]> git.karo-electronics.de Git - mv-sheeva.git/blobdiff - drivers/base/power/runtime.c
Merge branch 'master' into tk71
[mv-sheeva.git] / drivers / base / power / runtime.c
index b78c401ffa7380f67f29dbddf1fecf1396746e85..42615b419dfb5bc4f30e0f6dfb2ad607015d95b7 100644 (file)
@@ -2,17 +2,55 @@
  * drivers/base/power/runtime.c - Helper functions for device run-time PM
  *
  * Copyright (c) 2009 Rafael J. Wysocki <rjw@sisk.pl>, Novell Inc.
+ * Copyright (C) 2010 Alan Stern <stern@rowland.harvard.edu>
  *
  * This file is released under the GPLv2.
  */
 
 #include <linux/sched.h>
 #include <linux/pm_runtime.h>
-#include <linux/jiffies.h>
+#include "power.h"
 
-static int __pm_runtime_resume(struct device *dev, bool from_wq);
-static int __pm_request_idle(struct device *dev);
-static int __pm_request_resume(struct device *dev);
+static int rpm_resume(struct device *dev, int rpmflags);
+static int rpm_suspend(struct device *dev, int rpmflags);
+
+/**
+ * update_pm_runtime_accounting - Update the time accounting of power states
+ * @dev: Device to update the accounting for
+ *
+ * In order to be able to have time accounting of the various power states
+ * (as used by programs such as PowerTOP to show the effectiveness of runtime
+ * PM), we need to track the time spent in each state.
+ * update_pm_runtime_accounting must be called each time before the
+ * runtime_status field is updated, to account the time in the old state
+ * correctly.
+ */
+void update_pm_runtime_accounting(struct device *dev)
+{
+       unsigned long now = jiffies;
+       int delta;
+
+       delta = now - dev->power.accounting_timestamp;
+
+       if (delta < 0)
+               delta = 0;
+
+       dev->power.accounting_timestamp = now;
+
+       if (dev->power.disable_depth > 0)
+               return;
+
+       if (dev->power.runtime_status == RPM_SUSPENDED)
+               dev->power.suspended_jiffies += delta;
+       else
+               dev->power.active_jiffies += delta;
+}
+
+static void __update_runtime_status(struct device *dev, enum rpm_status status)
+{
+       update_pm_runtime_accounting(dev);
+       dev->power.runtime_status = status;
+}
 
 /**
  * pm_runtime_deactivate_timer - Deactivate given device's suspend timer.
@@ -40,62 +78,154 @@ static void pm_runtime_cancel_pending(struct device *dev)
        dev->power.request = RPM_REQ_NONE;
 }
 
-/**
- * __pm_runtime_idle - Notify device bus type if the device can be suspended.
- * @dev: Device to notify the bus type about.
+/*
+ * pm_runtime_autosuspend_expiration - Get a device's autosuspend-delay expiration time.
+ * @dev: Device to handle.
  *
- * This function must be called under dev->power.lock with interrupts disabled.
+ * Compute the autosuspend-delay expiration time based on the device's
+ * power.last_busy time.  If the delay has already expired or is disabled
+ * (negative) or the power.use_autosuspend flag isn't set, return 0.
+ * Otherwise return the expiration time in jiffies (adjusted to be nonzero).
+ *
+ * This function may be called either with or without dev->power.lock held.
+ * Either way it can be racy, since power.last_busy may be updated at any time.
  */
-static int __pm_runtime_idle(struct device *dev)
-       __releases(&dev->power.lock) __acquires(&dev->power.lock)
+unsigned long pm_runtime_autosuspend_expiration(struct device *dev)
+{
+       int autosuspend_delay;
+       long elapsed;
+       unsigned long last_busy;
+       unsigned long expires = 0;
+
+       if (!dev->power.use_autosuspend)
+               goto out;
+
+       autosuspend_delay = ACCESS_ONCE(dev->power.autosuspend_delay);
+       if (autosuspend_delay < 0)
+               goto out;
+
+       last_busy = ACCESS_ONCE(dev->power.last_busy);
+       elapsed = jiffies - last_busy;
+       if (elapsed < 0)
+               goto out;       /* jiffies has wrapped around. */
+
+       /*
+        * If the autosuspend_delay is >= 1 second, align the timer by rounding
+        * up to the nearest second.
+        */
+       expires = last_busy + msecs_to_jiffies(autosuspend_delay);
+       if (autosuspend_delay >= 1000)
+               expires = round_jiffies(expires);
+       expires += !expires;
+       if (elapsed >= expires - last_busy)
+               expires = 0;    /* Already expired. */
+
+ out:
+       return expires;
+}
+EXPORT_SYMBOL_GPL(pm_runtime_autosuspend_expiration);
+
+/**
+ * rpm_check_suspend_allowed - Test whether a device may be suspended.
+ * @dev: Device to test.
+ */
+static int rpm_check_suspend_allowed(struct device *dev)
 {
        int retval = 0;
 
        if (dev->power.runtime_error)
                retval = -EINVAL;
-       else if (dev->power.idle_notification)
-               retval = -EINPROGRESS;
        else if (atomic_read(&dev->power.usage_count) > 0
-           || dev->power.disable_depth > 0
-           || dev->power.runtime_status != RPM_ACTIVE)
+           || dev->power.disable_depth > 0)
                retval = -EAGAIN;
        else if (!pm_children_suspended(dev))
                retval = -EBUSY;
+
+       /* Pending resume requests take precedence over suspends. */
+       else if ((dev->power.deferred_resume
+                       && dev->power.runtime_status == RPM_SUSPENDING)
+           || (dev->power.request_pending
+                       && dev->power.request == RPM_REQ_RESUME))
+               retval = -EAGAIN;
+       else if (dev->power.runtime_status == RPM_SUSPENDED)
+               retval = 1;
+
+       return retval;
+}
+
+/**
+ * rpm_idle - Notify device bus type if the device can be suspended.
+ * @dev: Device to notify the bus type about.
+ * @rpmflags: Flag bits.
+ *
+ * Check if the device's run-time PM status allows it to be suspended.  If
+ * another idle notification has been started earlier, return immediately.  If
+ * the RPM_ASYNC flag is set then queue an idle-notification request; otherwise
+ * run the ->runtime_idle() callback directly.
+ *
+ * This function must be called under dev->power.lock with interrupts disabled.
+ */
+static int rpm_idle(struct device *dev, int rpmflags)
+{
+       int (*callback)(struct device *);
+       int retval;
+
+       retval = rpm_check_suspend_allowed(dev);
+       if (retval < 0)
+               ;       /* Conditions are wrong. */
+
+       /* Idle notifications are allowed only in the RPM_ACTIVE state. */
+       else if (dev->power.runtime_status != RPM_ACTIVE)
+               retval = -EAGAIN;
+
+       /*
+        * Any pending request other than an idle notification takes
+        * precedence over us, except that the timer may be running.
+        */
+       else if (dev->power.request_pending &&
+           dev->power.request > RPM_REQ_IDLE)
+               retval = -EAGAIN;
+
+       /* Act as though RPM_NOWAIT is always set. */
+       else if (dev->power.idle_notification)
+               retval = -EINPROGRESS;
        if (retval)
                goto out;
 
-       if (dev->power.request_pending) {
-               /*
-                * If an idle notification request is pending, cancel it.  Any
-                * other pending request takes precedence over us.
-                */
-               if (dev->power.request == RPM_REQ_IDLE) {
-                       dev->power.request = RPM_REQ_NONE;
-               } else if (dev->power.request != RPM_REQ_NONE) {
-                       retval = -EAGAIN;
-                       goto out;
+       /* Pending requests need to be canceled. */
+       dev->power.request = RPM_REQ_NONE;
+
+       if (dev->power.no_callbacks) {
+               /* Assume ->runtime_idle() callback would have suspended. */
+               retval = rpm_suspend(dev, rpmflags);
+               goto out;
+       }
+
+       /* Carry out an asynchronous or a synchronous idle notification. */
+       if (rpmflags & RPM_ASYNC) {
+               dev->power.request = RPM_REQ_IDLE;
+               if (!dev->power.request_pending) {
+                       dev->power.request_pending = true;
+                       queue_work(pm_wq, &dev->power.work);
                }
+               goto out;
        }
 
        dev->power.idle_notification = true;
 
-       if (dev->bus && dev->bus->pm && dev->bus->pm->runtime_idle) {
-               spin_unlock_irq(&dev->power.lock);
-
-               dev->bus->pm->runtime_idle(dev);
-
-               spin_lock_irq(&dev->power.lock);
-       } else if (dev->type && dev->type->pm && dev->type->pm->runtime_idle) {
-               spin_unlock_irq(&dev->power.lock);
-
-               dev->type->pm->runtime_idle(dev);
+       if (dev->bus && dev->bus->pm && dev->bus->pm->runtime_idle)
+               callback = dev->bus->pm->runtime_idle;
+       else if (dev->type && dev->type->pm && dev->type->pm->runtime_idle)
+               callback = dev->type->pm->runtime_idle;
+       else if (dev->class && dev->class->pm)
+               callback = dev->class->pm->runtime_idle;
+       else
+               callback = NULL;
 
-               spin_lock_irq(&dev->power.lock);
-       } else if (dev->class && dev->class->pm
-           && dev->class->pm->runtime_idle) {
+       if (callback) {
                spin_unlock_irq(&dev->power.lock);
 
-               dev->class->pm->runtime_idle(dev);
+               callback(dev);
 
                spin_lock_irq(&dev->power.lock);
        }
@@ -108,113 +238,102 @@ static int __pm_runtime_idle(struct device *dev)
 }
 
 /**
- * pm_runtime_idle - Notify device bus type if the device can be suspended.
- * @dev: Device to notify the bus type about.
+ * rpm_callback - Run a given runtime PM callback for a given device.
+ * @cb: Runtime PM callback to run.
+ * @dev: Device to run the callback for.
  */
-int pm_runtime_idle(struct device *dev)
+static int rpm_callback(int (*cb)(struct device *), struct device *dev)
+       __releases(&dev->power.lock) __acquires(&dev->power.lock)
 {
        int retval;
 
-       spin_lock_irq(&dev->power.lock);
-       retval = __pm_runtime_idle(dev);
-       spin_unlock_irq(&dev->power.lock);
-
-       return retval;
-}
-EXPORT_SYMBOL_GPL(pm_runtime_idle);
-
-
-/**
- * update_pm_runtime_accounting - Update the time accounting of power states
- * @dev: Device to update the accounting for
- *
- * In order to be able to have time accounting of the various power states
- * (as used by programs such as PowerTOP to show the effectiveness of runtime
- * PM), we need to track the time spent in each state.
- * update_pm_runtime_accounting must be called each time before the
- * runtime_status field is updated, to account the time in the old state
- * correctly.
- */
-void update_pm_runtime_accounting(struct device *dev)
-{
-       unsigned long now = jiffies;
-       int delta;
-
-       delta = now - dev->power.accounting_timestamp;
-
-       if (delta < 0)
-               delta = 0;
+       if (!cb)
+               return -ENOSYS;
 
-       dev->power.accounting_timestamp = now;
-
-       if (dev->power.disable_depth > 0)
-               return;
+       if (dev->power.irq_safe) {
+               retval = cb(dev);
+       } else {
+               spin_unlock_irq(&dev->power.lock);
 
-       if (dev->power.runtime_status == RPM_SUSPENDED)
-               dev->power.suspended_jiffies += delta;
-       else
-               dev->power.active_jiffies += delta;
-}
+               retval = cb(dev);
 
-static void __update_runtime_status(struct device *dev, enum rpm_status status)
-{
-       update_pm_runtime_accounting(dev);
-       dev->power.runtime_status = status;
+               spin_lock_irq(&dev->power.lock);
+       }
+       dev->power.runtime_error = retval;
+       return retval;
 }
 
 /**
- * __pm_runtime_suspend - Carry out run-time suspend of given device.
+ * rpm_suspend - Carry out run-time suspend of given device.
  * @dev: Device to suspend.
- * @from_wq: If set, the function has been called via pm_wq.
+ * @rpmflags: Flag bits.
  *
- * Check if the device can be suspended and run the ->runtime_suspend() callback
- * provided by its bus type.  If another suspend has been started earlier, wait
- * for it to finish.  If an idle notification or suspend request is pending or
- * scheduled, cancel it.
+ * Check if the device's run-time PM status allows it to be suspended.  If
+ * another suspend has been started earlier, either return immediately or wait
+ * for it to finish, depending on the RPM_NOWAIT and RPM_ASYNC flags.  Cancel a
+ * pending idle notification.  If the RPM_ASYNC flag is set then queue a
+ * suspend request; otherwise run the ->runtime_suspend() callback directly.
+ * If a deferred resume was requested while the callback was running then carry
+ * it out; otherwise send an idle notification for the device (if the suspend
+ * failed) or for its parent (if the suspend succeeded).
  *
  * This function must be called under dev->power.lock with interrupts disabled.
  */
-int __pm_runtime_suspend(struct device *dev, bool from_wq)
+static int rpm_suspend(struct device *dev, int rpmflags)
        __releases(&dev->power.lock) __acquires(&dev->power.lock)
 {
+       int (*callback)(struct device *);
        struct device *parent = NULL;
-       bool notify = false;
-       int retval = 0;
+       int retval;
 
-       dev_dbg(dev, "__pm_runtime_suspend()%s!\n",
-               from_wq ? " from workqueue" : "");
+       dev_dbg(dev, "%s flags 0x%x\n", __func__, rpmflags);
 
  repeat:
-       if (dev->power.runtime_error) {
-               retval = -EINVAL;
-               goto out;
-       }
+       retval = rpm_check_suspend_allowed(dev);
 
-       /* Pending resume requests take precedence over us. */
-       if (dev->power.request_pending
-           && dev->power.request == RPM_REQ_RESUME) {
+       if (retval < 0)
+               ;       /* Conditions are wrong. */
+
+       /* Synchronous suspends are not allowed in the RPM_RESUMING state. */
+       else if (dev->power.runtime_status == RPM_RESUMING &&
+           !(rpmflags & RPM_ASYNC))
                retval = -EAGAIN;
+       if (retval)
                goto out;
+
+       /* If the autosuspend_delay time hasn't expired yet, reschedule. */
+       if ((rpmflags & RPM_AUTO)
+           && dev->power.runtime_status != RPM_SUSPENDING) {
+               unsigned long expires = pm_runtime_autosuspend_expiration(dev);
+
+               if (expires != 0) {
+                       /* Pending requests need to be canceled. */
+                       dev->power.request = RPM_REQ_NONE;
+
+                       /*
+                        * Optimization: If the timer is already running and is
+                        * set to expire at or before the autosuspend delay,
+                        * avoid the overhead of resetting it.  Just let it
+                        * expire; pm_suspend_timer_fn() will take care of the
+                        * rest.
+                        */
+                       if (!(dev->power.timer_expires && time_before_eq(
+                           dev->power.timer_expires, expires))) {
+                               dev->power.timer_expires = expires;
+                               mod_timer(&dev->power.suspend_timer, expires);
+                       }
+                       dev->power.timer_autosuspends = 1;
+                       goto out;
+               }
        }
 
        /* Other scheduled or pending requests need to be canceled. */
        pm_runtime_cancel_pending(dev);
 
-       if (dev->power.runtime_status == RPM_SUSPENDED)
-               retval = 1;
-       else if (dev->power.runtime_status == RPM_RESUMING
-           || dev->power.disable_depth > 0
-           || atomic_read(&dev->power.usage_count) > 0)
-               retval = -EAGAIN;
-       else if (!pm_children_suspended(dev))
-               retval = -EBUSY;
-       if (retval)
-               goto out;
-
        if (dev->power.runtime_status == RPM_SUSPENDING) {
                DEFINE_WAIT(wait);
 
-               if (from_wq) {
+               if (rpmflags & (RPM_ASYNC | RPM_NOWAIT)) {
                        retval = -EINPROGRESS;
                        goto out;
                }
@@ -236,46 +355,42 @@ int __pm_runtime_suspend(struct device *dev, bool from_wq)
                goto repeat;
        }
 
-       __update_runtime_status(dev, RPM_SUSPENDING);
        dev->power.deferred_resume = false;
+       if (dev->power.no_callbacks)
+               goto no_callback;       /* Assume success. */
+
+       /* Carry out an asynchronous or a synchronous suspend. */
+       if (rpmflags & RPM_ASYNC) {
+               dev->power.request = (rpmflags & RPM_AUTO) ?
+                   RPM_REQ_AUTOSUSPEND : RPM_REQ_SUSPEND;
+               if (!dev->power.request_pending) {
+                       dev->power.request_pending = true;
+                       queue_work(pm_wq, &dev->power.work);
+               }
+               goto out;
+       }
 
-       if (dev->bus && dev->bus->pm && dev->bus->pm->runtime_suspend) {
-               spin_unlock_irq(&dev->power.lock);
-
-               retval = dev->bus->pm->runtime_suspend(dev);
-
-               spin_lock_irq(&dev->power.lock);
-               dev->power.runtime_error = retval;
-       } else if (dev->type && dev->type->pm
-           && dev->type->pm->runtime_suspend) {
-               spin_unlock_irq(&dev->power.lock);
-
-               retval = dev->type->pm->runtime_suspend(dev);
-
-               spin_lock_irq(&dev->power.lock);
-               dev->power.runtime_error = retval;
-       } else if (dev->class && dev->class->pm
-           && dev->class->pm->runtime_suspend) {
-               spin_unlock_irq(&dev->power.lock);
-
-               retval = dev->class->pm->runtime_suspend(dev);
+       __update_runtime_status(dev, RPM_SUSPENDING);
 
-               spin_lock_irq(&dev->power.lock);
-               dev->power.runtime_error = retval;
-       } else {
-               retval = -ENOSYS;
-       }
+       if (dev->bus && dev->bus->pm && dev->bus->pm->runtime_suspend)
+               callback = dev->bus->pm->runtime_suspend;
+       else if (dev->type && dev->type->pm && dev->type->pm->runtime_suspend)
+               callback = dev->type->pm->runtime_suspend;
+       else if (dev->class && dev->class->pm)
+               callback = dev->class->pm->runtime_suspend;
+       else
+               callback = NULL;
 
+       retval = rpm_callback(callback, dev);
        if (retval) {
                __update_runtime_status(dev, RPM_ACTIVE);
-               if (retval == -EAGAIN || retval == -EBUSY) {
-                       if (dev->power.timer_expires == 0)
-                               notify = true;
+               dev->power.deferred_resume = 0;
+               if (retval == -EAGAIN || retval == -EBUSY)
                        dev->power.runtime_error = 0;
-               } else {
+               else
                        pm_runtime_cancel_pending(dev);
-               }
        } else {
+ no_callback:
                __update_runtime_status(dev, RPM_SUSPENDED);
                pm_runtime_deactivate_timer(dev);
 
@@ -287,89 +402,86 @@ int __pm_runtime_suspend(struct device *dev, bool from_wq)
        wake_up_all(&dev->power.wait_queue);
 
        if (dev->power.deferred_resume) {
-               __pm_runtime_resume(dev, false);
+               rpm_resume(dev, 0);
                retval = -EAGAIN;
                goto out;
        }
 
-       if (notify)
-               __pm_runtime_idle(dev);
-
-       if (parent && !parent->power.ignore_children) {
-               spin_unlock_irq(&dev->power.lock);
+       /* Maybe the parent is now able to suspend. */
+       if (parent && !parent->power.ignore_children && !dev->power.irq_safe) {
+               spin_unlock(&dev->power.lock);
 
-               pm_request_idle(parent);
+               spin_lock(&parent->power.lock);
+               rpm_idle(parent, RPM_ASYNC);
+               spin_unlock(&parent->power.lock);
 
-               spin_lock_irq(&dev->power.lock);
+               spin_lock(&dev->power.lock);
        }
 
  out:
-       dev_dbg(dev, "__pm_runtime_suspend() returns %d!\n", retval);
+       dev_dbg(dev, "%s returns %d\n", __func__, retval);
 
        return retval;
 }
 
 /**
- * pm_runtime_suspend - Carry out run-time suspend of given device.
- * @dev: Device to suspend.
- */
-int pm_runtime_suspend(struct device *dev)
-{
-       int retval;
-
-       spin_lock_irq(&dev->power.lock);
-       retval = __pm_runtime_suspend(dev, false);
-       spin_unlock_irq(&dev->power.lock);
-
-       return retval;
-}
-EXPORT_SYMBOL_GPL(pm_runtime_suspend);
-
-/**
- * __pm_runtime_resume - Carry out run-time resume of given device.
+ * rpm_resume - Carry out run-time resume of given device.
  * @dev: Device to resume.
- * @from_wq: If set, the function has been called via pm_wq.
+ * @rpmflags: Flag bits.
  *
- * Check if the device can be woken up and run the ->runtime_resume() callback
- * provided by its bus type.  If another resume has been started earlier, wait
- * for it to finish.  If there's a suspend running in parallel with this
- * function, wait for it to finish and resume the device.  Cancel any scheduled
- * or pending requests.
+ * Check if the device's run-time PM status allows it to be resumed.  Cancel
+ * any scheduled or pending requests.  If another resume has been started
+ * earlier, either return imediately or wait for it to finish, depending on the
+ * RPM_NOWAIT and RPM_ASYNC flags.  Similarly, if there's a suspend running in
+ * parallel with this function, either tell the other process to resume after
+ * suspending (deferred_resume) or wait for it to finish.  If the RPM_ASYNC
+ * flag is set then queue a resume request; otherwise run the
+ * ->runtime_resume() callback directly.  Queue an idle notification for the
+ * device if the resume succeeded.
  *
  * This function must be called under dev->power.lock with interrupts disabled.
  */
-int __pm_runtime_resume(struct device *dev, bool from_wq)
+static int rpm_resume(struct device *dev, int rpmflags)
        __releases(&dev->power.lock) __acquires(&dev->power.lock)
 {
+       int (*callback)(struct device *);
        struct device *parent = NULL;
        int retval = 0;
 
-       dev_dbg(dev, "__pm_runtime_resume()%s!\n",
-               from_wq ? " from workqueue" : "");
+       dev_dbg(dev, "%s flags 0x%x\n", __func__, rpmflags);
 
  repeat:
-       if (dev->power.runtime_error) {
+       if (dev->power.runtime_error)
                retval = -EINVAL;
+       else if (dev->power.disable_depth > 0)
+               retval = -EAGAIN;
+       if (retval)
                goto out;
-       }
 
-       pm_runtime_cancel_pending(dev);
+       /*
+        * Other scheduled or pending requests need to be canceled.  Small
+        * optimization: If an autosuspend timer is running, leave it running
+        * rather than cancelling it now only to restart it again in the near
+        * future.
+        */
+       dev->power.request = RPM_REQ_NONE;
+       if (!dev->power.timer_autosuspends)
+               pm_runtime_deactivate_timer(dev);
 
-       if (dev->power.runtime_status == RPM_ACTIVE)
+       if (dev->power.runtime_status == RPM_ACTIVE) {
                retval = 1;
-       else if (dev->power.disable_depth > 0)
-               retval = -EAGAIN;
-       if (retval)
                goto out;
+       }
 
        if (dev->power.runtime_status == RPM_RESUMING
            || dev->power.runtime_status == RPM_SUSPENDING) {
                DEFINE_WAIT(wait);
 
-               if (from_wq) {
+               if (rpmflags & (RPM_ASYNC | RPM_NOWAIT)) {
                        if (dev->power.runtime_status == RPM_SUSPENDING)
                                dev->power.deferred_resume = true;
-                       retval = -EINPROGRESS;
+                       else
+                               retval = -EINPROGRESS;
                        goto out;
                }
 
@@ -391,12 +503,43 @@ int __pm_runtime_resume(struct device *dev, bool from_wq)
                goto repeat;
        }
 
+       /*
+        * See if we can skip waking up the parent.  This is safe only if
+        * power.no_callbacks is set, because otherwise we don't know whether
+        * the resume will actually succeed.
+        */
+       if (dev->power.no_callbacks && !parent && dev->parent) {
+               spin_lock_nested(&dev->parent->power.lock, SINGLE_DEPTH_NESTING);
+               if (dev->parent->power.disable_depth > 0
+                   || dev->parent->power.ignore_children
+                   || dev->parent->power.runtime_status == RPM_ACTIVE) {
+                       atomic_inc(&dev->parent->power.child_count);
+                       spin_unlock(&dev->parent->power.lock);
+                       goto no_callback;       /* Assume success. */
+               }
+               spin_unlock(&dev->parent->power.lock);
+       }
+
+       /* Carry out an asynchronous or a synchronous resume. */
+       if (rpmflags & RPM_ASYNC) {
+               dev->power.request = RPM_REQ_RESUME;
+               if (!dev->power.request_pending) {
+                       dev->power.request_pending = true;
+                       queue_work(pm_wq, &dev->power.work);
+               }
+               retval = 0;
+               goto out;
+       }
+
        if (!parent && dev->parent) {
                /*
-                * Increment the parent's resume counter and resume it if
-                * necessary.
+                * Increment the parent's usage counter and resume it if
+                * necessary.  Not needed if dev is irq-safe; then the
+                * parent is permanently resumed.
                 */
                parent = dev->parent;
+               if (dev->power.irq_safe)
+                       goto skip_parent;
                spin_unlock(&dev->power.lock);
 
                pm_runtime_get_noresume(parent);
@@ -408,7 +551,7 @@ int __pm_runtime_resume(struct device *dev, bool from_wq)
                 */
                if (!parent->power.disable_depth
                    && !parent->power.ignore_children) {
-                       __pm_runtime_resume(parent, false);
+                       rpm_resume(parent, 0);
                        if (parent->power.runtime_status != RPM_ACTIVE)
                                retval = -EBUSY;
                }
@@ -419,40 +562,28 @@ int __pm_runtime_resume(struct device *dev, bool from_wq)
                        goto out;
                goto repeat;
        }
+ skip_parent:
 
-       __update_runtime_status(dev, RPM_RESUMING);
-
-       if (dev->bus && dev->bus->pm && dev->bus->pm->runtime_resume) {
-               spin_unlock_irq(&dev->power.lock);
-
-               retval = dev->bus->pm->runtime_resume(dev);
-
-               spin_lock_irq(&dev->power.lock);
-               dev->power.runtime_error = retval;
-       } else if (dev->type && dev->type->pm
-           && dev->type->pm->runtime_resume) {
-               spin_unlock_irq(&dev->power.lock);
-
-               retval = dev->type->pm->runtime_resume(dev);
-
-               spin_lock_irq(&dev->power.lock);
-               dev->power.runtime_error = retval;
-       } else if (dev->class && dev->class->pm
-           && dev->class->pm->runtime_resume) {
-               spin_unlock_irq(&dev->power.lock);
+       if (dev->power.no_callbacks)
+               goto no_callback;       /* Assume success. */
 
-               retval = dev->class->pm->runtime_resume(dev);
+       __update_runtime_status(dev, RPM_RESUMING);
 
-               spin_lock_irq(&dev->power.lock);
-               dev->power.runtime_error = retval;
-       } else {
-               retval = -ENOSYS;
-       }
+       if (dev->bus && dev->bus->pm && dev->bus->pm->runtime_resume)
+               callback = dev->bus->pm->runtime_resume;
+       else if (dev->type && dev->type->pm && dev->type->pm->runtime_resume)
+               callback = dev->type->pm->runtime_resume;
+       else if (dev->class && dev->class->pm)
+               callback = dev->class->pm->runtime_resume;
+       else
+               callback = NULL;
 
+       retval = rpm_callback(callback, dev);
        if (retval) {
                __update_runtime_status(dev, RPM_SUSPENDED);
                pm_runtime_cancel_pending(dev);
        } else {
+ no_callback:
                __update_runtime_status(dev, RPM_ACTIVE);
                if (parent)
                        atomic_inc(&parent->power.child_count);
@@ -460,10 +591,10 @@ int __pm_runtime_resume(struct device *dev, bool from_wq)
        wake_up_all(&dev->power.wait_queue);
 
        if (!retval)
-               __pm_request_idle(dev);
+               rpm_idle(dev, RPM_ASYNC);
 
  out:
-       if (parent) {
+       if (parent && !dev->power.irq_safe) {
                spin_unlock_irq(&dev->power.lock);
 
                pm_runtime_put(parent);
@@ -471,27 +602,11 @@ int __pm_runtime_resume(struct device *dev, bool from_wq)
                spin_lock_irq(&dev->power.lock);
        }
 
-       dev_dbg(dev, "__pm_runtime_resume() returns %d!\n", retval);
+       dev_dbg(dev, "%s returns %d\n", __func__, retval);
 
        return retval;
 }
 
-/**
- * pm_runtime_resume - Carry out run-time resume of given device.
- * @dev: Device to suspend.
- */
-int pm_runtime_resume(struct device *dev)
-{
-       int retval;
-
-       spin_lock_irq(&dev->power.lock);
-       retval = __pm_runtime_resume(dev, false);
-       spin_unlock_irq(&dev->power.lock);
-
-       return retval;
-}
-EXPORT_SYMBOL_GPL(pm_runtime_resume);
-
 /**
  * pm_runtime_work - Universal run-time PM work function.
  * @work: Work structure used for scheduling the execution of this function.
@@ -517,13 +632,16 @@ static void pm_runtime_work(struct work_struct *work)
        case RPM_REQ_NONE:
                break;
        case RPM_REQ_IDLE:
-               __pm_runtime_idle(dev);
+               rpm_idle(dev, RPM_NOWAIT);
                break;
        case RPM_REQ_SUSPEND:
-               __pm_runtime_suspend(dev, true);
+               rpm_suspend(dev, RPM_NOWAIT);
+               break;
+       case RPM_REQ_AUTOSUSPEND:
+               rpm_suspend(dev, RPM_NOWAIT | RPM_AUTO);
                break;
        case RPM_REQ_RESUME:
-               __pm_runtime_resume(dev, true);
+               rpm_resume(dev, RPM_NOWAIT);
                break;
        }
 
@@ -531,118 +649,11 @@ static void pm_runtime_work(struct work_struct *work)
        spin_unlock_irq(&dev->power.lock);
 }
 
-/**
- * __pm_request_idle - Submit an idle notification request for given device.
- * @dev: Device to handle.
- *
- * Check if the device's run-time PM status is correct for suspending the device
- * and queue up a request to run __pm_runtime_idle() for it.
- *
- * This function must be called under dev->power.lock with interrupts disabled.
- */
-static int __pm_request_idle(struct device *dev)
-{
-       int retval = 0;
-
-       if (dev->power.runtime_error)
-               retval = -EINVAL;
-       else if (atomic_read(&dev->power.usage_count) > 0
-           || dev->power.disable_depth > 0
-           || dev->power.runtime_status == RPM_SUSPENDED
-           || dev->power.runtime_status == RPM_SUSPENDING)
-               retval = -EAGAIN;
-       else if (!pm_children_suspended(dev))
-               retval = -EBUSY;
-       if (retval)
-               return retval;
-
-       if (dev->power.request_pending) {
-               /* Any requests other then RPM_REQ_IDLE take precedence. */
-               if (dev->power.request == RPM_REQ_NONE)
-                       dev->power.request = RPM_REQ_IDLE;
-               else if (dev->power.request != RPM_REQ_IDLE)
-                       retval = -EAGAIN;
-               return retval;
-       }
-
-       dev->power.request = RPM_REQ_IDLE;
-       dev->power.request_pending = true;
-       queue_work(pm_wq, &dev->power.work);
-
-       return retval;
-}
-
-/**
- * pm_request_idle - Submit an idle notification request for given device.
- * @dev: Device to handle.
- */
-int pm_request_idle(struct device *dev)
-{
-       unsigned long flags;
-       int retval;
-
-       spin_lock_irqsave(&dev->power.lock, flags);
-       retval = __pm_request_idle(dev);
-       spin_unlock_irqrestore(&dev->power.lock, flags);
-
-       return retval;
-}
-EXPORT_SYMBOL_GPL(pm_request_idle);
-
-/**
- * __pm_request_suspend - Submit a suspend request for given device.
- * @dev: Device to suspend.
- *
- * This function must be called under dev->power.lock with interrupts disabled.
- */
-static int __pm_request_suspend(struct device *dev)
-{
-       int retval = 0;
-
-       if (dev->power.runtime_error)
-               return -EINVAL;
-
-       if (dev->power.runtime_status == RPM_SUSPENDED)
-               retval = 1;
-       else if (atomic_read(&dev->power.usage_count) > 0
-           || dev->power.disable_depth > 0)
-               retval = -EAGAIN;
-       else if (dev->power.runtime_status == RPM_SUSPENDING)
-               retval = -EINPROGRESS;
-       else if (!pm_children_suspended(dev))
-               retval = -EBUSY;
-       if (retval < 0)
-               return retval;
-
-       pm_runtime_deactivate_timer(dev);
-
-       if (dev->power.request_pending) {
-               /*
-                * Pending resume requests take precedence over us, but we can
-                * overtake any other pending request.
-                */
-               if (dev->power.request == RPM_REQ_RESUME)
-                       retval = -EAGAIN;
-               else if (dev->power.request != RPM_REQ_SUSPEND)
-                       dev->power.request = retval ?
-                                               RPM_REQ_NONE : RPM_REQ_SUSPEND;
-               return retval;
-       } else if (retval) {
-               return retval;
-       }
-
-       dev->power.request = RPM_REQ_SUSPEND;
-       dev->power.request_pending = true;
-       queue_work(pm_wq, &dev->power.work);
-
-       return 0;
-}
-
 /**
  * pm_suspend_timer_fn - Timer function for pm_schedule_suspend().
  * @data: Device pointer passed by pm_schedule_suspend().
  *
- * Check if the time is right and execute __pm_request_suspend() in that case.
+ * Check if the time is right and queue a suspend request.
  */
 static void pm_suspend_timer_fn(unsigned long data)
 {
@@ -656,7 +667,8 @@ static void pm_suspend_timer_fn(unsigned long data)
        /* If 'expire' is after 'jiffies' we've been called too early. */
        if (expires > 0 && !time_after(expires, jiffies)) {
                dev->power.timer_expires = 0;
-               __pm_request_suspend(dev);
+               rpm_suspend(dev, dev->power.timer_autosuspends ?
+                   (RPM_ASYNC | RPM_AUTO) : RPM_ASYNC);
        }
 
        spin_unlock_irqrestore(&dev->power.lock, flags);
@@ -670,47 +682,25 @@ static void pm_suspend_timer_fn(unsigned long data)
 int pm_schedule_suspend(struct device *dev, unsigned int delay)
 {
        unsigned long flags;
-       int retval = 0;
+       int retval;
 
        spin_lock_irqsave(&dev->power.lock, flags);
 
-       if (dev->power.runtime_error) {
-               retval = -EINVAL;
-               goto out;
-       }
-
        if (!delay) {
-               retval = __pm_request_suspend(dev);
+               retval = rpm_suspend(dev, RPM_ASYNC);
                goto out;
        }
 
-       pm_runtime_deactivate_timer(dev);
-
-       if (dev->power.request_pending) {
-               /*
-                * Pending resume requests take precedence over us, but any
-                * other pending requests have to be canceled.
-                */
-               if (dev->power.request == RPM_REQ_RESUME) {
-                       retval = -EAGAIN;
-                       goto out;
-               }
-               dev->power.request = RPM_REQ_NONE;
-       }
-
-       if (dev->power.runtime_status == RPM_SUSPENDED)
-               retval = 1;
-       else if (atomic_read(&dev->power.usage_count) > 0
-           || dev->power.disable_depth > 0)
-               retval = -EAGAIN;
-       else if (!pm_children_suspended(dev))
-               retval = -EBUSY;
+       retval = rpm_check_suspend_allowed(dev);
        if (retval)
                goto out;
 
+       /* Other scheduled or pending requests need to be canceled. */
+       pm_runtime_cancel_pending(dev);
+
        dev->power.timer_expires = jiffies + msecs_to_jiffies(delay);
-       if (!dev->power.timer_expires)
-               dev->power.timer_expires = 1;
+       dev->power.timer_expires += !dev->power.timer_expires;
+       dev->power.timer_autosuspends = 0;
        mod_timer(&dev->power.suspend_timer, dev->power.timer_expires);
 
  out:
@@ -721,103 +711,88 @@ int pm_schedule_suspend(struct device *dev, unsigned int delay)
 EXPORT_SYMBOL_GPL(pm_schedule_suspend);
 
 /**
- * pm_request_resume - Submit a resume request for given device.
- * @dev: Device to resume.
+ * __pm_runtime_idle - Entry point for run-time idle operations.
+ * @dev: Device to send idle notification for.
+ * @rpmflags: Flag bits.
  *
- * This function must be called under dev->power.lock with interrupts disabled.
+ * If the RPM_GET_PUT flag is set, decrement the device's usage count and
+ * return immediately if it is larger than zero.  Then carry out an idle
+ * notification, either synchronous or asynchronous.
+ *
+ * This routine may be called in atomic context if the RPM_ASYNC flag is set.
  */
-static int __pm_request_resume(struct device *dev)
+int __pm_runtime_idle(struct device *dev, int rpmflags)
 {
-       int retval = 0;
-
-       if (dev->power.runtime_error)
-               return -EINVAL;
-
-       if (dev->power.runtime_status == RPM_ACTIVE)
-               retval = 1;
-       else if (dev->power.runtime_status == RPM_RESUMING)
-               retval = -EINPROGRESS;
-       else if (dev->power.disable_depth > 0)
-               retval = -EAGAIN;
-       if (retval < 0)
-               return retval;
-
-       pm_runtime_deactivate_timer(dev);
+       unsigned long flags;
+       int retval;
 
-       if (dev->power.runtime_status == RPM_SUSPENDING) {
-               dev->power.deferred_resume = true;
-               return retval;
-       }
-       if (dev->power.request_pending) {
-               /* If non-resume request is pending, we can overtake it. */
-               dev->power.request = retval ? RPM_REQ_NONE : RPM_REQ_RESUME;
-               return retval;
+       if (rpmflags & RPM_GET_PUT) {
+               if (!atomic_dec_and_test(&dev->power.usage_count))
+                       return 0;
        }
-       if (retval)
-               return retval;
 
-       dev->power.request = RPM_REQ_RESUME;
-       dev->power.request_pending = true;
-       queue_work(pm_wq, &dev->power.work);
+       spin_lock_irqsave(&dev->power.lock, flags);
+       retval = rpm_idle(dev, rpmflags);
+       spin_unlock_irqrestore(&dev->power.lock, flags);
 
        return retval;
 }
+EXPORT_SYMBOL_GPL(__pm_runtime_idle);
 
 /**
- * pm_request_resume - Submit a resume request for given device.
- * @dev: Device to resume.
+ * __pm_runtime_suspend - Entry point for run-time put/suspend operations.
+ * @dev: Device to suspend.
+ * @rpmflags: Flag bits.
+ *
+ * If the RPM_GET_PUT flag is set, decrement the device's usage count and
+ * return immediately if it is larger than zero.  Then carry out a suspend,
+ * either synchronous or asynchronous.
+ *
+ * This routine may be called in atomic context if the RPM_ASYNC flag is set.
  */
-int pm_request_resume(struct device *dev)
+int __pm_runtime_suspend(struct device *dev, int rpmflags)
 {
        unsigned long flags;
        int retval;
 
+       if (rpmflags & RPM_GET_PUT) {
+               if (!atomic_dec_and_test(&dev->power.usage_count))
+                       return 0;
+       }
+
        spin_lock_irqsave(&dev->power.lock, flags);
-       retval = __pm_request_resume(dev);
+       retval = rpm_suspend(dev, rpmflags);
        spin_unlock_irqrestore(&dev->power.lock, flags);
 
        return retval;
 }
-EXPORT_SYMBOL_GPL(pm_request_resume);
+EXPORT_SYMBOL_GPL(__pm_runtime_suspend);
 
 /**
- * __pm_runtime_get - Reference count a device and wake it up, if necessary.
- * @dev: Device to handle.
- * @sync: If set and the device is suspended, resume it synchronously.
+ * __pm_runtime_resume - Entry point for run-time resume operations.
+ * @dev: Device to resume.
+ * @rpmflags: Flag bits.
+ *
+ * If the RPM_GET_PUT flag is set, increment the device's usage count.  Then
+ * carry out a resume, either synchronous or asynchronous.
  *
- * Increment the usage count of the device and resume it or submit a resume
- * request for it, depending on the value of @sync.
+ * This routine may be called in atomic context if the RPM_ASYNC flag is set.
  */
-int __pm_runtime_get(struct device *dev, bool sync)
+int __pm_runtime_resume(struct device *dev, int rpmflags)
 {
+       unsigned long flags;
        int retval;
 
-       atomic_inc(&dev->power.usage_count);
-       retval = sync ? pm_runtime_resume(dev) : pm_request_resume(dev);
-
-       return retval;
-}
-EXPORT_SYMBOL_GPL(__pm_runtime_get);
-
-/**
- * __pm_runtime_put - Decrement the device's usage counter and notify its bus.
- * @dev: Device to handle.
- * @sync: If the device's bus type is to be notified, do that synchronously.
- *
- * Decrement the usage count of the device and if it reaches zero, carry out a
- * synchronous idle notification or submit an idle notification request for it,
- * depending on the value of @sync.
- */
-int __pm_runtime_put(struct device *dev, bool sync)
-{
-       int retval = 0;
+       if (rpmflags & RPM_GET_PUT)
+               atomic_inc(&dev->power.usage_count);
 
-       if (atomic_dec_and_test(&dev->power.usage_count))
-               retval = sync ? pm_runtime_idle(dev) : pm_request_idle(dev);
+       spin_lock_irqsave(&dev->power.lock, flags);
+       retval = rpm_resume(dev, rpmflags);
+       spin_unlock_irqrestore(&dev->power.lock, flags);
 
        return retval;
 }
-EXPORT_SYMBOL_GPL(__pm_runtime_put);
+EXPORT_SYMBOL_GPL(__pm_runtime_resume);
 
 /**
  * __pm_runtime_set_status - Set run-time PM status of a device.
@@ -968,7 +943,7 @@ int pm_runtime_barrier(struct device *dev)
 
        if (dev->power.request_pending
            && dev->power.request == RPM_REQ_RESUME) {
-               __pm_runtime_resume(dev, false);
+               rpm_resume(dev, 0);
                retval = 1;
        }
 
@@ -1017,7 +992,7 @@ void __pm_runtime_disable(struct device *dev, bool check_resume)
                 */
                pm_runtime_get_noresume(dev);
 
-               __pm_runtime_resume(dev, false);
+               rpm_resume(dev, 0);
 
                pm_runtime_put_noidle(dev);
        }
@@ -1065,7 +1040,7 @@ void pm_runtime_forbid(struct device *dev)
 
        dev->power.runtime_auto = false;
        atomic_inc(&dev->power.usage_count);
-       __pm_runtime_resume(dev, false);
+       rpm_resume(dev, 0);
 
  out:
        spin_unlock_irq(&dev->power.lock);
@@ -1086,21 +1061,138 @@ void pm_runtime_allow(struct device *dev)
 
        dev->power.runtime_auto = true;
        if (atomic_dec_and_test(&dev->power.usage_count))
-               __pm_runtime_idle(dev);
+               rpm_idle(dev, RPM_AUTO);
 
  out:
        spin_unlock_irq(&dev->power.lock);
 }
 EXPORT_SYMBOL_GPL(pm_runtime_allow);
 
+/**
+ * pm_runtime_no_callbacks - Ignore run-time PM callbacks for a device.
+ * @dev: Device to handle.
+ *
+ * Set the power.no_callbacks flag, which tells the PM core that this
+ * device is power-managed through its parent and has no run-time PM
+ * callbacks of its own.  The run-time sysfs attributes will be removed.
+ */
+void pm_runtime_no_callbacks(struct device *dev)
+{
+       spin_lock_irq(&dev->power.lock);
+       dev->power.no_callbacks = 1;
+       spin_unlock_irq(&dev->power.lock);
+       if (device_is_registered(dev))
+               rpm_sysfs_remove(dev);
+}
+EXPORT_SYMBOL_GPL(pm_runtime_no_callbacks);
+
+/**
+ * pm_runtime_irq_safe - Leave interrupts disabled during callbacks.
+ * @dev: Device to handle
+ *
+ * Set the power.irq_safe flag, which tells the PM core that the
+ * ->runtime_suspend() and ->runtime_resume() callbacks for this device should
+ * always be invoked with the spinlock held and interrupts disabled.  It also
+ * causes the parent's usage counter to be permanently incremented, preventing
+ * the parent from runtime suspending -- otherwise an irq-safe child might have
+ * to wait for a non-irq-safe parent.
+ */
+void pm_runtime_irq_safe(struct device *dev)
+{
+       if (dev->parent)
+               pm_runtime_get_sync(dev->parent);
+       spin_lock_irq(&dev->power.lock);
+       dev->power.irq_safe = 1;
+       spin_unlock_irq(&dev->power.lock);
+}
+EXPORT_SYMBOL_GPL(pm_runtime_irq_safe);
+
+/**
+ * update_autosuspend - Handle a change to a device's autosuspend settings.
+ * @dev: Device to handle.
+ * @old_delay: The former autosuspend_delay value.
+ * @old_use: The former use_autosuspend value.
+ *
+ * Prevent runtime suspend if the new delay is negative and use_autosuspend is
+ * set; otherwise allow it.  Send an idle notification if suspends are allowed.
+ *
+ * This function must be called under dev->power.lock with interrupts disabled.
+ */
+static void update_autosuspend(struct device *dev, int old_delay, int old_use)
+{
+       int delay = dev->power.autosuspend_delay;
+
+       /* Should runtime suspend be prevented now? */
+       if (dev->power.use_autosuspend && delay < 0) {
+
+               /* If it used to be allowed then prevent it. */
+               if (!old_use || old_delay >= 0) {
+                       atomic_inc(&dev->power.usage_count);
+                       rpm_resume(dev, 0);
+               }
+       }
+
+       /* Runtime suspend should be allowed now. */
+       else {
+
+               /* If it used to be prevented then allow it. */
+               if (old_use && old_delay < 0)
+                       atomic_dec(&dev->power.usage_count);
+
+               /* Maybe we can autosuspend now. */
+               rpm_idle(dev, RPM_AUTO);
+       }
+}
+
+/**
+ * pm_runtime_set_autosuspend_delay - Set a device's autosuspend_delay value.
+ * @dev: Device to handle.
+ * @delay: Value of the new delay in milliseconds.
+ *
+ * Set the device's power.autosuspend_delay value.  If it changes to negative
+ * and the power.use_autosuspend flag is set, prevent run-time suspends.  If it
+ * changes the other way, allow run-time suspends.
+ */
+void pm_runtime_set_autosuspend_delay(struct device *dev, int delay)
+{
+       int old_delay, old_use;
+
+       spin_lock_irq(&dev->power.lock);
+       old_delay = dev->power.autosuspend_delay;
+       old_use = dev->power.use_autosuspend;
+       dev->power.autosuspend_delay = delay;
+       update_autosuspend(dev, old_delay, old_use);
+       spin_unlock_irq(&dev->power.lock);
+}
+EXPORT_SYMBOL_GPL(pm_runtime_set_autosuspend_delay);
+
+/**
+ * __pm_runtime_use_autosuspend - Set a device's use_autosuspend flag.
+ * @dev: Device to handle.
+ * @use: New value for use_autosuspend.
+ *
+ * Set the device's power.use_autosuspend flag, and allow or prevent run-time
+ * suspends as needed.
+ */
+void __pm_runtime_use_autosuspend(struct device *dev, bool use)
+{
+       int old_delay, old_use;
+
+       spin_lock_irq(&dev->power.lock);
+       old_delay = dev->power.autosuspend_delay;
+       old_use = dev->power.use_autosuspend;
+       dev->power.use_autosuspend = use;
+       update_autosuspend(dev, old_delay, old_use);
+       spin_unlock_irq(&dev->power.lock);
+}
+EXPORT_SYMBOL_GPL(__pm_runtime_use_autosuspend);
+
 /**
  * pm_runtime_init - Initialize run-time PM fields in given device object.
  * @dev: Device object to initialize.
  */
 void pm_runtime_init(struct device *dev)
 {
-       spin_lock_init(&dev->power.lock);
-
        dev->power.runtime_status = RPM_SUSPENDED;
        dev->power.idle_notification = false;
 
@@ -1137,4 +1229,6 @@ void pm_runtime_remove(struct device *dev)
        /* Change the status back to 'suspended' to match the initial status. */
        if (dev->power.runtime_status == RPM_ACTIVE)
                pm_runtime_set_suspended(dev);
+       if (dev->power.irq_safe && dev->parent)
+               pm_runtime_put_sync(dev->parent);
 }