drivers/base/power/runtime.c

   1 /*
   2  * drivers/base/power/runtime.c - Helper functions for device runtime PM
   3  *
   4  * Copyright (c) 2009 Rafael J. Wysocki <rjw@sisk.pl>, Novell Inc.
   5  * Copyright (C) 2010 Alan Stern <stern@rowland.harvard.edu>
   6  *
   7  * This file is released under the GPLv2.
   8  */
   9
  10 #include <linux/sched/mm.h>
  11 #include <linux/ktime.h>
  12 #include <linux/hrtimer.h>
  13 #include <linux/export.h>
  14 #include <linux/pm_runtime.h>
  15 #include <linux/pm_wakeirq.h>
  16 #include <trace/events/rpm.h>
  17
  18 #include "../base.h"
  19 #include "power.h"
  20
  21 typedef int (*pm_callback_t)(struct device *);
  22
  23 static pm_callback_t __rpm_get_callback(struct device *dev, size_t cb_offset)
  24 {
  25         pm_callback_t cb;
  26         const struct dev_pm_ops *ops;
  27
  28         if (dev->pm_domain)
  29                 ops = &dev->pm_domain->ops;
  30         else if (dev->type && dev->type->pm)
  31                 ops = dev->type->pm;
  32         else if (dev->class && dev->class->pm)
  33                 ops = dev->class->pm;
  34         else if (dev->bus && dev->bus->pm)
  35                 ops = dev->bus->pm;
  36         else
  37                 ops = NULL;
  38
  39         if (ops)
  40                 cb = *(pm_callback_t *)((void *)ops + cb_offset);
  41         else
  42                 cb = NULL;
  43
  44         if (!cb && dev->driver && dev->driver->pm)
  45                 cb = *(pm_callback_t *)((void *)dev->driver->pm + cb_offset);
  46
  47         return cb;
  48 }
  49
  50 #define RPM_GET_CALLBACK(dev, callback) \
  51                 __rpm_get_callback(dev, offsetof(struct dev_pm_ops, callback))
  52
  53 static int rpm_resume(struct device *dev, int rpmflags);
  54 static int rpm_suspend(struct device *dev, int rpmflags);
  55
  56 /**
  57  * update_pm_runtime_accounting - Update the time accounting of power states
  58  * @dev: Device to update the accounting for
  59  *
  60  * In order to be able to have time accounting of the various power states
  61  * (as used by programs such as PowerTOP to show the effectiveness of runtime
  62  * PM), we need to track the time spent in each state.
  63  * update_pm_runtime_accounting must be called each time before the
  64  * runtime_status field is updated, to account the time in the old state
  65  * correctly.
  66  */
  67 void update_pm_runtime_accounting(struct device *dev)
  68 {
  69         unsigned long now = jiffies;
  70         unsigned long delta;
  71
  72         delta = now - dev->power.accounting_timestamp;
  73
  74         dev->power.accounting_timestamp = now;
  75
  76         if (dev->power.disable_depth > 0)
  77                 return;
  78
  79         if (dev->power.runtime_status == RPM_SUSPENDED)
  80                 dev->power.suspended_jiffies += delta;
  81         else
  82                 dev->power.active_jiffies += delta;
  83 }
  84
  85 static void __update_runtime_status(struct device *dev, enum rpm_status status)
  86 {
  87         update_pm_runtime_accounting(dev);
  88         dev->power.runtime_status = status;
  89 }
  90
  91 /**
  92  * pm_runtime_deactivate_timer - Deactivate given device's suspend timer.
  93  * @dev: Device to handle.
  94  */
  95 static void pm_runtime_deactivate_timer(struct device *dev)
  96 {
  97         if (dev->power.timer_expires > 0) {
  98                 hrtimer_cancel(&dev->power.suspend_timer);
  99                 dev->power.timer_expires = 0;
 100         }
 101 }
 102
 103 /**
 104  * pm_runtime_cancel_pending - Deactivate suspend timer and cancel requests.
 105  * @dev: Device to handle.
 106  */
 107 static void pm_runtime_cancel_pending(struct device *dev)
 108 {
 109         pm_runtime_deactivate_timer(dev);
 110         /*
 111          * In case there's a request pending, make sure its work function will
 112          * return without doing anything.
 113          */
 114         dev->power.request = RPM_REQ_NONE;
 115 }
 116
 117 /*
 118  * pm_runtime_autosuspend_expiration - Get a device's autosuspend-delay expiration time.
 119  * @dev: Device to handle.
 120  *
 121  * Compute the autosuspend-delay expiration time based on the device's
 122  * power.last_busy time.  If the delay has already expired or is disabled
 123  * (negative) or the power.use_autosuspend flag isn't set, return 0.
 124  * Otherwise return the expiration time in nanoseconds (adjusted to be nonzero).
 125  *
 126  * This function may be called either with or without dev->power.lock held.
 127  * Either way it can be racy, since power.last_busy may be updated at any time.
 128  */
 129 u64 pm_runtime_autosuspend_expiration(struct device *dev)
 130 {
 131         int autosuspend_delay;
 132         u64 last_busy, expires = 0;
 133         u64 now = ktime_get_mono_fast_ns();
 134
 135         if (!dev->power.use_autosuspend)
 136                 goto out;
 137
 138         autosuspend_delay = READ_ONCE(dev->power.autosuspend_delay);
 139         if (autosuspend_delay < 0)
 140                 goto out;
 141
 142         last_busy = READ_ONCE(dev->power.last_busy);
 143
 144         expires = last_busy + (u64)autosuspend_delay * NSEC_PER_MSEC;
 145         if (expires <= now)
 146                 expires = 0;    /* Already expired. */
 147
 148  out:
 149         return expires;
 150 }
 151 EXPORT_SYMBOL_GPL(pm_runtime_autosuspend_expiration);
 152
 153 static int dev_memalloc_noio(struct device *dev, void *data)
 154 {
 155         return dev->power.memalloc_noio;
 156 }
 157
 158 /*
 159  * pm_runtime_set_memalloc_noio - Set a device's memalloc_noio flag.
 160  * @dev: Device to handle.
 161  * @enable: True for setting the flag and False for clearing the flag.
 162  *
 163  * Set the flag for all devices in the path from the device to the
 164  * root device in the device tree if @enable is true, otherwise clear
 165  * the flag for devices in the path whose siblings don't set the flag.
 166  *
 167  * The function should only be called by block device, or network
 168  * device driver for solving the deadlock problem during runtime
 169  * resume/suspend:
 170  *
 171  *     If memory allocation with GFP_KERNEL is called inside runtime
 172  *     resume/suspend callback of any one of its ancestors(or the
 173  *     block device itself), the deadlock may be triggered inside the
 174  *     memory allocation since it might not complete until the block
 175  *     device becomes active and the involed page I/O finishes. The
 176  *     situation is pointed out first by Alan Stern. Network device
 177  *     are involved in iSCSI kind of situation.
 178  *
 179  * The lock of dev_hotplug_mutex is held in the function for handling
 180  * hotplug race because pm_runtime_set_memalloc_noio() may be called
 181  * in async probe().
 182  *
 183  * The function should be called between device_add() and device_del()
 184  * on the affected device(block/network device).
 185  */
 186 void pm_runtime_set_memalloc_noio(struct device *dev, bool enable)
 187 {
 188         static DEFINE_MUTEX(dev_hotplug_mutex);
 189
 190         mutex_lock(&dev_hotplug_mutex);
 191         for (;;) {
 192                 bool enabled;
 193
 194                 /* hold power lock since bitfield is not SMP-safe. */
 195                 spin_lock_irq(&dev->power.lock);
 196                 enabled = dev->power.memalloc_noio;
 197                 dev->power.memalloc_noio = enable;
 198                 spin_unlock_irq(&dev->power.lock);
 199
 200                 /*
 201                  * not need to enable ancestors any more if the device
 202                  * has been enabled.
 203                  */
 204                 if (enabled && enable)
 205                         break;
 206
 207                 dev = dev->parent;
 208
 209                 /*
 210                  * clear flag of the parent device only if all the
 211                  * children don't set the flag because ancestor's
 212                  * flag was set by any one of the descendants.
 213                  */
 214                 if (!dev || (!enable &&
 215                              device_for_each_child(dev, NULL,
 216                                                    dev_memalloc_noio)))
 217                         break;
 218         }
 219         mutex_unlock(&dev_hotplug_mutex);
 220 }
 221 EXPORT_SYMBOL_GPL(pm_runtime_set_memalloc_noio);
 222
 223 /**
 224  * rpm_check_suspend_allowed - Test whether a device may be suspended.
 225  * @dev: Device to test.
 226  */
 227 static int rpm_check_suspend_allowed(struct device *dev)
 228 {
 229         int retval = 0;
 230
 231         if (dev->power.runtime_error)
 232                 retval = -EINVAL;
 233         else if (dev->power.disable_depth > 0)
 234                 retval = -EACCES;
 235         else if (atomic_read(&dev->power.usage_count) > 0)
 236                 retval = -EAGAIN;
 237         else if (!dev->power.ignore_children &&
 238                         atomic_read(&dev->power.child_count))
 239                 retval = -EBUSY;
 240
 241         /* Pending resume requests take precedence over suspends. */
 242         else if ((dev->power.deferred_resume
 243                         && dev->power.runtime_status == RPM_SUSPENDING)
 244             || (dev->power.request_pending
 245                         && dev->power.request == RPM_REQ_RESUME))
 246                 retval = -EAGAIN;
 247         else if (__dev_pm_qos_read_value(dev) == 0)
 248                 retval = -EPERM;
 249         else if (dev->power.runtime_status == RPM_SUSPENDED)
 250                 retval = 1;
 251
 252         return retval;
 253 }
 254
 255 static int rpm_get_suppliers(struct device *dev)
 256 {
 257         struct device_link *link;
 258
 259         list_for_each_entry_rcu(link, &dev->links.suppliers, c_node) {
 260                 int retval;
 261
 262                 if (!(link->flags & DL_FLAG_PM_RUNTIME) ||
 263                     READ_ONCE(link->status) == DL_STATE_SUPPLIER_UNBIND)
 264                         continue;
 265
 266                 retval = pm_runtime_get_sync(link->supplier);
 267                 /* Ignore suppliers with disabled runtime PM. */
 268                 if (retval < 0 && retval != -EACCES) {
 269                         pm_runtime_put_noidle(link->supplier);
 270                         return retval;
 271                 }
 272                 refcount_inc(&link->rpm_active);
 273         }
 274         return 0;
 275 }
 276
 277 static void rpm_put_suppliers(struct device *dev)
 278 {
 279         struct device_link *link;
 280
 281         list_for_each_entry_rcu(link, &dev->links.suppliers, c_node) {
 282                 if (READ_ONCE(link->status) == DL_STATE_SUPPLIER_UNBIND)
 283                         continue;
 284
 285                 while (refcount_dec_not_one(&link->rpm_active))
 286                         pm_runtime_put(link->supplier);
 287         }
 288 }
 289
 290 /**
 291  * __rpm_callback - Run a given runtime PM callback for a given device.
 292  * @cb: Runtime PM callback to run.
 293  * @dev: Device to run the callback for.
 294  */
 295 static int __rpm_callback(int (*cb)(struct device *), struct device *dev)
 296         __releases(&dev->power.lock) __acquires(&dev->power.lock)
 297 {
 298         int retval, idx;
 299         bool use_links = dev->power.links_count > 0;
 300
 301         if (dev->power.irq_safe) {
 302                 spin_unlock(&dev->power.lock);
 303         } else {
 304                 spin_unlock_irq(&dev->power.lock);
 305
 306                 /*
 307                  * Resume suppliers if necessary.
 308                  *
 309                  * The device's runtime PM status cannot change until this
 310                  * routine returns, so it is safe to read the status outside of
 311                  * the lock.
 312                  */
 313                 if (use_links && dev->power.runtime_status == RPM_RESUMING) {
 314                         idx = device_links_read_lock();
 315
 316                         retval = rpm_get_suppliers(dev);
 317                         if (retval)
 318                                 goto fail;
 319
 320                         device_links_read_unlock(idx);
 321                 }
 322         }
 323
 324         retval = cb(dev);
 325
 326         if (dev->power.irq_safe) {
 327                 spin_lock(&dev->power.lock);
 328         } else {
 329                 /*
 330                  * If the device is suspending and the callback has returned
 331                  * success, drop the usage counters of the suppliers that have
 332                  * been reference counted on its resume.
 333                  *
 334                  * Do that if resume fails too.
 335                  */
 336                 if (use_links
 337                     && ((dev->power.runtime_status == RPM_SUSPENDING && !retval)
 338                     || (dev->power.runtime_status == RPM_RESUMING && retval))) {
 339                         idx = device_links_read_lock();
 340
 341  fail:
 342                         rpm_put_suppliers(dev);
 343
 344                         device_links_read_unlock(idx);
 345                 }
 346
 347                 spin_lock_irq(&dev->power.lock);
 348         }
 349
 350         return retval;
 351 }
 352
 353 /**
 354  * rpm_idle - Notify device bus type if the device can be suspended.
 355  * @dev: Device to notify the bus type about.
 356  * @rpmflags: Flag bits.
 357  *
 358  * Check if the device's runtime PM status allows it to be suspended.  If
 359  * another idle notification has been started earlier, return immediately.  If
 360  * the RPM_ASYNC flag is set then queue an idle-notification request; otherwise
 361  * run the ->runtime_idle() callback directly. If the ->runtime_idle callback
 362  * doesn't exist or if it returns 0, call rpm_suspend with the RPM_AUTO flag.
 363  *
 364  * This function must be called under dev->power.lock with interrupts disabled.
 365  */
 366 static int rpm_idle(struct device *dev, int rpmflags)
 367 {
 368         int (*callback)(struct device *);
 369         int retval;
 370
 371         trace_rpm_idle_rcuidle(dev, rpmflags);
 372         retval = rpm_check_suspend_allowed(dev);
 373         if (retval < 0)
 374                 ;       /* Conditions are wrong. */
 375
 376         /* Idle notifications are allowed only in the RPM_ACTIVE state. */
 377         else if (dev->power.runtime_status != RPM_ACTIVE)
 378                 retval = -EAGAIN;
 379
 380         /*
 381          * Any pending request other than an idle notification takes
 382          * precedence over us, except that the timer may be running.
 383          */
 384         else if (dev->power.request_pending &&
 385             dev->power.request > RPM_REQ_IDLE)
 386                 retval = -EAGAIN;
 387
 388         /* Act as though RPM_NOWAIT is always set. */
 389         else if (dev->power.idle_notification)
 390                 retval = -EINPROGRESS;
 391         if (retval)
 392                 goto out;
 393
 394         /* Pending requests need to be canceled. */
 395         dev->power.request = RPM_REQ_NONE;
 396
 397         if (dev->power.no_callbacks)
 398                 goto out;
 399
 400         /* Carry out an asynchronous or a synchronous idle notification. */
 401         if (rpmflags & RPM_ASYNC) {
 402                 dev->power.request = RPM_REQ_IDLE;
 403                 if (!dev->power.request_pending) {
 404                         dev->power.request_pending = true;
 405                         queue_work(pm_wq, &dev->power.work);
 406                 }
 407                 trace_rpm_return_int_rcuidle(dev, _THIS_IP_, 0);
 408                 return 0;
 409         }
 410
 411         dev->power.idle_notification = true;
 412
 413         callback = RPM_GET_CALLBACK(dev, runtime_idle);
 414
 415         if (callback)
 416                 retval = __rpm_callback(callback, dev);
 417
 418         dev->power.idle_notification = false;
 419         wake_up_all(&dev->power.wait_queue);
 420
 421  out:
 422         trace_rpm_return_int_rcuidle(dev, _THIS_IP_, retval);
 423         return retval ? retval : rpm_suspend(dev, rpmflags | RPM_AUTO);
 424 }
 425
 426 /**
 427  * rpm_callback - Run a given runtime PM callback for a given device.
 428  * @cb: Runtime PM callback to run.
 429  * @dev: Device to run the callback for.
 430  */
 431 static int rpm_callback(int (*cb)(struct device *), struct device *dev)
 432 {
 433         int retval;
 434
 435         if (!cb)
 436                 return -ENOSYS;
 437
 438         if (dev->power.memalloc_noio) {
 439                 unsigned int noio_flag;
 440
 441                 /*
 442                  * Deadlock might be caused if memory allocation with
 443                  * GFP_KERNEL happens inside runtime_suspend and
 444                  * runtime_resume callbacks of one block device's
 445                  * ancestor or the block device itself. Network
 446                  * device might be thought as part of iSCSI block
 447                  * device, so network device and its ancestor should
 448                  * be marked as memalloc_noio too.
 449                  */
 450                 noio_flag = memalloc_noio_save();
 451                 retval = __rpm_callback(cb, dev);
 452                 memalloc_noio_restore(noio_flag);
 453         } else {
 454                 retval = __rpm_callback(cb, dev);
 455         }
 456
 457         dev->power.runtime_error = retval;
 458         return retval != -EACCES ? retval : -EIO;
 459 }
 460
 461 /**
 462  * rpm_suspend - Carry out runtime suspend of given device.
 463  * @dev: Device to suspend.
 464  * @rpmflags: Flag bits.
 465  *
 466  * Check if the device's runtime PM status allows it to be suspended.
 467  * Cancel a pending idle notification, autosuspend or suspend. If
 468  * another suspend has been started earlier, either return immediately
 469  * or wait for it to finish, depending on the RPM_NOWAIT and RPM_ASYNC
 470  * flags. If the RPM_ASYNC flag is set then queue a suspend request;
 471  * otherwise run the ->runtime_suspend() callback directly. When
 472  * ->runtime_suspend succeeded, if a deferred resume was requested while
 473  * the callback was running then carry it out, otherwise send an idle
 474  * notification for its parent (if the suspend succeeded and both
 475  * ignore_children of parent->power and irq_safe of dev->power are not set).
 476  * If ->runtime_suspend failed with -EAGAIN or -EBUSY, and if the RPM_AUTO
 477  * flag is set and the next autosuspend-delay expiration time is in the
 478  * future, schedule another autosuspend attempt.
 479  *
 480  * This function must be called under dev->power.lock with interrupts disabled.
 481  */
 482 static int rpm_suspend(struct device *dev, int rpmflags)
 483         __releases(&dev->power.lock) __acquires(&dev->power.lock)
 484 {
 485         int (*callback)(struct device *);
 486         struct device *parent = NULL;
 487         int retval;
 488
 489         trace_rpm_suspend_rcuidle(dev, rpmflags);
 490
 491  repeat:
 492         retval = rpm_check_suspend_allowed(dev);
 493
 494         if (retval < 0)
 495                 ;       /* Conditions are wrong. */
 496
 497         /* Synchronous suspends are not allowed in the RPM_RESUMING state. */
 498         else if (dev->power.runtime_status == RPM_RESUMING &&
 499             !(rpmflags & RPM_ASYNC))
 500                 retval = -EAGAIN;
 501         if (retval)
 502                 goto out;
 503
 504         /* If the autosuspend_delay time hasn't expired yet, reschedule. */
 505         if ((rpmflags & RPM_AUTO)
 506             && dev->power.runtime_status != RPM_SUSPENDING) {
 507                 u64 expires = pm_runtime_autosuspend_expiration(dev);
 508
 509                 if (expires != 0) {
 510                         /* Pending requests need to be canceled. */
 511                         dev->power.request = RPM_REQ_NONE;
 512
 513                         /*
 514                          * Optimization: If the timer is already running and is
 515                          * set to expire at or before the autosuspend delay,
 516                          * avoid the overhead of resetting it.  Just let it
 517                          * expire; pm_suspend_timer_fn() will take care of the
 518                          * rest.
 519                          */
 520                         if (!(dev->power.timer_expires &&
 521                                         dev->power.timer_expires <= expires)) {
 522                                 /*
 523                                  * We add a slack of 25% to gather wakeups
 524                                  * without sacrificing the granularity.
 525                                  */
 526                                 u64 slack = (u64)READ_ONCE(dev->power.autosuspend_delay) *
 527                                                     (NSEC_PER_MSEC >> 2);
 528
 529                                 dev->power.timer_expires = expires;
 530                                 hrtimer_start_range_ns(&dev->power.suspend_timer,
 531                                                 ns_to_ktime(expires),
 532                                                 slack,
 533                                                 HRTIMER_MODE_ABS);
 534                         }
 535                         dev->power.timer_autosuspends = 1;
 536                         goto out;
 537                 }
 538         }
 539
 540         /* Other scheduled or pending requests need to be canceled. */
 541         pm_runtime_cancel_pending(dev);
 542
 543         if (dev->power.runtime_status == RPM_SUSPENDING) {
 544                 DEFINE_WAIT(wait);
 545
 546                 if (rpmflags & (RPM_ASYNC | RPM_NOWAIT)) {
 547                         retval = -EINPROGRESS;
 548                         goto out;
 549                 }
 550
 551                 if (dev->power.irq_safe) {
 552                         spin_unlock(&dev->power.lock);
 553
 554                         cpu_relax();
 555
 556                         spin_lock(&dev->power.lock);
 557                         goto repeat;
 558                 }
 559
 560                 /* Wait for the other suspend running in parallel with us. */
 561                 for (;;) {
 562                         prepare_to_wait(&dev->power.wait_queue, &wait,
 563                                         TASK_UNINTERRUPTIBLE);
 564                         if (dev->power.runtime_status != RPM_SUSPENDING)
 565                                 break;
 566
 567                         spin_unlock_irq(&dev->power.lock);
 568
 569                         schedule();
 570
 571                         spin_lock_irq(&dev->power.lock);
 572                 }
 573                 finish_wait(&dev->power.wait_queue, &wait);
 574                 goto repeat;
 575         }
 576
 577         if (dev->power.no_callbacks)
 578                 goto no_callback;       /* Assume success. */
 579
 580         /* Carry out an asynchronous or a synchronous suspend. */
 581         if (rpmflags & RPM_ASYNC) {
 582                 dev->power.request = (rpmflags & RPM_AUTO) ?
 583                     RPM_REQ_AUTOSUSPEND : RPM_REQ_SUSPEND;
 584                 if (!dev->power.request_pending) {
 585                         dev->power.request_pending = true;
 586                         queue_work(pm_wq, &dev->power.work);
 587                 }
 588                 goto out;
 589         }
 590
 591         __update_runtime_status(dev, RPM_SUSPENDING);
 592
 593         callback = RPM_GET_CALLBACK(dev, runtime_suspend);
 594
 595         dev_pm_enable_wake_irq_check(dev, true);
 596         retval = rpm_callback(callback, dev);
 597         if (retval)
 598                 goto fail;
 599
 600  no_callback:
 601         __update_runtime_status(dev, RPM_SUSPENDED);
 602         pm_runtime_deactivate_timer(dev);
 603
 604         if (dev->parent) {
 605                 parent = dev->parent;
 606                 atomic_add_unless(&parent->power.child_count, -1, 0);
 607         }
 608         wake_up_all(&dev->power.wait_queue);
 609
 610         if (dev->power.deferred_resume) {
 611                 dev->power.deferred_resume = false;
 612                 rpm_resume(dev, 0);
 613                 retval = -EAGAIN;
 614                 goto out;
 615         }
 616
 617         /* Maybe the parent is now able to suspend. */
 618         if (parent && !parent->power.ignore_children && !dev->power.irq_safe) {
 619                 spin_unlock(&dev->power.lock);
 620
 621                 spin_lock(&parent->power.lock);
 622                 rpm_idle(parent, RPM_ASYNC);
 623                 spin_unlock(&parent->power.lock);
 624
 625                 spin_lock(&dev->power.lock);
 626         }
 627
 628  out:
 629         trace_rpm_return_int_rcuidle(dev, _THIS_IP_, retval);
 630
 631         return retval;
 632
 633  fail:
 634         dev_pm_disable_wake_irq_check(dev);
 635         __update_runtime_status(dev, RPM_ACTIVE);
 636         dev->power.deferred_resume = false;
 637         wake_up_all(&dev->power.wait_queue);
 638
 639         if (retval == -EAGAIN || retval == -EBUSY) {
 640                 dev->power.runtime_error = 0;
 641
 642                 /*
 643                  * If the callback routine failed an autosuspend, and
 644                  * if the last_busy time has been updated so that there
 645                  * is a new autosuspend expiration time, automatically
 646                  * reschedule another autosuspend.
 647                  */
 648                 if ((rpmflags & RPM_AUTO) &&
 649                     pm_runtime_autosuspend_expiration(dev) != 0)
 650                         goto repeat;
 651         } else {
 652                 pm_runtime_cancel_pending(dev);
 653         }
 654         goto out;
 655 }
 656
 657 /**
 658  * rpm_resume - Carry out runtime resume of given device.
 659  * @dev: Device to resume.
 660  * @rpmflags: Flag bits.
 661  *
 662  * Check if the device's runtime PM status allows it to be resumed.  Cancel
 663  * any scheduled or pending requests.  If another resume has been started
 664  * earlier, either return immediately or wait for it to finish, depending on the
 665  * RPM_NOWAIT and RPM_ASYNC flags.  Similarly, if there's a suspend running in
 666  * parallel with this function, either tell the other process to resume after
 667  * suspending (deferred_resume) or wait for it to finish.  If the RPM_ASYNC
 668  * flag is set then queue a resume request; otherwise run the
 669  * ->runtime_resume() callback directly.  Queue an idle notification for the
 670  * device if the resume succeeded.
 671  *
 672  * This function must be called under dev->power.lock with interrupts disabled.
 673  */
 674 static int rpm_resume(struct device *dev, int rpmflags)
 675         __releases(&dev->power.lock) __acquires(&dev->power.lock)
 676 {
 677         int (*callback)(struct device *);
 678         struct device *parent = NULL;
 679         int retval = 0;
 680
 681         trace_rpm_resume_rcuidle(dev, rpmflags);
 682
 683  repeat:
 684         if (dev->power.runtime_error)
 685                 retval = -EINVAL;
 686         else if (dev->power.disable_depth == 1 && dev->power.is_suspended
 687             && dev->power.runtime_status == RPM_ACTIVE)
 688                 retval = 1;
 689         else if (dev->power.disable_depth > 0)
 690                 retval = -EACCES;
 691         if (retval)
 692                 goto out;
 693
 694         /*
 695          * Other scheduled or pending requests need to be canceled.  Small
 696          * optimization: If an autosuspend timer is running, leave it running
 697          * rather than cancelling it now only to restart it again in the near
 698          * future.
 699          */
 700         dev->power.request = RPM_REQ_NONE;
 701         if (!dev->power.timer_autosuspends)
 702                 pm_runtime_deactivate_timer(dev);
 703
 704         if (dev->power.runtime_status == RPM_ACTIVE) {
 705                 retval = 1;
 706                 goto out;
 707         }
 708
 709         if (dev->power.runtime_status == RPM_RESUMING
 710             || dev->power.runtime_status == RPM_SUSPENDING) {
 711                 DEFINE_WAIT(wait);
 712
 713                 if (rpmflags & (RPM_ASYNC | RPM_NOWAIT)) {
 714                         if (dev->power.runtime_status == RPM_SUSPENDING)
 715                                 dev->power.deferred_resume = true;
 716                         else
 717                                 retval = -EINPROGRESS;
 718                         goto out;
 719                 }
 720
 721                 if (dev->power.irq_safe) {
 722                         spin_unlock(&dev->power.lock);
 723
 724                         cpu_relax();
 725
 726                         spin_lock(&dev->power.lock);
 727                         goto repeat;
 728                 }
 729
 730                 /* Wait for the operation carried out in parallel with us. */
 731                 for (;;) {
 732                         prepare_to_wait(&dev->power.wait_queue, &wait,
 733                                         TASK_UNINTERRUPTIBLE);
 734                         if (dev->power.runtime_status != RPM_RESUMING
 735                             && dev->power.runtime_status != RPM_SUSPENDING)
 736                                 break;
 737
 738                         spin_unlock_irq(&dev->power.lock);
 739
 740                         schedule();
 741
 742                         spin_lock_irq(&dev->power.lock);
 743                 }
 744                 finish_wait(&dev->power.wait_queue, &wait);
 745                 goto repeat;
 746         }
 747
 748         /*
 749          * See if we can skip waking up the parent.  This is safe only if
 750          * power.no_callbacks is set, because otherwise we don't know whether
 751          * the resume will actually succeed.
 752          */
 753         if (dev->power.no_callbacks && !parent && dev->parent) {
 754                 spin_lock_nested(&dev->parent->power.lock, SINGLE_DEPTH_NESTING);
 755                 if (dev->parent->power.disable_depth > 0
 756                     || dev->parent->power.ignore_children
 757                     || dev->parent->power.runtime_status == RPM_ACTIVE) {
 758                         atomic_inc(&dev->parent->power.child_count);
 759                         spin_unlock(&dev->parent->power.lock);
 760                         retval = 1;
 761                         goto no_callback;       /* Assume success. */
 762                 }
 763                 spin_unlock(&dev->parent->power.lock);
 764         }
 765
 766         /* Carry out an asynchronous or a synchronous resume. */
 767         if (rpmflags & RPM_ASYNC) {
 768                 dev->power.request = RPM_REQ_RESUME;
 769                 if (!dev->power.request_pending) {
 770                         dev->power.request_pending = true;
 771                         queue_work(pm_wq, &dev->power.work);
 772                 }
 773                 retval = 0;
 774                 goto out;
 775         }
 776
 777         if (!parent && dev->parent) {
 778                 /*
 779                  * Increment the parent's usage counter and resume it if
 780                  * necessary.  Not needed if dev is irq-safe; then the
 781                  * parent is permanently resumed.
 782                  */
 783                 parent = dev->parent;
 784                 if (dev->power.irq_safe)
 785                         goto skip_parent;
 786                 spin_unlock(&dev->power.lock);
 787
 788                 pm_runtime_get_noresume(parent);
 789
 790                 spin_lock(&parent->power.lock);
 791                 /*
 792                  * Resume the parent if it has runtime PM enabled and not been
 793                  * set to ignore its children.
 794                  */
 795                 if (!parent->power.disable_depth
 796                     && !parent->power.ignore_children) {
 797                         rpm_resume(parent, 0);
 798                         if (parent->power.runtime_status != RPM_ACTIVE)
 799                                 retval = -EBUSY;
 800                 }
 801                 spin_unlock(&parent->power.lock);
 802
 803                 spin_lock(&dev->power.lock);
 804                 if (retval)
 805                         goto out;
 806                 goto repeat;
 807         }
 808  skip_parent:
 809
 810         if (dev->power.no_callbacks)
 811                 goto no_callback;       /* Assume success. */
 812
 813         __update_runtime_status(dev, RPM_RESUMING);
 814
 815         callback = RPM_GET_CALLBACK(dev, runtime_resume);
 816
 817         dev_pm_disable_wake_irq_check(dev);
 818         retval = rpm_callback(callback, dev);
 819         if (retval) {
 820                 __update_runtime_status(dev, RPM_SUSPENDED);
 821                 pm_runtime_cancel_pending(dev);
 822                 dev_pm_enable_wake_irq_check(dev, false);
 823         } else {
 824  no_callback:
 825                 __update_runtime_status(dev, RPM_ACTIVE);
 826                 pm_runtime_mark_last_busy(dev);
 827                 if (parent)
 828                         atomic_inc(&parent->power.child_count);
 829         }
 830         wake_up_all(&dev->power.wait_queue);
 831
 832         if (retval >= 0)
 833                 rpm_idle(dev, RPM_ASYNC);
 834
 835  out:
 836         if (parent && !dev->power.irq_safe) {
 837                 spin_unlock_irq(&dev->power.lock);
 838
 839                 pm_runtime_put(parent);
 840
 841                 spin_lock_irq(&dev->power.lock);
 842         }
 843
 844         trace_rpm_return_int_rcuidle(dev, _THIS_IP_, retval);
 845
 846         return retval;
 847 }
 848
 849 /**
 850  * pm_runtime_work - Universal runtime PM work function.
 851  * @work: Work structure used for scheduling the execution of this function.
 852  *
 853  * Use @work to get the device object the work is to be done for, determine what
 854  * is to be done and execute the appropriate runtime PM function.
 855  */
 856 static void pm_runtime_work(struct work_struct *work)
 857 {
 858         struct device *dev = container_of(work, struct device, power.work);
 859         enum rpm_request req;
 860
 861         spin_lock_irq(&dev->power.lock);
 862
 863         if (!dev->power.request_pending)
 864                 goto out;
 865
 866         req = dev->power.request;
 867         dev->power.request = RPM_REQ_NONE;
 868         dev->power.request_pending = false;
 869
 870         switch (req) {
 871         case RPM_REQ_NONE:
 872                 break;
 873         case RPM_REQ_IDLE:
 874                 rpm_idle(dev, RPM_NOWAIT);
 875                 break;
 876         case RPM_REQ_SUSPEND:
 877                 rpm_suspend(dev, RPM_NOWAIT);
 878                 break;
 879         case RPM_REQ_AUTOSUSPEND:
 880                 rpm_suspend(dev, RPM_NOWAIT | RPM_AUTO);
 881                 break;
 882         case RPM_REQ_RESUME:
 883                 rpm_resume(dev, RPM_NOWAIT);
 884                 break;
 885         }
 886
 887  out:
 888         spin_unlock_irq(&dev->power.lock);
 889 }
 890
 891 /**
 892  * pm_suspend_timer_fn - Timer function for pm_schedule_suspend().
 893  * @data: Device pointer passed by pm_schedule_suspend().
 894  *
 895  * Check if the time is right and queue a suspend request.
 896  */
 897 static enum hrtimer_restart  pm_suspend_timer_fn(struct hrtimer *timer)
 898 {
 899         struct device *dev = container_of(timer, struct device, power.suspend_timer);
 900         unsigned long flags;
 901         u64 expires;
 902
 903         spin_lock_irqsave(&dev->power.lock, flags);
 904
 905         expires = dev->power.timer_expires;
 906         /*
 907          * If 'expires' is after the current time, we've been called
 908          * too early.
 909          */
 910         if (expires > 0 && expires < ktime_get_mono_fast_ns()) {
 911                 dev->power.timer_expires = 0;
 912                 rpm_suspend(dev, dev->power.timer_autosuspends ?
 913                     (RPM_ASYNC | RPM_AUTO) : RPM_ASYNC);
 914         }
 915
 916         spin_unlock_irqrestore(&dev->power.lock, flags);
 917
 918         return HRTIMER_NORESTART;
 919 }
 920
 921 /**
 922  * pm_schedule_suspend - Set up a timer to submit a suspend request in future.
 923  * @dev: Device to suspend.
 924  * @delay: Time to wait before submitting a suspend request, in milliseconds.
 925  */
 926 int pm_schedule_suspend(struct device *dev, unsigned int delay)
 927 {
 928         unsigned long flags;
 929         u64 expires;
 930         int retval;
 931
 932         spin_lock_irqsave(&dev->power.lock, flags);
 933
 934         if (!delay) {
 935                 retval = rpm_suspend(dev, RPM_ASYNC);
 936                 goto out;
 937         }
 938
 939         retval = rpm_check_suspend_allowed(dev);
 940         if (retval)
 941                 goto out;
 942
 943         /* Other scheduled or pending requests need to be canceled. */
 944         pm_runtime_cancel_pending(dev);
 945
 946         expires = ktime_get_mono_fast_ns() + (u64)delay * NSEC_PER_MSEC;
 947         dev->power.timer_expires = expires;
 948         dev->power.timer_autosuspends = 0;
 949         hrtimer_start(&dev->power.suspend_timer, expires, HRTIMER_MODE_ABS);
 950
 951  out:
 952         spin_unlock_irqrestore(&dev->power.lock, flags);
 953
 954         return retval;
 955 }
 956 EXPORT_SYMBOL_GPL(pm_schedule_suspend);
 957
 958 /**
 959  * __pm_runtime_idle - Entry point for runtime idle operations.
 960  * @dev: Device to send idle notification for.
 961  * @rpmflags: Flag bits.
 962  *
 963  * If the RPM_GET_PUT flag is set, decrement the device's usage count and
 964  * return immediately if it is larger than zero.  Then carry out an idle
 965  * notification, either synchronous or asynchronous.
 966  *
 967  * This routine may be called in atomic context if the RPM_ASYNC flag is set,
 968  * or if pm_runtime_irq_safe() has been called.
 969  */
 970 int __pm_runtime_idle(struct device *dev, int rpmflags)
 971 {
 972         unsigned long flags;
 973         int retval;
 974
 975         if (rpmflags & RPM_GET_PUT) {
 976                 if (!atomic_dec_and_test(&dev->power.usage_count))
 977                         return 0;
 978         }
 979
 980         might_sleep_if(!(rpmflags & RPM_ASYNC) && !dev->power.irq_safe);
 981
 982         spin_lock_irqsave(&dev->power.lock, flags);
 983         retval = rpm_idle(dev, rpmflags);
 984         spin_unlock_irqrestore(&dev->power.lock, flags);
 985
 986         return retval;
 987 }
 988 EXPORT_SYMBOL_GPL(__pm_runtime_idle);
 989
 990 /**
 991  * __pm_runtime_suspend - Entry point for runtime put/suspend operations.
 992  * @dev: Device to suspend.
 993  * @rpmflags: Flag bits.
 994  *
 995  * If the RPM_GET_PUT flag is set, decrement the device's usage count and
 996  * return immediately if it is larger than zero.  Then carry out a suspend,
 997  * either synchronous or asynchronous.
 998  *
 999  * This routine may be called in atomic context if the RPM_ASYNC flag is set,
1000  * or if pm_runtime_irq_safe() has been called.
1001  */
1002 int __pm_runtime_suspend(struct device *dev, int rpmflags)
1003 {
1004         unsigned long flags;
1005         int retval;
1006
1007         if (rpmflags & RPM_GET_PUT) {
1008                 if (!atomic_dec_and_test(&dev->power.usage_count))
1009                         return 0;
1010         }
1011
1012         might_sleep_if(!(rpmflags & RPM_ASYNC) && !dev->power.irq_safe);
1013
1014         spin_lock_irqsave(&dev->power.lock, flags);
1015         retval = rpm_suspend(dev, rpmflags);
1016         spin_unlock_irqrestore(&dev->power.lock, flags);
1017
1018         return retval;
1019 }
1020 EXPORT_SYMBOL_GPL(__pm_runtime_suspend);
1021
1022 /**
1023  * __pm_runtime_resume - Entry point for runtime resume operations.
1024  * @dev: Device to resume.
1025  * @rpmflags: Flag bits.
1026  *
1027  * If the RPM_GET_PUT flag is set, increment the device's usage count.  Then
1028  * carry out a resume, either synchronous or asynchronous.
1029  *
1030  * This routine may be called in atomic context if the RPM_ASYNC flag is set,
1031  * or if pm_runtime_irq_safe() has been called.
1032  */
1033 int __pm_runtime_resume(struct device *dev, int rpmflags)
1034 {
1035         unsigned long flags;
1036         int retval;
1037
1038         might_sleep_if(!(rpmflags & RPM_ASYNC) && !dev->power.irq_safe &&
1039                         dev->power.runtime_status != RPM_ACTIVE);
1040
1041         if (rpmflags & RPM_GET_PUT)
1042                 atomic_inc(&dev->power.usage_count);
1043
1044         spin_lock_irqsave(&dev->power.lock, flags);
1045         retval = rpm_resume(dev, rpmflags);
1046         spin_unlock_irqrestore(&dev->power.lock, flags);
1047
1048         return retval;
1049 }
1050 EXPORT_SYMBOL_GPL(__pm_runtime_resume);
1051
1052 /**
1053  * pm_runtime_get_if_in_use - Conditionally bump up the device's usage counter.
1054  * @dev: Device to handle.
1055  *
1056  * Return -EINVAL if runtime PM is disabled for the device.
1057  *
1058  * If that's not the case and if the device's runtime PM status is RPM_ACTIVE
1059  * and the runtime PM usage counter is nonzero, increment the counter and
1060  * return 1.  Otherwise return 0 without changing the counter.
1061  */
1062 int pm_runtime_get_if_in_use(struct device *dev)
1063 {
1064         unsigned long flags;
1065         int retval;
1066
1067         spin_lock_irqsave(&dev->power.lock, flags);
1068         retval = dev->power.disable_depth > 0 ? -EINVAL :
1069                 dev->power.runtime_status == RPM_ACTIVE
1070                         && atomic_inc_not_zero(&dev->power.usage_count);
1071         spin_unlock_irqrestore(&dev->power.lock, flags);
1072         return retval;
1073 }
1074 EXPORT_SYMBOL_GPL(pm_runtime_get_if_in_use);
1075
1076 /**
1077  * __pm_runtime_set_status - Set runtime PM status of a device.
1078  * @dev: Device to handle.
1079  * @status: New runtime PM status of the device.
1080  *
1081  * If runtime PM of the device is disabled or its power.runtime_error field is
1082  * different from zero, the status may be changed either to RPM_ACTIVE, or to
1083  * RPM_SUSPENDED, as long as that reflects the actual state of the device.
1084  * However, if the device has a parent and the parent is not active, and the
1085  * parent's power.ignore_children flag is unset, the device's status cannot be
1086  * set to RPM_ACTIVE, so -EBUSY is returned in that case.
1087  *
1088  * If successful, __pm_runtime_set_status() clears the power.runtime_error field
1089  * and the device parent's counter of unsuspended children is modified to
1090  * reflect the new status.  If the new status is RPM_SUSPENDED, an idle
1091  * notification request for the parent is submitted.
1092  *
1093  * If @dev has any suppliers (as reflected by device links to them), and @status
1094  * is RPM_ACTIVE, they will be activated upfront and if the activation of one
1095  * of them fails, the status of @dev will be changed to RPM_SUSPENDED (instead
1096  * of the @status value) and the suppliers will be deacticated on exit.  The
1097  * error returned by the failing supplier activation will be returned in that
1098  * case.
1099  */
1100 int __pm_runtime_set_status(struct device *dev, unsigned int status)
1101 {
1102         struct device *parent = dev->parent;
1103         bool notify_parent = false;
1104         int error = 0;
1105
1106         if (status != RPM_ACTIVE && status != RPM_SUSPENDED)
1107                 return -EINVAL;
1108
1109         spin_lock_irq(&dev->power.lock);
1110
1111         /*
1112          * Prevent PM-runtime from being enabled for the device or return an
1113          * error if it is enabled already and working.
1114          */
1115         if (dev->power.runtime_error || dev->power.disable_depth)
1116                 dev->power.disable_depth++;
1117         else
1118                 error = -EAGAIN;
1119
1120         spin_unlock_irq(&dev->power.lock);
1121
1122         if (error)
1123                 return error;
1124
1125         /*
1126          * If the new status is RPM_ACTIVE, the suppliers can be activated
1127          * upfront regardless of the current status, because next time
1128          * rpm_put_suppliers() runs, the rpm_active refcounts of the links
1129          * involved will be dropped down to one anyway.
1130          */
1131         if (status == RPM_ACTIVE) {
1132                 int idx = device_links_read_lock();
1133
1134                 error = rpm_get_suppliers(dev);
1135                 if (error)
1136                         status = RPM_SUSPENDED;
1137
1138                 device_links_read_unlock(idx);
1139         }
1140
1141         spin_lock_irq(&dev->power.lock);
1142
1143         if (dev->power.runtime_status == status || !parent)
1144                 goto out_set;
1145
1146         if (status == RPM_SUSPENDED) {
1147                 atomic_add_unless(&parent->power.child_count, -1, 0);
1148                 notify_parent = !parent->power.ignore_children;
1149         } else {
1150                 spin_lock_nested(&parent->power.lock, SINGLE_DEPTH_NESTING);
1151
1152                 /*
1153                  * It is invalid to put an active child under a parent that is
1154                  * not active, has runtime PM enabled and the
1155                  * 'power.ignore_children' flag unset.
1156                  */
1157                 if (!parent->power.disable_depth
1158                     && !parent->power.ignore_children
1159                     && parent->power.runtime_status != RPM_ACTIVE) {
1160                         dev_err(dev, "runtime PM trying to activate child device %s but parent (%s) is not active\n",
1161                                 dev_name(dev),
1162                                 dev_name(parent));
1163                         error = -EBUSY;
1164                 } else if (dev->power.runtime_status == RPM_SUSPENDED) {
1165                         atomic_inc(&parent->power.child_count);
1166                 }
1167
1168                 spin_unlock(&parent->power.lock);
1169
1170                 if (error) {
1171                         status = RPM_SUSPENDED;
1172                         goto out;
1173                 }
1174         }
1175
1176  out_set:
1177         __update_runtime_status(dev, status);
1178         if (!error)
1179                 dev->power.runtime_error = 0;
1180
1181  out:
1182         spin_unlock_irq(&dev->power.lock);
1183
1184         if (notify_parent)
1185                 pm_request_idle(parent);
1186
1187         if (status == RPM_SUSPENDED) {
1188                 int idx = device_links_read_lock();
1189
1190                 rpm_put_suppliers(dev);
1191
1192                 device_links_read_unlock(idx);
1193         }
1194
1195         pm_runtime_enable(dev);
1196
1197         return error;
1198 }
1199 EXPORT_SYMBOL_GPL(__pm_runtime_set_status);
1200
1201 /**
1202  * __pm_runtime_barrier - Cancel pending requests and wait for completions.
1203  * @dev: Device to handle.
1204  *
1205  * Flush all pending requests for the device from pm_wq and wait for all
1206  * runtime PM operations involving the device in progress to complete.
1207  *
1208  * Should be called under dev->power.lock with interrupts disabled.
1209  */
1210 static void __pm_runtime_barrier(struct device *dev)
1211 {
1212         pm_runtime_deactivate_timer(dev);
1213
1214         if (dev->power.request_pending) {
1215                 dev->power.request = RPM_REQ_NONE;
1216                 spin_unlock_irq(&dev->power.lock);
1217
1218                 cancel_work_sync(&dev->power.work);
1219
1220                 spin_lock_irq(&dev->power.lock);
1221                 dev->power.request_pending = false;
1222         }
1223
1224         if (dev->power.runtime_status == RPM_SUSPENDING
1225             || dev->power.runtime_status == RPM_RESUMING
1226             || dev->power.idle_notification) {
1227                 DEFINE_WAIT(wait);
1228
1229                 /* Suspend, wake-up or idle notification in progress. */
1230                 for (;;) {
1231                         prepare_to_wait(&dev->power.wait_queue, &wait,
1232                                         TASK_UNINTERRUPTIBLE);
1233                         if (dev->power.runtime_status != RPM_SUSPENDING
1234                             && dev->power.runtime_status != RPM_RESUMING
1235                             && !dev->power.idle_notification)
1236                                 break;
1237                         spin_unlock_irq(&dev->power.lock);
1238
1239                         schedule();
1240
1241                         spin_lock_irq(&dev->power.lock);
1242                 }
1243                 finish_wait(&dev->power.wait_queue, &wait);
1244         }
1245 }
1246
1247 /**
1248  * pm_runtime_barrier - Flush pending requests and wait for completions.
1249  * @dev: Device to handle.
1250  *
1251  * Prevent the device from being suspended by incrementing its usage counter and
1252  * if there's a pending resume request for the device, wake the device up.
1253  * Next, make sure that all pending requests for the device have been flushed
1254  * from pm_wq and wait for all runtime PM operations involving the device in
1255  * progress to complete.
1256  *
1257  * Return value:
1258  * 1, if there was a resume request pending and the device had to be woken up,
1259  * 0, otherwise
1260  */
1261 int pm_runtime_barrier(struct device *dev)
1262 {
1263         int retval = 0;
1264
1265         pm_runtime_get_noresume(dev);
1266         spin_lock_irq(&dev->power.lock);
1267
1268         if (dev->power.request_pending
1269             && dev->power.request == RPM_REQ_RESUME) {
1270                 rpm_resume(dev, 0);
1271                 retval = 1;
1272         }
1273
1274         __pm_runtime_barrier(dev);
1275
1276         spin_unlock_irq(&dev->power.lock);
1277         pm_runtime_put_noidle(dev);
1278
1279         return retval;
1280 }
1281 EXPORT_SYMBOL_GPL(pm_runtime_barrier);
1282
1283 /**
1284  * __pm_runtime_disable - Disable runtime PM of a device.
1285  * @dev: Device to handle.
1286  * @check_resume: If set, check if there's a resume request for the device.
1287  *
1288  * Increment power.disable_depth for the device and if it was zero previously,
1289  * cancel all pending runtime PM requests for the device and wait for all
1290  * operations in progress to complete.  The device can be either active or
1291  * suspended after its runtime PM has been disabled.
1292  *
1293  * If @check_resume is set and there's a resume request pending when
1294  * __pm_runtime_disable() is called and power.disable_depth is zero, the
1295  * function will wake up the device before disabling its runtime PM.
1296  */
1297 void __pm_runtime_disable(struct device *dev, bool check_resume)
1298 {
1299         spin_lock_irq(&dev->power.lock);
1300
1301         if (dev->power.disable_depth > 0) {
1302                 dev->power.disable_depth++;
1303                 goto out;
1304         }
1305
1306         /*
1307          * Wake up the device if there's a resume request pending, because that
1308          * means there probably is some I/O to process and disabling runtime PM
1309          * shouldn't prevent the device from processing the I/O.
1310          */
1311         if (check_resume && dev->power.request_pending
1312             && dev->power.request == RPM_REQ_RESUME) {
1313                 /*
1314                  * Prevent suspends and idle notifications from being carried
1315                  * out after we have woken up the device.
1316                  */
1317                 pm_runtime_get_noresume(dev);
1318
1319                 rpm_resume(dev, 0);
1320
1321                 pm_runtime_put_noidle(dev);
1322         }
1323
1324         if (!dev->power.disable_depth++)
1325                 __pm_runtime_barrier(dev);
1326
1327  out:
1328         spin_unlock_irq(&dev->power.lock);
1329 }
1330 EXPORT_SYMBOL_GPL(__pm_runtime_disable);
1331
1332 /**
1333  * pm_runtime_enable - Enable runtime PM of a device.
1334  * @dev: Device to handle.
1335  */
1336 void pm_runtime_enable(struct device *dev)
1337 {
1338         unsigned long flags;
1339
1340         spin_lock_irqsave(&dev->power.lock, flags);
1341
1342         if (dev->power.disable_depth > 0)
1343                 dev->power.disable_depth--;
1344         else
1345                 dev_warn(dev, "Unbalanced %s!\n", __func__);
1346
1347         WARN(!dev->power.disable_depth &&
1348              dev->power.runtime_status == RPM_SUSPENDED &&
1349              !dev->power.ignore_children &&
1350              atomic_read(&dev->power.child_count) > 0,
1351              "Enabling runtime PM for inactive device (%s) with active children\n",
1352              dev_name(dev));
1353
1354         spin_unlock_irqrestore(&dev->power.lock, flags);
1355 }
1356 EXPORT_SYMBOL_GPL(pm_runtime_enable);
1357
1358 /**
1359  * pm_runtime_forbid - Block runtime PM of a device.
1360  * @dev: Device to handle.
1361  *
1362  * Increase the device's usage count and clear its power.runtime_auto flag,
1363  * so that it cannot be suspended at run time until pm_runtime_allow() is called
1364  * for it.
1365  */
1366 void pm_runtime_forbid(struct device *dev)
1367 {
1368         spin_lock_irq(&dev->power.lock);
1369         if (!dev->power.runtime_auto)
1370                 goto out;
1371
1372         dev->power.runtime_auto = false;
1373         atomic_inc(&dev->power.usage_count);
1374         rpm_resume(dev, 0);
1375
1376  out:
1377         spin_unlock_irq(&dev->power.lock);
1378 }
1379 EXPORT_SYMBOL_GPL(pm_runtime_forbid);
1380
1381 /**
1382  * pm_runtime_allow - Unblock runtime PM of a device.
1383  * @dev: Device to handle.
1384  *
1385  * Decrease the device's usage count and set its power.runtime_auto flag.
1386  */
1387 void pm_runtime_allow(struct device *dev)
1388 {
1389         spin_lock_irq(&dev->power.lock);
1390         if (dev->power.runtime_auto)
1391                 goto out;
1392
1393         dev->power.runtime_auto = true;
1394         if (atomic_dec_and_test(&dev->power.usage_count))
1395                 rpm_idle(dev, RPM_AUTO | RPM_ASYNC);
1396
1397  out:
1398         spin_unlock_irq(&dev->power.lock);
1399 }
1400 EXPORT_SYMBOL_GPL(pm_runtime_allow);
1401
1402 /**
1403  * pm_runtime_no_callbacks - Ignore runtime PM callbacks for a device.
1404  * @dev: Device to handle.
1405  *
1406  * Set the power.no_callbacks flag, which tells the PM core that this
1407  * device is power-managed through its parent and has no runtime PM
1408  * callbacks of its own.  The runtime sysfs attributes will be removed.
1409  */
1410 void pm_runtime_no_callbacks(struct device *dev)
1411 {
1412         spin_lock_irq(&dev->power.lock);
1413         dev->power.no_callbacks = 1;
1414         spin_unlock_irq(&dev->power.lock);
1415         if (device_is_registered(dev))
1416                 rpm_sysfs_remove(dev);
1417 }
1418 EXPORT_SYMBOL_GPL(pm_runtime_no_callbacks);
1419
1420 /**
1421  * pm_runtime_irq_safe - Leave interrupts disabled during callbacks.
1422  * @dev: Device to handle
1423  *
1424  * Set the power.irq_safe flag, which tells the PM core that the
1425  * ->runtime_suspend() and ->runtime_resume() callbacks for this device should
1426  * always be invoked with the spinlock held and interrupts disabled.  It also
1427  * causes the parent's usage counter to be permanently incremented, preventing
1428  * the parent from runtime suspending -- otherwise an irq-safe child might have
1429  * to wait for a non-irq-safe parent.
1430  */
1431 void pm_runtime_irq_safe(struct device *dev)
1432 {
1433         if (dev->parent)
1434                 pm_runtime_get_sync(dev->parent);
1435         spin_lock_irq(&dev->power.lock);
1436         dev->power.irq_safe = 1;
1437         spin_unlock_irq(&dev->power.lock);
1438 }
1439 EXPORT_SYMBOL_GPL(pm_runtime_irq_safe);
1440
1441 /**
1442  * update_autosuspend - Handle a change to a device's autosuspend settings.
1443  * @dev: Device to handle.
1444  * @old_delay: The former autosuspend_delay value.
1445  * @old_use: The former use_autosuspend value.
1446  *
1447  * Prevent runtime suspend if the new delay is negative and use_autosuspend is
1448  * set; otherwise allow it.  Send an idle notification if suspends are allowed.
1449  *
1450  * This function must be called under dev->power.lock with interrupts disabled.
1451  */
1452 static void update_autosuspend(struct device *dev, int old_delay, int old_use)
1453 {
1454         int delay = dev->power.autosuspend_delay;
1455
1456         /* Should runtime suspend be prevented now? */
1457         if (dev->power.use_autosuspend && delay < 0) {
1458
1459                 /* If it used to be allowed then prevent it. */
1460                 if (!old_use || old_delay >= 0) {
1461                         atomic_inc(&dev->power.usage_count);
1462                         rpm_resume(dev, 0);
1463                 }
1464         }
1465
1466         /* Runtime suspend should be allowed now. */
1467         else {
1468
1469                 /* If it used to be prevented then allow it. */
1470                 if (old_use && old_delay < 0)
1471                         atomic_dec(&dev->power.usage_count);
1472
1473                 /* Maybe we can autosuspend now. */
1474                 rpm_idle(dev, RPM_AUTO);
1475         }
1476 }
1477
1478 /**
1479  * pm_runtime_set_autosuspend_delay - Set a device's autosuspend_delay value.
1480  * @dev: Device to handle.
1481  * @delay: Value of the new delay in milliseconds.
1482  *
1483  * Set the device's power.autosuspend_delay value.  If it changes to negative
1484  * and the power.use_autosuspend flag is set, prevent runtime suspends.  If it
1485  * changes the other way, allow runtime suspends.
1486  */
1487 void pm_runtime_set_autosuspend_delay(struct device *dev, int delay)
1488 {
1489         int old_delay, old_use;
1490
1491         spin_lock_irq(&dev->power.lock);
1492         old_delay = dev->power.autosuspend_delay;
1493         old_use = dev->power.use_autosuspend;
1494         dev->power.autosuspend_delay = delay;
1495         update_autosuspend(dev, old_delay, old_use);
1496         spin_unlock_irq(&dev->power.lock);
1497 }
1498 EXPORT_SYMBOL_GPL(pm_runtime_set_autosuspend_delay);
1499
1500 /**
1501  * __pm_runtime_use_autosuspend - Set a device's use_autosuspend flag.
1502  * @dev: Device to handle.
1503  * @use: New value for use_autosuspend.
1504  *
1505  * Set the device's power.use_autosuspend flag, and allow or prevent runtime
1506  * suspends as needed.
1507  */
1508 void __pm_runtime_use_autosuspend(struct device *dev, bool use)
1509 {
1510         int old_delay, old_use;
1511
1512         spin_lock_irq(&dev->power.lock);
1513         old_delay = dev->power.autosuspend_delay;
1514         old_use = dev->power.use_autosuspend;
1515         dev->power.use_autosuspend = use;
1516         update_autosuspend(dev, old_delay, old_use);
1517         spin_unlock_irq(&dev->power.lock);
1518 }
1519 EXPORT_SYMBOL_GPL(__pm_runtime_use_autosuspend);
1520
1521 /**
1522  * pm_runtime_init - Initialize runtime PM fields in given device object.
1523  * @dev: Device object to initialize.
1524  */
1525 void pm_runtime_init(struct device *dev)
1526 {
1527         dev->power.runtime_status = RPM_SUSPENDED;
1528         dev->power.idle_notification = false;
1529
1530         dev->power.disable_depth = 1;
1531         atomic_set(&dev->power.usage_count, 0);
1532
1533         dev->power.runtime_error = 0;
1534
1535         atomic_set(&dev->power.child_count, 0);
1536         pm_suspend_ignore_children(dev, false);
1537         dev->power.runtime_auto = true;
1538
1539         dev->power.request_pending = false;
1540         dev->power.request = RPM_REQ_NONE;
1541         dev->power.deferred_resume = false;
1542         dev->power.accounting_timestamp = jiffies;
1543         INIT_WORK(&dev->power.work, pm_runtime_work);
1544
1545         dev->power.timer_expires = 0;
1546         hrtimer_init(&dev->power.suspend_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1547         dev->power.suspend_timer.function = pm_suspend_timer_fn;
1548
1549         init_waitqueue_head(&dev->power.wait_queue);
1550 }
1551
1552 /**
1553  * pm_runtime_reinit - Re-initialize runtime PM fields in given device object.
1554  * @dev: Device object to re-initialize.
1555  */
1556 void pm_runtime_reinit(struct device *dev)
1557 {
1558         if (!pm_runtime_enabled(dev)) {
1559                 if (dev->power.runtime_status == RPM_ACTIVE)
1560                         pm_runtime_set_suspended(dev);
1561                 if (dev->power.irq_safe) {
1562                         spin_lock_irq(&dev->power.lock);
1563                         dev->power.irq_safe = 0;
1564                         spin_unlock_irq(&dev->power.lock);
1565                         if (dev->parent)
1566                                 pm_runtime_put(dev->parent);
1567                 }
1568         }
1569 }
1570
1571 /**
1572  * pm_runtime_remove - Prepare for removing a device from device hierarchy.
1573  * @dev: Device object being removed from device hierarchy.
1574  */
1575 void pm_runtime_remove(struct device *dev)
1576 {
1577         __pm_runtime_disable(dev, false);
1578         pm_runtime_reinit(dev);
1579 }
1580
1581 /**
1582  * pm_runtime_clean_up_links - Prepare links to consumers for driver removal.
1583  * @dev: Device whose driver is going to be removed.
1584  *
1585  * Check links from this device to any consumers and if any of them have active
1586  * runtime PM references to the device, drop the usage counter of the device
1587  * (as many times as needed).
1588  *
1589  * Links with the DL_FLAG_STATELESS flag set are ignored.
1590  *
1591  * Since the device is guaranteed to be runtime-active at the point this is
1592  * called, nothing else needs to be done here.
1593  *
1594  * Moreover, this is called after device_links_busy() has returned 'false', so
1595  * the status of each link is guaranteed to be DL_STATE_SUPPLIER_UNBIND and
1596  * therefore rpm_active can't be manipulated concurrently.
1597  */
1598 void pm_runtime_clean_up_links(struct device *dev)
1599 {
1600         struct device_link *link;
1601         int idx;
1602
1603         idx = device_links_read_lock();
1604
1605         list_for_each_entry_rcu(link, &dev->links.consumers, s_node) {
1606                 if (link->flags & DL_FLAG_STATELESS)
1607                         continue;
1608
1609                 while (refcount_dec_not_one(&link->rpm_active))
1610                         pm_runtime_put_noidle(dev);
1611         }
1612
1613         device_links_read_unlock(idx);
1614 }
1615
1616 /**
1617  * pm_runtime_get_suppliers - Resume and reference-count supplier devices.
1618  * @dev: Consumer device.
1619  */
1620 void pm_runtime_get_suppliers(struct device *dev)
1621 {
1622         struct device_link *link;
1623         int idx;
1624
1625         idx = device_links_read_lock();
1626
1627         list_for_each_entry_rcu(link, &dev->links.suppliers, c_node)
1628                 if (link->flags & DL_FLAG_PM_RUNTIME) {
1629                         refcount_inc(&link->rpm_active);
1630                         pm_runtime_get_sync(link->supplier);
1631                 }
1632
1633         device_links_read_unlock(idx);
1634 }
1635
1636 /**
1637  * pm_runtime_put_suppliers - Drop references to supplier devices.
1638  * @dev: Consumer device.
1639  */
1640 void pm_runtime_put_suppliers(struct device *dev)
1641 {
1642         struct device_link *link;
1643         int idx;
1644
1645         idx = device_links_read_lock();
1646
1647         list_for_each_entry_rcu(link, &dev->links.suppliers, c_node)
1648                 if (link->flags & DL_FLAG_PM_RUNTIME &&
1649                     refcount_dec_not_one(&link->rpm_active))
1650                         pm_runtime_put(link->supplier);
1651
1652         device_links_read_unlock(idx);
1653 }
1654
1655 void pm_runtime_new_link(struct device *dev)
1656 {
1657         spin_lock_irq(&dev->power.lock);
1658         dev->power.links_count++;
1659         spin_unlock_irq(&dev->power.lock);
1660 }
1661
1662 /**
1663  * pm_runtime_active_link - Set up new device link as active for PM-runtime.
1664  * @link: Device link to be set up as active.
1665  * @supplier: Supplier end of the link.
1666  *
1667  * Add 2 to the rpm_active refcount of @link and increment the PM-runtime
1668  * usage counter of @supplier once more in case the link is being added while
1669  * the consumer driver is probing and pm_runtime_put_suppliers() will be called
1670  * subsequently.
1671  *
1672  * Note that this doesn't prevent rpm_put_suppliers() from decreasing the link's
1673  * rpm_active refcount down to one, so runtime suspend of the consumer end of
1674  * @link is not affected.
1675  */
1676 void pm_runtime_active_link(struct device_link *link, struct device *supplier)
1677 {
1678         refcount_add(2, &link->rpm_active);
1679         pm_runtime_get_noresume(supplier);
1680 }
1681
1682 void pm_runtime_drop_link(struct device *dev)
1683 {
1684         spin_lock_irq(&dev->power.lock);
1685         WARN_ON(dev->power.links_count == 0);
1686         dev->power.links_count--;
1687         spin_unlock_irq(&dev->power.lock);
1688 }
1689
1690 static bool pm_runtime_need_not_resume(struct device *dev)
1691 {
1692         return atomic_read(&dev->power.usage_count) <= 1 &&
1693                 (atomic_read(&dev->power.child_count) == 0 ||
1694                  dev->power.ignore_children);
1695 }
1696
1697 /**
1698  * pm_runtime_force_suspend - Force a device into suspend state if needed.
1699  * @dev: Device to suspend.
1700  *
1701  * Disable runtime PM so we safely can check the device's runtime PM status and
1702  * if it is active, invoke its ->runtime_suspend callback to suspend it and
1703  * change its runtime PM status field to RPM_SUSPENDED.  Also, if the device's
1704  * usage and children counters don't indicate that the device was in use before
1705  * the system-wide transition under way, decrement its parent's children counter
1706  * (if there is a parent).  Keep runtime PM disabled to preserve the state
1707  * unless we encounter errors.
1708  *
1709  * Typically this function may be invoked from a system suspend callback to make
1710  * sure the device is put into low power state and it should only be used during
1711  * system-wide PM transitions to sleep states.  It assumes that the analogous
1712  * pm_runtime_force_resume() will be used to resume the device.
1713  */
1714 int pm_runtime_force_suspend(struct device *dev)
1715 {
1716         int (*callback)(struct device *);
1717         int ret;
1718
1719         pm_runtime_disable(dev);
1720         if (pm_runtime_status_suspended(dev))
1721                 return 0;
1722
1723         callback = RPM_GET_CALLBACK(dev, runtime_suspend);
1724
1725         ret = callback ? callback(dev) : 0;
1726         if (ret)
1727                 goto err;
1728
1729         /*
1730          * If the device can stay in suspend after the system-wide transition
1731          * to the working state that will follow, drop the children counter of
1732          * its parent, but set its status to RPM_SUSPENDED anyway in case this
1733          * function will be called again for it in the meantime.
1734          */
1735         if (pm_runtime_need_not_resume(dev))
1736                 pm_runtime_set_suspended(dev);
1737         else
1738                 __update_runtime_status(dev, RPM_SUSPENDED);
1739
1740         return 0;
1741
1742 err:
1743         pm_runtime_enable(dev);
1744         return ret;
1745 }
1746 EXPORT_SYMBOL_GPL(pm_runtime_force_suspend);
1747
1748 /**
1749  * pm_runtime_force_resume - Force a device into resume state if needed.
1750  * @dev: Device to resume.
1751  *
1752  * Prior invoking this function we expect the user to have brought the device
1753  * into low power state by a call to pm_runtime_force_suspend(). Here we reverse
1754  * those actions and bring the device into full power, if it is expected to be
1755  * used on system resume.  In the other case, we defer the resume to be managed
1756  * via runtime PM.
1757  *
1758  * Typically this function may be invoked from a system resume callback.
1759  */
1760 int pm_runtime_force_resume(struct device *dev)
1761 {
1762         int (*callback)(struct device *);
1763         int ret = 0;
1764
1765         if (!pm_runtime_status_suspended(dev) || pm_runtime_need_not_resume(dev))
1766                 goto out;
1767
1768         /*
1769          * The value of the parent's children counter is correct already, so
1770          * just update the status of the device.
1771          */
1772         __update_runtime_status(dev, RPM_ACTIVE);
1773
1774         callback = RPM_GET_CALLBACK(dev, runtime_resume);
1775
1776         ret = callback ? callback(dev) : 0;
1777         if (ret) {
1778                 pm_runtime_set_suspended(dev);
1779                 goto out;
1780         }
1781
1782         pm_runtime_mark_last_busy(dev);
1783 out:
1784         pm_runtime_enable(dev);
1785         return ret;
1786 }
1787 EXPORT_SYMBOL_GPL(pm_runtime_force_resume);