Merge tag 'lsk-v4.4-16.06-android'
[firefly-linux-kernel-4.4.55.git] / drivers / cpuidle / cpuidle.c
index 5b11022d7745669a2288dfd778601a2041e8c314..151971627757b0ce968636cb3100375ac68852e5 100644 (file)
@@ -19,6 +19,8 @@
 #include <linux/ktime.h>
 #include <linux/hrtimer.h>
 #include <linux/module.h>
+#include <linux/suspend.h>
+#include <linux/tick.h>
 #include <trace/events/power.h>
 
 #include "cpuidle.h"
@@ -42,7 +44,11 @@ void disable_cpuidle(void)
        off = 1;
 }
 
-static int __cpuidle_register_device(struct cpuidle_device *dev);
+bool cpuidle_not_available(struct cpuidle_driver *drv,
+                          struct cpuidle_device *dev)
+{
+       return off || !initialized || !drv || !dev || !dev->enabled;
+}
 
 /**
  * cpuidle_play_dead - cpu off-lining
@@ -59,18 +65,106 @@ int cpuidle_play_dead(void)
                return -ENODEV;
 
        /* Find lowest-power state that supports long-term idle */
-       for (i = drv->state_count - 1; i >= CPUIDLE_DRIVER_STATE_START; i--)
+       for (i = drv->state_count - 1; i >= 0; i--)
                if (drv->states[i].enter_dead)
                        return drv->states[i].enter_dead(dev, i);
 
        return -ENODEV;
 }
 
+static int find_deepest_state(struct cpuidle_driver *drv,
+                             struct cpuidle_device *dev,
+                             unsigned int max_latency,
+                             unsigned int forbidden_flags,
+                             bool freeze)
+{
+       unsigned int latency_req = 0;
+       int i, ret = -ENXIO;
+
+       for (i = 0; i < drv->state_count; i++) {
+               struct cpuidle_state *s = &drv->states[i];
+               struct cpuidle_state_usage *su = &dev->states_usage[i];
+
+               if (s->disabled || su->disable || s->exit_latency <= latency_req
+                   || s->exit_latency > max_latency
+                   || (s->flags & forbidden_flags)
+                   || (freeze && !s->enter_freeze))
+                       continue;
+
+               latency_req = s->exit_latency;
+               ret = i;
+       }
+       return ret;
+}
+
+#ifdef CONFIG_SUSPEND
+/**
+ * cpuidle_find_deepest_state - Find the deepest available idle state.
+ * @drv: cpuidle driver for the given CPU.
+ * @dev: cpuidle device for the given CPU.
+ */
+int cpuidle_find_deepest_state(struct cpuidle_driver *drv,
+                              struct cpuidle_device *dev)
+{
+       return find_deepest_state(drv, dev, UINT_MAX, 0, false);
+}
+
+static void enter_freeze_proper(struct cpuidle_driver *drv,
+                               struct cpuidle_device *dev, int index)
+{
+       /*
+        * trace_suspend_resume() called by tick_freeze() for the last CPU
+        * executing it contains RCU usage regarded as invalid in the idle
+        * context, so tell RCU about that.
+        */
+       RCU_NONIDLE(tick_freeze());
+       /*
+        * The state used here cannot be a "coupled" one, because the "coupled"
+        * cpuidle mechanism enables interrupts and doing that with timekeeping
+        * suspended is generally unsafe.
+        */
+       stop_critical_timings();
+       drv->states[index].enter_freeze(dev, drv, index);
+       WARN_ON(!irqs_disabled());
+       /*
+        * timekeeping_resume() that will be called by tick_unfreeze() for the
+        * first CPU executing it calls functions containing RCU read-side
+        * critical sections, so tell RCU about that.
+        */
+       RCU_NONIDLE(tick_unfreeze());
+       start_critical_timings();
+}
+
+/**
+ * cpuidle_enter_freeze - Enter an idle state suitable for suspend-to-idle.
+ * @drv: cpuidle driver for the given CPU.
+ * @dev: cpuidle device for the given CPU.
+ *
+ * If there are states with the ->enter_freeze callback, find the deepest of
+ * them and enter it with frozen tick.
+ */
+int cpuidle_enter_freeze(struct cpuidle_driver *drv, struct cpuidle_device *dev)
+{
+       int index;
+
+       /*
+        * Find the deepest state with ->enter_freeze present, which guarantees
+        * that interrupts won't be enabled when it exits and allows the tick to
+        * be frozen safely.
+        */
+       index = find_deepest_state(drv, dev, UINT_MAX, 0, true);
+       if (index >= 0)
+               enter_freeze_proper(drv, dev, index);
+
+       return index;
+}
+#endif /* CONFIG_SUSPEND */
+
 /**
  * cpuidle_enter_state - enter the state and update stats
  * @dev: cpuidle device for this cpu
  * @drv: cpuidle driver for this cpu
- * @next_state: index into drv->states of the state to enter
+ * @index: index into the states table in @drv of the state to enter
  */
 int cpuidle_enter_state(struct cpuidle_device *dev, struct cpuidle_driver *drv,
                        int index)
@@ -78,16 +172,50 @@ int cpuidle_enter_state(struct cpuidle_device *dev, struct cpuidle_driver *drv,
        int entered_state;
 
        struct cpuidle_state *target_state = &drv->states[index];
+       bool broadcast = !!(target_state->flags & CPUIDLE_FLAG_TIMER_STOP);
        ktime_t time_start, time_end;
        s64 diff;
 
+       /*
+        * Tell the time framework to switch to a broadcast timer because our
+        * local timer will be shut down.  If a local timer is used from another
+        * CPU as a broadcast timer, this call may fail if it is not available.
+        */
+       if (broadcast && tick_broadcast_enter()) {
+               index = find_deepest_state(drv, dev, target_state->exit_latency,
+                                          CPUIDLE_FLAG_TIMER_STOP, false);
+               if (index < 0) {
+                       default_idle_call();
+                       return -EBUSY;
+               }
+               target_state = &drv->states[index];
+       }
+
+       /* Take note of the planned idle state. */
+       sched_idle_set_state(target_state, index);
+
+       trace_cpu_idle_rcuidle(index, dev->cpu);
        time_start = ktime_get();
 
+       stop_critical_timings();
        entered_state = target_state->enter(dev, drv, index);
+       start_critical_timings();
 
        time_end = ktime_get();
+       trace_cpu_idle_rcuidle(PWR_EVENT_EXIT, dev->cpu);
+
+       /* The cpu is no longer idle or about to enter idle. */
+       sched_idle_set_state(NULL, -1);
+
+       if (broadcast) {
+               if (WARN_ON_ONCE(!irqs_disabled()))
+                       local_irq_disable();
+
+               tick_broadcast_exit();
+       }
 
-       local_irq_enable();
+       if (!cpuidle_state_is_coupled(drv, index))
+               local_irq_enable();
 
        diff = ktime_to_us(ktime_sub(time_end, time_start));
        if (diff > INT_MAX)
@@ -110,66 +238,48 @@ int cpuidle_enter_state(struct cpuidle_device *dev, struct cpuidle_driver *drv,
 }
 
 /**
- * cpuidle_idle_call - the main idle loop
+ * cpuidle_select - ask the cpuidle framework to choose an idle state
+ *
+ * @drv: the cpuidle driver
+ * @dev: the cpuidle device
  *
- * NOTE: no locks or semaphores should be used here
- * return non-zero on failure
+ * Returns the index of the idle state.
  */
-int cpuidle_idle_call(void)
+int cpuidle_select(struct cpuidle_driver *drv, struct cpuidle_device *dev)
 {
-       struct cpuidle_device *dev = __this_cpu_read(cpuidle_devices);
-       struct cpuidle_driver *drv;
-       int next_state, entered_state;
-
-       if (off)
-               return -ENODEV;
-
-       if (!initialized)
-               return -ENODEV;
-
-       /* check if the device is ready */
-       if (!dev || !dev->enabled)
-               return -EBUSY;
-
-       drv = cpuidle_get_cpu_driver(dev);
-
-       /* ask the governor for the next state */
-       next_state = cpuidle_curr_governor->select(drv, dev);
-       if (next_state < 0)
-               return -EBUSY;
-
-       if (need_resched()) {
-               dev->last_residency = 0;
-               /* give the governor an opportunity to reflect on the outcome */
-               if (cpuidle_curr_governor->reflect)
-                       cpuidle_curr_governor->reflect(dev, next_state);
-               local_irq_enable();
-               return 0;
-       }
-
-       trace_cpu_idle_rcuidle(next_state, dev->cpu);
-
-       if (drv->states[next_state].flags & CPUIDLE_FLAG_TIMER_STOP)
-               clockevents_notify(CLOCK_EVT_NOTIFY_BROADCAST_ENTER,
-                                  &dev->cpu);
-
-       if (cpuidle_state_is_coupled(dev, drv, next_state))
-               entered_state = cpuidle_enter_state_coupled(dev, drv,
-                                                           next_state);
-       else
-               entered_state = cpuidle_enter_state(dev, drv, next_state);
-
-       if (drv->states[next_state].flags & CPUIDLE_FLAG_TIMER_STOP)
-               clockevents_notify(CLOCK_EVT_NOTIFY_BROADCAST_EXIT,
-                                  &dev->cpu);
-
-       trace_cpu_idle_rcuidle(PWR_EVENT_EXIT, dev->cpu);
+       return cpuidle_curr_governor->select(drv, dev);
+}
 
-       /* give the governor an opportunity to reflect on the outcome */
-       if (cpuidle_curr_governor->reflect)
-               cpuidle_curr_governor->reflect(dev, entered_state);
+/**
+ * cpuidle_enter - enter into the specified idle state
+ *
+ * @drv:   the cpuidle driver tied with the cpu
+ * @dev:   the cpuidle device
+ * @index: the index in the idle state table
+ *
+ * Returns the index in the idle state, < 0 in case of error.
+ * The error code depends on the backend driver
+ */
+int cpuidle_enter(struct cpuidle_driver *drv, struct cpuidle_device *dev,
+                 int index)
+{
+       if (cpuidle_state_is_coupled(drv, index))
+               return cpuidle_enter_state_coupled(dev, drv, index);
+       return cpuidle_enter_state(dev, drv, index);
+}
 
-       return 0;
+/**
+ * cpuidle_reflect - tell the underlying governor what was the state
+ * we were in
+ *
+ * @dev  : the cpuidle device
+ * @index: the index in the idle state table
+ *
+ */
+void cpuidle_reflect(struct cpuidle_device *dev, int index)
+{
+       if (cpuidle_curr_governor->reflect && index >= 0)
+               cpuidle_curr_governor->reflect(dev, index);
 }
 
 /**
@@ -191,8 +301,14 @@ void cpuidle_uninstall_idle_handler(void)
 {
        if (enabled_devices) {
                initialized = 0;
-               kick_all_cpus_sync();
+               wake_up_all_idle_cpus();
        }
+
+       /*
+        * Make sure external observers (such as the scheduler)
+        * are done looking at pointed idle states.
+        */
+       synchronize_rcu();
 }
 
 /**
@@ -233,45 +349,6 @@ void cpuidle_resume(void)
        mutex_unlock(&cpuidle_lock);
 }
 
-#ifdef CONFIG_ARCH_HAS_CPU_RELAX
-static int poll_idle(struct cpuidle_device *dev,
-               struct cpuidle_driver *drv, int index)
-{
-       ktime_t t1, t2;
-       s64 diff;
-
-       t1 = ktime_get();
-       local_irq_enable();
-       while (!need_resched())
-               cpu_relax();
-
-       t2 = ktime_get();
-       diff = ktime_to_us(ktime_sub(t2, t1));
-       if (diff > INT_MAX)
-               diff = INT_MAX;
-
-       dev->last_residency = (int) diff;
-
-       return index;
-}
-
-static void poll_idle_init(struct cpuidle_driver *drv)
-{
-       struct cpuidle_state *state = &drv->states[0];
-
-       snprintf(state->name, CPUIDLE_NAME_LEN, "POLL");
-       snprintf(state->desc, CPUIDLE_DESC_LEN, "CPUIDLE CORE POLL IDLE");
-       state->exit_latency = 0;
-       state->target_residency = 0;
-       state->power_usage = -1;
-       state->flags = 0;
-       state->enter = poll_idle;
-       state->disabled = false;
-}
-#else
-static void poll_idle_init(struct cpuidle_driver *drv) {}
-#endif /* CONFIG_ARCH_HAS_CPU_RELAX */
-
 /**
  * cpuidle_enable_device - enables idle PM for a CPU
  * @dev: the CPU
@@ -281,7 +358,7 @@ static void poll_idle_init(struct cpuidle_driver *drv) {}
  */
 int cpuidle_enable_device(struct cpuidle_device *dev)
 {
-       int ret, i;
+       int ret;
        struct cpuidle_driver *drv;
 
        if (!dev)
@@ -295,16 +372,8 @@ int cpuidle_enable_device(struct cpuidle_device *dev)
        if (!drv || !cpuidle_curr_governor)
                return -EIO;
 
-       if (!dev->state_count)
-               dev->state_count = drv->state_count;
-
-       if (dev->registered == 0) {
-               ret = __cpuidle_register_device(dev);
-               if (ret)
-                       return ret;
-       }
-
-       poll_idle_init(drv);
+       if (!dev->registered)
+               return -EINVAL;
 
        ret = cpuidle_add_device_sysfs(dev);
        if (ret)
@@ -314,12 +383,6 @@ int cpuidle_enable_device(struct cpuidle_device *dev)
            (ret = cpuidle_curr_governor->enable(drv, dev)))
                goto fail_sysfs;
 
-       for (i = 0; i < dev->state_count; i++) {
-               dev->states_usage[i].usage = 0;
-               dev->states_usage[i].time = 0;
-       }
-       dev->last_residency = 0;
-
        smp_wmb();
 
        dev->enabled = 1;
@@ -363,6 +426,23 @@ void cpuidle_disable_device(struct cpuidle_device *dev)
 
 EXPORT_SYMBOL_GPL(cpuidle_disable_device);
 
+static void __cpuidle_unregister_device(struct cpuidle_device *dev)
+{
+       struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);
+
+       list_del(&dev->device_list);
+       per_cpu(cpuidle_devices, dev->cpu) = NULL;
+       module_put(drv->owner);
+
+       dev->registered = 0;
+}
+
+static void __cpuidle_device_init(struct cpuidle_device *dev)
+{
+       memset(dev->states_usage, 0, sizeof(dev->states_usage));
+       dev->last_residency = 0;
+}
+
 /**
  * __cpuidle_register_device - internal register function called before register
  * and enable routines
@@ -380,23 +460,13 @@ static int __cpuidle_register_device(struct cpuidle_device *dev)
 
        per_cpu(cpuidle_devices, dev->cpu) = dev;
        list_add(&dev->device_list, &cpuidle_detected_devices);
-       ret = cpuidle_add_sysfs(dev);
-       if (ret)
-               goto err_sysfs;
 
        ret = cpuidle_coupled_register_device(dev);
        if (ret)
-               goto err_coupled;
-
-       dev->registered = 1;
-       return 0;
+               __cpuidle_unregister_device(dev);
+       else
+               dev->registered = 1;
 
-err_coupled:
-       cpuidle_remove_sysfs(dev);
-err_sysfs:
-       list_del(&dev->device_list);
-       per_cpu(cpuidle_devices, dev->cpu) = NULL;
-       module_put(drv->owner);
        return ret;
 }
 
@@ -406,25 +476,42 @@ err_sysfs:
  */
 int cpuidle_register_device(struct cpuidle_device *dev)
 {
-       int ret;
+       int ret = -EBUSY;
 
        if (!dev)
                return -EINVAL;
 
        mutex_lock(&cpuidle_lock);
 
-       if ((ret = __cpuidle_register_device(dev))) {
-               mutex_unlock(&cpuidle_lock);
-               return ret;
-       }
+       if (dev->registered)
+               goto out_unlock;
+
+       __cpuidle_device_init(dev);
+
+       ret = __cpuidle_register_device(dev);
+       if (ret)
+               goto out_unlock;
+
+       ret = cpuidle_add_sysfs(dev);
+       if (ret)
+               goto out_unregister;
+
+       ret = cpuidle_enable_device(dev);
+       if (ret)
+               goto out_sysfs;
 
-       cpuidle_enable_device(dev);
        cpuidle_install_idle_handler();
 
+out_unlock:
        mutex_unlock(&cpuidle_lock);
 
-       return 0;
+       return ret;
 
+out_sysfs:
+       cpuidle_remove_sysfs(dev);
+out_unregister:
+       __cpuidle_unregister_device(dev);
+       goto out_unlock;
 }
 
 EXPORT_SYMBOL_GPL(cpuidle_register_device);
@@ -435,9 +522,7 @@ EXPORT_SYMBOL_GPL(cpuidle_register_device);
  */
 void cpuidle_unregister_device(struct cpuidle_device *dev)
 {
-       struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);
-
-       if (dev->registered == 0)
+       if (!dev || dev->registered == 0)
                return;
 
        cpuidle_pause_and_lock();
@@ -445,14 +530,12 @@ void cpuidle_unregister_device(struct cpuidle_device *dev)
        cpuidle_disable_device(dev);
 
        cpuidle_remove_sysfs(dev);
-       list_del(&dev->device_list);
-       per_cpu(cpuidle_devices, dev->cpu) = NULL;
+
+       __cpuidle_unregister_device(dev);
 
        cpuidle_coupled_unregister_device(dev);
 
        cpuidle_resume_and_unlock();
-
-       module_put(drv->owner);
 }
 
 EXPORT_SYMBOL_GPL(cpuidle_unregister_device);
@@ -469,7 +552,7 @@ void cpuidle_unregister(struct cpuidle_driver *drv)
        int cpu;
        struct cpuidle_device *device;
 
-       for_each_possible_cpu(cpu) {
+       for_each_cpu(cpu, drv->cpumask) {
                device = &per_cpu(cpuidle_dev, cpu);
                cpuidle_unregister_device(device);
        }
@@ -501,13 +584,13 @@ int cpuidle_register(struct cpuidle_driver *drv,
                return ret;
        }
 
-       for_each_possible_cpu(cpu) {
+       for_each_cpu(cpu, drv->cpumask) {
                device = &per_cpu(cpuidle_dev, cpu);
                device->cpu = cpu;
 
 #ifdef CONFIG_ARCH_NEEDS_CPU_IDLE_COUPLED
                /*
-                * On multiplatform for ARM, the coupled idle states could
+                * On multiplatform for ARM, the coupled idle states could be
                 * enabled in the kernel even if the cpuidle driver does not
                 * use it. Note, coupled_cpus is a struct copy.
                 */
@@ -530,11 +613,6 @@ EXPORT_SYMBOL_GPL(cpuidle_register);
 
 #ifdef CONFIG_SMP
 
-static void smp_callback(void *v)
-{
-       /* we already woke the CPU up, nothing more to do */
-}
-
 /*
  * This function gets called when a part of the kernel has a new latency
  * requirement.  This means we need to get all processors out of their C-state,
@@ -544,7 +622,7 @@ static void smp_callback(void *v)
 static int cpuidle_latency_notify(struct notifier_block *b,
                unsigned long l, void *v)
 {
-       smp_call_function(smp_callback, NULL, 1);
+       wake_up_all_idle_cpus();
        return NOTIFY_OK;
 }