Merge branch 'sched-fixes-for-linus' of git://git.kernel.org/pub/scm/linux/kernel...
[safe/jmp/linux-2.6] / kernel / hrtimer.c
index 052a0f5..3e1c36e 100644 (file)
@@ -48,6 +48,8 @@
 
 #include <asm/uaccess.h>
 
+#include <trace/events/timer.h>
+
 /*
  * The timer bases:
  *
@@ -135,6 +137,46 @@ struct hrtimer_clock_base *lock_hrtimer_base(const struct hrtimer *timer,
        }
 }
 
+
+/*
+ * Get the preferred target CPU for NOHZ
+ */
+static int hrtimer_get_target(int this_cpu, int pinned)
+{
+#ifdef CONFIG_NO_HZ
+       if (!pinned && get_sysctl_timer_migration() && idle_cpu(this_cpu)) {
+               int preferred_cpu = get_nohz_load_balancer();
+
+               if (preferred_cpu >= 0)
+                       return preferred_cpu;
+       }
+#endif
+       return this_cpu;
+}
+
+/*
+ * With HIGHRES=y we do not migrate the timer when it is expiring
+ * before the next event on the target cpu because we cannot reprogram
+ * the target cpu hardware and we would cause it to fire late.
+ *
+ * Called with cpu_base->lock of target cpu held.
+ */
+static int
+hrtimer_check_target(struct hrtimer *timer, struct hrtimer_clock_base *new_base)
+{
+#ifdef CONFIG_HIGH_RES_TIMERS
+       ktime_t expires;
+
+       if (!new_base->cpu_base->hres_active)
+               return 0;
+
+       expires = ktime_sub(hrtimer_get_expires(timer), new_base->offset);
+       return expires.tv64 <= new_base->cpu_base->expires_next.tv64;
+#else
+       return 0;
+#endif
+}
+
 /*
  * Switch the timer base to the current CPU when possible.
  */
@@ -144,16 +186,8 @@ switch_hrtimer_base(struct hrtimer *timer, struct hrtimer_clock_base *base,
 {
        struct hrtimer_clock_base *new_base;
        struct hrtimer_cpu_base *new_cpu_base;
-       int cpu, preferred_cpu = -1;
-
-       cpu = smp_processor_id();
-#if defined(CONFIG_NO_HZ) && defined(CONFIG_SMP)
-       if (!pinned && get_sysctl_timer_migration() && idle_cpu(cpu)) {
-               preferred_cpu = get_nohz_load_balancer();
-               if (preferred_cpu >= 0)
-                       cpu = preferred_cpu;
-       }
-#endif
+       int this_cpu = smp_processor_id();
+       int cpu = hrtimer_get_target(this_cpu, pinned);
 
 again:
        new_cpu_base = &per_cpu(hrtimer_bases, cpu);
@@ -161,7 +195,7 @@ again:
 
        if (base != new_base) {
                /*
-                * We are trying to schedule the timer on the local CPU.
+                * We are trying to move timer to new_base.
                 * However we can't change timer's base while it is running,
                 * so we keep it on the same CPU. No hassle vs. reprogramming
                 * the event source in the high resolution case. The softirq
@@ -177,38 +211,12 @@ again:
                spin_unlock(&base->cpu_base->lock);
                spin_lock(&new_base->cpu_base->lock);
 
-               /* Optimized away for NOHZ=n SMP=n */
-               if (cpu == preferred_cpu) {
-                       /* Calculate clock monotonic expiry time */
-#ifdef CONFIG_HIGH_RES_TIMERS
-                       ktime_t expires = ktime_sub(hrtimer_get_expires(timer),
-                                                       new_base->offset);
-#else
-                       ktime_t expires = hrtimer_get_expires(timer);
-#endif
-
-                       /*
-                        * Get the next event on target cpu from the
-                        * clock events layer.
-                        * This covers the highres=off nohz=on case as well.
-                        */
-                       ktime_t next = clockevents_get_next_event(cpu);
-
-                       ktime_t delta = ktime_sub(expires, next);
-
-                       /*
-                        * We do not migrate the timer when it is expiring
-                        * before the next event on the target cpu because
-                        * we cannot reprogram the target cpu hardware and
-                        * we would cause it to fire late.
-                        */
-                       if (delta.tv64 < 0) {
-                               cpu = smp_processor_id();
-                               spin_unlock(&new_base->cpu_base->lock);
-                               spin_lock(&base->cpu_base->lock);
-                               timer->base = base;
-                               goto again;
-                       }
+               if (cpu != this_cpu && hrtimer_check_target(timer, new_base)) {
+                       cpu = this_cpu;
+                       spin_unlock(&new_base->cpu_base->lock);
+                       spin_lock(&base->cpu_base->lock);
+                       timer->base = base;
+                       goto again;
                }
                timer->base = new_base;
        }
@@ -423,6 +431,7 @@ void hrtimer_init_on_stack(struct hrtimer *timer, clockid_t clock_id,
        debug_object_init_on_stack(timer, &hrtimer_debug_descr);
        __hrtimer_init(timer, clock_id, mode);
 }
+EXPORT_SYMBOL_GPL(hrtimer_init_on_stack);
 
 void destroy_hrtimer_on_stack(struct hrtimer *timer)
 {
@@ -435,6 +444,26 @@ static inline void debug_hrtimer_activate(struct hrtimer *timer) { }
 static inline void debug_hrtimer_deactivate(struct hrtimer *timer) { }
 #endif
 
+static inline void
+debug_init(struct hrtimer *timer, clockid_t clockid,
+          enum hrtimer_mode mode)
+{
+       debug_hrtimer_init(timer);
+       trace_hrtimer_init(timer, clockid, mode);
+}
+
+static inline void debug_activate(struct hrtimer *timer)
+{
+       debug_hrtimer_activate(timer);
+       trace_hrtimer_start(timer);
+}
+
+static inline void debug_deactivate(struct hrtimer *timer)
+{
+       debug_hrtimer_deactivate(timer);
+       trace_hrtimer_cancel(timer);
+}
+
 /* High resolution timer related functions */
 #ifdef CONFIG_HIGH_RES_TIMERS
 
@@ -480,13 +509,14 @@ static inline int hrtimer_hres_active(void)
  * next event
  * Called with interrupts disabled and base->lock held
  */
-static void hrtimer_force_reprogram(struct hrtimer_cpu_base *cpu_base)
+static void
+hrtimer_force_reprogram(struct hrtimer_cpu_base *cpu_base, int skip_equal)
 {
        int i;
        struct hrtimer_clock_base *base = cpu_base->clock_base;
-       ktime_t expires;
+       ktime_t expires, expires_next;
 
-       cpu_base->expires_next.tv64 = KTIME_MAX;
+       expires_next.tv64 = KTIME_MAX;
 
        for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++, base++) {
                struct hrtimer *timer;
@@ -502,10 +532,15 @@ static void hrtimer_force_reprogram(struct hrtimer_cpu_base *cpu_base)
                 */
                if (expires.tv64 < 0)
                        expires.tv64 = 0;
-               if (expires.tv64 < cpu_base->expires_next.tv64)
-                       cpu_base->expires_next = expires;
+               if (expires.tv64 < expires_next.tv64)
+                       expires_next = expires;
        }
 
+       if (skip_equal && expires_next.tv64 == cpu_base->expires_next.tv64)
+               return;
+
+       cpu_base->expires_next.tv64 = expires_next.tv64;
+
        if (cpu_base->expires_next.tv64 != KTIME_MAX)
                tick_program_event(cpu_base->expires_next, 1);
 }
@@ -588,7 +623,7 @@ static void retrigger_next_event(void *arg)
        base->clock_base[CLOCK_REALTIME].offset =
                timespec_to_ktime(realtime_offset);
 
-       hrtimer_force_reprogram(base);
+       hrtimer_force_reprogram(base, 0);
        spin_unlock(&base->lock);
 }
 
@@ -691,8 +726,6 @@ static int hrtimer_switch_to_hres(void)
        /* "Retrigger" the interrupt to get things going */
        retrigger_next_event(NULL);
        local_irq_restore(flags);
-       printk(KERN_DEBUG "Switched to high resolution mode on CPU %d\n",
-              smp_processor_id());
        return 1;
 }
 
@@ -701,7 +734,8 @@ static int hrtimer_switch_to_hres(void)
 static inline int hrtimer_hres_active(void) { return 0; }
 static inline int hrtimer_is_hres_enabled(void) { return 0; }
 static inline int hrtimer_switch_to_hres(void) { return 0; }
-static inline void hrtimer_force_reprogram(struct hrtimer_cpu_base *base) { }
+static inline void
+hrtimer_force_reprogram(struct hrtimer_cpu_base *base, int skip_equal) { }
 static inline int hrtimer_enqueue_reprogram(struct hrtimer *timer,
                                            struct hrtimer_clock_base *base,
                                            int wakeup)
@@ -791,7 +825,7 @@ static int enqueue_hrtimer(struct hrtimer *timer,
        struct hrtimer *entry;
        int leftmost = 1;
 
-       debug_hrtimer_activate(timer);
+       debug_activate(timer);
 
        /*
         * Find the right place in the rbtree:
@@ -844,19 +878,29 @@ static void __remove_hrtimer(struct hrtimer *timer,
                             struct hrtimer_clock_base *base,
                             unsigned long newstate, int reprogram)
 {
-       if (timer->state & HRTIMER_STATE_ENQUEUED) {
-               /*
-                * Remove the timer from the rbtree and replace the
-                * first entry pointer if necessary.
-                */
-               if (base->first == &timer->node) {
-                       base->first = rb_next(&timer->node);
-                       /* Reprogram the clock event device. if enabled */
-                       if (reprogram && hrtimer_hres_active())
-                               hrtimer_force_reprogram(base->cpu_base);
+       if (!(timer->state & HRTIMER_STATE_ENQUEUED))
+               goto out;
+
+       /*
+        * Remove the timer from the rbtree and replace the first
+        * entry pointer if necessary.
+        */
+       if (base->first == &timer->node) {
+               base->first = rb_next(&timer->node);
+#ifdef CONFIG_HIGH_RES_TIMERS
+               /* Reprogram the clock event device. if enabled */
+               if (reprogram && hrtimer_hres_active()) {
+                       ktime_t expires;
+
+                       expires = ktime_sub(hrtimer_get_expires(timer),
+                                           base->offset);
+                       if (base->cpu_base->expires_next.tv64 == expires.tv64)
+                               hrtimer_force_reprogram(base->cpu_base, 1);
                }
-               rb_erase(&timer->node, &base->active);
+#endif
        }
+       rb_erase(&timer->node, &base->active);
+out:
        timer->state = newstate;
 }
 
@@ -877,7 +921,7 @@ remove_hrtimer(struct hrtimer *timer, struct hrtimer_clock_base *base)
                 * reprogramming happens in the interrupt handler. This is a
                 * rare case and less expensive than a smp call.
                 */
-               debug_hrtimer_deactivate(timer);
+               debug_deactivate(timer);
                timer_stats_hrtimer_clear_start_info(timer);
                reprogram = base->cpu_base == &__get_cpu_var(hrtimer_bases);
                __remove_hrtimer(timer, base, HRTIMER_STATE_INACTIVE,
@@ -1110,7 +1154,7 @@ static void __hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
 void hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
                  enum hrtimer_mode mode)
 {
-       debug_hrtimer_init(timer);
+       debug_init(timer, clock_id, mode);
        __hrtimer_init(timer, clock_id, mode);
 }
 EXPORT_SYMBOL_GPL(hrtimer_init);
@@ -1134,7 +1178,7 @@ int hrtimer_get_res(const clockid_t which_clock, struct timespec *tp)
 }
 EXPORT_SYMBOL_GPL(hrtimer_get_res);
 
-static void __run_hrtimer(struct hrtimer *timer)
+static void __run_hrtimer(struct hrtimer *timer, ktime_t *now)
 {
        struct hrtimer_clock_base *base = timer->base;
        struct hrtimer_cpu_base *cpu_base = base->cpu_base;
@@ -1143,7 +1187,7 @@ static void __run_hrtimer(struct hrtimer *timer)
 
        WARN_ON(!irqs_disabled());
 
-       debug_hrtimer_deactivate(timer);
+       debug_deactivate(timer);
        __remove_hrtimer(timer, base, HRTIMER_STATE_CALLBACK, 0);
        timer_stats_account_hrtimer(timer);
        fn = timer->function;
@@ -1154,7 +1198,9 @@ static void __run_hrtimer(struct hrtimer *timer)
         * the timer base.
         */
        spin_unlock(&cpu_base->lock);
+       trace_hrtimer_expire_entry(timer, now);
        restart = fn(timer);
+       trace_hrtimer_expire_exit(timer);
        spin_lock(&cpu_base->lock);
 
        /*
@@ -1219,14 +1265,22 @@ void hrtimer_interrupt(struct clock_event_device *dev)
 
        expires_next.tv64 = KTIME_MAX;
 
+       spin_lock(&cpu_base->lock);
+       /*
+        * We set expires_next to KTIME_MAX here with cpu_base->lock
+        * held to prevent that a timer is enqueued in our queue via
+        * the migration code. This does not affect enqueueing of
+        * timers which run their callback and need to be requeued on
+        * this CPU.
+        */
+       cpu_base->expires_next.tv64 = KTIME_MAX;
+
        base = cpu_base->clock_base;
 
        for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++) {
                ktime_t basenow;
                struct rb_node *node;
 
-               spin_lock(&cpu_base->lock);
-
                basenow = ktime_add(now, base->offset);
 
                while ((node = base->first)) {
@@ -1257,13 +1311,17 @@ void hrtimer_interrupt(struct clock_event_device *dev)
                                break;
                        }
 
-                       __run_hrtimer(timer);
+                       __run_hrtimer(timer, &basenow);
                }
-               spin_unlock(&cpu_base->lock);
                base++;
        }
 
+       /*
+        * Store the new expiry value so the migration code can verify
+        * against it.
+        */
        cpu_base->expires_next = expires_next;
+       spin_unlock(&cpu_base->lock);
 
        /* Reprogramming necessary ? */
        if (expires_next.tv64 != KTIME_MAX) {
@@ -1375,7 +1433,7 @@ void hrtimer_run_queues(void)
                                        hrtimer_get_expires_tv64(timer))
                                break;
 
-                       __run_hrtimer(timer);
+                       __run_hrtimer(timer, &base->softirq_time);
                }
                spin_unlock(&cpu_base->lock);
        }
@@ -1402,6 +1460,7 @@ void hrtimer_init_sleeper(struct hrtimer_sleeper *sl, struct task_struct *task)
        sl->timer.function = hrtimer_wakeup;
        sl->task = task;
 }
+EXPORT_SYMBOL_GPL(hrtimer_init_sleeper);
 
 static int __sched do_nanosleep(struct hrtimer_sleeper *t, enum hrtimer_mode mode)
 {
@@ -1551,7 +1610,7 @@ static void migrate_hrtimer_list(struct hrtimer_clock_base *old_base,
        while ((node = rb_first(&old_base->active))) {
                timer = rb_entry(node, struct hrtimer, node);
                BUG_ON(hrtimer_callback_running(timer));
-               debug_hrtimer_deactivate(timer);
+               debug_deactivate(timer);
 
                /*
                 * Mark it as STATE_MIGRATE not INACTIVE otherwise the