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