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 = ACCESS_ONCE(dev->power.autosuspend_delay);
 138         if (autosuspend_delay < 0)
 139                 goto out;
 140
 141         last_busy = ACCESS_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                         continue;
 273
 274                 if (READ_ONCE(link->status) == DL_STATE_SUPPLIER_UNBIND ||
 275                     link->rpm_active)
 276                         continue;
 277
 278                 retval = pm_runtime_get_sync(link->supplier);
 279                 if (retval < 0) {
 280                         pm_runtime_put_noidle(link->supplier);
 281                         return retval;
 282                 }
 283                 link->rpm_active = true;
 284         }
 285         return 0;
 286 }
 287
 288 static void rpm_put_suppliers(struct device *dev)
 289 {
 290         struct device_link *link;
 291
 292         list_for_each_entry_rcu(link, &dev->links.suppliers, c_node)
 293                 if (link->rpm_active &&
 294                     READ_ONCE(link->status) != DL_STATE_SUPPLIER_UNBIND) {
 295                         pm_runtime_put(link->supplier);
 296                         link->rpm_active = false;
 297                 }
 298 }
 299
 300 /**
 301  * __rpm_callback - Run a given runtime PM callback for a given device.
 302  * @cb: Runtime PM callback to run.
 303  * @dev: Device to run the callback for.
 304  */
 305 static int __rpm_callback(int (*cb)(struct device *), struct device *dev)
 306         __releases(&dev->power.lock) __acquires(&dev->power.lock)
 307 {
 308         int retval, idx;
 309         bool use_links = dev->power.links_count > 0;
 310
 311         if (dev->power.irq_safe) {
 312                 spin_unlock(&dev->power.lock);
 313         } else {
 314                 spin_unlock_irq(&dev->power.lock);
 315
 316                 /*
 317                  * Resume suppliers if necessary.
 318                  *
 319                  * The device's runtime PM status cannot change until this
 320                  * routine returns, so it is safe to read the status outside of
 321                  * the lock.
 322                  */
 323                 if (use_links && dev->power.runtime_status == RPM_RESUMING) {
 324                         idx = device_links_read_lock();
 325
 326                         retval = rpm_get_suppliers(dev);
 327                         if (retval)
 328                                 goto fail;
 329
 330                         device_links_read_unlock(idx);
 331                 }
 332         }
 333
 334         retval = cb(dev);
 335
 336         if (dev->power.irq_safe) {
 337                 spin_lock(&dev->power.lock);
 338         } else {
 339                 /*
 340                  * If the device is suspending and the callback has returned
 341                  * success, drop the usage counters of the suppliers that have
 342                  * been reference counted on its resume.
 343                  *
 344                  * Do that if resume fails too.
 345                  */
 346                 if (use_links
 347                     && ((dev->power.runtime_status == RPM_SUSPENDING && !retval)
 348                     || (dev->power.runtime_status == RPM_RESUMING && retval))) {
 349                         idx = device_links_read_lock();
 350
 351  fail:
 352                         rpm_put_suppliers(dev);
 353
 354                         device_links_read_unlock(idx);
 355                 }
 356
 357                 spin_lock_irq(&dev->power.lock);
 358         }
 359
 360         return retval;
 361 }
 362
 363 /**
 364  * rpm_idle - Notify device bus type if the device can be suspended.
 365  * @dev: Device to notify the bus type about.
 366  * @rpmflags: Flag bits.
 367  *
 368  * Check if the device's runtime PM status allows it to be suspended.  If
 369  * another idle notification has been started earlier, return immediately.  If
 370  * the RPM_ASYNC flag is set then queue an idle-notification request; otherwise
 371  * run the ->runtime_idle() callback directly. If the ->runtime_idle callback
 372  * doesn't exist or if it returns 0, call rpm_suspend with the RPM_AUTO flag.
 373  *
 374  * This function must be called under dev->power.lock with interrupts disabled.
 375  */
 376 static int rpm_idle(struct device *dev, int rpmflags)
 377 {
 378         int (*callback)(struct device *);
 379         int retval;
 380
 381         trace_rpm_idle_rcuidle(dev, rpmflags);
 382         retval = rpm_check_suspend_allowed(dev);
 383         if (retval < 0)
 384                 ;       /* Conditions are wrong. */
 385
 386         /* Idle notifications are allowed only in the RPM_ACTIVE state. */
 387         else if (dev->power.runtime_status != RPM_ACTIVE)
 388                 retval = -EAGAIN;
 389
 390         /*
 391          * Any pending request other than an idle notification takes
 392          * precedence over us, except that the timer may be running.
 393          */
 394         else if (dev->power.request_pending &&
 395             dev->power.request > RPM_REQ_IDLE)
 396                 retval = -EAGAIN;
 397
 398         /* Act as though RPM_NOWAIT is always set. */
 399         else if (dev->power.idle_notification)
 400                 retval = -EINPROGRESS;
 401         if (retval)
 402                 goto out;
 403
 404         /* Pending requests need to be canceled. */
 405         dev->power.request = RPM_REQ_NONE;
 406
 407         if (dev->power.no_callbacks)
 408                 goto out;
 409
 410         /* Carry out an asynchronous or a synchronous idle notification. */
 411         if (rpmflags & RPM_ASYNC) {
 412                 dev->power.request = RPM_REQ_IDLE;
 413                 if (!dev->power.request_pending) {
 414                         dev->power.request_pending = true;
 415                         queue_work(pm_wq, &dev->power.work);
 416                 }
 417                 trace_rpm_return_int_rcuidle(dev, _THIS_IP_, 0);
 418                 return 0;
 419         }
 420
 421         dev->power.idle_notification = true;
 422
 423         callback = RPM_GET_CALLBACK(dev, runtime_idle);
 424
 425         if (callback)
 426                 retval = __rpm_callback(callback, dev);
 427
 428         dev->power.idle_notification = false;
 429         wake_up_all(&dev->power.wait_queue);
 430
 431  out:
 432         trace_rpm_return_int_rcuidle(dev, _THIS_IP_, retval);
 433         return retval ? retval : rpm_suspend(dev, rpmflags | RPM_AUTO);
 434 }
 435
 436 /**
 437  * rpm_callback - Run a given runtime PM callback for a given device.
 438  * @cb: Runtime PM callback to run.
 439  * @dev: Device to run the callback for.
 440  */
 441 static int rpm_callback(int (*cb)(struct device *), struct device *dev)
 442 {
 443         int retval;
 444
 445         if (!cb)
 446                 return -ENOSYS;
 447
 448         if (dev->power.memalloc_noio) {
 449                 unsigned int noio_flag;
 450
 451                 /*
 452                  * Deadlock might be caused if memory allocation with
 453                  * GFP_KERNEL happens inside runtime_suspend and
 454                  * runtime_resume callbacks of one block device's
 455                  * ancestor or the block device itself. Network
 456                  * device might be thought as part of iSCSI block
 457                  * device, so network device and its ancestor should
 458                  * be marked as memalloc_noio too.
 459                  */
 460                 noio_flag = memalloc_noio_save();
 461                 retval = __rpm_callback(cb, dev);
 462                 memalloc_noio_restore(noio_flag);
 463         } else {
 464                 retval = __rpm_callback(cb, dev);
 465         }
 466
 467         dev->power.runtime_error = retval;
 468         return retval != -EACCES ? retval : -EIO;
 469 }
 470
 471 /**
 472  * rpm_suspend - Carry out runtime suspend of given device.
 473  * @dev: Device to suspend.
 474  * @rpmflags: Flag bits.
 475  *
 476  * Check if the device's runtime PM status allows it to be suspended.
 477  * Cancel a pending idle notification, autosuspend or suspend. If
 478  * another suspend has been started earlier, either return immediately
 479  * or wait for it to finish, depending on the RPM_NOWAIT and RPM_ASYNC
 480  * flags. If the RPM_ASYNC flag is set then queue a suspend request;
 481  * otherwise run the ->runtime_suspend() callback directly. When
 482  * ->runtime_suspend succeeded, if a deferred resume was requested while
 483  * the callback was running then carry it out, otherwise send an idle
 484  * notification for its parent (if the suspend succeeded and both
 485  * ignore_children of parent->power and irq_safe of dev->power are not set).
 486  * If ->runtime_suspend failed with -EAGAIN or -EBUSY, and if the RPM_AUTO
 487  * flag is set and the next autosuspend-delay expiration time is in the
 488  * future, schedule another autosuspend attempt.
 489  *
 490  * This function must be called under dev->power.lock with interrupts disabled.
 491  */
 492 static int rpm_suspend(struct device *dev, int rpmflags)
 493         __releases(&dev->power.lock) __acquires(&dev->power.lock)
 494 {
 495         int (*callback)(struct device *);
 496         struct device *parent = NULL;
 497         int retval;
 498
 499         trace_rpm_suspend_rcuidle(dev, rpmflags);
 500
 501  repeat:
 502         retval = rpm_check_suspend_allowed(dev);
 503
 504         if (retval < 0)
 505                 ;       /* Conditions are wrong. */
 506
 507         /* Synchronous suspends are not allowed in the RPM_RESUMING state. */
 508         else if (dev->power.runtime_status == RPM_RESUMING &&
 509             !(rpmflags & RPM_ASYNC))
 510                 retval = -EAGAIN;
 511         if (retval)
 512                 goto out;
 513
 514         /* If the autosuspend_delay time hasn't expired yet, reschedule. */
 515         if ((rpmflags & RPM_AUTO)
 516             && dev->power.runtime_status != RPM_SUSPENDING) {
 517                 unsigned long expires = pm_runtime_autosuspend_expiration(dev);
 518
 519                 if (expires != 0) {
 520                         /* Pending requests need to be canceled. */
 521                         dev->power.request = RPM_REQ_NONE;
 522
 523                         /*
 524                          * Optimization: If the timer is already running and is
 525                          * set to expire at or before the autosuspend delay,
 526                          * avoid the overhead of resetting it.  Just let it
 527                          * expire; pm_suspend_timer_fn() will take care of the
 528                          * rest.
 529                          */
 530                         if (!(dev->power.timer_expires && time_before_eq(
 531                             dev->power.timer_expires, expires))) {
 532                                 dev->power.timer_expires = expires;
 533                                 mod_timer(&dev->power.suspend_timer, expires);
 534                         }
 535                         dev->power.timer_autosuspends = 1;
 536                         goto out;
 537                 }
 538         }
 539
 540         /* Other scheduled or pending requests need to be canceled. */
 541         pm_runtime_cancel_pending(dev);
 542
 543         if (dev->power.runtime_status == RPM_SUSPENDING) {
 544                 DEFINE_WAIT(wait);
 545
 546                 if (rpmflags & (RPM_ASYNC | RPM_NOWAIT)) {
 547                         retval = -EINPROGRESS;
 548                         goto out;
 549                 }
 550
 551                 if (dev->power.irq_safe) {
 552                         spin_unlock(&dev->power.lock);
 553
 554                         cpu_relax();
 555
 556                         spin_lock(&dev->power.lock);
 557                         goto repeat;
 558                 }
 559
 560                 /* Wait for the other suspend running in parallel with us. */
 561                 for (;;) {
 562                         prepare_to_wait(&dev->power.wait_queue, &wait,
 563                                         TASK_UNINTERRUPTIBLE);
 564                         if (dev->power.runtime_status != RPM_SUSPENDING)
 565                                 break;
 566
 567                         spin_unlock_irq(&dev->power.lock);
 568
 569                         schedule();
 570
 571                         spin_lock_irq(&dev->power.lock);
 572                 }
 573                 finish_wait(&dev->power.wait_queue, &wait);
 574                 goto repeat;
 575         }
 576
 577         if (dev->power.no_callbacks)
 578                 goto no_callback;       /* Assume success. */
 579
 580         /* Carry out an asynchronous or a synchronous suspend. */
 581         if (rpmflags & RPM_ASYNC) {
 582                 dev->power.request = (rpmflags & RPM_AUTO) ?
 583                     RPM_REQ_AUTOSUSPEND : RPM_REQ_SUSPEND;
 584                 if (!dev->power.request_pending) {
 585                         dev->power.request_pending = true;
 586                         queue_work(pm_wq, &dev->power.work);
 587                 }
 588                 goto out;
 589         }
 590
 591         __update_runtime_status(dev, RPM_SUSPENDING);
 592
 593         callback = RPM_GET_CALLBACK(dev, runtime_suspend);
 594
 595         dev_pm_enable_wake_irq_check(dev, true);
 596         retval = rpm_callback(callback, dev);
 597         if (retval)
 598                 goto fail;
 599
 600  no_callback:
 601         __update_runtime_status(dev, RPM_SUSPENDED);
 602         pm_runtime_deactivate_timer(dev);
 603
 604         if (dev->parent) {
 605                 parent = dev->parent;
 606                 atomic_add_unless(&parent->power.child_count, -1, 0);
 607         }
 608         wake_up_all(&dev->power.wait_queue);
 609
 610         if (dev->power.deferred_resume) {
 611                 dev->power.deferred_resume = false;
 612                 rpm_resume(dev, 0);
 613                 retval = -EAGAIN;
 614                 goto out;
 615         }
 616
 617         /* Maybe the parent is now able to suspend. */
 618         if (parent && !parent->power.ignore_children && !dev->power.irq_safe) {
 619                 spin_unlock(&dev->power.lock);
 620
 621                 spin_lock(&parent->power.lock);
 622                 rpm_idle(parent, RPM_ASYNC);
 623                 spin_unlock(&parent->power.lock);
 624
 625                 spin_lock(&dev->power.lock);
 626         }
 627
 628  out:
 629         trace_rpm_return_int_rcuidle(dev, _THIS_IP_, retval);
 630
 631         return retval;
 632
 633  fail:
 634         dev_pm_disable_wake_irq_check(dev);
 635         __update_runtime_status(dev, RPM_ACTIVE);
 636         dev->power.deferred_resume = false;
 637         wake_up_all(&dev->power.wait_queue);
 638
 639         if (retval == -EAGAIN || retval == -EBUSY) {
 640                 dev->power.runtime_error = 0;
 641
 642                 /*
 643                  * If the callback routine failed an autosuspend, and
 644                  * if the last_busy time has been updated so that there
 645                  * is a new autosuspend expiration time, automatically
 646                  * reschedule another autosuspend.
 647                  */
 648                 if ((rpmflags & RPM_AUTO) &&
 649                     pm_runtime_autosuspend_expiration(dev) != 0)
 650                         goto repeat;
 651         } else {
 652                 pm_runtime_cancel_pending(dev);
 653         }
 654         goto out;
 655 }
 656
 657 /**
 658  * rpm_resume - Carry out runtime resume of given device.
 659  * @dev: Device to resume.
 660  * @rpmflags: Flag bits.
 661  *
 662  * Check if the device's runtime PM status allows it to be resumed.  Cancel
 663  * any scheduled or pending requests.  If another resume has been started
 664  * earlier, either return immediately or wait for it to finish, depending on the
 665  * RPM_NOWAIT and RPM_ASYNC flags.  Similarly, if there's a suspend running in
 666  * parallel with this function, either tell the other process to resume after
 667  * suspending (deferred_resume) or wait for it to finish.  If the RPM_ASYNC
 668  * flag is set then queue a resume request; otherwise run the
 669  * ->runtime_resume() callback directly.  Queue an idle notification for the
 670  * device if the resume succeeded.
 671  *
 672  * This function must be called under dev->power.lock with interrupts disabled.
 673  */
 674 static int rpm_resume(struct device *dev, int rpmflags)
 675         __releases(&dev->power.lock) __acquires(&dev->power.lock)
 676 {
 677         int (*callback)(struct device *);
 678         struct device *parent = NULL;
 679         int retval = 0;
 680
 681         trace_rpm_resume_rcuidle(dev, rpmflags);
 682
 683  repeat:
 684         if (dev->power.runtime_error)
 685                 retval = -EINVAL;
 686         else if (dev->power.disable_depth == 1 && dev->power.is_suspended
 687             && dev->power.runtime_status == RPM_ACTIVE)
 688                 retval = 1;
 689         else if (dev->power.disable_depth > 0)
 690                 retval = -EACCES;
 691         if (retval)
 692                 goto out;
 693
 694         /*
 695          * Other scheduled or pending requests need to be canceled.  Small
 696          * optimization: If an autosuspend timer is running, leave it running
 697          * rather than cancelling it now only to restart it again in the near
 698          * future.
 699          */
 700         dev->power.request = RPM_REQ_NONE;
 701         if (!dev->power.timer_autosuspends)
 702                 pm_runtime_deactivate_timer(dev);
 703
 704         if (dev->power.runtime_status == RPM_ACTIVE) {
 705                 retval = 1;
 706                 goto out;
 707         }
 708
 709         if (dev->power.runtime_status == RPM_RESUMING
 710             || dev->power.runtime_status == RPM_SUSPENDING) {
 711                 DEFINE_WAIT(wait);
 712
 713                 if (rpmflags & (RPM_ASYNC | RPM_NOWAIT)) {
 714                         if (dev->power.runtime_status == RPM_SUSPENDING)
 715                                 dev->power.deferred_resume = true;
 716                         else
 717                                 retval = -EINPROGRESS;
 718                         goto out;
 719                 }
 720
 721                 if (dev->power.irq_safe) {
 722                         spin_unlock(&dev->power.lock);
 723
 724                         cpu_relax();
 725
 726                         spin_lock(&dev->power.lock);
 727                         goto repeat;
 728                 }
 729
 730                 /* Wait for the operation carried out in parallel with us. */
 731                 for (;;) {
 732                         prepare_to_wait(&dev->power.wait_queue, &wait,
 733                                         TASK_UNINTERRUPTIBLE);
 734                         if (dev->power.runtime_status != RPM_RESUMING
 735                             && dev->power.runtime_status != RPM_SUSPENDING)
 736                                 break;
 737
 738                         spin_unlock_irq(&dev->power.lock);
 739
 740                         schedule();
 741
 742                         spin_lock_irq(&dev->power.lock);
 743                 }
 744                 finish_wait(&dev->power.wait_queue, &wait);
 745                 goto repeat;
 746         }
 747
 748         /*
 749          * See if we can skip waking up the parent.  This is safe only if
 750          * power.no_callbacks is set, because otherwise we don't know whether
 751          * the resume will actually succeed.
 752          */
 753         if (dev->power.no_callbacks && !parent && dev->parent) {
 754                 spin_lock_nested(&dev->parent->power.lock, SINGLE_DEPTH_NESTING);
 755                 if (dev->parent->power.disable_depth > 0
 756                     || dev->parent->power.ignore_children
 757                     || dev->parent->power.runtime_status == RPM_ACTIVE) {
 758                         atomic_inc(&dev->parent->power.child_count);
 759                         spin_unlock(&dev->parent->power.lock);
 760                         retval = 1;
 761                         goto no_callback;       /* Assume success. */
 762                 }
 763                 spin_unlock(&dev->parent->power.lock);
 764         }
 765
 766         /* Carry out an asynchronous or a synchronous resume. */
 767         if (rpmflags & RPM_ASYNC) {
 768                 dev->power.request = RPM_REQ_RESUME;
 769                 if (!dev->power.request_pending) {
 770                         dev->power.request_pending = true;
 771                         queue_work(pm_wq, &dev->power.work);
 772                 }
 773                 retval = 0;
 774                 goto out;
 775         }
 776
 777         if (!parent && dev->parent) {
 778                 /*
 779                  * Increment the parent's usage counter and resume it if
 780                  * necessary.  Not needed if dev is irq-safe; then the
 781                  * parent is permanently resumed.
 782                  */
 783                 parent = dev->parent;
 784                 if (dev->power.irq_safe)
 785                         goto skip_parent;
 786                 spin_unlock(&dev->power.lock);
 787
 788                 pm_runtime_get_noresume(parent);
 789
 790                 spin_lock(&parent->power.lock);
 791                 /*
 792                  * Resume the parent if it has runtime PM enabled and not been
 793                  * set to ignore its children.
 794                  */
 795                 if (!parent->power.disable_depth
 796                     && !parent->power.ignore_children) {
 797                         rpm_resume(parent, 0);
 798                         if (parent->power.runtime_status != RPM_ACTIVE)
 799                                 retval = -EBUSY;
 800                 }
 801                 spin_unlock(&parent->power.lock);
 802
 803                 spin_lock(&dev->power.lock);
 804                 if (retval)
 805                         goto out;
 806                 goto repeat;
 807         }
 808  skip_parent:
 809
 810         if (dev->power.no_callbacks)
 811                 goto no_callback;       /* Assume success. */
 812
 813         __update_runtime_status(dev, RPM_RESUMING);
 814
 815         callback = RPM_GET_CALLBACK(dev, runtime_resume);
 816
 817         dev_pm_disable_wake_irq_check(dev);
 818         retval = rpm_callback(callback, dev);
 819         if (retval) {
 820                 __update_runtime_status(dev, RPM_SUSPENDED);
 821                 pm_runtime_cancel_pending(dev);
 822                 dev_pm_enable_wake_irq_check(dev, false);
 823         } else {
 824  no_callback:
 825                 __update_runtime_status(dev, RPM_ACTIVE);
 826                 pm_runtime_mark_last_busy(dev);
 827                 if (parent)
 828                         atomic_inc(&parent->power.child_count);
 829         }
 830         wake_up_all(&dev->power.wait_queue);
 831
 832         if (retval >= 0)
 833                 rpm_idle(dev, RPM_ASYNC);
 834
 835  out:
 836         if (parent && !dev->power.irq_safe) {
 837                 spin_unlock_irq(&dev->power.lock);
 838
 839                 pm_runtime_put(parent);
 840
 841                 spin_lock_irq(&dev->power.lock);
 842         }
 843
 844         trace_rpm_return_int_rcuidle(dev, _THIS_IP_, retval);
 845
 846         return retval;
 847 }
 848
 849 /**
 850  * pm_runtime_work - Universal runtime PM work function.
 851  * @work: Work structure used for scheduling the execution of this function.
 852  *
 853  * Use @work to get the device object the work is to be done for, determine what
 854  * is to be done and execute the appropriate runtime PM function.
 855  */
 856 static void pm_runtime_work(struct work_struct *work)
 857 {
 858         struct device *dev = container_of(work, struct device, power.work);
 859         enum rpm_request req;
 860
 861         spin_lock_irq(&dev->power.lock);
 862
 863         if (!dev->power.request_pending)
 864                 goto out;
 865
 866         req = dev->power.request;
 867         dev->power.request = RPM_REQ_NONE;
 868         dev->power.request_pending = false;
 869
 870         switch (req) {
 871         case RPM_REQ_NONE:
 872                 break;
 873         case RPM_REQ_IDLE:
 874                 rpm_idle(dev, RPM_NOWAIT);
 875                 break;
 876         case RPM_REQ_SUSPEND:
 877                 rpm_suspend(dev, RPM_NOWAIT);
 878                 break;
 879         case RPM_REQ_AUTOSUSPEND:
 880                 rpm_suspend(dev, RPM_NOWAIT | RPM_AUTO);
 881                 break;
 882         case RPM_REQ_RESUME:
 883                 rpm_resume(dev, RPM_NOWAIT);
 884                 break;
 885         }
 886
 887  out:
 888         spin_unlock_irq(&dev->power.lock);
 889 }
 890
 891 /**
 892  * pm_suspend_timer_fn - Timer function for pm_schedule_suspend().
 893  * @data: Device pointer passed by pm_schedule_suspend().
 894  *
 895  * Check if the time is right and queue a suspend request.
 896  */
 897 static void pm_suspend_timer_fn(unsigned long data)
 898 {
 899         struct device *dev = (struct device *)data;
 900         unsigned long flags;
 901         unsigned long expires;
 902
 903         spin_lock_irqsave(&dev->power.lock, flags);
 904
 905         expires = dev->power.timer_expires;
 906         /* If 'expire' is after 'jiffies' we've been called too early. */
 907         if (expires > 0 && !time_after(expires, jiffies)) {
 908                 dev->power.timer_expires = 0;
 909                 rpm_suspend(dev, dev->power.timer_autosuspends ?
 910                     (RPM_ASYNC | RPM_AUTO) : RPM_ASYNC);
 911         }
 912
 913         spin_unlock_irqrestore(&dev->power.lock, flags);
 914 }
 915
 916 /**
 917  * pm_schedule_suspend - Set up a timer to submit a suspend request in future.
 918  * @dev: Device to suspend.
 919  * @delay: Time to wait before submitting a suspend request, in milliseconds.
 920  */
 921 int pm_schedule_suspend(struct device *dev, unsigned int delay)
 922 {
 923         unsigned long flags;
 924         int retval;
 925
 926         spin_lock_irqsave(&dev->power.lock, flags);
 927
 928         if (!delay) {
 929                 retval = rpm_suspend(dev, RPM_ASYNC);
 930                 goto out;
 931         }
 932
 933         retval = rpm_check_suspend_allowed(dev);
 934         if (retval)
 935                 goto out;
 936
 937         /* Other scheduled or pending requests need to be canceled. */
 938         pm_runtime_cancel_pending(dev);
 939
 940         dev->power.timer_expires = jiffies + msecs_to_jiffies(delay);
 941         dev->power.timer_expires += !dev->power.timer_expires;
 942         dev->power.timer_autosuspends = 0;
 943         mod_timer(&dev->power.suspend_timer, dev->power.timer_expires);
 944
 945  out:
 946         spin_unlock_irqrestore(&dev->power.lock, flags);
 947
 948         return retval;
 949 }
 950 EXPORT_SYMBOL_GPL(pm_schedule_suspend);
 951
 952 /**
 953  * __pm_runtime_idle - Entry point for runtime idle operations.
 954  * @dev: Device to send idle notification for.
 955  * @rpmflags: Flag bits.
 956  *
 957  * If the RPM_GET_PUT flag is set, decrement the device's usage count and
 958  * return immediately if it is larger than zero.  Then carry out an idle
 959  * notification, either synchronous or asynchronous.
 960  *
 961  * This routine may be called in atomic context if the RPM_ASYNC flag is set,
 962  * or if pm_runtime_irq_safe() has been called.
 963  */
 964 int __pm_runtime_idle(struct device *dev, int rpmflags)
 965 {
 966         unsigned long flags;
 967         int retval;
 968
 969         if (rpmflags & RPM_GET_PUT) {
 970                 if (!atomic_dec_and_test(&dev->power.usage_count))
 971                         return 0;
 972         }
 973
 974         might_sleep_if(!(rpmflags & RPM_ASYNC) && !dev->power.irq_safe);
 975
 976         spin_lock_irqsave(&dev->power.lock, flags);
 977         retval = rpm_idle(dev, rpmflags);
 978         spin_unlock_irqrestore(&dev->power.lock, flags);
 979
 980         return retval;
 981 }
 982 EXPORT_SYMBOL_GPL(__pm_runtime_idle);
 983
 984 /**
 985  * __pm_runtime_suspend - Entry point for runtime put/suspend operations.
 986  * @dev: Device to suspend.
 987  * @rpmflags: Flag bits.
 988  *
 989  * If the RPM_GET_PUT flag is set, decrement the device's usage count and
 990  * return immediately if it is larger than zero.  Then carry out a suspend,
 991  * either synchronous or asynchronous.
 992  *
 993  * This routine may be called in atomic context if the RPM_ASYNC flag is set,
 994  * or if pm_runtime_irq_safe() has been called.
 995  */
 996 int __pm_runtime_suspend(struct device *dev, int rpmflags)
 997 {
 998         unsigned long flags;
 999         int retval;
1000
1001         if (rpmflags & RPM_GET_PUT) {
1002                 if (!atomic_dec_and_test(&dev->power.usage_count))
1003                         return 0;
1004         }
1005
1006         might_sleep_if(!(rpmflags & RPM_ASYNC) && !dev->power.irq_safe);
1007
1008         spin_lock_irqsave(&dev->power.lock, flags);
1009         retval = rpm_suspend(dev, rpmflags);
1010         spin_unlock_irqrestore(&dev->power.lock, flags);
1011
1012         return retval;
1013 }
1014 EXPORT_SYMBOL_GPL(__pm_runtime_suspend);
1015
1016 /**
1017  * __pm_runtime_resume - Entry point for runtime resume operations.
1018  * @dev: Device to resume.
1019  * @rpmflags: Flag bits.
1020  *
1021  * If the RPM_GET_PUT flag is set, increment the device's usage count.  Then
1022  * carry out a resume, either synchronous or asynchronous.
1023  *
1024  * This routine may be called in atomic context if the RPM_ASYNC flag is set,
1025  * or if pm_runtime_irq_safe() has been called.
1026  */
1027 int __pm_runtime_resume(struct device *dev, int rpmflags)
1028 {
1029         unsigned long flags;
1030         int retval;
1031
1032         might_sleep_if(!(rpmflags & RPM_ASYNC) && !dev->power.irq_safe &&
1033                         dev->power.runtime_status != RPM_ACTIVE);
1034
1035         if (rpmflags & RPM_GET_PUT)
1036                 atomic_inc(&dev->power.usage_count);
1037
1038         spin_lock_irqsave(&dev->power.lock, flags);
1039         retval = rpm_resume(dev, rpmflags);
1040         spin_unlock_irqrestore(&dev->power.lock, flags);
1041
1042         return retval;
1043 }
1044 EXPORT_SYMBOL_GPL(__pm_runtime_resume);
1045
1046 /**
1047  * pm_runtime_get_if_in_use - Conditionally bump up the device's usage counter.
1048  * @dev: Device to handle.
1049  *
1050  * Return -EINVAL if runtime PM is disabled for the device.
1051  *
1052  * If that's not the case and if the device's runtime PM status is RPM_ACTIVE
1053  * and the runtime PM usage counter is nonzero, increment the counter and
1054  * return 1.  Otherwise return 0 without changing the counter.
1055  */
1056 int pm_runtime_get_if_in_use(struct device *dev)
1057 {
1058         unsigned long flags;
1059         int retval;
1060
1061         spin_lock_irqsave(&dev->power.lock, flags);
1062         retval = dev->power.disable_depth > 0 ? -EINVAL :
1063                 dev->power.runtime_status == RPM_ACTIVE
1064                         && atomic_inc_not_zero(&dev->power.usage_count);
1065         spin_unlock_irqrestore(&dev->power.lock, flags);
1066         return retval;
1067 }
1068 EXPORT_SYMBOL_GPL(pm_runtime_get_if_in_use);
1069
1070 /**
1071  * __pm_runtime_set_status - Set runtime PM status of a device.
1072  * @dev: Device to handle.
1073  * @status: New runtime PM status of the device.
1074  *
1075  * If runtime PM of the device is disabled or its power.runtime_error field is
1076  * different from zero, the status may be changed either to RPM_ACTIVE, or to
1077  * RPM_SUSPENDED, as long as that reflects the actual state of the device.
1078  * However, if the device has a parent and the parent is not active, and the
1079  * parent's power.ignore_children flag is unset, the device's status cannot be
1080  * set to RPM_ACTIVE, so -EBUSY is returned in that case.
1081  *
1082  * If successful, __pm_runtime_set_status() clears the power.runtime_error field
1083  * and the device parent's counter of unsuspended children is modified to
1084  * reflect the new status.  If the new status is RPM_SUSPENDED, an idle
1085  * notification request for the parent is submitted.
1086  */
1087 int __pm_runtime_set_status(struct device *dev, unsigned int status)
1088 {
1089         struct device *parent = dev->parent;
1090         unsigned long flags;
1091         bool notify_parent = false;
1092         int error = 0;
1093
1094         if (status != RPM_ACTIVE && status != RPM_SUSPENDED)
1095                 return -EINVAL;
1096
1097         spin_lock_irqsave(&dev->power.lock, flags);
1098
1099         if (!dev->power.runtime_error && !dev->power.disable_depth) {
1100                 error = -EAGAIN;
1101                 goto out;
1102         }
1103
1104         if (dev->power.runtime_status == status)
1105                 goto out_set;
1106
1107         if (status == RPM_SUSPENDED) {
1108                 /*
1109                  * It is invalid to suspend a device with an active child,
1110                  * unless it has been set to ignore its children.
1111                  */
1112                 if (!dev->power.ignore_children &&
1113                         atomic_read(&dev->power.child_count)) {
1114                         dev_err(dev, "runtime PM trying to suspend device but active child\n");
1115                         error = -EBUSY;
1116                         goto out;
1117                 }
1118
1119                 if (parent) {
1120                         atomic_add_unless(&parent->power.child_count, -1, 0);
1121                         notify_parent = !parent->power.ignore_children;
1122                 }
1123                 goto out_set;
1124         }
1125
1126         if (parent) {
1127                 spin_lock_nested(&parent->power.lock, SINGLE_DEPTH_NESTING);
1128
1129                 /*
1130                  * It is invalid to put an active child under a parent that is
1131                  * not active, has runtime PM enabled and the
1132                  * 'power.ignore_children' flag unset.
1133                  */
1134                 if (!parent->power.disable_depth
1135                     && !parent->power.ignore_children
1136                     && parent->power.runtime_status != RPM_ACTIVE) {
1137                         dev_err(dev, "runtime PM trying to activate child device %s but parent (%s) is not active\n",
1138                                 dev_name(dev),
1139                                 dev_name(parent));
1140                         error = -EBUSY;
1141                 } else if (dev->power.runtime_status == RPM_SUSPENDED) {
1142                         atomic_inc(&parent->power.child_count);
1143                 }
1144
1145                 spin_unlock(&parent->power.lock);
1146
1147                 if (error)
1148                         goto out;
1149         }
1150
1151  out_set:
1152         __update_runtime_status(dev, status);
1153         dev->power.runtime_error = 0;
1154  out:
1155         spin_unlock_irqrestore(&dev->power.lock, flags);
1156
1157         if (notify_parent)
1158                 pm_request_idle(parent);
1159
1160         return error;
1161 }
1162 EXPORT_SYMBOL_GPL(__pm_runtime_set_status);
1163
1164 /**
1165  * __pm_runtime_barrier - Cancel pending requests and wait for completions.
1166  * @dev: Device to handle.
1167  *
1168  * Flush all pending requests for the device from pm_wq and wait for all
1169  * runtime PM operations involving the device in progress to complete.
1170  *
1171  * Should be called under dev->power.lock with interrupts disabled.
1172  */
1173 static void __pm_runtime_barrier(struct device *dev)
1174 {
1175         pm_runtime_deactivate_timer(dev);
1176
1177         if (dev->power.request_pending) {
1178                 dev->power.request = RPM_REQ_NONE;
1179                 spin_unlock_irq(&dev->power.lock);
1180
1181                 cancel_work_sync(&dev->power.work);
1182
1183                 spin_lock_irq(&dev->power.lock);
1184                 dev->power.request_pending = false;
1185         }
1186
1187         if (dev->power.runtime_status == RPM_SUSPENDING
1188             || dev->power.runtime_status == RPM_RESUMING
1189             || dev->power.idle_notification) {
1190                 DEFINE_WAIT(wait);
1191
1192                 /* Suspend, wake-up or idle notification in progress. */
1193                 for (;;) {
1194                         prepare_to_wait(&dev->power.wait_queue, &wait,
1195                                         TASK_UNINTERRUPTIBLE);
1196                         if (dev->power.runtime_status != RPM_SUSPENDING
1197                             && dev->power.runtime_status != RPM_RESUMING
1198                             && !dev->power.idle_notification)
1199                                 break;
1200                         spin_unlock_irq(&dev->power.lock);
1201
1202                         schedule();
1203
1204                         spin_lock_irq(&dev->power.lock);
1205                 }
1206                 finish_wait(&dev->power.wait_queue, &wait);
1207         }
1208 }
1209
1210 /**
1211  * pm_runtime_barrier - Flush pending requests and wait for completions.
1212  * @dev: Device to handle.
1213  *
1214  * Prevent the device from being suspended by incrementing its usage counter and
1215  * if there's a pending resume request for the device, wake the device up.
1216  * Next, make sure that all pending requests for the device have been flushed
1217  * from pm_wq and wait for all runtime PM operations involving the device in
1218  * progress to complete.
1219  *
1220  * Return value:
1221  * 1, if there was a resume request pending and the device had to be woken up,
1222  * 0, otherwise
1223  */
1224 int pm_runtime_barrier(struct device *dev)
1225 {
1226         int retval = 0;
1227
1228         pm_runtime_get_noresume(dev);
1229         spin_lock_irq(&dev->power.lock);
1230
1231         if (dev->power.request_pending
1232             && dev->power.request == RPM_REQ_RESUME) {
1233                 rpm_resume(dev, 0);
1234                 retval = 1;
1235         }
1236
1237         __pm_runtime_barrier(dev);
1238
1239         spin_unlock_irq(&dev->power.lock);
1240         pm_runtime_put_noidle(dev);
1241
1242         return retval;
1243 }
1244 EXPORT_SYMBOL_GPL(pm_runtime_barrier);
1245
1246 /**
1247  * __pm_runtime_disable - Disable runtime PM of a device.
1248  * @dev: Device to handle.
1249  * @check_resume: If set, check if there's a resume request for the device.
1250  *
1251  * Increment power.disable_depth for the device and if it was zero previously,
1252  * cancel all pending runtime PM requests for the device and wait for all
1253  * operations in progress to complete.  The device can be either active or
1254  * suspended after its runtime PM has been disabled.
1255  *
1256  * If @check_resume is set and there's a resume request pending when
1257  * __pm_runtime_disable() is called and power.disable_depth is zero, the
1258  * function will wake up the device before disabling its runtime PM.
1259  */
1260 void __pm_runtime_disable(struct device *dev, bool check_resume)
1261 {
1262         spin_lock_irq(&dev->power.lock);
1263
1264         if (dev->power.disable_depth > 0) {
1265                 dev->power.disable_depth++;
1266                 goto out;
1267         }
1268
1269         /*
1270          * Wake up the device if there's a resume request pending, because that
1271          * means there probably is some I/O to process and disabling runtime PM
1272          * shouldn't prevent the device from processing the I/O.
1273          */
1274         if (check_resume && dev->power.request_pending
1275             && dev->power.request == RPM_REQ_RESUME) {
1276                 /*
1277                  * Prevent suspends and idle notifications from being carried
1278                  * out after we have woken up the device.
1279                  */
1280                 pm_runtime_get_noresume(dev);
1281
1282                 rpm_resume(dev, 0);
1283
1284                 pm_runtime_put_noidle(dev);
1285         }
1286
1287         if (!dev->power.disable_depth++)
1288                 __pm_runtime_barrier(dev);
1289
1290  out:
1291         spin_unlock_irq(&dev->power.lock);
1292 }
1293 EXPORT_SYMBOL_GPL(__pm_runtime_disable);
1294
1295 /**
1296  * pm_runtime_enable - Enable runtime PM of a device.
1297  * @dev: Device to handle.
1298  */
1299 void pm_runtime_enable(struct device *dev)
1300 {
1301         unsigned long flags;
1302
1303         spin_lock_irqsave(&dev->power.lock, flags);
1304
1305         if (dev->power.disable_depth > 0)
1306                 dev->power.disable_depth--;
1307         else
1308                 dev_warn(dev, "Unbalanced %s!\n", __func__);
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         setup_timer(&dev->power.suspend_timer, pm_suspend_timer_fn,
1503                         (unsigned long)dev);
1504
1505         init_waitqueue_head(&dev->power.wait_queue);
1506 }
1507
1508 /**
1509  * pm_runtime_reinit - Re-initialize runtime PM fields in given device object.
1510  * @dev: Device object to re-initialize.
1511  */
1512 void pm_runtime_reinit(struct device *dev)
1513 {
1514         if (!pm_runtime_enabled(dev)) {
1515                 if (dev->power.runtime_status == RPM_ACTIVE)
1516                         pm_runtime_set_suspended(dev);
1517                 if (dev->power.irq_safe) {
1518                         spin_lock_irq(&dev->power.lock);
1519                         dev->power.irq_safe = 0;
1520                         spin_unlock_irq(&dev->power.lock);
1521                         if (dev->parent)
1522                                 pm_runtime_put(dev->parent);
1523                 }
1524         }
1525 }
1526
1527 /**
1528  * pm_runtime_remove - Prepare for removing a device from device hierarchy.
1529  * @dev: Device object being removed from device hierarchy.
1530  */
1531 void pm_runtime_remove(struct device *dev)
1532 {
1533         __pm_runtime_disable(dev, false);
1534         pm_runtime_reinit(dev);
1535 }
1536
1537 /**
1538  * pm_runtime_clean_up_links - Prepare links to consumers for driver removal.
1539  * @dev: Device whose driver is going to be removed.
1540  *
1541  * Check links from this device to any consumers and if any of them have active
1542  * runtime PM references to the device, drop the usage counter of the device
1543  * (once per link).
1544  *
1545  * Links with the DL_FLAG_STATELESS flag set are ignored.
1546  *
1547  * Since the device is guaranteed to be runtime-active at the point this is
1548  * called, nothing else needs to be done here.
1549  *
1550  * Moreover, this is called after device_links_busy() has returned 'false', so
1551  * the status of each link is guaranteed to be DL_STATE_SUPPLIER_UNBIND and
1552  * therefore rpm_active can't be manipulated concurrently.
1553  */
1554 void pm_runtime_clean_up_links(struct device *dev)
1555 {
1556         struct device_link *link;
1557         int idx;
1558
1559         idx = device_links_read_lock();
1560
1561         list_for_each_entry_rcu(link, &dev->links.consumers, s_node) {
1562                 if (link->flags & DL_FLAG_STATELESS)
1563                         continue;
1564
1565                 if (link->rpm_active) {
1566                         pm_runtime_put_noidle(dev);
1567                         link->rpm_active = false;
1568                 }
1569         }
1570
1571         device_links_read_unlock(idx);
1572 }
1573
1574 /**
1575  * pm_runtime_get_suppliers - Resume and reference-count supplier devices.
1576  * @dev: Consumer device.
1577  */
1578 void pm_runtime_get_suppliers(struct device *dev)
1579 {
1580         struct device_link *link;
1581         int idx;
1582
1583         idx = device_links_read_lock();
1584
1585         list_for_each_entry_rcu(link, &dev->links.suppliers, c_node)
1586                 if (link->flags & DL_FLAG_PM_RUNTIME)
1587                         pm_runtime_get_sync(link->supplier);
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         int idx;
1600
1601         idx = device_links_read_lock();
1602
1603         list_for_each_entry_rcu(link, &dev->links.suppliers, c_node)
1604                 if (link->flags & DL_FLAG_PM_RUNTIME)
1605                         pm_runtime_put(link->supplier);
1606
1607         device_links_read_unlock(idx);
1608 }
1609
1610 void pm_runtime_new_link(struct device *dev)
1611 {
1612         spin_lock_irq(&dev->power.lock);
1613         dev->power.links_count++;
1614         spin_unlock_irq(&dev->power.lock);
1615 }
1616
1617 void pm_runtime_drop_link(struct device *dev)
1618 {
1619         spin_lock_irq(&dev->power.lock);
1620         WARN_ON(dev->power.links_count == 0);
1621         dev->power.links_count--;
1622         spin_unlock_irq(&dev->power.lock);
1623 }
1624
1625 /**
1626  * pm_runtime_force_suspend - Force a device into suspend state if needed.
1627  * @dev: Device to suspend.
1628  *
1629  * Disable runtime PM so we safely can check the device's runtime PM status and
1630  * if it is active, invoke it's .runtime_suspend callback to bring it into
1631  * suspend state. Keep runtime PM disabled to preserve the state unless we
1632  * encounter errors.
1633  *
1634  * Typically this function may be invoked from a system suspend callback to make
1635  * sure the device is put into low power state.
1636  */
1637 int pm_runtime_force_suspend(struct device *dev)
1638 {
1639         int (*callback)(struct device *);
1640         int ret = 0;
1641
1642         pm_runtime_disable(dev);
1643         if (pm_runtime_status_suspended(dev))
1644                 return 0;
1645
1646         callback = RPM_GET_CALLBACK(dev, runtime_suspend);
1647
1648         if (!callback) {
1649                 ret = -ENOSYS;
1650                 goto err;
1651         }
1652
1653         ret = callback(dev);
1654         if (ret)
1655                 goto err;
1656
1657         /*
1658          * Increase the runtime PM usage count for the device's parent, in case
1659          * when we find the device being used when system suspend was invoked.
1660          * This informs pm_runtime_force_resume() to resume the parent
1661          * immediately, which is needed to be able to resume its children,
1662          * when not deferring the resume to be managed via runtime PM.
1663          */
1664         if (dev->parent && atomic_read(&dev->power.usage_count) > 1)
1665                 pm_runtime_get_noresume(dev->parent);
1666
1667         pm_runtime_set_suspended(dev);
1668         return 0;
1669 err:
1670         pm_runtime_enable(dev);
1671         return ret;
1672 }
1673 EXPORT_SYMBOL_GPL(pm_runtime_force_suspend);
1674
1675 /**
1676  * pm_runtime_force_resume - Force a device into resume state if needed.
1677  * @dev: Device to resume.
1678  *
1679  * Prior invoking this function we expect the user to have brought the device
1680  * into low power state by a call to pm_runtime_force_suspend(). Here we reverse
1681  * those actions and brings the device into full power, if it is expected to be
1682  * used on system resume. To distinguish that, we check whether the runtime PM
1683  * usage count is greater than 1 (the PM core increases the usage count in the
1684  * system PM prepare phase), as that indicates a real user (such as a subsystem,
1685  * driver, userspace, etc.) is using it. If that is the case, the device is
1686  * expected to be used on system resume as well, so then we resume it. In the
1687  * other case, we defer the resume to be managed via runtime PM.
1688  *
1689  * Typically this function may be invoked from a system resume callback.
1690  */
1691 int pm_runtime_force_resume(struct device *dev)
1692 {
1693         int (*callback)(struct device *);
1694         int ret = 0;
1695
1696         callback = RPM_GET_CALLBACK(dev, runtime_resume);
1697
1698         if (!callback) {
1699                 ret = -ENOSYS;
1700                 goto out;
1701         }
1702
1703         if (!pm_runtime_status_suspended(dev))
1704                 goto out;
1705
1706         /*
1707          * Decrease the parent's runtime PM usage count, if we increased it
1708          * during system suspend in pm_runtime_force_suspend().
1709         */
1710         if (atomic_read(&dev->power.usage_count) > 1) {
1711                 if (dev->parent)
1712                         pm_runtime_put_noidle(dev->parent);
1713         } else {
1714                 goto out;
1715         }
1716
1717         ret = pm_runtime_set_active(dev);
1718         if (ret)
1719                 goto out;
1720
1721         ret = callback(dev);
1722         if (ret) {
1723                 pm_runtime_set_suspended(dev);
1724                 goto out;
1725         }
1726
1727         pm_runtime_mark_last_busy(dev);
1728 out:
1729         pm_runtime_enable(dev);
1730         return ret;
1731 }
1732 EXPORT_SYMBOL_GPL(pm_runtime_force_resume);