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