cpuidle.c revision e8db0be1245de16a6cc6365506abc392c3c212d4 
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 <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;
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	/*
87	 * Call the device's prepare function before calling the
88	 * governor's select function.  ->prepare gives the device's
89	 * cpuidle driver a chance to update any dynamic information
90	 * of its cpuidle states for the current idle period, e.g.
91	 * state availability, latencies, residencies, etc.
92	 */
93	if (dev->prepare)
94		dev->prepare(dev);
95
96	/* ask the governor for the next state */
97	next_state = cpuidle_curr_governor->select(dev);
98	if (need_resched()) {
99		local_irq_enable();
100		return 0;
101	}
102
103	target_state = &dev->states[next_state];
104
105	/* enter the state and update stats */
106	dev->last_state = target_state;
107
108	trace_power_start(POWER_CSTATE, next_state, dev->cpu);
109	trace_cpu_idle(next_state, dev->cpu);
110
111	dev->last_residency = target_state->enter(dev, target_state);
112
113	trace_power_end(dev->cpu);
114	trace_cpu_idle(PWR_EVENT_EXIT, dev->cpu);
115
116	if (dev->last_state)
117		target_state = dev->last_state;
118
119	target_state->time += (unsigned long long)dev->last_residency;
120	target_state->usage++;
121
122	/* give the governor an opportunity to reflect on the outcome */
123	if (cpuidle_curr_governor->reflect)
124		cpuidle_curr_governor->reflect(dev);
125
126	return 0;
127}
128
129/**
130 * cpuidle_install_idle_handler - installs the cpuidle idle loop handler
131 */
132void cpuidle_install_idle_handler(void)
133{
134	if (enabled_devices) {
135		/* Make sure all changes finished before we switch to new idle */
136		smp_wmb();
137		initialized = 1;
138	}
139}
140
141/**
142 * cpuidle_uninstall_idle_handler - uninstalls the cpuidle idle loop handler
143 */
144void cpuidle_uninstall_idle_handler(void)
145{
146	if (enabled_devices) {
147		initialized = 0;
148		cpuidle_kick_cpus();
149	}
150}
151
152/**
153 * cpuidle_pause_and_lock - temporarily disables CPUIDLE
154 */
155void cpuidle_pause_and_lock(void)
156{
157	mutex_lock(&cpuidle_lock);
158	cpuidle_uninstall_idle_handler();
159}
160
161EXPORT_SYMBOL_GPL(cpuidle_pause_and_lock);
162
163/**
164 * cpuidle_resume_and_unlock - resumes CPUIDLE operation
165 */
166void cpuidle_resume_and_unlock(void)
167{
168	cpuidle_install_idle_handler();
169	mutex_unlock(&cpuidle_lock);
170}
171
172EXPORT_SYMBOL_GPL(cpuidle_resume_and_unlock);
173
174#ifdef CONFIG_ARCH_HAS_CPU_RELAX
175static int poll_idle(struct cpuidle_device *dev, struct cpuidle_state *st)
176{
177	ktime_t	t1, t2;
178	s64 diff;
179	int ret;
180
181	t1 = ktime_get();
182	local_irq_enable();
183	while (!need_resched())
184		cpu_relax();
185
186	t2 = ktime_get();
187	diff = ktime_to_us(ktime_sub(t2, t1));
188	if (diff > INT_MAX)
189		diff = INT_MAX;
190
191	ret = (int) diff;
192	return ret;
193}
194
195static void poll_idle_init(struct cpuidle_device *dev)
196{
197	struct cpuidle_state *state = &dev->states[0];
198
199	cpuidle_set_statedata(state, NULL);
200
201	snprintf(state->name, CPUIDLE_NAME_LEN, "POLL");
202	snprintf(state->desc, CPUIDLE_DESC_LEN, "CPUIDLE CORE POLL IDLE");
203	state->exit_latency = 0;
204	state->target_residency = 0;
205	state->power_usage = -1;
206	state->flags = 0;
207	state->enter = poll_idle;
208}
209#else
210static void poll_idle_init(struct cpuidle_device *dev) {}
211#endif /* CONFIG_ARCH_HAS_CPU_RELAX */
212
213/**
214 * cpuidle_enable_device - enables idle PM for a CPU
215 * @dev: the CPU
216 *
217 * This function must be called between cpuidle_pause_and_lock and
218 * cpuidle_resume_and_unlock when used externally.
219 */
220int cpuidle_enable_device(struct cpuidle_device *dev)
221{
222	int ret, i;
223
224	if (dev->enabled)
225		return 0;
226	if (!cpuidle_get_driver() || !cpuidle_curr_governor)
227		return -EIO;
228	if (!dev->state_count)
229		return -EINVAL;
230
231	if (dev->registered == 0) {
232		ret = __cpuidle_register_device(dev);
233		if (ret)
234			return ret;
235	}
236
237	poll_idle_init(dev);
238
239	if ((ret = cpuidle_add_state_sysfs(dev)))
240		return ret;
241
242	if (cpuidle_curr_governor->enable &&
243	    (ret = cpuidle_curr_governor->enable(dev)))
244		goto fail_sysfs;
245
246	for (i = 0; i < dev->state_count; i++) {
247		dev->states[i].usage = 0;
248		dev->states[i].time = 0;
249	}
250	dev->last_residency = 0;
251	dev->last_state = NULL;
252
253	smp_wmb();
254
255	dev->enabled = 1;
256
257	enabled_devices++;
258	return 0;
259
260fail_sysfs:
261	cpuidle_remove_state_sysfs(dev);
262
263	return ret;
264}
265
266EXPORT_SYMBOL_GPL(cpuidle_enable_device);
267
268/**
269 * cpuidle_disable_device - disables idle PM for a CPU
270 * @dev: the CPU
271 *
272 * This function must be called between cpuidle_pause_and_lock and
273 * cpuidle_resume_and_unlock when used externally.
274 */
275void cpuidle_disable_device(struct cpuidle_device *dev)
276{
277	if (!dev->enabled)
278		return;
279	if (!cpuidle_get_driver() || !cpuidle_curr_governor)
280		return;
281
282	dev->enabled = 0;
283
284	if (cpuidle_curr_governor->disable)
285		cpuidle_curr_governor->disable(dev);
286
287	cpuidle_remove_state_sysfs(dev);
288	enabled_devices--;
289}
290
291EXPORT_SYMBOL_GPL(cpuidle_disable_device);
292
293/**
294 * __cpuidle_register_device - internal register function called before register
295 * and enable routines
296 * @dev: the cpu
297 *
298 * cpuidle_lock mutex must be held before this is called
299 */
300static int __cpuidle_register_device(struct cpuidle_device *dev)
301{
302	int ret;
303	struct sys_device *sys_dev = get_cpu_sysdev((unsigned long)dev->cpu);
304	struct cpuidle_driver *cpuidle_driver = cpuidle_get_driver();
305
306	if (!sys_dev)
307		return -EINVAL;
308	if (!try_module_get(cpuidle_driver->owner))
309		return -EINVAL;
310
311	init_completion(&dev->kobj_unregister);
312
313	/*
314	 * cpuidle driver should set the dev->power_specified bit
315	 * before registering the device if the driver provides
316	 * power_usage numbers.
317	 *
318	 * For those devices whose ->power_specified is not set,
319	 * we fill in power_usage with decreasing values as the
320	 * cpuidle code has an implicit assumption that state Cn
321	 * uses less power than C(n-1).
322	 *
323	 * With CONFIG_ARCH_HAS_CPU_RELAX, C0 is already assigned
324	 * an power value of -1.  So we use -2, -3, etc, for other
325	 * c-states.
326	 */
327	if (!dev->power_specified) {
328		int i;
329		for (i = CPUIDLE_DRIVER_STATE_START; i < dev->state_count; i++)
330			dev->states[i].power_usage = -1 - i;
331	}
332
333	per_cpu(cpuidle_devices, dev->cpu) = dev;
334	list_add(&dev->device_list, &cpuidle_detected_devices);
335	if ((ret = cpuidle_add_sysfs(sys_dev))) {
336		module_put(cpuidle_driver->owner);
337		return ret;
338	}
339
340	dev->registered = 1;
341	return 0;
342}
343
344/**
345 * cpuidle_register_device - registers a CPU's idle PM feature
346 * @dev: the cpu
347 */
348int cpuidle_register_device(struct cpuidle_device *dev)
349{
350	int ret;
351
352	mutex_lock(&cpuidle_lock);
353
354	if ((ret = __cpuidle_register_device(dev))) {
355		mutex_unlock(&cpuidle_lock);
356		return ret;
357	}
358
359	cpuidle_enable_device(dev);
360	cpuidle_install_idle_handler();
361
362	mutex_unlock(&cpuidle_lock);
363
364	return 0;
365
366}
367
368EXPORT_SYMBOL_GPL(cpuidle_register_device);
369
370/**
371 * cpuidle_unregister_device - unregisters a CPU's idle PM feature
372 * @dev: the cpu
373 */
374void cpuidle_unregister_device(struct cpuidle_device *dev)
375{
376	struct sys_device *sys_dev = get_cpu_sysdev((unsigned long)dev->cpu);
377	struct cpuidle_driver *cpuidle_driver = cpuidle_get_driver();
378
379	if (dev->registered == 0)
380		return;
381
382	cpuidle_pause_and_lock();
383
384	cpuidle_disable_device(dev);
385
386	cpuidle_remove_sysfs(sys_dev);
387	list_del(&dev->device_list);
388	wait_for_completion(&dev->kobj_unregister);
389	per_cpu(cpuidle_devices, dev->cpu) = NULL;
390
391	cpuidle_resume_and_unlock();
392
393	module_put(cpuidle_driver->owner);
394}
395
396EXPORT_SYMBOL_GPL(cpuidle_unregister_device);
397
398#ifdef CONFIG_SMP
399
400static void smp_callback(void *v)
401{
402	/* we already woke the CPU up, nothing more to do */
403}
404
405/*
406 * This function gets called when a part of the kernel has a new latency
407 * requirement.  This means we need to get all processors out of their C-state,
408 * and then recalculate a new suitable C-state. Just do a cross-cpu IPI; that
409 * wakes them all right up.
410 */
411static int cpuidle_latency_notify(struct notifier_block *b,
412		unsigned long l, void *v)
413{
414	smp_call_function(smp_callback, NULL, 1);
415	return NOTIFY_OK;
416}
417
418static struct notifier_block cpuidle_latency_notifier = {
419	.notifier_call = cpuidle_latency_notify,
420};
421
422static inline void latency_notifier_init(struct notifier_block *n)
423{
424	pm_qos_add_notifier(PM_QOS_CPU_DMA_LATENCY, n);
425}
426
427#else /* CONFIG_SMP */
428
429#define latency_notifier_init(x) do { } while (0)
430
431#endif /* CONFIG_SMP */
432
433/**
434 * cpuidle_init - core initializer
435 */
436static int __init cpuidle_init(void)
437{
438	int ret;
439
440	if (cpuidle_disabled())
441		return -ENODEV;
442
443	ret = cpuidle_add_class_sysfs(&cpu_sysdev_class);
444	if (ret)
445		return ret;
446
447	latency_notifier_init(&cpuidle_latency_notifier);
448
449	return 0;
450}
451
452module_param(off, int, 0444);
453core_initcall(cpuidle_init);
454