstalls the cpuidle idle loop handler
134 */
135void cpuidle_uninstall_idle_handler(void)
136{
137	if (enabled_devices) {
138		initialized = 0;
139		cpuidle_kick_cpus();
140	}
141}
142
143/**
144 * cpuidle_pause_and_lock - temporarily disables CPUIDLE
145 */
146void cpuidle_pause_and_lock(void)
147{
148	mutex_lock(&cpuidle_lock);
149	cpuidle_uninstall_idle_handler();
150}
151
152EXPORT_SYMBOL_GPL(cpuidle_pause_and_lock);
153
154/**
155 * cpuidle_resume_and_unlock - resumes CPUIDLE operation
156 */
157void cpuidle_resume_and_unlock(void)
158{
159	cpuidle_install_idle_handler();
160	mutex_unlock(&cpuidle_lock);
161}
162
163EXPORT_SYMBOL_GPL(cpuidle_resume_and_unlock);
164
165#ifdef CONFIG_ARCH_HAS_CPU_RELAX
166static int poll_idle(struct cpuidle_device *dev, int index)
167{
168	ktime_t	t1, t2;
169	s64 diff;
170
171	t1 = ktime_get();
172	local_irq_enable();
173	while (!need_resched())
174		cpu_relax();
175
176	t2 = ktime_get();
177	diff = ktime_to_us(ktime_sub(t2, t1));
178	if (diff > INT_MAX)
179		diff = INT_MAX;
180
181	dev->last_residency = (int) diff;
182
183	return index;
184}
185
186static void poll_idle_init(struct cpuidle_device *dev)
187{
188	struct cpuidle_state *state = &dev->states[0];
189
190	cpuidle_set_statedata(state, NULL);
191
192	snprintf(state->name, CPUIDLE_NAME_LEN, "POLL");
193	snprintf(state->desc, CPUIDLE_DESC_LEN, "CPUIDLE CORE POLL IDLE");
194	state->exit_latency = 0;
195	state->target_residency = 0;
196	state->power_usage = -1;
197	state->flags = 0;
198	state->enter = poll_idle;
199}
200#else
201static void poll_idle_init(struct cpuidle_device *dev) {}
202#endif /* CONFIG_ARCH_HAS_CPU_RELAX */
203
204/**
205 * cpuidle_enable_device - enables idle PM for a CPU
206 * @dev: the CPU
207 *
208 * This function must be called between cpuidle_pause_and_lock and
209 * cpuidle_resume_and_unlock when used externally.
210 */
211int cpuidle_enable_device(struct cpuidle_device *dev)
212{
213	int ret, i;
214
215	if (dev->enabled)
216		return 0;
217	if (!cpuidle_get_driver() || !cpuidle_curr_governor)
218		return -EIO;
219	if (!dev->state_count)
220		return -EINVAL;
221
222	if (dev->registered == 0) {
223		ret = __cpuidle_register_device(dev);
224		if (ret)
225			return ret;
226	}
227
228	poll_idle_init(dev);
229
230	if ((ret = cpuidle_add_state_sysfs(dev)))
231		return ret;
232
233	if (cpuidle_curr_governor->enable &&
234	    (ret = cpuidle_curr_governor->enable(dev)))
235		goto fail_sysfs;
236
237	for (i = 0; i < dev->state_count; i++) {
238		dev->states[i].usage = 0;
239		dev->states[i].time = 0;
240	}
241	dev->last_residency = 0;
242
243	smp_wmb();
244
245	dev->enabled = 1;
246
247	enabled_devices++;
248	return 0;
249
250fail_sysfs:
251	cpuidle_remove_state_sysfs(dev);
252
253	return ret;
254}
255
256EXPORT_SYMBOL_GPL(cpuidle_enable_device);
257
258/**
259 * cpuidle_disable_device - disables idle PM for a CPU
260 * @dev: the CPU
261 *
262 * This function must be called between cpuidle_pause_and_lock and
263 * cpuidle_resume_and_unlock when used externally.
264 */
265void cpuidle_disable_device(struct cpuidle_device *dev)
266{
267	if (!dev->enabled)
268		return;
269	if (!cpuidle_get_driver() || !cpuidle_curr_governor)
270		return;
271
272	dev->enabled = 0;
273
274	if (cpuidle_curr_governor->disable)
275		cpuidle_curr_governor->disable(dev);
276
277	cpuidle_remove_state_sysfs(dev);
278	enabled_devices--;
279}
280
281EXPORT_SYMBOL_GPL(cpuidle_disable_device);
282
283/**
284 * __cpuidle_register_device - internal register function called before register
285 * and enable routines
286 * @dev: the cpu
287 *
288 * cpuidle_lock mutex must be held before this is called
289 */
290static int __cpuidle_register_device(struct cpuidle_device *dev)
291{
292	int ret;
293	struct sys_device *sys_dev = get_cpu_sysdev((unsigned long)dev->cpu);
294	struct cpuidle_driver *cpuidle_driver = cpuidle_get_driver();
295
296	if (!sys_dev)
297		return -EINVAL;
298	if (!try_module_get(cpuidle_driver->owner))
299		return -EINVAL;
300
301	init_completion(&dev->kobj_unregister);
302
303	/*
304	 * cpuidle driver should set the dev->power_specified bit
305	 * before registering the device if the driver provides
306	 * power_usage numbers.
307	 *
308	 * For those devices whose ->power_specified is not set,
309	 * we fill in power_usage with decreasing values as the
310	 * cpuidle code has an implicit assumption that state Cn
311	 * uses less power than C(n-1).
312	 *
313	 * With CONFIG_ARCH_HAS_CPU_RELAX, C0 is already assigned
314	 * an power value of -1.  So we use -2, -3, etc, for other
315	 * c-states.
316	 */
317	if (!dev->power_specified) {
318		int i;
319		for (i = CPUIDLE_DRIVER_STATE_START; i < dev->state_count; i++)
320			dev->states[i].power_usage = -1 - i;
321	}
322
323	per_cpu(cpuidle_devices, dev->cpu) = dev;
324	list_add(&dev->device_list, &cpuidle_detected_devices);
325	if ((ret = cpuidle_add_sysfs(sys_dev))) {
326		module_put(cpuidle_driver->owner);
327		return ret;
328	}
329
330	dev->registered = 1;
331	return 0;
332}
333
334/**
335 * cpuidle_register_device - registers a CPU's idle PM feature
336 * @dev: the cpu
337 */
338int cpuidle_register_device(struct cpuidle_device *dev)
339{
340	int ret;
341
342	mutex_lock(&cpuidle_lock);
343
344	if ((ret = __cpuidle_register_device(dev))) {
345		mutex_unlock(&cpuidle_lock);
346		return ret;
347	}
348
349	cpuidle_enable_device(dev);
350	cpuidle_install_idle_handler();
351
352	mutex_unlock(&cpuidle_lock);
353
354	return 0;
355
356}
357
358EXPORT_SYMBOL_GPL(cpuidle_register_device);
359
360/**
361 * cpuidle_unregister_device - unregisters a CPU's idle PM feature
362 * @dev: the cpu
363 */
364void cpuidle_unregister_device(struct cpuidle_device *dev)
365{
366	struct sys_device *sys_dev = get_cpu_sysdev((unsigned long)dev->cpu);
367	struct cpuidle_driver *cpuidle_driver = cpuidle_get_driver();
368
369	if (dev->registered == 0)
370		return;
371
372	cpuidle_pause_and_lock();
373
374	cpuidle_disable_device(dev);
375
376	cpuidle_remove_sysfs(sys_dev);
377	list_del(&dev->device_list);
378	wait_for_completion(&dev->kobj_unregister);
379	per_cpu(cpuidle_devices, dev->cpu) = NULL;
380
381	cpuidle_resume_and_unlock();
382
383	module_put(cpuidle_driver->owner);
384}
385
386EXPORT_SYMBOL_GPL(cpuidle_unregister_device);
387
388#ifdef CONFIG_SMP
389
390static void smp_callback(void *v)
391{
392	/* we already woke the CPU up, nothing more to do */
393}
394
395/*
396 * This function gets called when a part of the kernel has a new latency
397 * requirement.  This means we need to get all processors out of their C-state,
398 * and then recalculate a new suitable C-state. Just do a cross-cpu IPI; that
399 * wakes them all right up.
400 */
401static int cpuidle_latency_notify(struct notifier_block *b,
402		unsigned long l, void *v)
403{
404	smp_call_function(smp_callback, NULL, 1);
405	return NOTIFY_OK;
406}
407
408static struct notifier_block cpuidle_latency_notifier = {
409	.notifier_call = cpuidle_latency_notify,
410};
411
412static inline void latency_notifier_init(struct notifier_block *n)
413{
414	pm_qos_add_notifier(PM_QOS_CPU_DMA_LATENCY, n);
415}
416
417#else /* CONFIG_SMP */
418
419#define latency_notifier_init(x) do { } while (0)
420
421#endif /* CONFIG_SMP */
422
423/**
424 * cpuidle_init - core initializer
425 */
426static int __init cpuidle_init(void)
427{
428	int ret;
429
430	if (cpuidle_disabled())
431		return -ENODEV;
432
433	ret = cpuidle_add_class_sysfs(&cpu_sysdev_class);
434	if (ret)
435		return ret;
436
437	latency_notifier_init(&cpuidle_latency_notifier);
438
439	return 0;
440}
441
442module_param(off, int, 0444);
443core_initcall(cpuidle_init);
444