]> asedeno.scripts.mit.edu Git - linux.git/blob - drivers/base/power/runtime.c
Linux 5.6-rc7
[linux.git] / drivers / base / power / runtime.c
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * drivers/base/power/runtime.c - Helper functions for device runtime PM
4  *
5  * Copyright (c) 2009 Rafael J. Wysocki <rjw@sisk.pl>, Novell Inc.
6  * Copyright (C) 2010 Alan Stern <stern@rowland.harvard.edu>
7  */
8 #include <linux/sched/mm.h>
9 #include <linux/ktime.h>
10 #include <linux/hrtimer.h>
11 #include <linux/export.h>
12 #include <linux/pm_runtime.h>
13 #include <linux/pm_wakeirq.h>
14 #include <trace/events/rpm.h>
15
16 #include "../base.h"
17 #include "power.h"
18
19 typedef int (*pm_callback_t)(struct device *);
20
21 static pm_callback_t __rpm_get_callback(struct device *dev, size_t cb_offset)
22 {
23         pm_callback_t cb;
24         const struct dev_pm_ops *ops;
25
26         if (dev->pm_domain)
27                 ops = &dev->pm_domain->ops;
28         else if (dev->type && dev->type->pm)
29                 ops = dev->type->pm;
30         else if (dev->class && dev->class->pm)
31                 ops = dev->class->pm;
32         else if (dev->bus && dev->bus->pm)
33                 ops = dev->bus->pm;
34         else
35                 ops = NULL;
36
37         if (ops)
38                 cb = *(pm_callback_t *)((void *)ops + cb_offset);
39         else
40                 cb = NULL;
41
42         if (!cb && dev->driver && dev->driver->pm)
43                 cb = *(pm_callback_t *)((void *)dev->driver->pm + cb_offset);
44
45         return cb;
46 }
47
48 #define RPM_GET_CALLBACK(dev, callback) \
49                 __rpm_get_callback(dev, offsetof(struct dev_pm_ops, callback))
50
51 static int rpm_resume(struct device *dev, int rpmflags);
52 static int rpm_suspend(struct device *dev, int rpmflags);
53
54 /**
55  * update_pm_runtime_accounting - Update the time accounting of power states
56  * @dev: Device to update the accounting for
57  *
58  * In order to be able to have time accounting of the various power states
59  * (as used by programs such as PowerTOP to show the effectiveness of runtime
60  * PM), we need to track the time spent in each state.
61  * update_pm_runtime_accounting must be called each time before the
62  * runtime_status field is updated, to account the time in the old state
63  * correctly.
64  */
65 static void update_pm_runtime_accounting(struct device *dev)
66 {
67         u64 now, last, delta;
68
69         if (dev->power.disable_depth > 0)
70                 return;
71
72         last = dev->power.accounting_timestamp;
73
74         now = ktime_get_mono_fast_ns();
75         dev->power.accounting_timestamp = now;
76
77         /*
78          * Because ktime_get_mono_fast_ns() is not monotonic during
79          * timekeeping updates, ensure that 'now' is after the last saved
80          * timesptamp.
81          */
82         if (now < last)
83                 return;
84
85         delta = now - last;
86
87         if (dev->power.runtime_status == RPM_SUSPENDED)
88                 dev->power.suspended_time += delta;
89         else
90                 dev->power.active_time += delta;
91 }
92
93 static void __update_runtime_status(struct device *dev, enum rpm_status status)
94 {
95         update_pm_runtime_accounting(dev);
96         dev->power.runtime_status = status;
97 }
98
99 static u64 rpm_get_accounted_time(struct device *dev, bool suspended)
100 {
101         u64 time;
102         unsigned long flags;
103
104         spin_lock_irqsave(&dev->power.lock, flags);
105
106         update_pm_runtime_accounting(dev);
107         time = suspended ? dev->power.suspended_time : dev->power.active_time;
108
109         spin_unlock_irqrestore(&dev->power.lock, flags);
110
111         return time;
112 }
113
114 u64 pm_runtime_active_time(struct device *dev)
115 {
116         return rpm_get_accounted_time(dev, false);
117 }
118
119 u64 pm_runtime_suspended_time(struct device *dev)
120 {
121         return rpm_get_accounted_time(dev, true);
122 }
123 EXPORT_SYMBOL_GPL(pm_runtime_suspended_time);
124
125 /**
126  * pm_runtime_deactivate_timer - Deactivate given device's suspend timer.
127  * @dev: Device to handle.
128  */
129 static void pm_runtime_deactivate_timer(struct device *dev)
130 {
131         if (dev->power.timer_expires > 0) {
132                 hrtimer_try_to_cancel(&dev->power.suspend_timer);
133                 dev->power.timer_expires = 0;
134         }
135 }
136
137 /**
138  * pm_runtime_cancel_pending - Deactivate suspend timer and cancel requests.
139  * @dev: Device to handle.
140  */
141 static void pm_runtime_cancel_pending(struct device *dev)
142 {
143         pm_runtime_deactivate_timer(dev);
144         /*
145          * In case there's a request pending, make sure its work function will
146          * return without doing anything.
147          */
148         dev->power.request = RPM_REQ_NONE;
149 }
150
151 /*
152  * pm_runtime_autosuspend_expiration - Get a device's autosuspend-delay expiration time.
153  * @dev: Device to handle.
154  *
155  * Compute the autosuspend-delay expiration time based on the device's
156  * power.last_busy time.  If the delay has already expired or is disabled
157  * (negative) or the power.use_autosuspend flag isn't set, return 0.
158  * Otherwise return the expiration time in nanoseconds (adjusted to be nonzero).
159  *
160  * This function may be called either with or without dev->power.lock held.
161  * Either way it can be racy, since power.last_busy may be updated at any time.
162  */
163 u64 pm_runtime_autosuspend_expiration(struct device *dev)
164 {
165         int autosuspend_delay;
166         u64 expires;
167
168         if (!dev->power.use_autosuspend)
169                 return 0;
170
171         autosuspend_delay = READ_ONCE(dev->power.autosuspend_delay);
172         if (autosuspend_delay < 0)
173                 return 0;
174
175         expires  = READ_ONCE(dev->power.last_busy);
176         expires += (u64)autosuspend_delay * NSEC_PER_MSEC;
177         if (expires > ktime_get_mono_fast_ns())
178                 return expires; /* Expires in the future */
179
180         return 0;
181 }
182 EXPORT_SYMBOL_GPL(pm_runtime_autosuspend_expiration);
183
184 static int dev_memalloc_noio(struct device *dev, void *data)
185 {
186         return dev->power.memalloc_noio;
187 }
188
189 /*
190  * pm_runtime_set_memalloc_noio - Set a device's memalloc_noio flag.
191  * @dev: Device to handle.
192  * @enable: True for setting the flag and False for clearing the flag.
193  *
194  * Set the flag for all devices in the path from the device to the
195  * root device in the device tree if @enable is true, otherwise clear
196  * the flag for devices in the path whose siblings don't set the flag.
197  *
198  * The function should only be called by block device, or network
199  * device driver for solving the deadlock problem during runtime
200  * resume/suspend:
201  *
202  *     If memory allocation with GFP_KERNEL is called inside runtime
203  *     resume/suspend callback of any one of its ancestors(or the
204  *     block device itself), the deadlock may be triggered inside the
205  *     memory allocation since it might not complete until the block
206  *     device becomes active and the involed page I/O finishes. The
207  *     situation is pointed out first by Alan Stern. Network device
208  *     are involved in iSCSI kind of situation.
209  *
210  * The lock of dev_hotplug_mutex is held in the function for handling
211  * hotplug race because pm_runtime_set_memalloc_noio() may be called
212  * in async probe().
213  *
214  * The function should be called between device_add() and device_del()
215  * on the affected device(block/network device).
216  */
217 void pm_runtime_set_memalloc_noio(struct device *dev, bool enable)
218 {
219         static DEFINE_MUTEX(dev_hotplug_mutex);
220
221         mutex_lock(&dev_hotplug_mutex);
222         for (;;) {
223                 bool enabled;
224
225                 /* hold power lock since bitfield is not SMP-safe. */
226                 spin_lock_irq(&dev->power.lock);
227                 enabled = dev->power.memalloc_noio;
228                 dev->power.memalloc_noio = enable;
229                 spin_unlock_irq(&dev->power.lock);
230
231                 /*
232                  * not need to enable ancestors any more if the device
233                  * has been enabled.
234                  */
235                 if (enabled && enable)
236                         break;
237
238                 dev = dev->parent;
239
240                 /*
241                  * clear flag of the parent device only if all the
242                  * children don't set the flag because ancestor's
243                  * flag was set by any one of the descendants.
244                  */
245                 if (!dev || (!enable &&
246                              device_for_each_child(dev, NULL,
247                                                    dev_memalloc_noio)))
248                         break;
249         }
250         mutex_unlock(&dev_hotplug_mutex);
251 }
252 EXPORT_SYMBOL_GPL(pm_runtime_set_memalloc_noio);
253
254 /**
255  * rpm_check_suspend_allowed - Test whether a device may be suspended.
256  * @dev: Device to test.
257  */
258 static int rpm_check_suspend_allowed(struct device *dev)
259 {
260         int retval = 0;
261
262         if (dev->power.runtime_error)
263                 retval = -EINVAL;
264         else if (dev->power.disable_depth > 0)
265                 retval = -EACCES;
266         else if (atomic_read(&dev->power.usage_count) > 0)
267                 retval = -EAGAIN;
268         else if (!dev->power.ignore_children &&
269                         atomic_read(&dev->power.child_count))
270                 retval = -EBUSY;
271
272         /* Pending resume requests take precedence over suspends. */
273         else if ((dev->power.deferred_resume
274                         && dev->power.runtime_status == RPM_SUSPENDING)
275             || (dev->power.request_pending
276                         && dev->power.request == RPM_REQ_RESUME))
277                 retval = -EAGAIN;
278         else if (__dev_pm_qos_resume_latency(dev) == 0)
279                 retval = -EPERM;
280         else if (dev->power.runtime_status == RPM_SUSPENDED)
281                 retval = 1;
282
283         return retval;
284 }
285
286 static int rpm_get_suppliers(struct device *dev)
287 {
288         struct device_link *link;
289
290         list_for_each_entry_rcu(link, &dev->links.suppliers, c_node,
291                                 device_links_read_lock_held()) {
292                 int retval;
293
294                 if (!(link->flags & DL_FLAG_PM_RUNTIME) ||
295                     READ_ONCE(link->status) == DL_STATE_SUPPLIER_UNBIND)
296                         continue;
297
298                 retval = pm_runtime_get_sync(link->supplier);
299                 /* Ignore suppliers with disabled runtime PM. */
300                 if (retval < 0 && retval != -EACCES) {
301                         pm_runtime_put_noidle(link->supplier);
302                         return retval;
303                 }
304                 refcount_inc(&link->rpm_active);
305         }
306         return 0;
307 }
308
309 static void rpm_put_suppliers(struct device *dev)
310 {
311         struct device_link *link;
312
313         list_for_each_entry_rcu(link, &dev->links.suppliers, c_node,
314                                 device_links_read_lock_held()) {
315                 if (READ_ONCE(link->status) == DL_STATE_SUPPLIER_UNBIND)
316                         continue;
317
318                 while (refcount_dec_not_one(&link->rpm_active))
319                         pm_runtime_put(link->supplier);
320         }
321 }
322
323 /**
324  * __rpm_callback - Run a given runtime PM callback for a given device.
325  * @cb: Runtime PM callback to run.
326  * @dev: Device to run the callback for.
327  */
328 static int __rpm_callback(int (*cb)(struct device *), struct device *dev)
329         __releases(&dev->power.lock) __acquires(&dev->power.lock)
330 {
331         int retval, idx;
332         bool use_links = dev->power.links_count > 0;
333
334         if (dev->power.irq_safe) {
335                 spin_unlock(&dev->power.lock);
336         } else {
337                 spin_unlock_irq(&dev->power.lock);
338
339                 /*
340                  * Resume suppliers if necessary.
341                  *
342                  * The device's runtime PM status cannot change until this
343                  * routine returns, so it is safe to read the status outside of
344                  * the lock.
345                  */
346                 if (use_links && dev->power.runtime_status == RPM_RESUMING) {
347                         idx = device_links_read_lock();
348
349                         retval = rpm_get_suppliers(dev);
350                         if (retval)
351                                 goto fail;
352
353                         device_links_read_unlock(idx);
354                 }
355         }
356
357         retval = cb(dev);
358
359         if (dev->power.irq_safe) {
360                 spin_lock(&dev->power.lock);
361         } else {
362                 /*
363                  * If the device is suspending and the callback has returned
364                  * success, drop the usage counters of the suppliers that have
365                  * been reference counted on its resume.
366                  *
367                  * Do that if resume fails too.
368                  */
369                 if (use_links
370                     && ((dev->power.runtime_status == RPM_SUSPENDING && !retval)
371                     || (dev->power.runtime_status == RPM_RESUMING && retval))) {
372                         idx = device_links_read_lock();
373
374  fail:
375                         rpm_put_suppliers(dev);
376
377                         device_links_read_unlock(idx);
378                 }
379
380                 spin_lock_irq(&dev->power.lock);
381         }
382
383         return retval;
384 }
385
386 /**
387  * rpm_idle - Notify device bus type if the device can be suspended.
388  * @dev: Device to notify the bus type about.
389  * @rpmflags: Flag bits.
390  *
391  * Check if the device's runtime PM status allows it to be suspended.  If
392  * another idle notification has been started earlier, return immediately.  If
393  * the RPM_ASYNC flag is set then queue an idle-notification request; otherwise
394  * run the ->runtime_idle() callback directly. If the ->runtime_idle callback
395  * doesn't exist or if it returns 0, call rpm_suspend with the RPM_AUTO flag.
396  *
397  * This function must be called under dev->power.lock with interrupts disabled.
398  */
399 static int rpm_idle(struct device *dev, int rpmflags)
400 {
401         int (*callback)(struct device *);
402         int retval;
403
404         trace_rpm_idle_rcuidle(dev, rpmflags);
405         retval = rpm_check_suspend_allowed(dev);
406         if (retval < 0)
407                 ;       /* Conditions are wrong. */
408
409         /* Idle notifications are allowed only in the RPM_ACTIVE state. */
410         else if (dev->power.runtime_status != RPM_ACTIVE)
411                 retval = -EAGAIN;
412
413         /*
414          * Any pending request other than an idle notification takes
415          * precedence over us, except that the timer may be running.
416          */
417         else if (dev->power.request_pending &&
418             dev->power.request > RPM_REQ_IDLE)
419                 retval = -EAGAIN;
420
421         /* Act as though RPM_NOWAIT is always set. */
422         else if (dev->power.idle_notification)
423                 retval = -EINPROGRESS;
424         if (retval)
425                 goto out;
426
427         /* Pending requests need to be canceled. */
428         dev->power.request = RPM_REQ_NONE;
429
430         if (dev->power.no_callbacks)
431                 goto out;
432
433         /* Carry out an asynchronous or a synchronous idle notification. */
434         if (rpmflags & RPM_ASYNC) {
435                 dev->power.request = RPM_REQ_IDLE;
436                 if (!dev->power.request_pending) {
437                         dev->power.request_pending = true;
438                         queue_work(pm_wq, &dev->power.work);
439                 }
440                 trace_rpm_return_int_rcuidle(dev, _THIS_IP_, 0);
441                 return 0;
442         }
443
444         dev->power.idle_notification = true;
445
446         callback = RPM_GET_CALLBACK(dev, runtime_idle);
447
448         if (callback)
449                 retval = __rpm_callback(callback, dev);
450
451         dev->power.idle_notification = false;
452         wake_up_all(&dev->power.wait_queue);
453
454  out:
455         trace_rpm_return_int_rcuidle(dev, _THIS_IP_, retval);
456         return retval ? retval : rpm_suspend(dev, rpmflags | RPM_AUTO);
457 }
458
459 /**
460  * rpm_callback - Run a given runtime PM callback for a given device.
461  * @cb: Runtime PM callback to run.
462  * @dev: Device to run the callback for.
463  */
464 static int rpm_callback(int (*cb)(struct device *), struct device *dev)
465 {
466         int retval;
467
468         if (!cb)
469                 return -ENOSYS;
470
471         if (dev->power.memalloc_noio) {
472                 unsigned int noio_flag;
473
474                 /*
475                  * Deadlock might be caused if memory allocation with
476                  * GFP_KERNEL happens inside runtime_suspend and
477                  * runtime_resume callbacks of one block device's
478                  * ancestor or the block device itself. Network
479                  * device might be thought as part of iSCSI block
480                  * device, so network device and its ancestor should
481                  * be marked as memalloc_noio too.
482                  */
483                 noio_flag = memalloc_noio_save();
484                 retval = __rpm_callback(cb, dev);
485                 memalloc_noio_restore(noio_flag);
486         } else {
487                 retval = __rpm_callback(cb, dev);
488         }
489
490         dev->power.runtime_error = retval;
491         return retval != -EACCES ? retval : -EIO;
492 }
493
494 /**
495  * rpm_suspend - Carry out runtime suspend of given device.
496  * @dev: Device to suspend.
497  * @rpmflags: Flag bits.
498  *
499  * Check if the device's runtime PM status allows it to be suspended.
500  * Cancel a pending idle notification, autosuspend or suspend. If
501  * another suspend has been started earlier, either return immediately
502  * or wait for it to finish, depending on the RPM_NOWAIT and RPM_ASYNC
503  * flags. If the RPM_ASYNC flag is set then queue a suspend request;
504  * otherwise run the ->runtime_suspend() callback directly. When
505  * ->runtime_suspend succeeded, if a deferred resume was requested while
506  * the callback was running then carry it out, otherwise send an idle
507  * notification for its parent (if the suspend succeeded and both
508  * ignore_children of parent->power and irq_safe of dev->power are not set).
509  * If ->runtime_suspend failed with -EAGAIN or -EBUSY, and if the RPM_AUTO
510  * flag is set and the next autosuspend-delay expiration time is in the
511  * future, schedule another autosuspend attempt.
512  *
513  * This function must be called under dev->power.lock with interrupts disabled.
514  */
515 static int rpm_suspend(struct device *dev, int rpmflags)
516         __releases(&dev->power.lock) __acquires(&dev->power.lock)
517 {
518         int (*callback)(struct device *);
519         struct device *parent = NULL;
520         int retval;
521
522         trace_rpm_suspend_rcuidle(dev, rpmflags);
523
524  repeat:
525         retval = rpm_check_suspend_allowed(dev);
526
527         if (retval < 0)
528                 ;       /* Conditions are wrong. */
529
530         /* Synchronous suspends are not allowed in the RPM_RESUMING state. */
531         else if (dev->power.runtime_status == RPM_RESUMING &&
532             !(rpmflags & RPM_ASYNC))
533                 retval = -EAGAIN;
534         if (retval)
535                 goto out;
536
537         /* If the autosuspend_delay time hasn't expired yet, reschedule. */
538         if ((rpmflags & RPM_AUTO)
539             && dev->power.runtime_status != RPM_SUSPENDING) {
540                 u64 expires = pm_runtime_autosuspend_expiration(dev);
541
542                 if (expires != 0) {
543                         /* Pending requests need to be canceled. */
544                         dev->power.request = RPM_REQ_NONE;
545
546                         /*
547                          * Optimization: If the timer is already running and is
548                          * set to expire at or before the autosuspend delay,
549                          * avoid the overhead of resetting it.  Just let it
550                          * expire; pm_suspend_timer_fn() will take care of the
551                          * rest.
552                          */
553                         if (!(dev->power.timer_expires &&
554                                         dev->power.timer_expires <= expires)) {
555                                 /*
556                                  * We add a slack of 25% to gather wakeups
557                                  * without sacrificing the granularity.
558                                  */
559                                 u64 slack = (u64)READ_ONCE(dev->power.autosuspend_delay) *
560                                                     (NSEC_PER_MSEC >> 2);
561
562                                 dev->power.timer_expires = expires;
563                                 hrtimer_start_range_ns(&dev->power.suspend_timer,
564                                                 ns_to_ktime(expires),
565                                                 slack,
566                                                 HRTIMER_MODE_ABS);
567                         }
568                         dev->power.timer_autosuspends = 1;
569                         goto out;
570                 }
571         }
572
573         /* Other scheduled or pending requests need to be canceled. */
574         pm_runtime_cancel_pending(dev);
575
576         if (dev->power.runtime_status == RPM_SUSPENDING) {
577                 DEFINE_WAIT(wait);
578
579                 if (rpmflags & (RPM_ASYNC | RPM_NOWAIT)) {
580                         retval = -EINPROGRESS;
581                         goto out;
582                 }
583
584                 if (dev->power.irq_safe) {
585                         spin_unlock(&dev->power.lock);
586
587                         cpu_relax();
588
589                         spin_lock(&dev->power.lock);
590                         goto repeat;
591                 }
592
593                 /* Wait for the other suspend running in parallel with us. */
594                 for (;;) {
595                         prepare_to_wait(&dev->power.wait_queue, &wait,
596                                         TASK_UNINTERRUPTIBLE);
597                         if (dev->power.runtime_status != RPM_SUSPENDING)
598                                 break;
599
600                         spin_unlock_irq(&dev->power.lock);
601
602                         schedule();
603
604                         spin_lock_irq(&dev->power.lock);
605                 }
606                 finish_wait(&dev->power.wait_queue, &wait);
607                 goto repeat;
608         }
609
610         if (dev->power.no_callbacks)
611                 goto no_callback;       /* Assume success. */
612
613         /* Carry out an asynchronous or a synchronous suspend. */
614         if (rpmflags & RPM_ASYNC) {
615                 dev->power.request = (rpmflags & RPM_AUTO) ?
616                     RPM_REQ_AUTOSUSPEND : RPM_REQ_SUSPEND;
617                 if (!dev->power.request_pending) {
618                         dev->power.request_pending = true;
619                         queue_work(pm_wq, &dev->power.work);
620                 }
621                 goto out;
622         }
623
624         __update_runtime_status(dev, RPM_SUSPENDING);
625
626         callback = RPM_GET_CALLBACK(dev, runtime_suspend);
627
628         dev_pm_enable_wake_irq_check(dev, true);
629         retval = rpm_callback(callback, dev);
630         if (retval)
631                 goto fail;
632
633  no_callback:
634         __update_runtime_status(dev, RPM_SUSPENDED);
635         pm_runtime_deactivate_timer(dev);
636
637         if (dev->parent) {
638                 parent = dev->parent;
639                 atomic_add_unless(&parent->power.child_count, -1, 0);
640         }
641         wake_up_all(&dev->power.wait_queue);
642
643         if (dev->power.deferred_resume) {
644                 dev->power.deferred_resume = false;
645                 rpm_resume(dev, 0);
646                 retval = -EAGAIN;
647                 goto out;
648         }
649
650         /* Maybe the parent is now able to suspend. */
651         if (parent && !parent->power.ignore_children && !dev->power.irq_safe) {
652                 spin_unlock(&dev->power.lock);
653
654                 spin_lock(&parent->power.lock);
655                 rpm_idle(parent, RPM_ASYNC);
656                 spin_unlock(&parent->power.lock);
657
658                 spin_lock(&dev->power.lock);
659         }
660
661  out:
662         trace_rpm_return_int_rcuidle(dev, _THIS_IP_, retval);
663
664         return retval;
665
666  fail:
667         dev_pm_disable_wake_irq_check(dev);
668         __update_runtime_status(dev, RPM_ACTIVE);
669         dev->power.deferred_resume = false;
670         wake_up_all(&dev->power.wait_queue);
671
672         if (retval == -EAGAIN || retval == -EBUSY) {
673                 dev->power.runtime_error = 0;
674
675                 /*
676                  * If the callback routine failed an autosuspend, and
677                  * if the last_busy time has been updated so that there
678                  * is a new autosuspend expiration time, automatically
679                  * reschedule another autosuspend.
680                  */
681                 if ((rpmflags & RPM_AUTO) &&
682                     pm_runtime_autosuspend_expiration(dev) != 0)
683                         goto repeat;
684         } else {
685                 pm_runtime_cancel_pending(dev);
686         }
687         goto out;
688 }
689
690 /**
691  * rpm_resume - Carry out runtime resume of given device.
692  * @dev: Device to resume.
693  * @rpmflags: Flag bits.
694  *
695  * Check if the device's runtime PM status allows it to be resumed.  Cancel
696  * any scheduled or pending requests.  If another resume has been started
697  * earlier, either return immediately or wait for it to finish, depending on the
698  * RPM_NOWAIT and RPM_ASYNC flags.  Similarly, if there's a suspend running in
699  * parallel with this function, either tell the other process to resume after
700  * suspending (deferred_resume) or wait for it to finish.  If the RPM_ASYNC
701  * flag is set then queue a resume request; otherwise run the
702  * ->runtime_resume() callback directly.  Queue an idle notification for the
703  * device if the resume succeeded.
704  *
705  * This function must be called under dev->power.lock with interrupts disabled.
706  */
707 static int rpm_resume(struct device *dev, int rpmflags)
708         __releases(&dev->power.lock) __acquires(&dev->power.lock)
709 {
710         int (*callback)(struct device *);
711         struct device *parent = NULL;
712         int retval = 0;
713
714         trace_rpm_resume_rcuidle(dev, rpmflags);
715
716  repeat:
717         if (dev->power.runtime_error)
718                 retval = -EINVAL;
719         else if (dev->power.disable_depth == 1 && dev->power.is_suspended
720             && dev->power.runtime_status == RPM_ACTIVE)
721                 retval = 1;
722         else if (dev->power.disable_depth > 0)
723                 retval = -EACCES;
724         if (retval)
725                 goto out;
726
727         /*
728          * Other scheduled or pending requests need to be canceled.  Small
729          * optimization: If an autosuspend timer is running, leave it running
730          * rather than cancelling it now only to restart it again in the near
731          * future.
732          */
733         dev->power.request = RPM_REQ_NONE;
734         if (!dev->power.timer_autosuspends)
735                 pm_runtime_deactivate_timer(dev);
736
737         if (dev->power.runtime_status == RPM_ACTIVE) {
738                 retval = 1;
739                 goto out;
740         }
741
742         if (dev->power.runtime_status == RPM_RESUMING
743             || dev->power.runtime_status == RPM_SUSPENDING) {
744                 DEFINE_WAIT(wait);
745
746                 if (rpmflags & (RPM_ASYNC | RPM_NOWAIT)) {
747                         if (dev->power.runtime_status == RPM_SUSPENDING)
748                                 dev->power.deferred_resume = true;
749                         else
750                                 retval = -EINPROGRESS;
751                         goto out;
752                 }
753
754                 if (dev->power.irq_safe) {
755                         spin_unlock(&dev->power.lock);
756
757                         cpu_relax();
758
759                         spin_lock(&dev->power.lock);
760                         goto repeat;
761                 }
762
763                 /* Wait for the operation carried out in parallel with us. */
764                 for (;;) {
765                         prepare_to_wait(&dev->power.wait_queue, &wait,
766                                         TASK_UNINTERRUPTIBLE);
767                         if (dev->power.runtime_status != RPM_RESUMING
768                             && dev->power.runtime_status != RPM_SUSPENDING)
769                                 break;
770
771                         spin_unlock_irq(&dev->power.lock);
772
773                         schedule();
774
775                         spin_lock_irq(&dev->power.lock);
776                 }
777                 finish_wait(&dev->power.wait_queue, &wait);
778                 goto repeat;
779         }
780
781         /*
782          * See if we can skip waking up the parent.  This is safe only if
783          * power.no_callbacks is set, because otherwise we don't know whether
784          * the resume will actually succeed.
785          */
786         if (dev->power.no_callbacks && !parent && dev->parent) {
787                 spin_lock_nested(&dev->parent->power.lock, SINGLE_DEPTH_NESTING);
788                 if (dev->parent->power.disable_depth > 0
789                     || dev->parent->power.ignore_children
790                     || dev->parent->power.runtime_status == RPM_ACTIVE) {
791                         atomic_inc(&dev->parent->power.child_count);
792                         spin_unlock(&dev->parent->power.lock);
793                         retval = 1;
794                         goto no_callback;       /* Assume success. */
795                 }
796                 spin_unlock(&dev->parent->power.lock);
797         }
798
799         /* Carry out an asynchronous or a synchronous resume. */
800         if (rpmflags & RPM_ASYNC) {
801                 dev->power.request = RPM_REQ_RESUME;
802                 if (!dev->power.request_pending) {
803                         dev->power.request_pending = true;
804                         queue_work(pm_wq, &dev->power.work);
805                 }
806                 retval = 0;
807                 goto out;
808         }
809
810         if (!parent && dev->parent) {
811                 /*
812                  * Increment the parent's usage counter and resume it if
813                  * necessary.  Not needed if dev is irq-safe; then the
814                  * parent is permanently resumed.
815                  */
816                 parent = dev->parent;
817                 if (dev->power.irq_safe)
818                         goto skip_parent;
819                 spin_unlock(&dev->power.lock);
820
821                 pm_runtime_get_noresume(parent);
822
823                 spin_lock(&parent->power.lock);
824                 /*
825                  * Resume the parent if it has runtime PM enabled and not been
826                  * set to ignore its children.
827                  */
828                 if (!parent->power.disable_depth
829                     && !parent->power.ignore_children) {
830                         rpm_resume(parent, 0);
831                         if (parent->power.runtime_status != RPM_ACTIVE)
832                                 retval = -EBUSY;
833                 }
834                 spin_unlock(&parent->power.lock);
835
836                 spin_lock(&dev->power.lock);
837                 if (retval)
838                         goto out;
839                 goto repeat;
840         }
841  skip_parent:
842
843         if (dev->power.no_callbacks)
844                 goto no_callback;       /* Assume success. */
845
846         __update_runtime_status(dev, RPM_RESUMING);
847
848         callback = RPM_GET_CALLBACK(dev, runtime_resume);
849
850         dev_pm_disable_wake_irq_check(dev);
851         retval = rpm_callback(callback, dev);
852         if (retval) {
853                 __update_runtime_status(dev, RPM_SUSPENDED);
854                 pm_runtime_cancel_pending(dev);
855                 dev_pm_enable_wake_irq_check(dev, false);
856         } else {
857  no_callback:
858                 __update_runtime_status(dev, RPM_ACTIVE);
859                 pm_runtime_mark_last_busy(dev);
860                 if (parent)
861                         atomic_inc(&parent->power.child_count);
862         }
863         wake_up_all(&dev->power.wait_queue);
864
865         if (retval >= 0)
866                 rpm_idle(dev, RPM_ASYNC);
867
868  out:
869         if (parent && !dev->power.irq_safe) {
870                 spin_unlock_irq(&dev->power.lock);
871
872                 pm_runtime_put(parent);
873
874                 spin_lock_irq(&dev->power.lock);
875         }
876
877         trace_rpm_return_int_rcuidle(dev, _THIS_IP_, retval);
878
879         return retval;
880 }
881
882 /**
883  * pm_runtime_work - Universal runtime PM work function.
884  * @work: Work structure used for scheduling the execution of this function.
885  *
886  * Use @work to get the device object the work is to be done for, determine what
887  * is to be done and execute the appropriate runtime PM function.
888  */
889 static void pm_runtime_work(struct work_struct *work)
890 {
891         struct device *dev = container_of(work, struct device, power.work);
892         enum rpm_request req;
893
894         spin_lock_irq(&dev->power.lock);
895
896         if (!dev->power.request_pending)
897                 goto out;
898
899         req = dev->power.request;
900         dev->power.request = RPM_REQ_NONE;
901         dev->power.request_pending = false;
902
903         switch (req) {
904         case RPM_REQ_NONE:
905                 break;
906         case RPM_REQ_IDLE:
907                 rpm_idle(dev, RPM_NOWAIT);
908                 break;
909         case RPM_REQ_SUSPEND:
910                 rpm_suspend(dev, RPM_NOWAIT);
911                 break;
912         case RPM_REQ_AUTOSUSPEND:
913                 rpm_suspend(dev, RPM_NOWAIT | RPM_AUTO);
914                 break;
915         case RPM_REQ_RESUME:
916                 rpm_resume(dev, RPM_NOWAIT);
917                 break;
918         }
919
920  out:
921         spin_unlock_irq(&dev->power.lock);
922 }
923
924 /**
925  * pm_suspend_timer_fn - Timer function for pm_schedule_suspend().
926  * @data: Device pointer passed by pm_schedule_suspend().
927  *
928  * Check if the time is right and queue a suspend request.
929  */
930 static enum hrtimer_restart  pm_suspend_timer_fn(struct hrtimer *timer)
931 {
932         struct device *dev = container_of(timer, struct device, power.suspend_timer);
933         unsigned long flags;
934         u64 expires;
935
936         spin_lock_irqsave(&dev->power.lock, flags);
937
938         expires = dev->power.timer_expires;
939         /*
940          * If 'expires' is after the current time, we've been called
941          * too early.
942          */
943         if (expires > 0 && expires < ktime_get_mono_fast_ns()) {
944                 dev->power.timer_expires = 0;
945                 rpm_suspend(dev, dev->power.timer_autosuspends ?
946                     (RPM_ASYNC | RPM_AUTO) : RPM_ASYNC);
947         }
948
949         spin_unlock_irqrestore(&dev->power.lock, flags);
950
951         return HRTIMER_NORESTART;
952 }
953
954 /**
955  * pm_schedule_suspend - Set up a timer to submit a suspend request in future.
956  * @dev: Device to suspend.
957  * @delay: Time to wait before submitting a suspend request, in milliseconds.
958  */
959 int pm_schedule_suspend(struct device *dev, unsigned int delay)
960 {
961         unsigned long flags;
962         u64 expires;
963         int retval;
964
965         spin_lock_irqsave(&dev->power.lock, flags);
966
967         if (!delay) {
968                 retval = rpm_suspend(dev, RPM_ASYNC);
969                 goto out;
970         }
971
972         retval = rpm_check_suspend_allowed(dev);
973         if (retval)
974                 goto out;
975
976         /* Other scheduled or pending requests need to be canceled. */
977         pm_runtime_cancel_pending(dev);
978
979         expires = ktime_get_mono_fast_ns() + (u64)delay * NSEC_PER_MSEC;
980         dev->power.timer_expires = expires;
981         dev->power.timer_autosuspends = 0;
982         hrtimer_start(&dev->power.suspend_timer, expires, HRTIMER_MODE_ABS);
983
984  out:
985         spin_unlock_irqrestore(&dev->power.lock, flags);
986
987         return retval;
988 }
989 EXPORT_SYMBOL_GPL(pm_schedule_suspend);
990
991 /**
992  * __pm_runtime_idle - Entry point for runtime idle operations.
993  * @dev: Device to send idle notification for.
994  * @rpmflags: Flag bits.
995  *
996  * If the RPM_GET_PUT flag is set, decrement the device's usage count and
997  * return immediately if it is larger than zero.  Then carry out an idle
998  * notification, either synchronous or asynchronous.
999  *
1000  * This routine may be called in atomic context if the RPM_ASYNC flag is set,
1001  * or if pm_runtime_irq_safe() has been called.
1002  */
1003 int __pm_runtime_idle(struct device *dev, int rpmflags)
1004 {
1005         unsigned long flags;
1006         int retval;
1007
1008         if (rpmflags & RPM_GET_PUT) {
1009                 if (!atomic_dec_and_test(&dev->power.usage_count)) {
1010                         trace_rpm_usage_rcuidle(dev, rpmflags);
1011                         return 0;
1012                 }
1013         }
1014
1015         might_sleep_if(!(rpmflags & RPM_ASYNC) && !dev->power.irq_safe);
1016
1017         spin_lock_irqsave(&dev->power.lock, flags);
1018         retval = rpm_idle(dev, rpmflags);
1019         spin_unlock_irqrestore(&dev->power.lock, flags);
1020
1021         return retval;
1022 }
1023 EXPORT_SYMBOL_GPL(__pm_runtime_idle);
1024
1025 /**
1026  * __pm_runtime_suspend - Entry point for runtime put/suspend operations.
1027  * @dev: Device to suspend.
1028  * @rpmflags: Flag bits.
1029  *
1030  * If the RPM_GET_PUT flag is set, decrement the device's usage count and
1031  * return immediately if it is larger than zero.  Then carry out a suspend,
1032  * either synchronous or asynchronous.
1033  *
1034  * This routine may be called in atomic context if the RPM_ASYNC flag is set,
1035  * or if pm_runtime_irq_safe() has been called.
1036  */
1037 int __pm_runtime_suspend(struct device *dev, int rpmflags)
1038 {
1039         unsigned long flags;
1040         int retval;
1041
1042         if (rpmflags & RPM_GET_PUT) {
1043                 if (!atomic_dec_and_test(&dev->power.usage_count)) {
1044                         trace_rpm_usage_rcuidle(dev, rpmflags);
1045                         return 0;
1046                 }
1047         }
1048
1049         might_sleep_if(!(rpmflags & RPM_ASYNC) && !dev->power.irq_safe);
1050
1051         spin_lock_irqsave(&dev->power.lock, flags);
1052         retval = rpm_suspend(dev, rpmflags);
1053         spin_unlock_irqrestore(&dev->power.lock, flags);
1054
1055         return retval;
1056 }
1057 EXPORT_SYMBOL_GPL(__pm_runtime_suspend);
1058
1059 /**
1060  * __pm_runtime_resume - Entry point for runtime resume operations.
1061  * @dev: Device to resume.
1062  * @rpmflags: Flag bits.
1063  *
1064  * If the RPM_GET_PUT flag is set, increment the device's usage count.  Then
1065  * carry out a resume, either synchronous or asynchronous.
1066  *
1067  * This routine may be called in atomic context if the RPM_ASYNC flag is set,
1068  * or if pm_runtime_irq_safe() has been called.
1069  */
1070 int __pm_runtime_resume(struct device *dev, int rpmflags)
1071 {
1072         unsigned long flags;
1073         int retval;
1074
1075         might_sleep_if(!(rpmflags & RPM_ASYNC) && !dev->power.irq_safe &&
1076                         dev->power.runtime_status != RPM_ACTIVE);
1077
1078         if (rpmflags & RPM_GET_PUT)
1079                 atomic_inc(&dev->power.usage_count);
1080
1081         spin_lock_irqsave(&dev->power.lock, flags);
1082         retval = rpm_resume(dev, rpmflags);
1083         spin_unlock_irqrestore(&dev->power.lock, flags);
1084
1085         return retval;
1086 }
1087 EXPORT_SYMBOL_GPL(__pm_runtime_resume);
1088
1089 /**
1090  * pm_runtime_get_if_in_use - Conditionally bump up the device's usage counter.
1091  * @dev: Device to handle.
1092  *
1093  * Return -EINVAL if runtime PM is disabled for the device.
1094  *
1095  * If that's not the case and if the device's runtime PM status is RPM_ACTIVE
1096  * and the runtime PM usage counter is nonzero, increment the counter and
1097  * return 1.  Otherwise return 0 without changing the counter.
1098  */
1099 int pm_runtime_get_if_in_use(struct device *dev)
1100 {
1101         unsigned long flags;
1102         int retval;
1103
1104         spin_lock_irqsave(&dev->power.lock, flags);
1105         retval = dev->power.disable_depth > 0 ? -EINVAL :
1106                 dev->power.runtime_status == RPM_ACTIVE
1107                         && atomic_inc_not_zero(&dev->power.usage_count);
1108         trace_rpm_usage_rcuidle(dev, 0);
1109         spin_unlock_irqrestore(&dev->power.lock, flags);
1110         return retval;
1111 }
1112 EXPORT_SYMBOL_GPL(pm_runtime_get_if_in_use);
1113
1114 /**
1115  * __pm_runtime_set_status - Set runtime PM status of a device.
1116  * @dev: Device to handle.
1117  * @status: New runtime PM status of the device.
1118  *
1119  * If runtime PM of the device is disabled or its power.runtime_error field is
1120  * different from zero, the status may be changed either to RPM_ACTIVE, or to
1121  * RPM_SUSPENDED, as long as that reflects the actual state of the device.
1122  * However, if the device has a parent and the parent is not active, and the
1123  * parent's power.ignore_children flag is unset, the device's status cannot be
1124  * set to RPM_ACTIVE, so -EBUSY is returned in that case.
1125  *
1126  * If successful, __pm_runtime_set_status() clears the power.runtime_error field
1127  * and the device parent's counter of unsuspended children is modified to
1128  * reflect the new status.  If the new status is RPM_SUSPENDED, an idle
1129  * notification request for the parent is submitted.
1130  *
1131  * If @dev has any suppliers (as reflected by device links to them), and @status
1132  * is RPM_ACTIVE, they will be activated upfront and if the activation of one
1133  * of them fails, the status of @dev will be changed to RPM_SUSPENDED (instead
1134  * of the @status value) and the suppliers will be deacticated on exit.  The
1135  * error returned by the failing supplier activation will be returned in that
1136  * case.
1137  */
1138 int __pm_runtime_set_status(struct device *dev, unsigned int status)
1139 {
1140         struct device *parent = dev->parent;
1141         bool notify_parent = false;
1142         int error = 0;
1143
1144         if (status != RPM_ACTIVE && status != RPM_SUSPENDED)
1145                 return -EINVAL;
1146
1147         spin_lock_irq(&dev->power.lock);
1148
1149         /*
1150          * Prevent PM-runtime from being enabled for the device or return an
1151          * error if it is enabled already and working.
1152          */
1153         if (dev->power.runtime_error || dev->power.disable_depth)
1154                 dev->power.disable_depth++;
1155         else
1156                 error = -EAGAIN;
1157
1158         spin_unlock_irq(&dev->power.lock);
1159
1160         if (error)
1161                 return error;
1162
1163         /*
1164          * If the new status is RPM_ACTIVE, the suppliers can be activated
1165          * upfront regardless of the current status, because next time
1166          * rpm_put_suppliers() runs, the rpm_active refcounts of the links
1167          * involved will be dropped down to one anyway.
1168          */
1169         if (status == RPM_ACTIVE) {
1170                 int idx = device_links_read_lock();
1171
1172                 error = rpm_get_suppliers(dev);
1173                 if (error)
1174                         status = RPM_SUSPENDED;
1175
1176                 device_links_read_unlock(idx);
1177         }
1178
1179         spin_lock_irq(&dev->power.lock);
1180
1181         if (dev->power.runtime_status == status || !parent)
1182                 goto out_set;
1183
1184         if (status == RPM_SUSPENDED) {
1185                 atomic_add_unless(&parent->power.child_count, -1, 0);
1186                 notify_parent = !parent->power.ignore_children;
1187         } else {
1188                 spin_lock_nested(&parent->power.lock, SINGLE_DEPTH_NESTING);
1189
1190                 /*
1191                  * It is invalid to put an active child under a parent that is
1192                  * not active, has runtime PM enabled and the
1193                  * 'power.ignore_children' flag unset.
1194                  */
1195                 if (!parent->power.disable_depth
1196                     && !parent->power.ignore_children
1197                     && parent->power.runtime_status != RPM_ACTIVE) {
1198                         dev_err(dev, "runtime PM trying to activate child device %s but parent (%s) is not active\n",
1199                                 dev_name(dev),
1200                                 dev_name(parent));
1201                         error = -EBUSY;
1202                 } else if (dev->power.runtime_status == RPM_SUSPENDED) {
1203                         atomic_inc(&parent->power.child_count);
1204                 }
1205
1206                 spin_unlock(&parent->power.lock);
1207
1208                 if (error) {
1209                         status = RPM_SUSPENDED;
1210                         goto out;
1211                 }
1212         }
1213
1214  out_set:
1215         __update_runtime_status(dev, status);
1216         if (!error)
1217                 dev->power.runtime_error = 0;
1218
1219  out:
1220         spin_unlock_irq(&dev->power.lock);
1221
1222         if (notify_parent)
1223                 pm_request_idle(parent);
1224
1225         if (status == RPM_SUSPENDED) {
1226                 int idx = device_links_read_lock();
1227
1228                 rpm_put_suppliers(dev);
1229
1230                 device_links_read_unlock(idx);
1231         }
1232
1233         pm_runtime_enable(dev);
1234
1235         return error;
1236 }
1237 EXPORT_SYMBOL_GPL(__pm_runtime_set_status);
1238
1239 /**
1240  * __pm_runtime_barrier - Cancel pending requests and wait for completions.
1241  * @dev: Device to handle.
1242  *
1243  * Flush all pending requests for the device from pm_wq and wait for all
1244  * runtime PM operations involving the device in progress to complete.
1245  *
1246  * Should be called under dev->power.lock with interrupts disabled.
1247  */
1248 static void __pm_runtime_barrier(struct device *dev)
1249 {
1250         pm_runtime_deactivate_timer(dev);
1251
1252         if (dev->power.request_pending) {
1253                 dev->power.request = RPM_REQ_NONE;
1254                 spin_unlock_irq(&dev->power.lock);
1255
1256                 cancel_work_sync(&dev->power.work);
1257
1258                 spin_lock_irq(&dev->power.lock);
1259                 dev->power.request_pending = false;
1260         }
1261
1262         if (dev->power.runtime_status == RPM_SUSPENDING
1263             || dev->power.runtime_status == RPM_RESUMING
1264             || dev->power.idle_notification) {
1265                 DEFINE_WAIT(wait);
1266
1267                 /* Suspend, wake-up or idle notification in progress. */
1268                 for (;;) {
1269                         prepare_to_wait(&dev->power.wait_queue, &wait,
1270                                         TASK_UNINTERRUPTIBLE);
1271                         if (dev->power.runtime_status != RPM_SUSPENDING
1272                             && dev->power.runtime_status != RPM_RESUMING
1273                             && !dev->power.idle_notification)
1274                                 break;
1275                         spin_unlock_irq(&dev->power.lock);
1276
1277                         schedule();
1278
1279                         spin_lock_irq(&dev->power.lock);
1280                 }
1281                 finish_wait(&dev->power.wait_queue, &wait);
1282         }
1283 }
1284
1285 /**
1286  * pm_runtime_barrier - Flush pending requests and wait for completions.
1287  * @dev: Device to handle.
1288  *
1289  * Prevent the device from being suspended by incrementing its usage counter and
1290  * if there's a pending resume request for the device, wake the device up.
1291  * Next, make sure that all pending requests for the device have been flushed
1292  * from pm_wq and wait for all runtime PM operations involving the device in
1293  * progress to complete.
1294  *
1295  * Return value:
1296  * 1, if there was a resume request pending and the device had to be woken up,
1297  * 0, otherwise
1298  */
1299 int pm_runtime_barrier(struct device *dev)
1300 {
1301         int retval = 0;
1302
1303         pm_runtime_get_noresume(dev);
1304         spin_lock_irq(&dev->power.lock);
1305
1306         if (dev->power.request_pending
1307             && dev->power.request == RPM_REQ_RESUME) {
1308                 rpm_resume(dev, 0);
1309                 retval = 1;
1310         }
1311
1312         __pm_runtime_barrier(dev);
1313
1314         spin_unlock_irq(&dev->power.lock);
1315         pm_runtime_put_noidle(dev);
1316
1317         return retval;
1318 }
1319 EXPORT_SYMBOL_GPL(pm_runtime_barrier);
1320
1321 /**
1322  * __pm_runtime_disable - Disable runtime PM of a device.
1323  * @dev: Device to handle.
1324  * @check_resume: If set, check if there's a resume request for the device.
1325  *
1326  * Increment power.disable_depth for the device and if it was zero previously,
1327  * cancel all pending runtime PM requests for the device and wait for all
1328  * operations in progress to complete.  The device can be either active or
1329  * suspended after its runtime PM has been disabled.
1330  *
1331  * If @check_resume is set and there's a resume request pending when
1332  * __pm_runtime_disable() is called and power.disable_depth is zero, the
1333  * function will wake up the device before disabling its runtime PM.
1334  */
1335 void __pm_runtime_disable(struct device *dev, bool check_resume)
1336 {
1337         spin_lock_irq(&dev->power.lock);
1338
1339         if (dev->power.disable_depth > 0) {
1340                 dev->power.disable_depth++;
1341                 goto out;
1342         }
1343
1344         /*
1345          * Wake up the device if there's a resume request pending, because that
1346          * means there probably is some I/O to process and disabling runtime PM
1347          * shouldn't prevent the device from processing the I/O.
1348          */
1349         if (check_resume && dev->power.request_pending
1350             && dev->power.request == RPM_REQ_RESUME) {
1351                 /*
1352                  * Prevent suspends and idle notifications from being carried
1353                  * out after we have woken up the device.
1354                  */
1355                 pm_runtime_get_noresume(dev);
1356
1357                 rpm_resume(dev, 0);
1358
1359                 pm_runtime_put_noidle(dev);
1360         }
1361
1362         /* Update time accounting before disabling PM-runtime. */
1363         update_pm_runtime_accounting(dev);
1364
1365         if (!dev->power.disable_depth++)
1366                 __pm_runtime_barrier(dev);
1367
1368  out:
1369         spin_unlock_irq(&dev->power.lock);
1370 }
1371 EXPORT_SYMBOL_GPL(__pm_runtime_disable);
1372
1373 /**
1374  * pm_runtime_enable - Enable runtime PM of a device.
1375  * @dev: Device to handle.
1376  */
1377 void pm_runtime_enable(struct device *dev)
1378 {
1379         unsigned long flags;
1380
1381         spin_lock_irqsave(&dev->power.lock, flags);
1382
1383         if (dev->power.disable_depth > 0) {
1384                 dev->power.disable_depth--;
1385
1386                 /* About to enable runtime pm, set accounting_timestamp to now */
1387                 if (!dev->power.disable_depth)
1388                         dev->power.accounting_timestamp = ktime_get_mono_fast_ns();
1389         } else {
1390                 dev_warn(dev, "Unbalanced %s!\n", __func__);
1391         }
1392
1393         WARN(!dev->power.disable_depth &&
1394              dev->power.runtime_status == RPM_SUSPENDED &&
1395              !dev->power.ignore_children &&
1396              atomic_read(&dev->power.child_count) > 0,
1397              "Enabling runtime PM for inactive device (%s) with active children\n",
1398              dev_name(dev));
1399
1400         spin_unlock_irqrestore(&dev->power.lock, flags);
1401 }
1402 EXPORT_SYMBOL_GPL(pm_runtime_enable);
1403
1404 /**
1405  * pm_runtime_forbid - Block runtime PM of a device.
1406  * @dev: Device to handle.
1407  *
1408  * Increase the device's usage count and clear its power.runtime_auto flag,
1409  * so that it cannot be suspended at run time until pm_runtime_allow() is called
1410  * for it.
1411  */
1412 void pm_runtime_forbid(struct device *dev)
1413 {
1414         spin_lock_irq(&dev->power.lock);
1415         if (!dev->power.runtime_auto)
1416                 goto out;
1417
1418         dev->power.runtime_auto = false;
1419         atomic_inc(&dev->power.usage_count);
1420         rpm_resume(dev, 0);
1421
1422  out:
1423         spin_unlock_irq(&dev->power.lock);
1424 }
1425 EXPORT_SYMBOL_GPL(pm_runtime_forbid);
1426
1427 /**
1428  * pm_runtime_allow - Unblock runtime PM of a device.
1429  * @dev: Device to handle.
1430  *
1431  * Decrease the device's usage count and set its power.runtime_auto flag.
1432  */
1433 void pm_runtime_allow(struct device *dev)
1434 {
1435         spin_lock_irq(&dev->power.lock);
1436         if (dev->power.runtime_auto)
1437                 goto out;
1438
1439         dev->power.runtime_auto = true;
1440         if (atomic_dec_and_test(&dev->power.usage_count))
1441                 rpm_idle(dev, RPM_AUTO | RPM_ASYNC);
1442         else
1443                 trace_rpm_usage_rcuidle(dev, RPM_AUTO | RPM_ASYNC);
1444
1445  out:
1446         spin_unlock_irq(&dev->power.lock);
1447 }
1448 EXPORT_SYMBOL_GPL(pm_runtime_allow);
1449
1450 /**
1451  * pm_runtime_no_callbacks - Ignore runtime PM callbacks for a device.
1452  * @dev: Device to handle.
1453  *
1454  * Set the power.no_callbacks flag, which tells the PM core that this
1455  * device is power-managed through its parent and has no runtime PM
1456  * callbacks of its own.  The runtime sysfs attributes will be removed.
1457  */
1458 void pm_runtime_no_callbacks(struct device *dev)
1459 {
1460         spin_lock_irq(&dev->power.lock);
1461         dev->power.no_callbacks = 1;
1462         spin_unlock_irq(&dev->power.lock);
1463         if (device_is_registered(dev))
1464                 rpm_sysfs_remove(dev);
1465 }
1466 EXPORT_SYMBOL_GPL(pm_runtime_no_callbacks);
1467
1468 /**
1469  * pm_runtime_irq_safe - Leave interrupts disabled during callbacks.
1470  * @dev: Device to handle
1471  *
1472  * Set the power.irq_safe flag, which tells the PM core that the
1473  * ->runtime_suspend() and ->runtime_resume() callbacks for this device should
1474  * always be invoked with the spinlock held and interrupts disabled.  It also
1475  * causes the parent's usage counter to be permanently incremented, preventing
1476  * the parent from runtime suspending -- otherwise an irq-safe child might have
1477  * to wait for a non-irq-safe parent.
1478  */
1479 void pm_runtime_irq_safe(struct device *dev)
1480 {
1481         if (dev->parent)
1482                 pm_runtime_get_sync(dev->parent);
1483         spin_lock_irq(&dev->power.lock);
1484         dev->power.irq_safe = 1;
1485         spin_unlock_irq(&dev->power.lock);
1486 }
1487 EXPORT_SYMBOL_GPL(pm_runtime_irq_safe);
1488
1489 /**
1490  * update_autosuspend - Handle a change to a device's autosuspend settings.
1491  * @dev: Device to handle.
1492  * @old_delay: The former autosuspend_delay value.
1493  * @old_use: The former use_autosuspend value.
1494  *
1495  * Prevent runtime suspend if the new delay is negative and use_autosuspend is
1496  * set; otherwise allow it.  Send an idle notification if suspends are allowed.
1497  *
1498  * This function must be called under dev->power.lock with interrupts disabled.
1499  */
1500 static void update_autosuspend(struct device *dev, int old_delay, int old_use)
1501 {
1502         int delay = dev->power.autosuspend_delay;
1503
1504         /* Should runtime suspend be prevented now? */
1505         if (dev->power.use_autosuspend && delay < 0) {
1506
1507                 /* If it used to be allowed then prevent it. */
1508                 if (!old_use || old_delay >= 0) {
1509                         atomic_inc(&dev->power.usage_count);
1510                         rpm_resume(dev, 0);
1511                 } else {
1512                         trace_rpm_usage_rcuidle(dev, 0);
1513                 }
1514         }
1515
1516         /* Runtime suspend should be allowed now. */
1517         else {
1518
1519                 /* If it used to be prevented then allow it. */
1520                 if (old_use && old_delay < 0)
1521                         atomic_dec(&dev->power.usage_count);
1522
1523                 /* Maybe we can autosuspend now. */
1524                 rpm_idle(dev, RPM_AUTO);
1525         }
1526 }
1527
1528 /**
1529  * pm_runtime_set_autosuspend_delay - Set a device's autosuspend_delay value.
1530  * @dev: Device to handle.
1531  * @delay: Value of the new delay in milliseconds.
1532  *
1533  * Set the device's power.autosuspend_delay value.  If it changes to negative
1534  * and the power.use_autosuspend flag is set, prevent runtime suspends.  If it
1535  * changes the other way, allow runtime suspends.
1536  */
1537 void pm_runtime_set_autosuspend_delay(struct device *dev, int delay)
1538 {
1539         int old_delay, old_use;
1540
1541         spin_lock_irq(&dev->power.lock);
1542         old_delay = dev->power.autosuspend_delay;
1543         old_use = dev->power.use_autosuspend;
1544         dev->power.autosuspend_delay = delay;
1545         update_autosuspend(dev, old_delay, old_use);
1546         spin_unlock_irq(&dev->power.lock);
1547 }
1548 EXPORT_SYMBOL_GPL(pm_runtime_set_autosuspend_delay);
1549
1550 /**
1551  * __pm_runtime_use_autosuspend - Set a device's use_autosuspend flag.
1552  * @dev: Device to handle.
1553  * @use: New value for use_autosuspend.
1554  *
1555  * Set the device's power.use_autosuspend flag, and allow or prevent runtime
1556  * suspends as needed.
1557  */
1558 void __pm_runtime_use_autosuspend(struct device *dev, bool use)
1559 {
1560         int old_delay, old_use;
1561
1562         spin_lock_irq(&dev->power.lock);
1563         old_delay = dev->power.autosuspend_delay;
1564         old_use = dev->power.use_autosuspend;
1565         dev->power.use_autosuspend = use;
1566         update_autosuspend(dev, old_delay, old_use);
1567         spin_unlock_irq(&dev->power.lock);
1568 }
1569 EXPORT_SYMBOL_GPL(__pm_runtime_use_autosuspend);
1570
1571 /**
1572  * pm_runtime_init - Initialize runtime PM fields in given device object.
1573  * @dev: Device object to initialize.
1574  */
1575 void pm_runtime_init(struct device *dev)
1576 {
1577         dev->power.runtime_status = RPM_SUSPENDED;
1578         dev->power.idle_notification = false;
1579
1580         dev->power.disable_depth = 1;
1581         atomic_set(&dev->power.usage_count, 0);
1582
1583         dev->power.runtime_error = 0;
1584
1585         atomic_set(&dev->power.child_count, 0);
1586         pm_suspend_ignore_children(dev, false);
1587         dev->power.runtime_auto = true;
1588
1589         dev->power.request_pending = false;
1590         dev->power.request = RPM_REQ_NONE;
1591         dev->power.deferred_resume = false;
1592         INIT_WORK(&dev->power.work, pm_runtime_work);
1593
1594         dev->power.timer_expires = 0;
1595         hrtimer_init(&dev->power.suspend_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1596         dev->power.suspend_timer.function = pm_suspend_timer_fn;
1597
1598         init_waitqueue_head(&dev->power.wait_queue);
1599 }
1600
1601 /**
1602  * pm_runtime_reinit - Re-initialize runtime PM fields in given device object.
1603  * @dev: Device object to re-initialize.
1604  */
1605 void pm_runtime_reinit(struct device *dev)
1606 {
1607         if (!pm_runtime_enabled(dev)) {
1608                 if (dev->power.runtime_status == RPM_ACTIVE)
1609                         pm_runtime_set_suspended(dev);
1610                 if (dev->power.irq_safe) {
1611                         spin_lock_irq(&dev->power.lock);
1612                         dev->power.irq_safe = 0;
1613                         spin_unlock_irq(&dev->power.lock);
1614                         if (dev->parent)
1615                                 pm_runtime_put(dev->parent);
1616                 }
1617         }
1618 }
1619
1620 /**
1621  * pm_runtime_remove - Prepare for removing a device from device hierarchy.
1622  * @dev: Device object being removed from device hierarchy.
1623  */
1624 void pm_runtime_remove(struct device *dev)
1625 {
1626         __pm_runtime_disable(dev, false);
1627         pm_runtime_reinit(dev);
1628 }
1629
1630 /**
1631  * pm_runtime_clean_up_links - Prepare links to consumers for driver removal.
1632  * @dev: Device whose driver is going to be removed.
1633  *
1634  * Check links from this device to any consumers and if any of them have active
1635  * runtime PM references to the device, drop the usage counter of the device
1636  * (as many times as needed).
1637  *
1638  * Links with the DL_FLAG_MANAGED flag unset are ignored.
1639  *
1640  * Since the device is guaranteed to be runtime-active at the point this is
1641  * called, nothing else needs to be done here.
1642  *
1643  * Moreover, this is called after device_links_busy() has returned 'false', so
1644  * the status of each link is guaranteed to be DL_STATE_SUPPLIER_UNBIND and
1645  * therefore rpm_active can't be manipulated concurrently.
1646  */
1647 void pm_runtime_clean_up_links(struct device *dev)
1648 {
1649         struct device_link *link;
1650         int idx;
1651
1652         idx = device_links_read_lock();
1653
1654         list_for_each_entry_rcu(link, &dev->links.consumers, s_node,
1655                                 device_links_read_lock_held()) {
1656                 if (!(link->flags & DL_FLAG_MANAGED))
1657                         continue;
1658
1659                 while (refcount_dec_not_one(&link->rpm_active))
1660                         pm_runtime_put_noidle(dev);
1661         }
1662
1663         device_links_read_unlock(idx);
1664 }
1665
1666 /**
1667  * pm_runtime_get_suppliers - Resume and reference-count supplier devices.
1668  * @dev: Consumer device.
1669  */
1670 void pm_runtime_get_suppliers(struct device *dev)
1671 {
1672         struct device_link *link;
1673         int idx;
1674
1675         idx = device_links_read_lock();
1676
1677         list_for_each_entry_rcu(link, &dev->links.suppliers, c_node,
1678                                 device_links_read_lock_held())
1679                 if (link->flags & DL_FLAG_PM_RUNTIME) {
1680                         link->supplier_preactivated = true;
1681                         refcount_inc(&link->rpm_active);
1682                         pm_runtime_get_sync(link->supplier);
1683                 }
1684
1685         device_links_read_unlock(idx);
1686 }
1687
1688 /**
1689  * pm_runtime_put_suppliers - Drop references to supplier devices.
1690  * @dev: Consumer device.
1691  */
1692 void pm_runtime_put_suppliers(struct device *dev)
1693 {
1694         struct device_link *link;
1695         int idx;
1696
1697         idx = device_links_read_lock();
1698
1699         list_for_each_entry_rcu(link, &dev->links.suppliers, c_node,
1700                                 device_links_read_lock_held())
1701                 if (link->supplier_preactivated) {
1702                         link->supplier_preactivated = false;
1703                         if (refcount_dec_not_one(&link->rpm_active))
1704                                 pm_runtime_put(link->supplier);
1705                 }
1706
1707         device_links_read_unlock(idx);
1708 }
1709
1710 void pm_runtime_new_link(struct device *dev)
1711 {
1712         spin_lock_irq(&dev->power.lock);
1713         dev->power.links_count++;
1714         spin_unlock_irq(&dev->power.lock);
1715 }
1716
1717 void pm_runtime_drop_link(struct device *dev)
1718 {
1719         spin_lock_irq(&dev->power.lock);
1720         WARN_ON(dev->power.links_count == 0);
1721         dev->power.links_count--;
1722         spin_unlock_irq(&dev->power.lock);
1723 }
1724
1725 static bool pm_runtime_need_not_resume(struct device *dev)
1726 {
1727         return atomic_read(&dev->power.usage_count) <= 1 &&
1728                 (atomic_read(&dev->power.child_count) == 0 ||
1729                  dev->power.ignore_children);
1730 }
1731
1732 /**
1733  * pm_runtime_force_suspend - Force a device into suspend state if needed.
1734  * @dev: Device to suspend.
1735  *
1736  * Disable runtime PM so we safely can check the device's runtime PM status and
1737  * if it is active, invoke its ->runtime_suspend callback to suspend it and
1738  * change its runtime PM status field to RPM_SUSPENDED.  Also, if the device's
1739  * usage and children counters don't indicate that the device was in use before
1740  * the system-wide transition under way, decrement its parent's children counter
1741  * (if there is a parent).  Keep runtime PM disabled to preserve the state
1742  * unless we encounter errors.
1743  *
1744  * Typically this function may be invoked from a system suspend callback to make
1745  * sure the device is put into low power state and it should only be used during
1746  * system-wide PM transitions to sleep states.  It assumes that the analogous
1747  * pm_runtime_force_resume() will be used to resume the device.
1748  */
1749 int pm_runtime_force_suspend(struct device *dev)
1750 {
1751         int (*callback)(struct device *);
1752         int ret;
1753
1754         pm_runtime_disable(dev);
1755         if (pm_runtime_status_suspended(dev))
1756                 return 0;
1757
1758         callback = RPM_GET_CALLBACK(dev, runtime_suspend);
1759
1760         ret = callback ? callback(dev) : 0;
1761         if (ret)
1762                 goto err;
1763
1764         /*
1765          * If the device can stay in suspend after the system-wide transition
1766          * to the working state that will follow, drop the children counter of
1767          * its parent, but set its status to RPM_SUSPENDED anyway in case this
1768          * function will be called again for it in the meantime.
1769          */
1770         if (pm_runtime_need_not_resume(dev))
1771                 pm_runtime_set_suspended(dev);
1772         else
1773                 __update_runtime_status(dev, RPM_SUSPENDED);
1774
1775         return 0;
1776
1777 err:
1778         pm_runtime_enable(dev);
1779         return ret;
1780 }
1781 EXPORT_SYMBOL_GPL(pm_runtime_force_suspend);
1782
1783 /**
1784  * pm_runtime_force_resume - Force a device into resume state if needed.
1785  * @dev: Device to resume.
1786  *
1787  * Prior invoking this function we expect the user to have brought the device
1788  * into low power state by a call to pm_runtime_force_suspend(). Here we reverse
1789  * those actions and bring the device into full power, if it is expected to be
1790  * used on system resume.  In the other case, we defer the resume to be managed
1791  * via runtime PM.
1792  *
1793  * Typically this function may be invoked from a system resume callback.
1794  */
1795 int pm_runtime_force_resume(struct device *dev)
1796 {
1797         int (*callback)(struct device *);
1798         int ret = 0;
1799
1800         if (!pm_runtime_status_suspended(dev) || pm_runtime_need_not_resume(dev))
1801                 goto out;
1802
1803         /*
1804          * The value of the parent's children counter is correct already, so
1805          * just update the status of the device.
1806          */
1807         __update_runtime_status(dev, RPM_ACTIVE);
1808
1809         callback = RPM_GET_CALLBACK(dev, runtime_resume);
1810
1811         ret = callback ? callback(dev) : 0;
1812         if (ret) {
1813                 pm_runtime_set_suspended(dev);
1814                 goto out;
1815         }
1816
1817         pm_runtime_mark_last_busy(dev);
1818 out:
1819         pm_runtime_enable(dev);
1820         return ret;
1821 }
1822 EXPORT_SYMBOL_GPL(pm_runtime_force_resume);