xref: /openbmc/linux/drivers/cpuidle/cpuidle.c (revision 54cbac81)
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 
25 DEFINE_PER_CPU(struct cpuidle_device *, cpuidle_devices);
26 
27 DEFINE_MUTEX(cpuidle_lock);
28 LIST_HEAD(cpuidle_detected_devices);
29 
30 static int enabled_devices;
31 static int off __read_mostly;
32 static int initialized __read_mostly;
33 
34 int cpuidle_disabled(void)
35 {
36 	return off;
37 }
38 void disable_cpuidle(void)
39 {
40 	off = 1;
41 }
42 
43 static int __cpuidle_register_device(struct cpuidle_device *dev);
44 
45 static inline int cpuidle_enter(struct cpuidle_device *dev,
46 				struct cpuidle_driver *drv, int index)
47 {
48 	struct cpuidle_state *target_state = &drv->states[index];
49 	return target_state->enter(dev, drv, index);
50 }
51 
52 static inline int cpuidle_enter_tk(struct cpuidle_device *dev,
53 			       struct cpuidle_driver *drv, int index)
54 {
55 	return cpuidle_wrap_enter(dev, drv, index, cpuidle_enter);
56 }
57 
58 typedef int (*cpuidle_enter_t)(struct cpuidle_device *dev,
59 			       struct cpuidle_driver *drv, int index);
60 
61 static cpuidle_enter_t cpuidle_enter_ops;
62 
63 /**
64  * cpuidle_play_dead - cpu off-lining
65  *
66  * Returns in case of an error or no driver
67  */
68 int cpuidle_play_dead(void)
69 {
70 	struct cpuidle_device *dev = __this_cpu_read(cpuidle_devices);
71 	struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);
72 	int i, dead_state = -1;
73 	int power_usage = INT_MAX;
74 
75 	if (!drv)
76 		return -ENODEV;
77 
78 	/* Find lowest-power state that supports long-term idle */
79 	for (i = CPUIDLE_DRIVER_STATE_START; i < drv->state_count; i++) {
80 		struct cpuidle_state *s = &drv->states[i];
81 
82 		if (s->power_usage < power_usage && s->enter_dead) {
83 			power_usage = s->power_usage;
84 			dead_state = i;
85 		}
86 	}
87 
88 	if (dead_state != -1)
89 		return drv->states[dead_state].enter_dead(dev, dead_state);
90 
91 	return -ENODEV;
92 }
93 
94 /**
95  * cpuidle_enter_state - enter the state and update stats
96  * @dev: cpuidle device for this cpu
97  * @drv: cpuidle driver for this cpu
98  * @next_state: index into drv->states of the state to enter
99  */
100 int cpuidle_enter_state(struct cpuidle_device *dev, struct cpuidle_driver *drv,
101 		int next_state)
102 {
103 	int entered_state;
104 
105 	entered_state = cpuidle_enter_ops(dev, drv, next_state);
106 
107 	if (entered_state >= 0) {
108 		/* Update cpuidle counters */
109 		/* This can be moved to within driver enter routine
110 		 * but that results in multiple copies of same code.
111 		 */
112 		dev->states_usage[entered_state].time += dev->last_residency;
113 		dev->states_usage[entered_state].usage++;
114 	} else {
115 		dev->last_residency = 0;
116 	}
117 
118 	return entered_state;
119 }
120 
121 /**
122  * cpuidle_idle_call - the main idle loop
123  *
124  * NOTE: no locks or semaphores should be used here
125  * return non-zero on failure
126  */
127 int cpuidle_idle_call(void)
128 {
129 	struct cpuidle_device *dev = __this_cpu_read(cpuidle_devices);
130 	struct cpuidle_driver *drv;
131 	int next_state, entered_state;
132 
133 	if (off)
134 		return -ENODEV;
135 
136 	if (!initialized)
137 		return -ENODEV;
138 
139 	/* check if the device is ready */
140 	if (!dev || !dev->enabled)
141 		return -EBUSY;
142 
143 	drv = cpuidle_get_cpu_driver(dev);
144 
145 	/* ask the governor for the next state */
146 	next_state = cpuidle_curr_governor->select(drv, dev);
147 	if (need_resched()) {
148 		dev->last_residency = 0;
149 		/* give the governor an opportunity to reflect on the outcome */
150 		if (cpuidle_curr_governor->reflect)
151 			cpuidle_curr_governor->reflect(dev, next_state);
152 		local_irq_enable();
153 		return 0;
154 	}
155 
156 	trace_power_start_rcuidle(POWER_CSTATE, next_state, dev->cpu);
157 	trace_cpu_idle_rcuidle(next_state, dev->cpu);
158 
159 	if (cpuidle_state_is_coupled(dev, drv, next_state))
160 		entered_state = cpuidle_enter_state_coupled(dev, drv,
161 							    next_state);
162 	else
163 		entered_state = cpuidle_enter_state(dev, drv, next_state);
164 
165 	trace_power_end_rcuidle(dev->cpu);
166 	trace_cpu_idle_rcuidle(PWR_EVENT_EXIT, dev->cpu);
167 
168 	/* give the governor an opportunity to reflect on the outcome */
169 	if (cpuidle_curr_governor->reflect)
170 		cpuidle_curr_governor->reflect(dev, entered_state);
171 
172 	return 0;
173 }
174 
175 /**
176  * cpuidle_install_idle_handler - installs the cpuidle idle loop handler
177  */
178 void cpuidle_install_idle_handler(void)
179 {
180 	if (enabled_devices) {
181 		/* Make sure all changes finished before we switch to new idle */
182 		smp_wmb();
183 		initialized = 1;
184 	}
185 }
186 
187 /**
188  * cpuidle_uninstall_idle_handler - uninstalls the cpuidle idle loop handler
189  */
190 void cpuidle_uninstall_idle_handler(void)
191 {
192 	if (enabled_devices) {
193 		initialized = 0;
194 		kick_all_cpus_sync();
195 	}
196 }
197 
198 /**
199  * cpuidle_pause_and_lock - temporarily disables CPUIDLE
200  */
201 void cpuidle_pause_and_lock(void)
202 {
203 	mutex_lock(&cpuidle_lock);
204 	cpuidle_uninstall_idle_handler();
205 }
206 
207 EXPORT_SYMBOL_GPL(cpuidle_pause_and_lock);
208 
209 /**
210  * cpuidle_resume_and_unlock - resumes CPUIDLE operation
211  */
212 void cpuidle_resume_and_unlock(void)
213 {
214 	cpuidle_install_idle_handler();
215 	mutex_unlock(&cpuidle_lock);
216 }
217 
218 EXPORT_SYMBOL_GPL(cpuidle_resume_and_unlock);
219 
220 /* Currently used in suspend/resume path to suspend cpuidle */
221 void cpuidle_pause(void)
222 {
223 	mutex_lock(&cpuidle_lock);
224 	cpuidle_uninstall_idle_handler();
225 	mutex_unlock(&cpuidle_lock);
226 }
227 
228 /* Currently used in suspend/resume path to resume cpuidle */
229 void cpuidle_resume(void)
230 {
231 	mutex_lock(&cpuidle_lock);
232 	cpuidle_install_idle_handler();
233 	mutex_unlock(&cpuidle_lock);
234 }
235 
236 /**
237  * cpuidle_wrap_enter - performs timekeeping and irqen around enter function
238  * @dev: pointer to a valid cpuidle_device object
239  * @drv: pointer to a valid cpuidle_driver object
240  * @index: index of the target cpuidle state.
241  */
242 int cpuidle_wrap_enter(struct cpuidle_device *dev,
243 				struct cpuidle_driver *drv, int index,
244 				int (*enter)(struct cpuidle_device *dev,
245 					struct cpuidle_driver *drv, int index))
246 {
247 	ktime_t time_start, time_end;
248 	s64 diff;
249 
250 	time_start = ktime_get();
251 
252 	index = enter(dev, drv, index);
253 
254 	time_end = ktime_get();
255 
256 	local_irq_enable();
257 
258 	diff = ktime_to_us(ktime_sub(time_end, time_start));
259 	if (diff > INT_MAX)
260 		diff = INT_MAX;
261 
262 	dev->last_residency = (int) diff;
263 
264 	return index;
265 }
266 
267 #ifdef CONFIG_ARCH_HAS_CPU_RELAX
268 static int poll_idle(struct cpuidle_device *dev,
269 		struct cpuidle_driver *drv, int index)
270 {
271 	ktime_t	t1, t2;
272 	s64 diff;
273 
274 	t1 = ktime_get();
275 	local_irq_enable();
276 	while (!need_resched())
277 		cpu_relax();
278 
279 	t2 = ktime_get();
280 	diff = ktime_to_us(ktime_sub(t2, t1));
281 	if (diff > INT_MAX)
282 		diff = INT_MAX;
283 
284 	dev->last_residency = (int) diff;
285 
286 	return index;
287 }
288 
289 static void poll_idle_init(struct cpuidle_driver *drv)
290 {
291 	struct cpuidle_state *state = &drv->states[0];
292 
293 	snprintf(state->name, CPUIDLE_NAME_LEN, "POLL");
294 	snprintf(state->desc, CPUIDLE_DESC_LEN, "CPUIDLE CORE POLL IDLE");
295 	state->exit_latency = 0;
296 	state->target_residency = 0;
297 	state->power_usage = -1;
298 	state->flags = 0;
299 	state->enter = poll_idle;
300 	state->disabled = false;
301 }
302 #else
303 static void poll_idle_init(struct cpuidle_driver *drv) {}
304 #endif /* CONFIG_ARCH_HAS_CPU_RELAX */
305 
306 /**
307  * cpuidle_enable_device - enables idle PM for a CPU
308  * @dev: the CPU
309  *
310  * This function must be called between cpuidle_pause_and_lock and
311  * cpuidle_resume_and_unlock when used externally.
312  */
313 int cpuidle_enable_device(struct cpuidle_device *dev)
314 {
315 	int ret, i;
316 	struct cpuidle_driver *drv;
317 
318 	if (!dev)
319 		return -EINVAL;
320 
321 	if (dev->enabled)
322 		return 0;
323 
324 	drv = cpuidle_get_cpu_driver(dev);
325 
326 	if (!drv || !cpuidle_curr_governor)
327 		return -EIO;
328 
329 	if (!dev->state_count)
330 		dev->state_count = drv->state_count;
331 
332 	if (dev->registered == 0) {
333 		ret = __cpuidle_register_device(dev);
334 		if (ret)
335 			return ret;
336 	}
337 
338 	cpuidle_enter_ops = drv->en_core_tk_irqen ?
339 		cpuidle_enter_tk : cpuidle_enter;
340 
341 	poll_idle_init(drv);
342 
343 	ret = cpuidle_add_device_sysfs(dev);
344 	if (ret)
345 		return ret;
346 
347 	if (cpuidle_curr_governor->enable &&
348 	    (ret = cpuidle_curr_governor->enable(drv, dev)))
349 		goto fail_sysfs;
350 
351 	for (i = 0; i < dev->state_count; i++) {
352 		dev->states_usage[i].usage = 0;
353 		dev->states_usage[i].time = 0;
354 	}
355 	dev->last_residency = 0;
356 
357 	smp_wmb();
358 
359 	dev->enabled = 1;
360 
361 	enabled_devices++;
362 	return 0;
363 
364 fail_sysfs:
365 	cpuidle_remove_device_sysfs(dev);
366 
367 	return ret;
368 }
369 
370 EXPORT_SYMBOL_GPL(cpuidle_enable_device);
371 
372 /**
373  * cpuidle_disable_device - disables idle PM for a CPU
374  * @dev: the CPU
375  *
376  * This function must be called between cpuidle_pause_and_lock and
377  * cpuidle_resume_and_unlock when used externally.
378  */
379 void cpuidle_disable_device(struct cpuidle_device *dev)
380 {
381 	struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);
382 
383 	if (!dev || !dev->enabled)
384 		return;
385 
386 	if (!drv || !cpuidle_curr_governor)
387 		return;
388 
389 	dev->enabled = 0;
390 
391 	if (cpuidle_curr_governor->disable)
392 		cpuidle_curr_governor->disable(drv, dev);
393 
394 	cpuidle_remove_device_sysfs(dev);
395 	enabled_devices--;
396 }
397 
398 EXPORT_SYMBOL_GPL(cpuidle_disable_device);
399 
400 /**
401  * __cpuidle_register_device - internal register function called before register
402  * and enable routines
403  * @dev: the cpu
404  *
405  * cpuidle_lock mutex must be held before this is called
406  */
407 static int __cpuidle_register_device(struct cpuidle_device *dev)
408 {
409 	int ret;
410 	struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);
411 
412 	if (!try_module_get(drv->owner))
413 		return -EINVAL;
414 
415 	per_cpu(cpuidle_devices, dev->cpu) = dev;
416 	list_add(&dev->device_list, &cpuidle_detected_devices);
417 	ret = cpuidle_add_sysfs(dev);
418 	if (ret)
419 		goto err_sysfs;
420 
421 	ret = cpuidle_coupled_register_device(dev);
422 	if (ret)
423 		goto err_coupled;
424 
425 	dev->registered = 1;
426 	return 0;
427 
428 err_coupled:
429 	cpuidle_remove_sysfs(dev);
430 err_sysfs:
431 	list_del(&dev->device_list);
432 	per_cpu(cpuidle_devices, dev->cpu) = NULL;
433 	module_put(drv->owner);
434 	return ret;
435 }
436 
437 /**
438  * cpuidle_register_device - registers a CPU's idle PM feature
439  * @dev: the cpu
440  */
441 int cpuidle_register_device(struct cpuidle_device *dev)
442 {
443 	int ret;
444 
445 	if (!dev)
446 		return -EINVAL;
447 
448 	mutex_lock(&cpuidle_lock);
449 
450 	if ((ret = __cpuidle_register_device(dev))) {
451 		mutex_unlock(&cpuidle_lock);
452 		return ret;
453 	}
454 
455 	cpuidle_enable_device(dev);
456 	cpuidle_install_idle_handler();
457 
458 	mutex_unlock(&cpuidle_lock);
459 
460 	return 0;
461 
462 }
463 
464 EXPORT_SYMBOL_GPL(cpuidle_register_device);
465 
466 /**
467  * cpuidle_unregister_device - unregisters a CPU's idle PM feature
468  * @dev: the cpu
469  */
470 void cpuidle_unregister_device(struct cpuidle_device *dev)
471 {
472 	struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);
473 
474 	if (dev->registered == 0)
475 		return;
476 
477 	cpuidle_pause_and_lock();
478 
479 	cpuidle_disable_device(dev);
480 
481 	cpuidle_remove_sysfs(dev);
482 	list_del(&dev->device_list);
483 	per_cpu(cpuidle_devices, dev->cpu) = NULL;
484 
485 	cpuidle_coupled_unregister_device(dev);
486 
487 	cpuidle_resume_and_unlock();
488 
489 	module_put(drv->owner);
490 }
491 
492 EXPORT_SYMBOL_GPL(cpuidle_unregister_device);
493 
494 #ifdef CONFIG_SMP
495 
496 static void smp_callback(void *v)
497 {
498 	/* we already woke the CPU up, nothing more to do */
499 }
500 
501 /*
502  * This function gets called when a part of the kernel has a new latency
503  * requirement.  This means we need to get all processors out of their C-state,
504  * and then recalculate a new suitable C-state. Just do a cross-cpu IPI; that
505  * wakes them all right up.
506  */
507 static int cpuidle_latency_notify(struct notifier_block *b,
508 		unsigned long l, void *v)
509 {
510 	smp_call_function(smp_callback, NULL, 1);
511 	return NOTIFY_OK;
512 }
513 
514 static struct notifier_block cpuidle_latency_notifier = {
515 	.notifier_call = cpuidle_latency_notify,
516 };
517 
518 static inline void latency_notifier_init(struct notifier_block *n)
519 {
520 	pm_qos_add_notifier(PM_QOS_CPU_DMA_LATENCY, n);
521 }
522 
523 #else /* CONFIG_SMP */
524 
525 #define latency_notifier_init(x) do { } while (0)
526 
527 #endif /* CONFIG_SMP */
528 
529 /**
530  * cpuidle_init - core initializer
531  */
532 static int __init cpuidle_init(void)
533 {
534 	int ret;
535 
536 	if (cpuidle_disabled())
537 		return -ENODEV;
538 
539 	ret = cpuidle_add_interface(cpu_subsys.dev_root);
540 	if (ret)
541 		return ret;
542 
543 	latency_notifier_init(&cpuidle_latency_notifier);
544 
545 	return 0;
546 }
547 
548 module_param(off, int, 0444);
549 core_initcall(cpuidle_init);
550