cpu-hotplug: CPUx should be active before it is marked online
[linux-2.6.git] / kernel / hrtimer.c
index 29de5ae..2043c08 100644 (file)
 /*
  * 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) =
 {
@@ -65,39 +64,55 @@ 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,
                },
        }
 };
 
+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,
+};
+
+static inline int hrtimer_clockid_to_base(clockid_t clock_id)
+{
+       return hrtimer_clock_to_base_table[clock_id];
+}
+
+
 /*
  * Get the coarse grained time at the softirq based on xtime and
  * wall_to_monotonic.
  */
 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 = __get_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;
 }
 
 /*
@@ -184,10 +199,11 @@ switch_hrtimer_base(struct hrtimer *timer, struct hrtimer_clock_base *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;
 
 again:
        new_cpu_base = &per_cpu(hrtimer_bases, cpu);
-       new_base = &new_cpu_base->clock_base[base->index];
+       new_base = &new_cpu_base->clock_base[basenum];
 
        if (base != new_base) {
                /*
@@ -334,6 +350,11 @@ EXPORT_SYMBOL_GPL(ktime_add_safe);
 
 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
@@ -393,6 +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,
@@ -516,10 +538,13 @@ hrtimer_force_reprogram(struct hrtimer_cpu_base *cpu_base, int skip_equal)
 
        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);
                /*
                 * clock_was_set() has changed base->offset so the
@@ -599,67 +624,6 @@ static int hrtimer_reprogram(struct hrtimer *timer,
        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, wtm;
-       unsigned long seq;
-
-       if (!hrtimer_hres_active())
-               return;
-
-       do {
-               seq = read_seqbegin(&xtime_lock);
-               wtm = __get_wall_to_monotonic();
-       } while (read_seqretry(&xtime_lock, seq));
-       set_normalized_timespec(&realtime_offset, -wtm.tv_sec, -wtm.tv_nsec);
-
-       base = &__get_cpu_var(hrtimer_bases);
-
-       /* Adjust CLOCK_REALTIME offset */
-       raw_spin_lock(&base->lock);
-       base->clock_base[CLOCK_REALTIME].offset =
-               timespec_to_ktime(realtime_offset);
-
-       hrtimer_force_reprogram(base, 0);
-       raw_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)
-{
-       WARN_ONCE(!irqs_disabled(),
-                 KERN_INFO "hres_timers_resume() called with IRQs enabled!");
-
-       retrigger_next_event(NULL);
-}
-
 /*
  * Initialize the high resolution related parts of cpu_base
  */
@@ -670,14 +634,6 @@ 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
@@ -702,11 +658,39 @@ static inline int hrtimer_enqueue_reprogram(struct hrtimer *timer,
 }
 
 /*
+ * 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;
 
@@ -722,8 +706,8 @@ 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();
 
@@ -747,10 +731,43 @@ static inline int hrtimer_enqueue_reprogram(struct hrtimer *timer,
        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 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)
+{
+#ifdef CONFIG_HIGH_RES_TIMERS
+       /* Retrigger the CPU local events everywhere */
+       on_each_cpu(retrigger_next_event, NULL, 1);
+#endif
+       timerfd_clock_was_set();
+}
+
+/*
+ * 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!");
+
+       retrigger_next_event(NULL);
+       timerfd_clock_was_set();
+}
+
 static inline void timer_stats_hrtimer_set_start_info(struct hrtimer *timer)
 {
 #ifdef CONFIG_TIMER_STATS
@@ -840,48 +857,18 @@ EXPORT_SYMBOL_GPL(hrtimer_forward);
 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_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)
-               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 leftmost;
+       return (&timer->node == base->active.next);
 }
 
 /*
@@ -898,15 +885,13 @@ static void __remove_hrtimer(struct hrtimer *timer,
                             struct hrtimer_clock_base *base,
                             unsigned long newstate, int reprogram)
 {
+       struct timerqueue_node *next_timer;
        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);
+       next_timer = timerqueue_getnext(&base->active);
+       timerqueue_del(&base->active, &timer->node);
+       if (&timer->node == next_timer) {
 #ifdef CONFIG_HIGH_RES_TIMERS
                /* Reprogram the clock event device. if enabled */
                if (reprogram && hrtimer_hres_active()) {
@@ -919,7 +904,8 @@ static void __remove_hrtimer(struct hrtimer *timer,
                }
 #endif
        }
-       rb_erase(&timer->node, &base->active);
+       if (!timerqueue_getnext(&base->active))
+               base->cpu_base->active_bases &= ~(1 << base->index);
 out:
        timer->state = newstate;
 }
@@ -1128,11 +1114,13 @@ ktime_t hrtimer_get_next_event(void)
        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)
@@ -1152,6 +1140,7 @@ 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));
 
@@ -1160,8 +1149,9 @@ 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];
-       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;
@@ -1195,9 +1185,10 @@ 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;
 }
@@ -1252,7 +1243,6 @@ static void __run_hrtimer(struct hrtimer *timer, ktime_t *now)
 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, entry_time, delta;
        int i, retries = 0;
 
@@ -1274,18 +1264,21 @@ retry:
         */
        cpu_base->expires_next.tv64 = KTIME_MAX;
 
-       base = cpu_base->clock_base;
-
        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;
 
+               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
@@ -1312,7 +1305,6 @@ retry:
 
                        __run_hrtimer(timer, &basenow);
                }
-               base++;
        }
 
        /*
@@ -1441,7 +1433,7 @@ 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;
@@ -1451,8 +1443,7 @@ 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 (gettime) {
@@ -1462,10 +1453,10 @@ void hrtimer_run_queues(void)
 
                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;
@@ -1544,7 +1535,7 @@ 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);
 
@@ -1596,7 +1587,7 @@ 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);
 
@@ -1630,8 +1621,10 @@ static void __cpuinit init_hrtimers_cpu(int cpu)
 
        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);
+       }
 
        hrtimer_init_hres(cpu_base);
 }
@@ -1642,10 +1635,10 @@ 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_deactivate(timer);
 
@@ -1774,7 +1767,7 @@ schedule_hrtimeout_range_clock(ktime_t *expires, unsigned long delta,
        }
 
        /*
-        * A NULL parameter means "inifinte"
+        * A NULL parameter means "infinite"
         */
        if (!expires) {
                schedule();