Merge branch 'sched-core-for-linus' of git://git.kernel.org/pub/scm/linux/kernel...
[linux-2.6.git] / kernel / hrtimer.c
index 0148300..c541ee5 100644 (file)
@@ -32,7 +32,6 @@
  */
 
 #include <linux/cpu.h>
-#include <linux/irq.h>
 #include <linux/module.h>
 #include <linux/percpu.h>
 #include <linux/hrtimer.h>
 #include <linux/seq_file.h>
 #include <linux/err.h>
 #include <linux/debugobjects.h>
+#include <linux/sched.h>
+#include <linux/timer.h>
 
 #include <asm/uaccess.h>
 
-/**
- * ktime_get - get the monotonic time in ktime_t format
- *
- * returns the time in ktime_t format
- */
-ktime_t ktime_get(void)
-{
-       struct timespec now;
-
-       ktime_get_ts(&now);
-
-       return timespec_to_ktime(now);
-}
-EXPORT_SYMBOL_GPL(ktime_get);
-
-/**
- * ktime_get_real - get the real (wall-) time in ktime_t format
- *
- * returns the time in ktime_t format
- */
-ktime_t ktime_get_real(void)
-{
-       struct timespec now;
-
-       getnstimeofday(&now);
-
-       return timespec_to_ktime(now);
-}
-
-EXPORT_SYMBOL_GPL(ktime_get_real);
+#include <trace/events/timer.h>
 
 /*
  * The timer bases:
  *
- * Note: If we want to add new timer bases, we have to skip the two
- * clock ids captured by the cpu-timers. We do this by holding empty
- * entries rather than doing math adjustment of the clock ids.
- * This ensures that we capture erroneous accesses to these clock ids
- * rather than moving them into the range of valid clock id's.
+ * There are more clockids then hrtimer bases. Thus, we index
+ * into the timer bases by the hrtimer_base_type enum. When trying
+ * to reach a base using a clockid, hrtimer_clockid_to_base()
+ * is used to convert from clockid to the proper hrtimer_base_type.
  */
 DEFINE_PER_CPU(struct hrtimer_cpu_base, hrtimer_bases) =
 {
@@ -93,42 +64,37 @@ DEFINE_PER_CPU(struct hrtimer_cpu_base, hrtimer_bases) =
        .clock_base =
        {
                {
-                       .index = CLOCK_REALTIME,
+                       .index = HRTIMER_BASE_MONOTONIC,
+                       .clockid = CLOCK_MONOTONIC,
+                       .get_time = &ktime_get,
+                       .resolution = KTIME_LOW_RES,
+               },
+               {
+                       .index = HRTIMER_BASE_REALTIME,
+                       .clockid = CLOCK_REALTIME,
                        .get_time = &ktime_get_real,
                        .resolution = KTIME_LOW_RES,
                },
                {
-                       .index = CLOCK_MONOTONIC,
-                       .get_time = &ktime_get,
+                       .index = HRTIMER_BASE_BOOTTIME,
+                       .clockid = CLOCK_BOOTTIME,
+                       .get_time = &ktime_get_boottime,
                        .resolution = KTIME_LOW_RES,
                },
        }
 };
 
-/**
- * ktime_get_ts - get the monotonic clock in timespec format
- * @ts:                pointer to timespec variable
- *
- * The function calculates the monotonic clock from the realtime
- * clock and the wall_to_monotonic offset and stores the result
- * in normalized timespec format in the variable pointed to by @ts.
- */
-void ktime_get_ts(struct timespec *ts)
-{
-       struct timespec tomono;
-       unsigned long seq;
-
-       do {
-               seq = read_seqbegin(&xtime_lock);
-               getnstimeofday(ts);
-               tomono = wall_to_monotonic;
-
-       } while (read_seqretry(&xtime_lock, seq));
+static const int hrtimer_clock_to_base_table[MAX_CLOCKS] = {
+       [CLOCK_REALTIME]        = HRTIMER_BASE_REALTIME,
+       [CLOCK_MONOTONIC]       = HRTIMER_BASE_MONOTONIC,
+       [CLOCK_BOOTTIME]        = HRTIMER_BASE_BOOTTIME,
+};
 
-       set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec,
-                               ts->tv_nsec + tomono.tv_nsec);
+static inline int hrtimer_clockid_to_base(clockid_t clock_id)
+{
+       return hrtimer_clock_to_base_table[clock_id];
 }
-EXPORT_SYMBOL_GPL(ktime_get_ts);
+
 
 /*
  * Get the coarse grained time at the softirq based on xtime and
@@ -136,21 +102,17 @@ EXPORT_SYMBOL_GPL(ktime_get_ts);
  */
 static void hrtimer_get_softirq_time(struct hrtimer_cpu_base *base)
 {
-       ktime_t xtim, tomono;
-       struct timespec xts, tom;
-       unsigned long seq;
+       ktime_t xtim, mono, boot;
+       struct timespec xts, tom, slp;
 
-       do {
-               seq = read_seqbegin(&xtime_lock);
-               xts = current_kernel_time();
-               tom = wall_to_monotonic;
-       } while (read_seqretry(&xtime_lock, seq));
+       get_xtime_and_monotonic_and_sleep_offset(&xts, &tom, &slp);
 
        xtim = timespec_to_ktime(xts);
-       tomono = timespec_to_ktime(tom);
-       base->clock_base[CLOCK_REALTIME].softirq_time = xtim;
-       base->clock_base[CLOCK_MONOTONIC].softirq_time =
-               ktime_add(xtim, tomono);
+       mono = ktime_add(xtim, timespec_to_ktime(tom));
+       boot = ktime_add(mono, timespec_to_ktime(slp));
+       base->clock_base[HRTIMER_BASE_REALTIME].softirq_time = xtim;
+       base->clock_base[HRTIMER_BASE_MONOTONIC].softirq_time = mono;
+       base->clock_base[HRTIMER_BASE_BOOTTIME].softirq_time = boot;
 }
 
 /*
@@ -180,31 +142,72 @@ struct hrtimer_clock_base *lock_hrtimer_base(const struct hrtimer *timer,
        for (;;) {
                base = timer->base;
                if (likely(base != NULL)) {
-                       spin_lock_irqsave(&base->cpu_base->lock, *flags);
+                       raw_spin_lock_irqsave(&base->cpu_base->lock, *flags);
                        if (likely(base == timer->base))
                                return base;
                        /* The timer has migrated to another CPU: */
-                       spin_unlock_irqrestore(&base->cpu_base->lock, *flags);
+                       raw_spin_unlock_irqrestore(&base->cpu_base->lock, *flags);
                }
                cpu_relax();
        }
 }
 
+
+/*
+ * 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))
+               return get_nohz_timer_target();
+#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.
  */
 static inline struct hrtimer_clock_base *
-switch_hrtimer_base(struct hrtimer *timer, struct hrtimer_clock_base *base)
+switch_hrtimer_base(struct hrtimer *timer, struct hrtimer_clock_base *base,
+                   int pinned)
 {
        struct hrtimer_clock_base *new_base;
        struct hrtimer_cpu_base *new_cpu_base;
+       int this_cpu = smp_processor_id();
+       int cpu = hrtimer_get_target(this_cpu, pinned);
+       int basenum = base->index;
 
-       new_cpu_base = &__get_cpu_var(hrtimer_bases);
-       new_base = &new_cpu_base->clock_base[base->index];
+again:
+       new_cpu_base = &per_cpu(hrtimer_bases, cpu);
+       new_base = &new_cpu_base->clock_base[basenum];
 
        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
@@ -217,8 +220,16 @@ switch_hrtimer_base(struct hrtimer *timer, struct hrtimer_clock_base *base)
 
                /* See the comment in lock_timer_base() */
                timer->base = NULL;
-               spin_unlock(&base->cpu_base->lock);
-               spin_lock(&new_base->cpu_base->lock);
+               raw_spin_unlock(&base->cpu_base->lock);
+               raw_spin_lock(&new_base->cpu_base->lock);
+
+               if (cpu != this_cpu && hrtimer_check_target(timer, new_base)) {
+                       cpu = this_cpu;
+                       raw_spin_unlock(&new_base->cpu_base->lock);
+                       raw_spin_lock(&base->cpu_base->lock);
+                       timer->base = base;
+                       goto again;
+               }
                timer->base = new_base;
        }
        return new_base;
@@ -231,12 +242,12 @@ lock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
 {
        struct hrtimer_clock_base *base = timer->base;
 
-       spin_lock_irqsave(&base->cpu_base->lock, *flags);
+       raw_spin_lock_irqsave(&base->cpu_base->lock, *flags);
 
        return base;
 }
 
-# define switch_hrtimer_base(t, b)     (b)
+# define switch_hrtimer_base(t, b, p)  (b)
 
 #endif /* !CONFIG_SMP */
 
@@ -333,10 +344,17 @@ ktime_t ktime_add_safe(const ktime_t lhs, const ktime_t rhs)
        return res;
 }
 
+EXPORT_SYMBOL_GPL(ktime_add_safe);
+
 #ifdef CONFIG_DEBUG_OBJECTS_TIMERS
 
 static struct debug_obj_descr hrtimer_debug_descr;
 
+static void *hrtimer_debug_hint(void *addr)
+{
+       return ((struct hrtimer *) addr)->function;
+}
+
 /*
  * fixup_init is called when:
  * - an active object is initialized
@@ -396,6 +414,7 @@ static int hrtimer_fixup_free(void *addr, enum debug_obj_state state)
 
 static struct debug_obj_descr hrtimer_debug_descr = {
        .name           = "hrtimer",
+       .debug_hint     = hrtimer_debug_hint,
        .fixup_init     = hrtimer_fixup_init,
        .fixup_activate = hrtimer_fixup_activate,
        .fixup_free     = hrtimer_fixup_free,
@@ -430,6 +449,7 @@ 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)
 {
@@ -442,20 +462,24 @@ static inline void debug_hrtimer_activate(struct hrtimer *timer) { }
 static inline void debug_hrtimer_deactivate(struct hrtimer *timer) { }
 #endif
 
-/*
- * Check, whether the timer is on the callback pending list
- */
-static inline int hrtimer_cb_pending(const struct hrtimer *timer)
+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)
 {
-       return timer->state & HRTIMER_STATE_PENDING;
+       debug_hrtimer_activate(timer);
+       trace_hrtimer_start(timer);
 }
 
-/*
- * Remove a timer from the callback pending list
- */
-static inline void hrtimer_remove_cb_pending(struct hrtimer *timer)
+static inline void debug_deactivate(struct hrtimer *timer)
 {
-       list_del_init(&timer->cb_entry);
+       debug_hrtimer_deactivate(timer);
+       trace_hrtimer_cancel(timer);
 }
 
 /* High resolution timer related functions */
@@ -495,7 +519,7 @@ static inline int hrtimer_is_hres_enabled(void)
  */
 static inline int hrtimer_hres_active(void)
 {
-       return __get_cpu_var(hrtimer_bases).hres_active;
+       return __this_cpu_read(hrtimer_bases.hres_active);
 }
 
 /*
@@ -503,25 +527,41 @@ 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;
+               struct timerqueue_node *next;
 
-               if (!base->first)
+               next = timerqueue_getnext(&base->active);
+               if (!next)
                        continue;
-               timer = rb_entry(base->first, struct hrtimer, node);
+               timer = container_of(next, struct hrtimer, node);
+
                expires = ktime_sub(hrtimer_get_expires(timer), base->offset);
-               if (expires.tv64 < cpu_base->expires_next.tv64)
-                       cpu_base->expires_next = expires;
+               /*
+                * clock_was_set() has changed base->offset so the
+                * result might be negative. Fix it up to prevent a
+                * false positive in clockevents_program_event()
+                */
+               if (expires.tv64 < 0)
+                       expires.tv64 = 0;
+               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);
 }
@@ -538,7 +578,7 @@ static void hrtimer_force_reprogram(struct hrtimer_cpu_base *cpu_base)
 static int hrtimer_reprogram(struct hrtimer *timer,
                             struct hrtimer_clock_base *base)
 {
-       ktime_t *expires_next = &__get_cpu_var(hrtimer_bases).expires_next;
+       struct hrtimer_cpu_base *cpu_base = &__get_cpu_var(hrtimer_bases);
        ktime_t expires = ktime_sub(hrtimer_get_expires(timer), base->offset);
        int res;
 
@@ -563,7 +603,16 @@ static int hrtimer_reprogram(struct hrtimer *timer,
        if (expires.tv64 < 0)
                return -ETIME;
 
-       if (expires.tv64 >= expires_next->tv64)
+       if (expires.tv64 >= cpu_base->expires_next.tv64)
+               return 0;
+
+       /*
+        * If a hang was detected in the last timer interrupt then we
+        * do not schedule a timer which is earlier than the expiry
+        * which we enforced in the hang detection. We want the system
+        * to make progress.
+        */
+       if (cpu_base->hang_detected)
                return 0;
 
        /*
@@ -571,70 +620,10 @@ static int hrtimer_reprogram(struct hrtimer *timer,
         */
        res = tick_program_event(expires, 0);
        if (!IS_ERR_VALUE(res))
-               *expires_next = expires;
+               cpu_base->expires_next = expires;
        return res;
 }
 
-
-/*
- * Retrigger next event is called after clock was set
- *
- * Called with interrupts disabled via on_each_cpu()
- */
-static void retrigger_next_event(void *arg)
-{
-       struct hrtimer_cpu_base *base;
-       struct timespec realtime_offset;
-       unsigned long seq;
-
-       if (!hrtimer_hres_active())
-               return;
-
-       do {
-               seq = read_seqbegin(&xtime_lock);
-               set_normalized_timespec(&realtime_offset,
-                                       -wall_to_monotonic.tv_sec,
-                                       -wall_to_monotonic.tv_nsec);
-       } while (read_seqretry(&xtime_lock, seq));
-
-       base = &__get_cpu_var(hrtimer_bases);
-
-       /* Adjust CLOCK_REALTIME offset */
-       spin_lock(&base->lock);
-       base->clock_base[CLOCK_REALTIME].offset =
-               timespec_to_ktime(realtime_offset);
-
-       hrtimer_force_reprogram(base);
-       spin_unlock(&base->lock);
-}
-
-/*
- * Clock realtime was set
- *
- * Change the offset of the realtime clock vs. the monotonic
- * clock.
- *
- * We might have to reprogram the high resolution timer interrupt. On
- * SMP we call the architecture specific code to retrigger _all_ high
- * resolution timer interrupts. On UP we just disable interrupts and
- * call the high resolution interrupt code.
- */
-void clock_was_set(void)
-{
-       /* Retrigger the CPU local events everywhere */
-       on_each_cpu(retrigger_next_event, NULL, 1);
-}
-
-/*
- * During resume we might have to reprogram the high resolution timer
- * interrupt (on the local CPU):
- */
-void hres_timers_resume(void)
-{
-       /* Retrigger the CPU local events: */
-       retrigger_next_event(NULL);
-}
-
 /*
  * Initialize the high resolution related parts of cpu_base
  */
@@ -645,65 +634,63 @@ static inline void hrtimer_init_hres(struct hrtimer_cpu_base *base)
 }
 
 /*
- * Initialize the high resolution related parts of a hrtimer
- */
-static inline void hrtimer_init_timer_hres(struct hrtimer *timer)
-{
-}
-
-/*
  * When High resolution timers are active, try to reprogram. Note, that in case
  * the state has HRTIMER_STATE_CALLBACK set, no reprogramming and no expiry
  * check happens. The timer gets enqueued into the rbtree. The reprogramming
  * and expiry check is done in the hrtimer_interrupt or in the softirq.
  */
 static inline int hrtimer_enqueue_reprogram(struct hrtimer *timer,
-                                           struct hrtimer_clock_base *base)
+                                           struct hrtimer_clock_base *base,
+                                           int wakeup)
 {
        if (base->cpu_base->hres_active && hrtimer_reprogram(timer, base)) {
+               if (wakeup) {
+                       raw_spin_unlock(&base->cpu_base->lock);
+                       raise_softirq_irqoff(HRTIMER_SOFTIRQ);
+                       raw_spin_lock(&base->cpu_base->lock);
+               } else
+                       __raise_softirq_irqoff(HRTIMER_SOFTIRQ);
 
-               /* Timer is expired, act upon the callback mode */
-               switch(timer->cb_mode) {
-               case HRTIMER_CB_IRQSAFE_NO_RESTART:
-                       debug_hrtimer_deactivate(timer);
-                       /*
-                        * We can call the callback from here. No restart
-                        * happens, so no danger of recursion
-                        */
-                       BUG_ON(timer->function(timer) != HRTIMER_NORESTART);
-                       return 1;
-               case HRTIMER_CB_IRQSAFE_NO_SOFTIRQ:
-                       /*
-                        * This is solely for the sched tick emulation with
-                        * dynamic tick support to ensure that we do not
-                        * restart the tick right on the edge and end up with
-                        * the tick timer in the softirq ! The calling site
-                        * takes care of this.
-                        */
-                       debug_hrtimer_deactivate(timer);
-                       return 1;
-               case HRTIMER_CB_IRQSAFE:
-               case HRTIMER_CB_SOFTIRQ:
-                       /*
-                        * Move everything else into the softirq pending list !
-                        */
-                       list_add_tail(&timer->cb_entry,
-                                     &base->cpu_base->cb_pending);
-                       timer->state = HRTIMER_STATE_PENDING;
-                       return 1;
-               default:
-                       BUG();
-               }
+               return 1;
        }
+
        return 0;
 }
 
 /*
+ * Retrigger next event is called after clock was set
+ *
+ * Called with interrupts disabled via on_each_cpu()
+ */
+static void retrigger_next_event(void *arg)
+{
+       struct hrtimer_cpu_base *base = &__get_cpu_var(hrtimer_bases);
+       struct timespec realtime_offset, xtim, wtm, sleep;
+
+       if (!hrtimer_hres_active())
+               return;
+
+       /* Optimized out for !HIGH_RES */
+       get_xtime_and_monotonic_and_sleep_offset(&xtim, &wtm, &sleep);
+       set_normalized_timespec(&realtime_offset, -wtm.tv_sec, -wtm.tv_nsec);
+
+       /* Adjust CLOCK_REALTIME offset */
+       raw_spin_lock(&base->lock);
+       base->clock_base[HRTIMER_BASE_REALTIME].offset =
+               timespec_to_ktime(realtime_offset);
+       base->clock_base[HRTIMER_BASE_BOOTTIME].offset =
+               timespec_to_ktime(sleep);
+
+       hrtimer_force_reprogram(base, 0);
+       raw_spin_unlock(&base->lock);
+}
+
+/*
  * Switch to high resolution mode
  */
 static int hrtimer_switch_to_hres(void)
 {
-       int cpu = smp_processor_id();
+       int i, cpu = smp_processor_id();
        struct hrtimer_cpu_base *base = &per_cpu(hrtimer_bases, cpu);
        unsigned long flags;
 
@@ -719,57 +706,95 @@ static int hrtimer_switch_to_hres(void)
                return 0;
        }
        base->hres_active = 1;
-       base->clock_base[CLOCK_REALTIME].resolution = KTIME_HIGH_RES;
-       base->clock_base[CLOCK_MONOTONIC].resolution = KTIME_HIGH_RES;
+       for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++)
+               base->clock_base[i].resolution = KTIME_HIGH_RES;
 
        tick_setup_sched_timer();
 
        /* "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;
 }
 
-static inline void hrtimer_raise_softirq(void)
-{
-       raise_softirq(HRTIMER_SOFTIRQ);
-}
-
 #else
 
 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)
+                                           struct hrtimer_clock_base *base,
+                                           int wakeup)
 {
        return 0;
 }
 static inline void hrtimer_init_hres(struct hrtimer_cpu_base *base) { }
-static inline void hrtimer_init_timer_hres(struct hrtimer *timer) { }
-static inline int hrtimer_reprogram(struct hrtimer *timer,
-                                   struct hrtimer_clock_base *base)
+static inline void retrigger_next_event(void *arg) { }
+
+#endif /* CONFIG_HIGH_RES_TIMERS */
+
+/*
+ * Clock realtime was set
+ *
+ * Change the offset of the realtime clock vs. the monotonic
+ * clock.
+ *
+ * We might have to reprogram the high resolution timer interrupt. On
+ * SMP we call the architecture specific code to retrigger _all_ high
+ * resolution timer interrupts. On UP we just disable interrupts and
+ * call the high resolution interrupt code.
+ */
+void clock_was_set(void)
 {
-       return 0;
+#ifdef CONFIG_HIGHRES_TIMERS
+       /* Retrigger the CPU local events everywhere */
+       on_each_cpu(retrigger_next_event, NULL, 1);
+#endif
+       timerfd_clock_was_set();
 }
-static inline void hrtimer_raise_softirq(void) { }
 
-#endif /* CONFIG_HIGH_RES_TIMERS */
+/*
+ * During resume we might have to reprogram the high resolution timer
+ * interrupt (on the local CPU):
+ */
+void hrtimers_resume(void)
+{
+       WARN_ONCE(!irqs_disabled(),
+                 KERN_INFO "hrtimers_resume() called with IRQs enabled!");
 
-#ifdef CONFIG_TIMER_STATS
-void __timer_stats_hrtimer_set_start_info(struct hrtimer *timer, void *addr)
+       retrigger_next_event(NULL);
+       timerfd_clock_was_set();
+}
+
+static inline void timer_stats_hrtimer_set_start_info(struct hrtimer *timer)
 {
+#ifdef CONFIG_TIMER_STATS
        if (timer->start_site)
                return;
-
-       timer->start_site = addr;
+       timer->start_site = __builtin_return_address(0);
        memcpy(timer->start_comm, current->comm, TASK_COMM_LEN);
        timer->start_pid = current->pid;
+#endif
 }
+
+static inline void timer_stats_hrtimer_clear_start_info(struct hrtimer *timer)
+{
+#ifdef CONFIG_TIMER_STATS
+       timer->start_site = NULL;
 #endif
+}
+
+static inline void timer_stats_account_hrtimer(struct hrtimer *timer)
+{
+#ifdef CONFIG_TIMER_STATS
+       if (likely(!timer_stats_active))
+               return;
+       timer_stats_update_stats(timer, timer->start_pid, timer->start_site,
+                                timer->function, timer->start_comm, 0);
+#endif
+}
 
 /*
  * Counterpart to lock_hrtimer_base above:
@@ -777,7 +802,7 @@ void __timer_stats_hrtimer_set_start_info(struct hrtimer *timer, void *addr)
 static inline
 void unlock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
 {
-       spin_unlock_irqrestore(&timer->base->cpu_base->lock, *flags);
+       raw_spin_unlock_irqrestore(&timer->base->cpu_base->lock, *flags);
 }
 
 /**
@@ -826,62 +851,24 @@ EXPORT_SYMBOL_GPL(hrtimer_forward);
  *
  * The timer is inserted in expiry order. Insertion into the
  * red black tree is O(log(n)). Must hold the base lock.
+ *
+ * Returns 1 when the new timer is the leftmost timer in the tree.
  */
-static void enqueue_hrtimer(struct hrtimer *timer,
-                           struct hrtimer_clock_base *base, int reprogram)
+static int enqueue_hrtimer(struct hrtimer *timer,
+                          struct hrtimer_clock_base *base)
 {
-       struct rb_node **link = &base->active.rb_node;
-       struct rb_node *parent = NULL;
-       struct hrtimer *entry;
-       int leftmost = 1;
-
-       debug_hrtimer_activate(timer);
-
-       /*
-        * Find the right place in the rbtree:
-        */
-       while (*link) {
-               parent = *link;
-               entry = rb_entry(parent, struct hrtimer, node);
-               /*
-                * We dont care about collisions. Nodes with
-                * the same expiry time stay together.
-                */
-               if (hrtimer_get_expires_tv64(timer) <
-                               hrtimer_get_expires_tv64(entry)) {
-                       link = &(*link)->rb_left;
-               } else {
-                       link = &(*link)->rb_right;
-                       leftmost = 0;
-               }
-       }
-
-       /*
-        * Insert the timer to the rbtree and check whether it
-        * replaces the first pending timer
-        */
-       if (leftmost) {
-               /*
-                * Reprogram the clock event device. When the timer is already
-                * expired hrtimer_enqueue_reprogram has either called the
-                * callback or added it to the pending list and raised the
-                * softirq.
-                *
-                * This is a NOP for !HIGHRES
-                */
-               if (reprogram && hrtimer_enqueue_reprogram(timer, base))
-                       return;
+       debug_activate(timer);
 
-               base->first = &timer->node;
-       }
+       timerqueue_add(&base->active, &timer->node);
+       base->cpu_base->active_bases |= 1 << base->index;
 
-       rb_link_node(&timer->node, parent, link);
-       rb_insert_color(&timer->node, &base->active);
        /*
         * HRTIMER_STATE_ENQUEUED is or'ed to the current state to preserve the
         * state of a possibly running callback.
         */
        timer->state |= HRTIMER_STATE_ENQUEUED;
+
+       return (&timer->node == base->active.next);
 }
 
 /*
@@ -898,22 +885,26 @@ static void __remove_hrtimer(struct hrtimer *timer,
                             struct hrtimer_clock_base *base,
                             unsigned long newstate, int reprogram)
 {
-       /* High res. callback list. NOP for !HIGHRES */
-       if (hrtimer_cb_pending(timer))
-               hrtimer_remove_cb_pending(timer);
-       else {
-               /*
-                * 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;
+
+       if (&timer->node == timerqueue_getnext(&base->active)) {
+#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
        }
+       timerqueue_del(&base->active, &timer->node);
+       if (!timerqueue_getnext(&base->active))
+               base->cpu_base->active_bases &= ~(1 << base->index);
+out:
        timer->state = newstate;
 }
 
@@ -924,6 +915,7 @@ static inline int
 remove_hrtimer(struct hrtimer *timer, struct hrtimer_clock_base *base)
 {
        if (hrtimer_is_queued(timer)) {
+               unsigned long state;
                int reprogram;
 
                /*
@@ -934,32 +926,28 @@ 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,
-                                reprogram);
+               /*
+                * We must preserve the CALLBACK state flag here,
+                * otherwise we could move the timer base in
+                * switch_hrtimer_base.
+                */
+               state = timer->state & HRTIMER_STATE_CALLBACK;
+               __remove_hrtimer(timer, base, state, reprogram);
                return 1;
        }
        return 0;
 }
 
-/**
- * hrtimer_start - (re)start an relative timer on the current CPU
- * @timer:     the timer to be added
- * @tim:       expiry time
- * @mode:      expiry mode: absolute (HRTIMER_ABS) or relative (HRTIMER_REL)
- *
- * Returns:
- *  0 on success
- *  1 when the timer was active
- */
-int
-hrtimer_start(struct hrtimer *timer, ktime_t tim, const enum hrtimer_mode mode)
+int __hrtimer_start_range_ns(struct hrtimer *timer, ktime_t tim,
+               unsigned long delta_ns, const enum hrtimer_mode mode,
+               int wakeup)
 {
        struct hrtimer_clock_base *base, *new_base;
        unsigned long flags;
-       int ret, raise;
+       int ret, leftmost;
 
        base = lock_hrtimer_base(timer, &flags);
 
@@ -967,9 +955,9 @@ hrtimer_start(struct hrtimer *timer, ktime_t tim, const enum hrtimer_mode mode)
        ret = remove_hrtimer(timer, base);
 
        /* Switch the timer base, if necessary: */
-       new_base = switch_hrtimer_base(timer, base);
+       new_base = switch_hrtimer_base(timer, base, mode & HRTIMER_MODE_PINNED);
 
-       if (mode == HRTIMER_MODE_REL) {
+       if (mode & HRTIMER_MODE_REL) {
                tim = ktime_add_safe(tim, new_base->get_time());
                /*
                 * CONFIG_TIME_LOW_RES is a temporary way for architectures
@@ -983,41 +971,62 @@ hrtimer_start(struct hrtimer *timer, ktime_t tim, const enum hrtimer_mode mode)
 #endif
        }
 
-       hrtimer_set_expires(timer, tim);
+       hrtimer_set_expires_range_ns(timer, tim, delta_ns);
 
        timer_stats_hrtimer_set_start_info(timer);
 
+       leftmost = enqueue_hrtimer(timer, new_base);
+
        /*
         * Only allow reprogramming if the new base is on this CPU.
         * (it might still be on another CPU if the timer was pending)
+        *
+        * XXX send_remote_softirq() ?
         */
-       enqueue_hrtimer(timer, new_base,
-                       new_base->cpu_base == &__get_cpu_var(hrtimer_bases));
-
-       /*
-        * The timer may be expired and moved to the cb_pending
-        * list. We can not raise the softirq with base lock held due
-        * to a possible deadlock with runqueue lock.
-        */
-       raise = timer->state == HRTIMER_STATE_PENDING;
-
-       /*
-        * We use preempt_disable to prevent this task from migrating after
-        * setting up the softirq and raising it. Otherwise, if me migrate
-        * we will raise the softirq on the wrong CPU.
-        */
-       preempt_disable();
+       if (leftmost && new_base->cpu_base == &__get_cpu_var(hrtimer_bases))
+               hrtimer_enqueue_reprogram(timer, new_base, wakeup);
 
        unlock_hrtimer_base(timer, &flags);
 
-       if (raise)
-               hrtimer_raise_softirq();
-       preempt_enable();
-
        return ret;
 }
+
+/**
+ * hrtimer_start_range_ns - (re)start an hrtimer on the current CPU
+ * @timer:     the timer to be added
+ * @tim:       expiry time
+ * @delta_ns:  "slack" range for the timer
+ * @mode:      expiry mode: absolute (HRTIMER_ABS) or relative (HRTIMER_REL)
+ *
+ * Returns:
+ *  0 on success
+ *  1 when the timer was active
+ */
+int hrtimer_start_range_ns(struct hrtimer *timer, ktime_t tim,
+               unsigned long delta_ns, const enum hrtimer_mode mode)
+{
+       return __hrtimer_start_range_ns(timer, tim, delta_ns, mode, 1);
+}
+EXPORT_SYMBOL_GPL(hrtimer_start_range_ns);
+
+/**
+ * hrtimer_start - (re)start an hrtimer on the current CPU
+ * @timer:     the timer to be added
+ * @tim:       expiry time
+ * @mode:      expiry mode: absolute (HRTIMER_ABS) or relative (HRTIMER_REL)
+ *
+ * Returns:
+ *  0 on success
+ *  1 when the timer was active
+ */
+int
+hrtimer_start(struct hrtimer *timer, ktime_t tim, const enum hrtimer_mode mode)
+{
+       return __hrtimer_start_range_ns(timer, tim, 0, mode, 1);
+}
 EXPORT_SYMBOL_GPL(hrtimer_start);
 
+
 /**
  * hrtimer_try_to_cancel - try to deactivate a timer
  * @timer:     hrtimer to stop
@@ -1072,11 +1081,10 @@ EXPORT_SYMBOL_GPL(hrtimer_cancel);
  */
 ktime_t hrtimer_get_remaining(const struct hrtimer *timer)
 {
-       struct hrtimer_clock_base *base;
        unsigned long flags;
        ktime_t rem;
 
-       base = lock_hrtimer_base(timer, &flags);
+       lock_hrtimer_base(timer, &flags);
        rem = hrtimer_expires_remaining(timer);
        unlock_hrtimer_base(timer, &flags);
 
@@ -1099,16 +1107,18 @@ ktime_t hrtimer_get_next_event(void)
        unsigned long flags;
        int i;
 
-       spin_lock_irqsave(&cpu_base->lock, flags);
+       raw_spin_lock_irqsave(&cpu_base->lock, flags);
 
        if (!hrtimer_hres_active()) {
                for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++, base++) {
                        struct hrtimer *timer;
+                       struct timerqueue_node *next;
 
-                       if (!base->first)
+                       next = timerqueue_getnext(&base->active);
+                       if (!next)
                                continue;
 
-                       timer = rb_entry(base->first, struct hrtimer, node);
+                       timer = container_of(next, struct hrtimer, node);
                        delta.tv64 = hrtimer_get_expires_tv64(timer);
                        delta = ktime_sub(delta, base->get_time());
                        if (delta.tv64 < mindelta.tv64)
@@ -1116,7 +1126,7 @@ ktime_t hrtimer_get_next_event(void)
                }
        }
 
-       spin_unlock_irqrestore(&cpu_base->lock, flags);
+       raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
 
        if (mindelta.tv64 < 0)
                mindelta.tv64 = 0;
@@ -1128,6 +1138,7 @@ static void __hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
                           enum hrtimer_mode mode)
 {
        struct hrtimer_cpu_base *cpu_base;
+       int base;
 
        memset(timer, 0, sizeof(struct hrtimer));
 
@@ -1136,9 +1147,9 @@ static void __hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
        if (clock_id == CLOCK_REALTIME && mode != HRTIMER_MODE_ABS)
                clock_id = CLOCK_MONOTONIC;
 
-       timer->base = &cpu_base->clock_base[clock_id];
-       INIT_LIST_HEAD(&timer->cb_entry);
-       hrtimer_init_timer_hres(timer);
+       base = hrtimer_clockid_to_base(clock_id);
+       timer->base = &cpu_base->clock_base[base];
+       timerqueue_init(&timer->node);
 
 #ifdef CONFIG_TIMER_STATS
        timer->start_site = NULL;
@@ -1156,7 +1167,7 @@ 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);
@@ -1172,103 +1183,52 @@ EXPORT_SYMBOL_GPL(hrtimer_init);
 int hrtimer_get_res(const clockid_t which_clock, struct timespec *tp)
 {
        struct hrtimer_cpu_base *cpu_base;
+       int base = hrtimer_clockid_to_base(which_clock);
 
        cpu_base = &__raw_get_cpu_var(hrtimer_bases);
-       *tp = ktime_to_timespec(cpu_base->clock_base[which_clock].resolution);
+       *tp = ktime_to_timespec(cpu_base->clock_base[base].resolution);
 
        return 0;
 }
 EXPORT_SYMBOL_GPL(hrtimer_get_res);
 
-static void run_hrtimer_pending(struct hrtimer_cpu_base *cpu_base)
-{
-       spin_lock_irq(&cpu_base->lock);
-
-       while (!list_empty(&cpu_base->cb_pending)) {
-               enum hrtimer_restart (*fn)(struct hrtimer *);
-               struct hrtimer *timer;
-               int restart;
-
-               timer = list_entry(cpu_base->cb_pending.next,
-                                  struct hrtimer, cb_entry);
-
-               debug_hrtimer_deactivate(timer);
-               timer_stats_account_hrtimer(timer);
-
-               fn = timer->function;
-               __remove_hrtimer(timer, timer->base, HRTIMER_STATE_CALLBACK, 0);
-               spin_unlock_irq(&cpu_base->lock);
-
-               restart = fn(timer);
-
-               spin_lock_irq(&cpu_base->lock);
-
-               timer->state &= ~HRTIMER_STATE_CALLBACK;
-               if (restart == HRTIMER_RESTART) {
-                       BUG_ON(hrtimer_active(timer));
-                       /*
-                        * Enqueue the timer, allow reprogramming of the event
-                        * device
-                        */
-                       enqueue_hrtimer(timer, timer->base, 1);
-               } else if (hrtimer_active(timer)) {
-                       /*
-                        * If the timer was rearmed on another CPU, reprogram
-                        * the event device.
-                        */
-                       struct hrtimer_clock_base *base = timer->base;
-
-                       if (base->first == &timer->node &&
-                           hrtimer_reprogram(timer, base)) {
-                               /*
-                                * Timer is expired. Thus move it from tree to
-                                * pending list again.
-                                */
-                               __remove_hrtimer(timer, base,
-                                                HRTIMER_STATE_PENDING, 0);
-                               list_add_tail(&timer->cb_entry,
-                                             &base->cpu_base->cb_pending);
-                       }
-               }
-       }
-       spin_unlock_irq(&cpu_base->lock);
-}
-
-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;
        enum hrtimer_restart (*fn)(struct hrtimer *);
        int restart;
 
-       debug_hrtimer_deactivate(timer);
+       WARN_ON(!irqs_disabled());
+
+       debug_deactivate(timer);
        __remove_hrtimer(timer, base, HRTIMER_STATE_CALLBACK, 0);
        timer_stats_account_hrtimer(timer);
-
        fn = timer->function;
-       if (timer->cb_mode == HRTIMER_CB_IRQSAFE_NO_SOFTIRQ) {
-               /*
-                * Used for scheduler timers, avoid lock inversion with
-                * rq->lock and tasklist_lock.
-                *
-                * These timers are required to deal with enqueue expiry
-                * themselves and are not allowed to migrate.
-                */
-               spin_unlock(&cpu_base->lock);
-               restart = fn(timer);
-               spin_lock(&cpu_base->lock);
-       } else
-               restart = fn(timer);
 
        /*
-        * Note: We clear the CALLBACK bit after enqueue_hrtimer to avoid
-        * reprogramming of the event hardware. This happens at the end of this
-        * function anyway.
+        * Because we run timers from hardirq context, there is no chance
+        * they get migrated to another cpu, therefore its safe to unlock
+        * the timer base.
+        */
+       raw_spin_unlock(&cpu_base->lock);
+       trace_hrtimer_expire_entry(timer, now);
+       restart = fn(timer);
+       trace_hrtimer_expire_exit(timer);
+       raw_spin_lock(&cpu_base->lock);
+
+       /*
+        * Note: We clear the CALLBACK bit after enqueue_hrtimer and
+        * we do not reprogramm the event hardware. Happens either in
+        * hrtimer_start_range_ns() or in hrtimer_interrupt()
         */
        if (restart != HRTIMER_NORESTART) {
                BUG_ON(timer->state != HRTIMER_STATE_CALLBACK);
-               enqueue_hrtimer(timer, base, 0);
+               enqueue_hrtimer(timer, base);
        }
+
+       WARN_ON_ONCE(!(timer->state & HRTIMER_STATE_CALLBACK));
+
        timer->state &= ~HRTIMER_STATE_CALLBACK;
 }
 
@@ -1281,33 +1241,42 @@ static void __run_hrtimer(struct hrtimer *timer)
 void hrtimer_interrupt(struct clock_event_device *dev)
 {
        struct hrtimer_cpu_base *cpu_base = &__get_cpu_var(hrtimer_bases);
-       struct hrtimer_clock_base *base;
-       ktime_t expires_next, now;
-       int i, raise = 0;
+       ktime_t expires_next, now, entry_time, delta;
+       int i, retries = 0;
 
        BUG_ON(!cpu_base->hres_active);
        cpu_base->nr_events++;
        dev->next_event.tv64 = KTIME_MAX;
 
- retry:
-       now = ktime_get();
-
+       entry_time = now = ktime_get();
+retry:
        expires_next.tv64 = KTIME_MAX;
 
-       base = cpu_base->clock_base;
+       raw_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;
 
        for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++) {
+               struct hrtimer_clock_base *base;
+               struct timerqueue_node *node;
                ktime_t basenow;
-               struct rb_node *node;
 
-               spin_lock(&cpu_base->lock);
+               if (!(cpu_base->active_bases & (1 << i)))
+                       continue;
 
+               base = cpu_base->clock_base + i;
                basenow = ktime_add(now, base->offset);
 
-               while ((node = base->first)) {
+               while ((node = timerqueue_getnext(&base->active))) {
                        struct hrtimer *timer;
 
-                       timer = rb_entry(node, struct hrtimer, node);
+                       timer = container_of(node, struct hrtimer, node);
 
                        /*
                         * The immediate goal for using the softexpires is
@@ -1332,41 +1301,106 @@ void hrtimer_interrupt(struct clock_event_device *dev)
                                break;
                        }
 
-                       /* Move softirq callbacks to the pending list */
-                       if (timer->cb_mode == HRTIMER_CB_SOFTIRQ) {
-                               __remove_hrtimer(timer, base,
-                                                HRTIMER_STATE_PENDING, 0);
-                               list_add_tail(&timer->cb_entry,
-                                             &base->cpu_base->cb_pending);
-                               raise = 1;
-                               continue;
-                       }
-
-                       __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;
+       raw_spin_unlock(&cpu_base->lock);
 
        /* Reprogramming necessary ? */
-       if (expires_next.tv64 != KTIME_MAX) {
-               if (tick_program_event(expires_next, 0))
-                       goto retry;
+       if (expires_next.tv64 == KTIME_MAX ||
+           !tick_program_event(expires_next, 0)) {
+               cpu_base->hang_detected = 0;
+               return;
        }
 
-       /* Raise softirq ? */
-       if (raise)
-               raise_softirq(HRTIMER_SOFTIRQ);
+       /*
+        * The next timer was already expired due to:
+        * - tracing
+        * - long lasting callbacks
+        * - being scheduled away when running in a VM
+        *
+        * We need to prevent that we loop forever in the hrtimer
+        * interrupt routine. We give it 3 attempts to avoid
+        * overreacting on some spurious event.
+        */
+       now = ktime_get();
+       cpu_base->nr_retries++;
+       if (++retries < 3)
+               goto retry;
+       /*
+        * Give the system a chance to do something else than looping
+        * here. We stored the entry time, so we know exactly how long
+        * we spent here. We schedule the next event this amount of
+        * time away.
+        */
+       cpu_base->nr_hangs++;
+       cpu_base->hang_detected = 1;
+       delta = ktime_sub(now, entry_time);
+       if (delta.tv64 > cpu_base->max_hang_time.tv64)
+               cpu_base->max_hang_time = delta;
+       /*
+        * Limit it to a sensible value as we enforce a longer
+        * delay. Give the CPU at least 100ms to catch up.
+        */
+       if (delta.tv64 > 100 * NSEC_PER_MSEC)
+               expires_next = ktime_add_ns(now, 100 * NSEC_PER_MSEC);
+       else
+               expires_next = ktime_add(now, delta);
+       tick_program_event(expires_next, 1);
+       printk_once(KERN_WARNING "hrtimer: interrupt took %llu ns\n",
+                   ktime_to_ns(delta));
+}
+
+/*
+ * local version of hrtimer_peek_ahead_timers() called with interrupts
+ * disabled.
+ */
+static void __hrtimer_peek_ahead_timers(void)
+{
+       struct tick_device *td;
+
+       if (!hrtimer_hres_active())
+               return;
+
+       td = &__get_cpu_var(tick_cpu_device);
+       if (td && td->evtdev)
+               hrtimer_interrupt(td->evtdev);
+}
+
+/**
+ * hrtimer_peek_ahead_timers -- run soft-expired timers now
+ *
+ * hrtimer_peek_ahead_timers will peek at the timer queue of
+ * the current cpu and check if there are any timers for which
+ * the soft expires time has passed. If any such timers exist,
+ * they are run immediately and then removed from the timer queue.
+ *
+ */
+void hrtimer_peek_ahead_timers(void)
+{
+       unsigned long flags;
+
+       local_irq_save(flags);
+       __hrtimer_peek_ahead_timers();
+       local_irq_restore(flags);
 }
 
 static void run_hrtimer_softirq(struct softirq_action *h)
 {
-       run_hrtimer_pending(&__get_cpu_var(hrtimer_bases));
+       hrtimer_peek_ahead_timers();
 }
 
-#endif /* CONFIG_HIGH_RES_TIMERS */
+#else /* CONFIG_HIGH_RES_TIMERS */
+
+static inline void __hrtimer_peek_ahead_timers(void) { }
+
+#endif /* !CONFIG_HIGH_RES_TIMERS */
 
 /*
  * Called from timer softirq every jiffy, expire hrtimers:
@@ -1377,8 +1411,6 @@ static void run_hrtimer_softirq(struct softirq_action *h)
  */
 void hrtimer_run_pending(void)
 {
-       struct hrtimer_cpu_base *cpu_base = &__get_cpu_var(hrtimer_bases);
-
        if (hrtimer_hres_active())
                return;
 
@@ -1392,8 +1424,6 @@ void hrtimer_run_pending(void)
         */
        if (tick_check_oneshot_change(!hrtimer_is_hres_enabled()))
                hrtimer_switch_to_hres();
-
-       run_hrtimer_pending(cpu_base);
 }
 
 /*
@@ -1401,7 +1431,7 @@ void hrtimer_run_pending(void)
  */
 void hrtimer_run_queues(void)
 {
-       struct rb_node *node;
+       struct timerqueue_node *node;
        struct hrtimer_cpu_base *cpu_base = &__get_cpu_var(hrtimer_bases);
        struct hrtimer_clock_base *base;
        int index, gettime = 1;
@@ -1411,38 +1441,27 @@ void hrtimer_run_queues(void)
 
        for (index = 0; index < HRTIMER_MAX_CLOCK_BASES; index++) {
                base = &cpu_base->clock_base[index];
-
-               if (!base->first)
+               if (!timerqueue_getnext(&base->active))
                        continue;
 
-               if (base->get_softirq_time)
-                       base->softirq_time = base->get_softirq_time();
-               else if (gettime) {
+               if (gettime) {
                        hrtimer_get_softirq_time(cpu_base);
                        gettime = 0;
                }
 
-               spin_lock(&cpu_base->lock);
+               raw_spin_lock(&cpu_base->lock);
 
-               while ((node = base->first)) {
+               while ((node = timerqueue_getnext(&base->active))) {
                        struct hrtimer *timer;
 
-                       timer = rb_entry(node, struct hrtimer, node);
+                       timer = container_of(node, struct hrtimer, node);
                        if (base->softirq_time.tv64 <=
                                        hrtimer_get_expires_tv64(timer))
                                break;
 
-                       if (timer->cb_mode == HRTIMER_CB_SOFTIRQ) {
-                               __remove_hrtimer(timer, base,
-                                       HRTIMER_STATE_PENDING, 0);
-                               list_add_tail(&timer->cb_entry,
-                                       &base->cpu_base->cb_pending);
-                               continue;
-                       }
-
-                       __run_hrtimer(timer);
+                       __run_hrtimer(timer, &base->softirq_time);
                }
-               spin_unlock(&cpu_base->lock);
+               raw_spin_unlock(&cpu_base->lock);
        }
 }
 
@@ -1466,10 +1485,8 @@ void hrtimer_init_sleeper(struct hrtimer_sleeper *sl, struct task_struct *task)
 {
        sl->timer.function = hrtimer_wakeup;
        sl->task = task;
-#ifdef CONFIG_HIGH_RES_TIMERS
-       sl->timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
-#endif
 }
+EXPORT_SYMBOL_GPL(hrtimer_init_sleeper);
 
 static int __sched do_nanosleep(struct hrtimer_sleeper *t, enum hrtimer_mode mode)
 {
@@ -1516,7 +1533,7 @@ long __sched hrtimer_nanosleep_restart(struct restart_block *restart)
        struct timespec __user  *rmtp;
        int ret = 0;
 
-       hrtimer_init_on_stack(&t.timer, restart->nanosleep.index,
+       hrtimer_init_on_stack(&t.timer, restart->nanosleep.clockid,
                                HRTIMER_MODE_ABS);
        hrtimer_set_expires_tv64(&t.timer, restart->nanosleep.expires);
 
@@ -1543,9 +1560,14 @@ long hrtimer_nanosleep(struct timespec *rqtp, struct timespec __user *rmtp,
        struct restart_block *restart;
        struct hrtimer_sleeper t;
        int ret = 0;
+       unsigned long slack;
+
+       slack = current->timer_slack_ns;
+       if (rt_task(current))
+               slack = 0;
 
        hrtimer_init_on_stack(&t.timer, clockid, mode);
-       hrtimer_set_expires(&t.timer, timespec_to_ktime(*rqtp));
+       hrtimer_set_expires_range_ns(&t.timer, timespec_to_ktime(*rqtp), slack);
        if (do_nanosleep(&t, mode))
                goto out;
 
@@ -1563,7 +1585,7 @@ long hrtimer_nanosleep(struct timespec *rqtp, struct timespec __user *rmtp,
 
        restart = &current_thread_info()->restart_block;
        restart->fn = hrtimer_nanosleep_restart;
-       restart->nanosleep.index = t.timer.base->index;
+       restart->nanosleep.clockid = t.timer.base->clockid;
        restart->nanosleep.rmtp = rmtp;
        restart->nanosleep.expires = hrtimer_get_expires_tv64(&t.timer);
 
@@ -1573,8 +1595,8 @@ out:
        return ret;
 }
 
-asmlinkage long
-sys_nanosleep(struct timespec __user *rqtp, struct timespec __user *rmtp)
+SYSCALL_DEFINE2(nanosleep, struct timespec __user *, rqtp,
+               struct timespec __user *, rmtp)
 {
        struct timespec tu;
 
@@ -1595,12 +1617,13 @@ static void __cpuinit init_hrtimers_cpu(int cpu)
        struct hrtimer_cpu_base *cpu_base = &per_cpu(hrtimer_bases, cpu);
        int i;
 
-       spin_lock_init(&cpu_base->lock);
+       raw_spin_lock_init(&cpu_base->lock);
 
-       for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++)
+       for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++) {
                cpu_base->clock_base[i].cpu_base = cpu_base;
+               timerqueue_init_head(&cpu_base->clock_base[i].active);
+       }
 
-       INIT_LIST_HEAD(&cpu_base->cb_pending);
        hrtimer_init_hres(cpu_base);
 }
 
@@ -1610,66 +1633,92 @@ static void migrate_hrtimer_list(struct hrtimer_clock_base *old_base,
                                struct hrtimer_clock_base *new_base)
 {
        struct hrtimer *timer;
-       struct rb_node *node;
+       struct timerqueue_node *node;
 
-       while ((node = rb_first(&old_base->active))) {
-               timer = rb_entry(node, struct hrtimer, node);
+       while ((node = timerqueue_getnext(&old_base->active))) {
+               timer = container_of(node, struct hrtimer, node);
                BUG_ON(hrtimer_callback_running(timer));
-               debug_hrtimer_deactivate(timer);
-               __remove_hrtimer(timer, old_base, HRTIMER_STATE_INACTIVE, 0);
+               debug_deactivate(timer);
+
+               /*
+                * Mark it as STATE_MIGRATE not INACTIVE otherwise the
+                * timer could be seen as !active and just vanish away
+                * under us on another CPU
+                */
+               __remove_hrtimer(timer, old_base, HRTIMER_STATE_MIGRATE, 0);
                timer->base = new_base;
                /*
-                * Enqueue the timer. Allow reprogramming of the event device
+                * Enqueue the timers on the new cpu. This does not
+                * reprogram the event device in case the timer
+                * expires before the earliest on this CPU, but we run
+                * hrtimer_interrupt after we migrated everything to
+                * sort out already expired timers and reprogram the
+                * event device.
                 */
-               enqueue_hrtimer(timer, new_base, 1);
+               enqueue_hrtimer(timer, new_base);
+
+               /* Clear the migration state bit */
+               timer->state &= ~HRTIMER_STATE_MIGRATE;
        }
 }
 
-static void migrate_hrtimers(int cpu)
+static void migrate_hrtimers(int scpu)
 {
        struct hrtimer_cpu_base *old_base, *new_base;
        int i;
 
-       BUG_ON(cpu_online(cpu));
-       old_base = &per_cpu(hrtimer_bases, cpu);
-       new_base = &get_cpu_var(hrtimer_bases);
-
-       tick_cancel_sched_timer(cpu);
+       BUG_ON(cpu_online(scpu));
+       tick_cancel_sched_timer(scpu);
 
        local_irq_disable();
-       spin_lock(&new_base->lock);
-       spin_lock_nested(&old_base->lock, SINGLE_DEPTH_NESTING);
+       old_base = &per_cpu(hrtimer_bases, scpu);
+       new_base = &__get_cpu_var(hrtimer_bases);
+       /*
+        * The caller is globally serialized and nobody else
+        * takes two locks at once, deadlock is not possible.
+        */
+       raw_spin_lock(&new_base->lock);
+       raw_spin_lock_nested(&old_base->lock, SINGLE_DEPTH_NESTING);
 
        for (i = 0; i < HRTIMER_MAX_CLOCK_BASES; i++) {
                migrate_hrtimer_list(&old_base->clock_base[i],
                                     &new_base->clock_base[i]);
        }
 
-       spin_unlock(&old_base->lock);
-       spin_unlock(&new_base->lock);
+       raw_spin_unlock(&old_base->lock);
+       raw_spin_unlock(&new_base->lock);
+
+       /* Check, if we got expired work to do */
+       __hrtimer_peek_ahead_timers();
        local_irq_enable();
-       put_cpu_var(hrtimer_bases);
 }
+
 #endif /* CONFIG_HOTPLUG_CPU */
 
 static int __cpuinit hrtimer_cpu_notify(struct notifier_block *self,
                                        unsigned long action, void *hcpu)
 {
-       unsigned int cpu = (long)hcpu;
+       int scpu = (long)hcpu;
 
        switch (action) {
 
        case CPU_UP_PREPARE:
        case CPU_UP_PREPARE_FROZEN:
-               init_hrtimers_cpu(cpu);
+               init_hrtimers_cpu(scpu);
                break;
 
 #ifdef CONFIG_HOTPLUG_CPU
+       case CPU_DYING:
+       case CPU_DYING_FROZEN:
+               clockevents_notify(CLOCK_EVT_NOTIFY_CPU_DYING, &scpu);
+               break;
        case CPU_DEAD:
        case CPU_DEAD_FROZEN:
-               clockevents_notify(CLOCK_EVT_NOTIFY_CPU_DEAD, &cpu);
-               migrate_hrtimers(cpu);
+       {
+               clockevents_notify(CLOCK_EVT_NOTIFY_CPU_DEAD, &scpu);
+               migrate_hrtimers(scpu);
                break;
+       }
 #endif
 
        default:
@@ -1694,35 +1743,15 @@ void __init hrtimers_init(void)
 }
 
 /**
- * schedule_hrtimeout_range - sleep until timeout
+ * schedule_hrtimeout_range_clock - sleep until timeout
  * @expires:   timeout value (ktime_t)
  * @delta:     slack in expires timeout (ktime_t)
  * @mode:      timer mode, HRTIMER_MODE_ABS or HRTIMER_MODE_REL
- *
- * Make the current task sleep until the given expiry time has
- * elapsed. The routine will return immediately unless
- * the current task state has been set (see set_current_state()).
- *
- * The @delta argument gives the kernel the freedom to schedule the
- * actual wakeup to a time that is both power and performance friendly.
- * The kernel give the normal best effort behavior for "@expires+@delta",
- * but may decide to fire the timer earlier, but no earlier than @expires.
- *
- * You can set the task state as follows -
- *
- * %TASK_UNINTERRUPTIBLE - at least @timeout time is guaranteed to
- * pass before the routine returns.
- *
- * %TASK_INTERRUPTIBLE - the routine may return early if a signal is
- * delivered to the current task.
- *
- * The current task state is guaranteed to be TASK_RUNNING when this
- * routine returns.
- *
- * Returns 0 when the timer has expired otherwise -EINTR
+ * @clock:     timer clock, CLOCK_MONOTONIC or CLOCK_REALTIME
  */
-int __sched schedule_hrtimeout_range(ktime_t *expires, unsigned long delta,
-                              const enum hrtimer_mode mode)
+int __sched
+schedule_hrtimeout_range_clock(ktime_t *expires, unsigned long delta,
+                              const enum hrtimer_mode mode, int clock)
 {
        struct hrtimer_sleeper t;
 
@@ -1736,7 +1765,7 @@ int __sched schedule_hrtimeout_range(ktime_t *expires, unsigned long delta,
        }
 
        /*
-        * A NULL parameter means "inifinte"
+        * A NULL parameter means "infinite"
         */
        if (!expires) {
                schedule();
@@ -1744,7 +1773,7 @@ int __sched schedule_hrtimeout_range(ktime_t *expires, unsigned long delta,
                return -EINTR;
        }
 
-       hrtimer_init_on_stack(&t.timer, CLOCK_MONOTONIC, mode);
+       hrtimer_init_on_stack(&t.timer, clock, mode);
        hrtimer_set_expires_range_ns(&t.timer, *expires, delta);
 
        hrtimer_init_sleeper(&t, current);
@@ -1763,6 +1792,41 @@ int __sched schedule_hrtimeout_range(ktime_t *expires, unsigned long delta,
 
        return !t.task ? 0 : -EINTR;
 }
+
+/**
+ * schedule_hrtimeout_range - sleep until timeout
+ * @expires:   timeout value (ktime_t)
+ * @delta:     slack in expires timeout (ktime_t)
+ * @mode:      timer mode, HRTIMER_MODE_ABS or HRTIMER_MODE_REL
+ *
+ * Make the current task sleep until the given expiry time has
+ * elapsed. The routine will return immediately unless
+ * the current task state has been set (see set_current_state()).
+ *
+ * The @delta argument gives the kernel the freedom to schedule the
+ * actual wakeup to a time that is both power and performance friendly.
+ * The kernel give the normal best effort behavior for "@expires+@delta",
+ * but may decide to fire the timer earlier, but no earlier than @expires.
+ *
+ * You can set the task state as follows -
+ *
+ * %TASK_UNINTERRUPTIBLE - at least @timeout time is guaranteed to
+ * pass before the routine returns.
+ *
+ * %TASK_INTERRUPTIBLE - the routine may return early if a signal is
+ * delivered to the current task.
+ *
+ * The current task state is guaranteed to be TASK_RUNNING when this
+ * routine returns.
+ *
+ * Returns 0 when the timer has expired otherwise -EINTR
+ */
+int __sched schedule_hrtimeout_range(ktime_t *expires, unsigned long delta,
+                                    const enum hrtimer_mode mode)
+{
+       return schedule_hrtimeout_range_clock(expires, delta, mode,
+                                             CLOCK_MONOTONIC);
+}
 EXPORT_SYMBOL_GPL(schedule_hrtimeout_range);
 
 /**