Merge branch 'timers-nohz-for-linus' of git://git.kernel.org/pub/scm/linux/kernel...
[linux-3.10.git] / kernel / time / tick-broadcast.c
index 3044a88..206bbfb 100644 (file)
@@ -18,7 +18,7 @@
 #include <linux/percpu.h>
 #include <linux/profile.h>
 #include <linux/sched.h>
-#include <linux/tick.h>
+#include <linux/smp.h>
 
 #include "tick-internal.h"
 
  * timer stops in C3 state.
  */
 
-struct tick_device tick_broadcast_device;
-static cpumask_t tick_broadcast_mask;
-static DEFINE_SPINLOCK(tick_broadcast_lock);
+static struct tick_device tick_broadcast_device;
+static cpumask_var_t tick_broadcast_mask;
+static cpumask_var_t tmpmask;
+static DEFINE_RAW_SPINLOCK(tick_broadcast_lock);
 static int tick_broadcast_force;
 
 #ifdef CONFIG_TICK_ONESHOT
@@ -46,9 +47,9 @@ struct tick_device *tick_get_broadcast_device(void)
        return &tick_broadcast_device;
 }
 
-cpumask_t *tick_get_broadcast_mask(void)
+struct cpumask *tick_get_broadcast_mask(void)
 {
-       return &tick_broadcast_mask;
+       return tick_broadcast_mask;
 }
 
 /*
@@ -65,15 +66,30 @@ static void tick_broadcast_start_periodic(struct clock_event_device *bc)
  */
 int tick_check_broadcast_device(struct clock_event_device *dev)
 {
-       if ((tick_broadcast_device.evtdev &&
+       struct clock_event_device *cur = tick_broadcast_device.evtdev;
+
+       if ((dev->features & CLOCK_EVT_FEAT_DUMMY) ||
+           (tick_broadcast_device.evtdev &&
             tick_broadcast_device.evtdev->rating >= dev->rating) ||
             (dev->features & CLOCK_EVT_FEAT_C3STOP))
                return 0;
 
-       clockevents_exchange_device(NULL, dev);
+       clockevents_exchange_device(tick_broadcast_device.evtdev, dev);
+       if (cur)
+               cur->event_handler = clockevents_handle_noop;
        tick_broadcast_device.evtdev = dev;
-       if (!cpus_empty(tick_broadcast_mask))
+       if (!cpumask_empty(tick_broadcast_mask))
                tick_broadcast_start_periodic(dev);
+       /*
+        * Inform all cpus about this. We might be in a situation
+        * where we did not switch to oneshot mode because the per cpu
+        * devices are affected by CLOCK_EVT_FEAT_C3STOP and the lack
+        * of a oneshot capable broadcast device. Without that
+        * notification the systems stays stuck in periodic mode
+        * forever.
+        */
+       if (dev->features & CLOCK_EVT_FEAT_ONESHOT)
+               tick_clock_notify();
        return 1;
 }
 
@@ -85,6 +101,22 @@ int tick_is_broadcast_device(struct clock_event_device *dev)
        return (dev && tick_broadcast_device.evtdev == dev);
 }
 
+static void err_broadcast(const struct cpumask *mask)
+{
+       pr_crit_once("Failed to broadcast timer tick. Some CPUs may be unresponsive.\n");
+}
+
+static void tick_device_setup_broadcast_func(struct clock_event_device *dev)
+{
+       if (!dev->broadcast)
+               dev->broadcast = tick_broadcast;
+       if (!dev->broadcast) {
+               pr_warn_once("%s depends on broadcast, but no broadcast function available\n",
+                            dev->name);
+               dev->broadcast = err_broadcast;
+       }
+}
+
 /*
  * Check, if the device is disfunctional and a place holder, which
  * needs to be handled by the broadcast device.
@@ -94,7 +126,7 @@ int tick_device_uses_broadcast(struct clock_event_device *dev, int cpu)
        unsigned long flags;
        int ret = 0;
 
-       spin_lock_irqsave(&tick_broadcast_lock, flags);
+       raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 
        /*
         * Devices might be registered with both periodic and oneshot
@@ -104,7 +136,8 @@ int tick_device_uses_broadcast(struct clock_event_device *dev, int cpu)
         */
        if (!tick_device_is_functional(dev)) {
                dev->event_handler = tick_handle_periodic;
-               cpu_set(cpu, tick_broadcast_mask);
+               tick_device_setup_broadcast_func(dev);
+               cpumask_set_cpu(cpu, tick_broadcast_mask);
                tick_broadcast_start_periodic(tick_broadcast_device.evtdev);
                ret = 1;
        } else {
@@ -115,19 +148,37 @@ int tick_device_uses_broadcast(struct clock_event_device *dev, int cpu)
                 */
                if (!(dev->features & CLOCK_EVT_FEAT_C3STOP)) {
                        int cpu = smp_processor_id();
-
-                       cpu_clear(cpu, tick_broadcast_mask);
+                       cpumask_clear_cpu(cpu, tick_broadcast_mask);
                        tick_broadcast_clear_oneshot(cpu);
+               } else {
+                       tick_device_setup_broadcast_func(dev);
                }
        }
-       spin_unlock_irqrestore(&tick_broadcast_lock, flags);
+       raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
        return ret;
 }
 
+#ifdef CONFIG_GENERIC_CLOCKEVENTS_BROADCAST
+int tick_receive_broadcast(void)
+{
+       struct tick_device *td = this_cpu_ptr(&tick_cpu_device);
+       struct clock_event_device *evt = td->evtdev;
+
+       if (!evt)
+               return -ENODEV;
+
+       if (!evt->event_handler)
+               return -EINVAL;
+
+       evt->event_handler(evt);
+       return 0;
+}
+#endif
+
 /*
- * Broadcast the event to the cpus, which are set in the mask
+ * Broadcast the event to the cpus, which are set in the mask (mangled).
  */
-static void tick_do_broadcast(cpumask_t mask)
+static void tick_do_broadcast(struct cpumask *mask)
 {
        int cpu = smp_processor_id();
        struct tick_device *td;
@@ -135,21 +186,20 @@ static void tick_do_broadcast(cpumask_t mask)
        /*
         * Check, if the current cpu is in the mask
         */
-       if (cpu_isset(cpu, mask)) {
-               cpu_clear(cpu, mask);
+       if (cpumask_test_cpu(cpu, mask)) {
+               cpumask_clear_cpu(cpu, mask);
                td = &per_cpu(tick_cpu_device, cpu);
                td->evtdev->event_handler(td->evtdev);
        }
 
-       if (!cpus_empty(mask)) {
+       if (!cpumask_empty(mask)) {
                /*
                 * It might be necessary to actually check whether the devices
                 * have different broadcast functions. For now, just use the
                 * one of the first device. This works as long as we have this
                 * misfeature only on x86 (lapic)
                 */
-               cpu = first_cpu(mask);
-               td = &per_cpu(tick_cpu_device, cpu);
+               td = &per_cpu(tick_cpu_device, cpumask_first(mask));
                td->evtdev->broadcast(mask);
        }
 }
@@ -160,14 +210,12 @@ static void tick_do_broadcast(cpumask_t mask)
  */
 static void tick_do_periodic_broadcast(void)
 {
-       cpumask_t mask;
+       raw_spin_lock(&tick_broadcast_lock);
 
-       spin_lock(&tick_broadcast_lock);
+       cpumask_and(tmpmask, cpu_online_mask, tick_broadcast_mask);
+       tick_do_broadcast(tmpmask);
 
-       cpus_and(mask, cpu_online_map, tick_broadcast_mask);
-       tick_do_broadcast(mask);
-
-       spin_unlock(&tick_broadcast_lock);
+       raw_spin_unlock(&tick_broadcast_lock);
 }
 
 /*
@@ -188,14 +236,14 @@ static void tick_handle_periodic_broadcast(struct clock_event_device *dev)
        /*
         * Setup the next period for devices, which do not have
         * periodic mode. We read dev->next_event first and add to it
-        * when the event alrady expired. clockevents_program_event()
+        * when the event already expired. clockevents_program_event()
         * sets dev->next_event only when the event is really
         * programmed to the device.
         */
        for (next = dev->next_event; ;) {
                next = ktime_add(next, tick_period);
 
-               if (!clockevents_program_event(dev, next, ktime_get()))
+               if (!clockevents_program_event(dev, next, false))
                        return;
                tick_do_periodic_broadcast();
        }
@@ -205,14 +253,14 @@ static void tick_handle_periodic_broadcast(struct clock_event_device *dev)
  * Powerstate information: The system enters/leaves a state, where
  * affected devices might stop
  */
-static void tick_do_broadcast_on_off(void *why)
+static void tick_do_broadcast_on_off(unsigned long *reason)
 {
        struct clock_event_device *bc, *dev;
        struct tick_device *td;
-       unsigned long flags, *reason = why;
-       int cpu;
+       unsigned long flags;
+       int cpu, bc_stopped;
 
-       spin_lock_irqsave(&tick_broadcast_lock, flags);
+       raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 
        cpu = smp_processor_id();
        td = &per_cpu(tick_cpu_device, cpu);
@@ -228,38 +276,40 @@ static void tick_do_broadcast_on_off(void *why)
        if (!tick_device_is_functional(dev))
                goto out;
 
+       bc_stopped = cpumask_empty(tick_broadcast_mask);
+
        switch (*reason) {
        case CLOCK_EVT_NOTIFY_BROADCAST_ON:
        case CLOCK_EVT_NOTIFY_BROADCAST_FORCE:
-               if (!cpu_isset(cpu, tick_broadcast_mask)) {
-                       cpu_set(cpu, tick_broadcast_mask);
-                       if (td->mode == TICKDEV_MODE_PERIODIC)
-                               clockevents_set_mode(dev,
-                                                    CLOCK_EVT_MODE_SHUTDOWN);
+               if (!cpumask_test_and_set_cpu(cpu, tick_broadcast_mask)) {
+                       if (tick_broadcast_device.mode ==
+                           TICKDEV_MODE_PERIODIC)
+                               clockevents_shutdown(dev);
                }
                if (*reason == CLOCK_EVT_NOTIFY_BROADCAST_FORCE)
                        tick_broadcast_force = 1;
                break;
        case CLOCK_EVT_NOTIFY_BROADCAST_OFF:
                if (!tick_broadcast_force &&
-                   cpu_isset(cpu, tick_broadcast_mask)) {
-                       cpu_clear(cpu, tick_broadcast_mask);
-                       if (td->mode == TICKDEV_MODE_PERIODIC)
+                   cpumask_test_and_clear_cpu(cpu, tick_broadcast_mask)) {
+                       if (tick_broadcast_device.mode ==
+                           TICKDEV_MODE_PERIODIC)
                                tick_setup_periodic(dev, 0);
                }
                break;
        }
 
-       if (cpus_empty(tick_broadcast_mask))
-               clockevents_set_mode(bc, CLOCK_EVT_MODE_SHUTDOWN);
-       else {
+       if (cpumask_empty(tick_broadcast_mask)) {
+               if (!bc_stopped)
+                       clockevents_shutdown(bc);
+       } else if (bc_stopped) {
                if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
                        tick_broadcast_start_periodic(bc);
                else
                        tick_broadcast_setup_oneshot(bc);
        }
 out:
-       spin_unlock_irqrestore(&tick_broadcast_lock, flags);
+       raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 }
 
 /*
@@ -268,12 +318,11 @@ out:
  */
 void tick_broadcast_on_off(unsigned long reason, int *oncpu)
 {
-       if (!cpu_isset(*oncpu, cpu_online_map))
+       if (!cpumask_test_cpu(*oncpu, cpu_online_mask))
                printk(KERN_ERR "tick-broadcast: ignoring broadcast for "
                       "offline CPU #%d\n", *oncpu);
        else
-               smp_call_function_single(*oncpu, tick_do_broadcast_on_off,
-                                        &reason, 1);
+               tick_do_broadcast_on_off(&reason);
 }
 
 /*
@@ -296,17 +345,17 @@ void tick_shutdown_broadcast(unsigned int *cpup)
        unsigned long flags;
        unsigned int cpu = *cpup;
 
-       spin_lock_irqsave(&tick_broadcast_lock, flags);
+       raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 
        bc = tick_broadcast_device.evtdev;
-       cpu_clear(cpu, tick_broadcast_mask);
+       cpumask_clear_cpu(cpu, tick_broadcast_mask);
 
        if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC) {
-               if (bc && cpus_empty(tick_broadcast_mask))
-                       clockevents_set_mode(bc, CLOCK_EVT_MODE_SHUTDOWN);
+               if (bc && cpumask_empty(tick_broadcast_mask))
+                       clockevents_shutdown(bc);
        }
 
-       spin_unlock_irqrestore(&tick_broadcast_lock, flags);
+       raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 }
 
 void tick_suspend_broadcast(void)
@@ -314,13 +363,13 @@ void tick_suspend_broadcast(void)
        struct clock_event_device *bc;
        unsigned long flags;
 
-       spin_lock_irqsave(&tick_broadcast_lock, flags);
+       raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 
        bc = tick_broadcast_device.evtdev;
        if (bc)
-               clockevents_set_mode(bc, CLOCK_EVT_MODE_SHUTDOWN);
+               clockevents_shutdown(bc);
 
-       spin_unlock_irqrestore(&tick_broadcast_lock, flags);
+       raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 }
 
 int tick_resume_broadcast(void)
@@ -329,7 +378,7 @@ int tick_resume_broadcast(void)
        unsigned long flags;
        int broadcast = 0;
 
-       spin_lock_irqsave(&tick_broadcast_lock, flags);
+       raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 
        bc = tick_broadcast_device.evtdev;
 
@@ -338,17 +387,18 @@ int tick_resume_broadcast(void)
 
                switch (tick_broadcast_device.mode) {
                case TICKDEV_MODE_PERIODIC:
-                       if(!cpus_empty(tick_broadcast_mask))
+                       if (!cpumask_empty(tick_broadcast_mask))
                                tick_broadcast_start_periodic(bc);
-                       broadcast = cpu_isset(smp_processor_id(),
-                                             tick_broadcast_mask);
+                       broadcast = cpumask_test_cpu(smp_processor_id(),
+                                                    tick_broadcast_mask);
                        break;
                case TICKDEV_MODE_ONESHOT:
-                       broadcast = tick_resume_broadcast_oneshot(bc);
+                       if (!cpumask_empty(tick_broadcast_mask))
+                               broadcast = tick_resume_broadcast_oneshot(bc);
                        break;
                }
        }
-       spin_unlock_irqrestore(&tick_broadcast_lock, flags);
+       raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 
        return broadcast;
 }
@@ -356,29 +406,58 @@ int tick_resume_broadcast(void)
 
 #ifdef CONFIG_TICK_ONESHOT
 
-static cpumask_t tick_broadcast_oneshot_mask;
+static cpumask_var_t tick_broadcast_oneshot_mask;
+static cpumask_var_t tick_broadcast_pending_mask;
+static cpumask_var_t tick_broadcast_force_mask;
 
 /*
- * Debugging: see timer_list.c
+ * Exposed for debugging: see timer_list.c
+ */
+struct cpumask *tick_get_broadcast_oneshot_mask(void)
+{
+       return tick_broadcast_oneshot_mask;
+}
+
+/*
+ * Called before going idle with interrupts disabled. Checks whether a
+ * broadcast event from the other core is about to happen. We detected
+ * that in tick_broadcast_oneshot_control(). The callsite can use this
+ * to avoid a deep idle transition as we are about to get the
+ * broadcast IPI right away.
  */
-cpumask_t *tick_get_broadcast_oneshot_mask(void)
+int tick_check_broadcast_expired(void)
 {
-       return &tick_broadcast_oneshot_mask;
+       return cpumask_test_cpu(smp_processor_id(), tick_broadcast_force_mask);
 }
 
-static int tick_broadcast_set_event(ktime_t expires, int force)
+/*
+ * Set broadcast interrupt affinity
+ */
+static void tick_broadcast_set_affinity(struct clock_event_device *bc,
+                                       const struct cpumask *cpumask)
 {
-       struct clock_event_device *bc = tick_broadcast_device.evtdev;
-       ktime_t now = ktime_get();
-       int res;
-
-       for(;;) {
-               res = clockevents_program_event(bc, expires, now);
-               if (!res || !force)
-                       return res;
-               now = ktime_get();
-               expires = ktime_add(now, ktime_set(0, bc->min_delta_ns));
-       }
+       if (!(bc->features & CLOCK_EVT_FEAT_DYNIRQ))
+               return;
+
+       if (cpumask_equal(bc->cpumask, cpumask))
+               return;
+
+       bc->cpumask = cpumask;
+       irq_set_affinity(bc->irq, bc->cpumask);
+}
+
+static int tick_broadcast_set_event(struct clock_event_device *bc, int cpu,
+                                   ktime_t expires, int force)
+{
+       int ret;
+
+       if (bc->mode != CLOCK_EVT_MODE_ONESHOT)
+               clockevents_set_mode(bc, CLOCK_EVT_MODE_ONESHOT);
+
+       ret = clockevents_program_event(bc, expires, force);
+       if (!ret)
+               tick_broadcast_set_affinity(bc, cpumask_of(cpu));
+       return ret;
 }
 
 int tick_resume_broadcast_oneshot(struct clock_event_device *bc)
@@ -388,34 +467,58 @@ int tick_resume_broadcast_oneshot(struct clock_event_device *bc)
 }
 
 /*
+ * Called from irq_enter() when idle was interrupted to reenable the
+ * per cpu device.
+ */
+void tick_check_oneshot_broadcast(int cpu)
+{
+       if (cpumask_test_cpu(cpu, tick_broadcast_oneshot_mask)) {
+               struct tick_device *td = &per_cpu(tick_cpu_device, cpu);
+
+               clockevents_set_mode(td->evtdev, CLOCK_EVT_MODE_ONESHOT);
+       }
+}
+
+/*
  * Handle oneshot mode broadcasting
  */
 static void tick_handle_oneshot_broadcast(struct clock_event_device *dev)
 {
        struct tick_device *td;
-       cpumask_t mask;
        ktime_t now, next_event;
-       int cpu;
+       int cpu, next_cpu = 0;
 
-       spin_lock(&tick_broadcast_lock);
+       raw_spin_lock(&tick_broadcast_lock);
 again:
        dev->next_event.tv64 = KTIME_MAX;
        next_event.tv64 = KTIME_MAX;
-       mask = CPU_MASK_NONE;
+       cpumask_clear(tmpmask);
        now = ktime_get();
        /* Find all expired events */
-       for_each_cpu_mask_nr(cpu, tick_broadcast_oneshot_mask) {
+       for_each_cpu(cpu, tick_broadcast_oneshot_mask) {
                td = &per_cpu(tick_cpu_device, cpu);
-               if (td->evtdev->next_event.tv64 <= now.tv64)
-                       cpu_set(cpu, mask);
-               else if (td->evtdev->next_event.tv64 < next_event.tv64)
+               if (td->evtdev->next_event.tv64 <= now.tv64) {
+                       cpumask_set_cpu(cpu, tmpmask);
+                       /*
+                        * Mark the remote cpu in the pending mask, so
+                        * it can avoid reprogramming the cpu local
+                        * timer in tick_broadcast_oneshot_control().
+                        */
+                       cpumask_set_cpu(cpu, tick_broadcast_pending_mask);
+               } else if (td->evtdev->next_event.tv64 < next_event.tv64) {
                        next_event.tv64 = td->evtdev->next_event.tv64;
+                       next_cpu = cpu;
+               }
        }
 
+       /* Take care of enforced broadcast requests */
+       cpumask_or(tmpmask, tmpmask, tick_broadcast_force_mask);
+       cpumask_clear(tick_broadcast_force_mask);
+
        /*
         * Wakeup the cpus which have an expired event.
         */
-       tick_do_broadcast(mask);
+       tick_do_broadcast(tmpmask);
 
        /*
         * Two reasons for reprogram:
@@ -432,10 +535,10 @@ again:
                 * Rearm the broadcast device. If event expired,
                 * repeat the above
                 */
-               if (tick_broadcast_set_event(next_event, 0))
+               if (tick_broadcast_set_event(dev, next_cpu, next_event, 0))
                        goto again;
        }
-       spin_unlock(&tick_broadcast_lock);
+       raw_spin_unlock(&tick_broadcast_lock);
 }
 
 /*
@@ -447,43 +550,110 @@ void tick_broadcast_oneshot_control(unsigned long reason)
        struct clock_event_device *bc, *dev;
        struct tick_device *td;
        unsigned long flags;
+       ktime_t now;
        int cpu;
 
-       spin_lock_irqsave(&tick_broadcast_lock, flags);
-
        /*
         * Periodic mode does not care about the enter/exit of power
         * states
         */
        if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
-               goto out;
+               return;
 
-       bc = tick_broadcast_device.evtdev;
+       /*
+        * We are called with preemtion disabled from the depth of the
+        * idle code, so we can't be moved away.
+        */
        cpu = smp_processor_id();
        td = &per_cpu(tick_cpu_device, cpu);
        dev = td->evtdev;
 
        if (!(dev->features & CLOCK_EVT_FEAT_C3STOP))
-               goto out;
+               return;
 
+       bc = tick_broadcast_device.evtdev;
+
+       raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
        if (reason == CLOCK_EVT_NOTIFY_BROADCAST_ENTER) {
-               if (!cpu_isset(cpu, tick_broadcast_oneshot_mask)) {
-                       cpu_set(cpu, tick_broadcast_oneshot_mask);
+               WARN_ON_ONCE(cpumask_test_cpu(cpu, tick_broadcast_pending_mask));
+               if (!cpumask_test_and_set_cpu(cpu, tick_broadcast_oneshot_mask)) {
                        clockevents_set_mode(dev, CLOCK_EVT_MODE_SHUTDOWN);
-                       if (dev->next_event.tv64 < bc->next_event.tv64)
-                               tick_broadcast_set_event(dev->next_event, 1);
+                       /*
+                        * We only reprogram the broadcast timer if we
+                        * did not mark ourself in the force mask and
+                        * if the cpu local event is earlier than the
+                        * broadcast event. If the current CPU is in
+                        * the force mask, then we are going to be
+                        * woken by the IPI right away.
+                        */
+                       if (!cpumask_test_cpu(cpu, tick_broadcast_force_mask) &&
+                           dev->next_event.tv64 < bc->next_event.tv64)
+                               tick_broadcast_set_event(bc, cpu, dev->next_event, 1);
                }
        } else {
-               if (cpu_isset(cpu, tick_broadcast_oneshot_mask)) {
-                       cpu_clear(cpu, tick_broadcast_oneshot_mask);
+               if (cpumask_test_and_clear_cpu(cpu, tick_broadcast_oneshot_mask)) {
                        clockevents_set_mode(dev, CLOCK_EVT_MODE_ONESHOT);
-                       if (dev->next_event.tv64 != KTIME_MAX)
-                               tick_program_event(dev->next_event, 1);
+                       if (dev->next_event.tv64 == KTIME_MAX)
+                               goto out;
+                       /*
+                        * The cpu which was handling the broadcast
+                        * timer marked this cpu in the broadcast
+                        * pending mask and fired the broadcast
+                        * IPI. So we are going to handle the expired
+                        * event anyway via the broadcast IPI
+                        * handler. No need to reprogram the timer
+                        * with an already expired event.
+                        */
+                       if (cpumask_test_and_clear_cpu(cpu,
+                                      tick_broadcast_pending_mask))
+                               goto out;
+
+                       /*
+                        * If the pending bit is not set, then we are
+                        * either the CPU handling the broadcast
+                        * interrupt or we got woken by something else.
+                        *
+                        * We are not longer in the broadcast mask, so
+                        * if the cpu local expiry time is already
+                        * reached, we would reprogram the cpu local
+                        * timer with an already expired event.
+                        *
+                        * This can lead to a ping-pong when we return
+                        * to idle and therefor rearm the broadcast
+                        * timer before the cpu local timer was able
+                        * to fire. This happens because the forced
+                        * reprogramming makes sure that the event
+                        * will happen in the future and depending on
+                        * the min_delta setting this might be far
+                        * enough out that the ping-pong starts.
+                        *
+                        * If the cpu local next_event has expired
+                        * then we know that the broadcast timer
+                        * next_event has expired as well and
+                        * broadcast is about to be handled. So we
+                        * avoid reprogramming and enforce that the
+                        * broadcast handler, which did not run yet,
+                        * will invoke the cpu local handler.
+                        *
+                        * We cannot call the handler directly from
+                        * here, because we might be in a NOHZ phase
+                        * and we did not go through the irq_enter()
+                        * nohz fixups.
+                        */
+                       now = ktime_get();
+                       if (dev->next_event.tv64 <= now.tv64) {
+                               cpumask_set_cpu(cpu, tick_broadcast_force_mask);
+                               goto out;
+                       }
+                       /*
+                        * We got woken by something else. Reprogram
+                        * the cpu local timer device.
+                        */
+                       tick_program_event(dev->next_event, 1);
                }
        }
-
 out:
-       spin_unlock_irqrestore(&tick_broadcast_lock, flags);
+       raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 }
 
 /*
@@ -493,7 +663,20 @@ out:
  */
 static void tick_broadcast_clear_oneshot(int cpu)
 {
-       cpu_clear(cpu, tick_broadcast_oneshot_mask);
+       cpumask_clear_cpu(cpu, tick_broadcast_oneshot_mask);
+}
+
+static void tick_broadcast_init_next_event(struct cpumask *mask,
+                                          ktime_t expires)
+{
+       struct tick_device *td;
+       int cpu;
+
+       for_each_cpu(cpu, mask) {
+               td = &per_cpu(tick_cpu_device, cpu);
+               if (td->evtdev)
+                       td->evtdev->next_event = expires;
+       }
 }
 
 /**
@@ -501,9 +684,46 @@ static void tick_broadcast_clear_oneshot(int cpu)
  */
 void tick_broadcast_setup_oneshot(struct clock_event_device *bc)
 {
-       bc->event_handler = tick_handle_oneshot_broadcast;
-       clockevents_set_mode(bc, CLOCK_EVT_MODE_ONESHOT);
-       bc->next_event.tv64 = KTIME_MAX;
+       int cpu = smp_processor_id();
+
+       /* Set it up only once ! */
+       if (bc->event_handler != tick_handle_oneshot_broadcast) {
+               int was_periodic = bc->mode == CLOCK_EVT_MODE_PERIODIC;
+
+               bc->event_handler = tick_handle_oneshot_broadcast;
+
+               /* Take the do_timer update */
+               if (!tick_nohz_full_cpu(cpu))
+                       tick_do_timer_cpu = cpu;
+
+               /*
+                * We must be careful here. There might be other CPUs
+                * waiting for periodic broadcast. We need to set the
+                * oneshot_mask bits for those and program the
+                * broadcast device to fire.
+                */
+               cpumask_copy(tmpmask, tick_broadcast_mask);
+               cpumask_clear_cpu(cpu, tmpmask);
+               cpumask_or(tick_broadcast_oneshot_mask,
+                          tick_broadcast_oneshot_mask, tmpmask);
+
+               if (was_periodic && !cpumask_empty(tmpmask)) {
+                       clockevents_set_mode(bc, CLOCK_EVT_MODE_ONESHOT);
+                       tick_broadcast_init_next_event(tmpmask,
+                                                      tick_next_period);
+                       tick_broadcast_set_event(bc, cpu, tick_next_period, 1);
+               } else
+                       bc->next_event.tv64 = KTIME_MAX;
+       } else {
+               /*
+                * The first cpu which switches to oneshot mode sets
+                * the bit for all other cpus which are in the general
+                * (periodic) broadcast mask. So the bit is set and
+                * would prevent the first broadcast enter after this
+                * to program the bc device.
+                */
+               tick_broadcast_clear_oneshot(cpu);
+       }
 }
 
 /*
@@ -514,13 +734,14 @@ void tick_broadcast_switch_to_oneshot(void)
        struct clock_event_device *bc;
        unsigned long flags;
 
-       spin_lock_irqsave(&tick_broadcast_lock, flags);
+       raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 
        tick_broadcast_device.mode = TICKDEV_MODE_ONESHOT;
        bc = tick_broadcast_device.evtdev;
        if (bc)
                tick_broadcast_setup_oneshot(bc);
-       spin_unlock_irqrestore(&tick_broadcast_lock, flags);
+
+       raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 }
 
 
@@ -532,15 +753,44 @@ void tick_shutdown_broadcast_oneshot(unsigned int *cpup)
        unsigned long flags;
        unsigned int cpu = *cpup;
 
-       spin_lock_irqsave(&tick_broadcast_lock, flags);
+       raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 
        /*
         * Clear the broadcast mask flag for the dead cpu, but do not
         * stop the broadcast device!
         */
-       cpu_clear(cpu, tick_broadcast_oneshot_mask);
+       cpumask_clear_cpu(cpu, tick_broadcast_oneshot_mask);
+
+       raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
+}
 
-       spin_unlock_irqrestore(&tick_broadcast_lock, flags);
+/*
+ * Check, whether the broadcast device is in one shot mode
+ */
+int tick_broadcast_oneshot_active(void)
+{
+       return tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT;
 }
 
+/*
+ * Check whether the broadcast device supports oneshot.
+ */
+bool tick_broadcast_oneshot_available(void)
+{
+       struct clock_event_device *bc = tick_broadcast_device.evtdev;
+
+       return bc ? bc->features & CLOCK_EVT_FEAT_ONESHOT : false;
+}
+
+#endif
+
+void __init tick_broadcast_init(void)
+{
+       alloc_cpumask_var(&tick_broadcast_mask, GFP_NOWAIT);
+       alloc_cpumask_var(&tmpmask, GFP_NOWAIT);
+#ifdef CONFIG_TICK_ONESHOT
+       alloc_cpumask_var(&tick_broadcast_oneshot_mask, GFP_NOWAIT);
+       alloc_cpumask_var(&tick_broadcast_pending_mask, GFP_NOWAIT);
+       alloc_cpumask_var(&tick_broadcast_force_mask, GFP_NOWAIT);
 #endif
+}