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 2fb8cb8..206bbfb 100644 (file)
@@ -28,9 +28,8 @@
  */
 
 static struct tick_device tick_broadcast_device;
-/* FIXME: Use cpumask_var_t. */
-static DECLARE_BITMAP(tick_broadcast_mask, NR_CPUS);
-static DECLARE_BITMAP(tmpmask, NR_CPUS);
+static cpumask_var_t tick_broadcast_mask;
+static cpumask_var_t tmpmask;
 static DEFINE_RAW_SPINLOCK(tick_broadcast_lock);
 static int tick_broadcast_force;
 
@@ -50,7 +49,7 @@ struct tick_device *tick_get_broadcast_device(void)
 
 struct cpumask *tick_get_broadcast_mask(void)
 {
-       return to_cpumask(tick_broadcast_mask);
+       return tick_broadcast_mask;
 }
 
 /*
@@ -67,15 +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(tick_broadcast_device.evtdev, dev);
+       if (cur)
+               cur->event_handler = clockevents_handle_noop;
        tick_broadcast_device.evtdev = dev;
-       if (!cpumask_empty(tick_get_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;
 }
 
@@ -123,7 +137,7 @@ int tick_device_uses_broadcast(struct clock_event_device *dev, int cpu)
        if (!tick_device_is_functional(dev)) {
                dev->event_handler = tick_handle_periodic;
                tick_device_setup_broadcast_func(dev);
-               cpumask_set_cpu(cpu, tick_get_broadcast_mask());
+               cpumask_set_cpu(cpu, tick_broadcast_mask);
                tick_broadcast_start_periodic(tick_broadcast_device.evtdev);
                ret = 1;
        } else {
@@ -134,7 +148,7 @@ int tick_device_uses_broadcast(struct clock_event_device *dev, int cpu)
                 */
                if (!(dev->features & CLOCK_EVT_FEAT_C3STOP)) {
                        int cpu = smp_processor_id();
-                       cpumask_clear_cpu(cpu, tick_get_broadcast_mask());
+                       cpumask_clear_cpu(cpu, tick_broadcast_mask);
                        tick_broadcast_clear_oneshot(cpu);
                } else {
                        tick_device_setup_broadcast_func(dev);
@@ -198,9 +212,8 @@ static void tick_do_periodic_broadcast(void)
 {
        raw_spin_lock(&tick_broadcast_lock);
 
-       cpumask_and(to_cpumask(tmpmask),
-                   cpu_online_mask, tick_get_broadcast_mask());
-       tick_do_broadcast(to_cpumask(tmpmask));
+       cpumask_and(tmpmask, cpu_online_mask, tick_broadcast_mask);
+       tick_do_broadcast(tmpmask);
 
        raw_spin_unlock(&tick_broadcast_lock);
 }
@@ -263,13 +276,12 @@ static void tick_do_broadcast_on_off(unsigned long *reason)
        if (!tick_device_is_functional(dev))
                goto out;
 
-       bc_stopped = cpumask_empty(tick_get_broadcast_mask());
+       bc_stopped = cpumask_empty(tick_broadcast_mask);
 
        switch (*reason) {
        case CLOCK_EVT_NOTIFY_BROADCAST_ON:
        case CLOCK_EVT_NOTIFY_BROADCAST_FORCE:
-               if (!cpumask_test_cpu(cpu, tick_get_broadcast_mask())) {
-                       cpumask_set_cpu(cpu, tick_get_broadcast_mask());
+               if (!cpumask_test_and_set_cpu(cpu, tick_broadcast_mask)) {
                        if (tick_broadcast_device.mode ==
                            TICKDEV_MODE_PERIODIC)
                                clockevents_shutdown(dev);
@@ -279,8 +291,7 @@ static void tick_do_broadcast_on_off(unsigned long *reason)
                break;
        case CLOCK_EVT_NOTIFY_BROADCAST_OFF:
                if (!tick_broadcast_force &&
-                   cpumask_test_cpu(cpu, tick_get_broadcast_mask())) {
-                       cpumask_clear_cpu(cpu, tick_get_broadcast_mask());
+                   cpumask_test_and_clear_cpu(cpu, tick_broadcast_mask)) {
                        if (tick_broadcast_device.mode ==
                            TICKDEV_MODE_PERIODIC)
                                tick_setup_periodic(dev, 0);
@@ -288,7 +299,7 @@ static void tick_do_broadcast_on_off(unsigned long *reason)
                break;
        }
 
-       if (cpumask_empty(tick_get_broadcast_mask())) {
+       if (cpumask_empty(tick_broadcast_mask)) {
                if (!bc_stopped)
                        clockevents_shutdown(bc);
        } else if (bc_stopped) {
@@ -337,10 +348,10 @@ void tick_shutdown_broadcast(unsigned int *cpup)
        raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
 
        bc = tick_broadcast_device.evtdev;
-       cpumask_clear_cpu(cpu, tick_get_broadcast_mask());
+       cpumask_clear_cpu(cpu, tick_broadcast_mask);
 
        if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC) {
-               if (bc && cpumask_empty(tick_get_broadcast_mask()))
+               if (bc && cpumask_empty(tick_broadcast_mask))
                        clockevents_shutdown(bc);
        }
 
@@ -376,13 +387,13 @@ int tick_resume_broadcast(void)
 
                switch (tick_broadcast_device.mode) {
                case TICKDEV_MODE_PERIODIC:
-                       if (!cpumask_empty(tick_get_broadcast_mask()))
+                       if (!cpumask_empty(tick_broadcast_mask))
                                tick_broadcast_start_periodic(bc);
                        broadcast = cpumask_test_cpu(smp_processor_id(),
-                                                    tick_get_broadcast_mask());
+                                                    tick_broadcast_mask);
                        break;
                case TICKDEV_MODE_ONESHOT:
-                       if (!cpumask_empty(tick_get_broadcast_mask()))
+                       if (!cpumask_empty(tick_broadcast_mask))
                                broadcast = tick_resume_broadcast_oneshot(bc);
                        break;
                }
@@ -395,25 +406,58 @@ int tick_resume_broadcast(void)
 
 #ifdef CONFIG_TICK_ONESHOT
 
-/* FIXME: use cpumask_var_t. */
-static DECLARE_BITMAP(tick_broadcast_oneshot_mask, NR_CPUS);
+static cpumask_var_t tick_broadcast_oneshot_mask;
+static cpumask_var_t tick_broadcast_pending_mask;
+static cpumask_var_t tick_broadcast_force_mask;
 
 /*
  * Exposed for debugging: see timer_list.c
  */
 struct cpumask *tick_get_broadcast_oneshot_mask(void)
 {
-       return to_cpumask(tick_broadcast_oneshot_mask);
+       return tick_broadcast_oneshot_mask;
 }
 
-static int tick_broadcast_set_event(ktime_t expires, int force)
+/*
+ * 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.
+ */
+int tick_check_broadcast_expired(void)
 {
-       struct clock_event_device *bc = tick_broadcast_device.evtdev;
+       return cpumask_test_cpu(smp_processor_id(), tick_broadcast_force_mask);
+}
+
+/*
+ * Set broadcast interrupt affinity
+ */
+static void tick_broadcast_set_affinity(struct clock_event_device *bc,
+                                       const struct cpumask *cpumask)
+{
+       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);
 
-       return clockevents_program_event(bc, expires, force);
+       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)
@@ -428,7 +472,7 @@ int tick_resume_broadcast_oneshot(struct clock_event_device *bc)
  */
 void tick_check_oneshot_broadcast(int cpu)
 {
-       if (cpumask_test_cpu(cpu, to_cpumask(tick_broadcast_oneshot_mask))) {
+       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);
@@ -442,27 +486,39 @@ static void tick_handle_oneshot_broadcast(struct clock_event_device *dev)
 {
        struct tick_device *td;
        ktime_t now, next_event;
-       int cpu;
+       int cpu, next_cpu = 0;
 
        raw_spin_lock(&tick_broadcast_lock);
 again:
        dev->next_event.tv64 = KTIME_MAX;
        next_event.tv64 = KTIME_MAX;
-       cpumask_clear(to_cpumask(tmpmask));
+       cpumask_clear(tmpmask);
        now = ktime_get();
        /* Find all expired events */
-       for_each_cpu(cpu, tick_get_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)
-                       cpumask_set_cpu(cpu, to_cpumask(tmpmask));
-               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(to_cpumask(tmpmask));
+       tick_do_broadcast(tmpmask);
 
        /*
         * Two reasons for reprogram:
@@ -479,7 +535,7 @@ 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;
        }
        raw_spin_unlock(&tick_broadcast_lock);
@@ -494,6 +550,7 @@ 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;
 
        /*
@@ -518,21 +575,84 @@ void tick_broadcast_oneshot_control(unsigned long reason)
 
        raw_spin_lock_irqsave(&tick_broadcast_lock, flags);
        if (reason == CLOCK_EVT_NOTIFY_BROADCAST_ENTER) {
-               if (!cpumask_test_cpu(cpu, tick_get_broadcast_oneshot_mask())) {
-                       cpumask_set_cpu(cpu, tick_get_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 (cpumask_test_cpu(cpu, tick_get_broadcast_oneshot_mask())) {
-                       cpumask_clear_cpu(cpu,
-                                         tick_get_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:
        raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 }
 
@@ -543,7 +663,7 @@ void tick_broadcast_oneshot_control(unsigned long reason)
  */
 static void tick_broadcast_clear_oneshot(int cpu)
 {
-       cpumask_clear_cpu(cpu, tick_get_broadcast_oneshot_mask());
+       cpumask_clear_cpu(cpu, tick_broadcast_oneshot_mask);
 }
 
 static void tick_broadcast_init_next_event(struct cpumask *mask,
@@ -573,7 +693,8 @@ void tick_broadcast_setup_oneshot(struct clock_event_device *bc)
                bc->event_handler = tick_handle_oneshot_broadcast;
 
                /* Take the do_timer update */
-               tick_do_timer_cpu = cpu;
+               if (!tick_nohz_full_cpu(cpu))
+                       tick_do_timer_cpu = cpu;
 
                /*
                 * We must be careful here. There might be other CPUs
@@ -581,17 +702,16 @@ void tick_broadcast_setup_oneshot(struct clock_event_device *bc)
                 * oneshot_mask bits for those and program the
                 * broadcast device to fire.
                 */
-               cpumask_copy(to_cpumask(tmpmask), tick_get_broadcast_mask());
-               cpumask_clear_cpu(cpu, to_cpumask(tmpmask));
-               cpumask_or(tick_get_broadcast_oneshot_mask(),
-                          tick_get_broadcast_oneshot_mask(),
-                          to_cpumask(tmpmask));
+               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(to_cpumask(tmpmask))) {
+               if (was_periodic && !cpumask_empty(tmpmask)) {
                        clockevents_set_mode(bc, CLOCK_EVT_MODE_ONESHOT);
-                       tick_broadcast_init_next_event(to_cpumask(tmpmask),
+                       tick_broadcast_init_next_event(tmpmask,
                                                       tick_next_period);
-                       tick_broadcast_set_event(tick_next_period, 1);
+                       tick_broadcast_set_event(bc, cpu, tick_next_period, 1);
                } else
                        bc->next_event.tv64 = KTIME_MAX;
        } else {
@@ -639,7 +759,7 @@ void tick_shutdown_broadcast_oneshot(unsigned int *cpup)
         * Clear the broadcast mask flag for the dead cpu, but do not
         * stop the broadcast device!
         */
-       cpumask_clear_cpu(cpu, tick_get_broadcast_oneshot_mask());
+       cpumask_clear_cpu(cpu, tick_broadcast_oneshot_mask);
 
        raw_spin_unlock_irqrestore(&tick_broadcast_lock, flags);
 }
@@ -663,3 +783,14 @@ bool tick_broadcast_oneshot_available(void)
 }
 
 #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
+}