This commit was manufactured by cvs2svn to create branch 'vserver'.
[linux-2.6.git] / kernel / hrtimer.c
diff --git a/kernel/hrtimer.c b/kernel/hrtimer.c
new file mode 100644 (file)
index 0000000..01fa2ae
--- /dev/null
@@ -0,0 +1,879 @@
+/*
+ *  linux/kernel/hrtimer.c
+ *
+ *  Copyright(C) 2005, Thomas Gleixner <tglx@linutronix.de>
+ *  Copyright(C) 2005, Red Hat, Inc., Ingo Molnar
+ *
+ *  High-resolution kernel timers
+ *
+ *  In contrast to the low-resolution timeout API implemented in
+ *  kernel/timer.c, hrtimers provide finer resolution and accuracy
+ *  depending on system configuration and capabilities.
+ *
+ *  These timers are currently used for:
+ *   - itimers
+ *   - POSIX timers
+ *   - nanosleep
+ *   - precise in-kernel timing
+ *
+ *  Started by: Thomas Gleixner and Ingo Molnar
+ *
+ *  Credits:
+ *     based on kernel/timer.c
+ *
+ *     Help, testing, suggestions, bugfixes, improvements were
+ *     provided by:
+ *
+ *     George Anzinger, Andrew Morton, Steven Rostedt, Roman Zippel
+ *     et. al.
+ *
+ *  For licencing details see kernel-base/COPYING
+ */
+
+#include <linux/cpu.h>
+#include <linux/module.h>
+#include <linux/percpu.h>
+#include <linux/hrtimer.h>
+#include <linux/notifier.h>
+#include <linux/syscalls.h>
+#include <linux/interrupt.h>
+
+#include <asm/uaccess.h>
+
+/**
+ * ktime_get - get the monotonic time in ktime_t format
+ *
+ * returns the time in ktime_t format
+ */
+static ktime_t ktime_get(void)
+{
+       struct timespec now;
+
+       ktime_get_ts(&now);
+
+       return timespec_to_ktime(now);
+}
+
+/**
+ * ktime_get_real - get the real (wall-) time in ktime_t format
+ *
+ * returns the time in ktime_t format
+ */
+static ktime_t ktime_get_real(void)
+{
+       struct timespec now;
+
+       getnstimeofday(&now);
+
+       return timespec_to_ktime(now);
+}
+
+EXPORT_SYMBOL_GPL(ktime_get_real);
+
+/*
+ * 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.
+ */
+
+#define MAX_HRTIMER_BASES 2
+
+static DEFINE_PER_CPU(struct hrtimer_base, hrtimer_bases[MAX_HRTIMER_BASES]) =
+{
+       {
+               .index = CLOCK_REALTIME,
+               .get_time = &ktime_get_real,
+               .resolution = KTIME_REALTIME_RES,
+       },
+       {
+               .index = CLOCK_MONOTONIC,
+               .get_time = &ktime_get,
+               .resolution = KTIME_MONOTONIC_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));
+
+       set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec,
+                               ts->tv_nsec + tomono.tv_nsec);
+}
+EXPORT_SYMBOL_GPL(ktime_get_ts);
+
+/*
+ * Get the coarse grained time at the softirq based on xtime and
+ * wall_to_monotonic.
+ */
+static void hrtimer_get_softirq_time(struct hrtimer_base *base)
+{
+       ktime_t xtim, tomono;
+       unsigned long seq;
+
+       do {
+               seq = read_seqbegin(&xtime_lock);
+               xtim = timespec_to_ktime(xtime);
+               tomono = timespec_to_ktime(wall_to_monotonic);
+
+       } while (read_seqretry(&xtime_lock, seq));
+
+       base[CLOCK_REALTIME].softirq_time = xtim;
+       base[CLOCK_MONOTONIC].softirq_time = ktime_add(xtim, tomono);
+}
+
+/*
+ * Functions and macros which are different for UP/SMP systems are kept in a
+ * single place
+ */
+#ifdef CONFIG_SMP
+
+#define set_curr_timer(b, t)           do { (b)->curr_timer = (t); } while (0)
+
+/*
+ * We are using hashed locking: holding per_cpu(hrtimer_bases)[n].lock
+ * means that all timers which are tied to this base via timer->base are
+ * locked, and the base itself is locked too.
+ *
+ * So __run_timers/migrate_timers can safely modify all timers which could
+ * be found on the lists/queues.
+ *
+ * When the timer's base is locked, and the timer removed from list, it is
+ * possible to set timer->base = NULL and drop the lock: the timer remains
+ * locked.
+ */
+static struct hrtimer_base *lock_hrtimer_base(const struct hrtimer *timer,
+                                             unsigned long *flags)
+{
+       struct hrtimer_base *base;
+
+       for (;;) {
+               base = timer->base;
+               if (likely(base != NULL)) {
+                       spin_lock_irqsave(&base->lock, *flags);
+                       if (likely(base == timer->base))
+                               return base;
+                       /* The timer has migrated to another CPU: */
+                       spin_unlock_irqrestore(&base->lock, *flags);
+               }
+               cpu_relax();
+       }
+}
+
+/*
+ * Switch the timer base to the current CPU when possible.
+ */
+static inline struct hrtimer_base *
+switch_hrtimer_base(struct hrtimer *timer, struct hrtimer_base *base)
+{
+       struct hrtimer_base *new_base;
+
+       new_base = &__get_cpu_var(hrtimer_bases[base->index]);
+
+       if (base != new_base) {
+               /*
+                * We are trying to schedule the timer on the local CPU.
+                * 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
+                * code will take care of this when the timer function has
+                * completed. There is no conflict as we hold the lock until
+                * the timer is enqueued.
+                */
+               if (unlikely(base->curr_timer == timer))
+                       return base;
+
+               /* See the comment in lock_timer_base() */
+               timer->base = NULL;
+               spin_unlock(&base->lock);
+               spin_lock(&new_base->lock);
+               timer->base = new_base;
+       }
+       return new_base;
+}
+
+#else /* CONFIG_SMP */
+
+#define set_curr_timer(b, t)           do { } while (0)
+
+static inline struct hrtimer_base *
+lock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
+{
+       struct hrtimer_base *base = timer->base;
+
+       spin_lock_irqsave(&base->lock, *flags);
+
+       return base;
+}
+
+#define switch_hrtimer_base(t, b)      (b)
+
+#endif /* !CONFIG_SMP */
+
+/*
+ * Functions for the union type storage format of ktime_t which are
+ * too large for inlining:
+ */
+#if BITS_PER_LONG < 64
+# ifndef CONFIG_KTIME_SCALAR
+/**
+ * ktime_add_ns - Add a scalar nanoseconds value to a ktime_t variable
+ *
+ * @kt:                addend
+ * @nsec:      the scalar nsec value to add
+ *
+ * Returns the sum of kt and nsec in ktime_t format
+ */
+ktime_t ktime_add_ns(const ktime_t kt, u64 nsec)
+{
+       ktime_t tmp;
+
+       if (likely(nsec < NSEC_PER_SEC)) {
+               tmp.tv64 = nsec;
+       } else {
+               unsigned long rem = do_div(nsec, NSEC_PER_SEC);
+
+               tmp = ktime_set((long)nsec, rem);
+       }
+
+       return ktime_add(kt, tmp);
+}
+
+#else /* CONFIG_KTIME_SCALAR */
+
+# endif /* !CONFIG_KTIME_SCALAR */
+
+/*
+ * Divide a ktime value by a nanosecond value
+ */
+static unsigned long ktime_divns(const ktime_t kt, s64 div)
+{
+       u64 dclc, inc, dns;
+       int sft = 0;
+
+       dclc = dns = ktime_to_ns(kt);
+       inc = div;
+       /* Make sure the divisor is less than 2^32: */
+       while (div >> 32) {
+               sft++;
+               div >>= 1;
+       }
+       dclc >>= sft;
+       do_div(dclc, (unsigned long) div);
+
+       return (unsigned long) dclc;
+}
+
+#else /* BITS_PER_LONG < 64 */
+# define ktime_divns(kt, div)          (unsigned long)((kt).tv64 / (div))
+#endif /* BITS_PER_LONG >= 64 */
+
+/*
+ * Counterpart to lock_timer_base above:
+ */
+static inline
+void unlock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
+{
+       spin_unlock_irqrestore(&timer->base->lock, *flags);
+}
+
+/**
+ * hrtimer_forward - forward the timer expiry
+ *
+ * @timer:     hrtimer to forward
+ * @now:       forward past this time
+ * @interval:  the interval to forward
+ *
+ * Forward the timer expiry so it will expire in the future.
+ * Returns the number of overruns.
+ */
+unsigned long
+hrtimer_forward(struct hrtimer *timer, ktime_t now, ktime_t interval)
+{
+       unsigned long orun = 1;
+       ktime_t delta;
+
+       delta = ktime_sub(now, timer->expires);
+
+       if (delta.tv64 < 0)
+               return 0;
+
+       if (interval.tv64 < timer->base->resolution.tv64)
+               interval.tv64 = timer->base->resolution.tv64;
+
+       if (unlikely(delta.tv64 >= interval.tv64)) {
+               s64 incr = ktime_to_ns(interval);
+
+               orun = ktime_divns(delta, incr);
+               timer->expires = ktime_add_ns(timer->expires, incr * orun);
+               if (timer->expires.tv64 > now.tv64)
+                       return orun;
+               /*
+                * This (and the ktime_add() below) is the
+                * correction for exact:
+                */
+               orun++;
+       }
+       timer->expires = ktime_add(timer->expires, interval);
+
+       return orun;
+}
+
+/*
+ * enqueue_hrtimer - internal function to (re)start a timer
+ *
+ * The timer is inserted in expiry order. Insertion into the
+ * red black tree is O(log(n)). Must hold the base lock.
+ */
+static void enqueue_hrtimer(struct hrtimer *timer, struct hrtimer_base *base)
+{
+       struct rb_node **link = &base->active.rb_node;
+       struct rb_node *parent = NULL;
+       struct hrtimer *entry;
+
+       /*
+        * 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 (timer->expires.tv64 < entry->expires.tv64)
+                       link = &(*link)->rb_left;
+               else
+                       link = &(*link)->rb_right;
+       }
+
+       /*
+        * Insert the timer to the rbtree and check whether it
+        * replaces the first pending timer
+        */
+       rb_link_node(&timer->node, parent, link);
+       rb_insert_color(&timer->node, &base->active);
+
+       if (!base->first || timer->expires.tv64 <
+           rb_entry(base->first, struct hrtimer, node)->expires.tv64)
+               base->first = &timer->node;
+}
+
+/*
+ * __remove_hrtimer - internal function to remove a timer
+ *
+ * Caller must hold the base lock.
+ */
+static void __remove_hrtimer(struct hrtimer *timer, struct hrtimer_base *base)
+{
+       /*
+        * 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);
+       rb_erase(&timer->node, &base->active);
+       timer->node.rb_parent = HRTIMER_INACTIVE;
+}
+
+/*
+ * remove hrtimer, called with base lock held
+ */
+static inline int
+remove_hrtimer(struct hrtimer *timer, struct hrtimer_base *base)
+{
+       if (hrtimer_active(timer)) {
+               __remove_hrtimer(timer, base);
+               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)
+{
+       struct hrtimer_base *base, *new_base;
+       unsigned long flags;
+       int ret;
+
+       base = lock_hrtimer_base(timer, &flags);
+
+       /* Remove an active timer from the queue: */
+       ret = remove_hrtimer(timer, base);
+
+       /* Switch the timer base, if necessary: */
+       new_base = switch_hrtimer_base(timer, base);
+
+       if (mode == HRTIMER_REL) {
+               tim = ktime_add(tim, new_base->get_time());
+               /*
+                * CONFIG_TIME_LOW_RES is a temporary way for architectures
+                * to signal that they simply return xtime in
+                * do_gettimeoffset(). In this case we want to round up by
+                * resolution when starting a relative timer, to avoid short
+                * timeouts. This will go away with the GTOD framework.
+                */
+#ifdef CONFIG_TIME_LOW_RES
+               tim = ktime_add(tim, base->resolution);
+#endif
+       }
+       timer->expires = tim;
+
+       enqueue_hrtimer(timer, new_base);
+
+       unlock_hrtimer_base(timer, &flags);
+
+       return ret;
+}
+EXPORT_SYMBOL_GPL(hrtimer_start);
+
+/**
+ * hrtimer_try_to_cancel - try to deactivate a timer
+ *
+ * @timer:     hrtimer to stop
+ *
+ * Returns:
+ *  0 when the timer was not active
+ *  1 when the timer was active
+ * -1 when the timer is currently excuting the callback function and
+ *    can not be stopped
+ */
+int hrtimer_try_to_cancel(struct hrtimer *timer)
+{
+       struct hrtimer_base *base;
+       unsigned long flags;
+       int ret = -1;
+
+       base = lock_hrtimer_base(timer, &flags);
+
+       if (base->curr_timer != timer)
+               ret = remove_hrtimer(timer, base);
+
+       unlock_hrtimer_base(timer, &flags);
+
+       return ret;
+
+}
+EXPORT_SYMBOL_GPL(hrtimer_try_to_cancel);
+
+/**
+ * hrtimer_cancel - cancel a timer and wait for the handler to finish.
+ *
+ * @timer:     the timer to be cancelled
+ *
+ * Returns:
+ *  0 when the timer was not active
+ *  1 when the timer was active
+ */
+int hrtimer_cancel(struct hrtimer *timer)
+{
+       for (;;) {
+               int ret = hrtimer_try_to_cancel(timer);
+
+               if (ret >= 0)
+                       return ret;
+               cpu_relax();
+       }
+}
+EXPORT_SYMBOL_GPL(hrtimer_cancel);
+
+/**
+ * hrtimer_get_remaining - get remaining time for the timer
+ *
+ * @timer:     the timer to read
+ */
+ktime_t hrtimer_get_remaining(const struct hrtimer *timer)
+{
+       struct hrtimer_base *base;
+       unsigned long flags;
+       ktime_t rem;
+
+       base = lock_hrtimer_base(timer, &flags);
+       rem = ktime_sub(timer->expires, timer->base->get_time());
+       unlock_hrtimer_base(timer, &flags);
+
+       return rem;
+}
+EXPORT_SYMBOL_GPL(hrtimer_get_remaining);
+
+#ifdef CONFIG_NO_IDLE_HZ
+/**
+ * hrtimer_get_next_event - get the time until next expiry event
+ *
+ * Returns the delta to the next expiry event or KTIME_MAX if no timer
+ * is pending.
+ */
+ktime_t hrtimer_get_next_event(void)
+{
+       struct hrtimer_base *base = __get_cpu_var(hrtimer_bases);
+       ktime_t delta, mindelta = { .tv64 = KTIME_MAX };
+       unsigned long flags;
+       int i;
+
+       for (i = 0; i < MAX_HRTIMER_BASES; i++, base++) {
+               struct hrtimer *timer;
+
+               spin_lock_irqsave(&base->lock, flags);
+               if (!base->first) {
+                       spin_unlock_irqrestore(&base->lock, flags);
+                       continue;
+               }
+               timer = rb_entry(base->first, struct hrtimer, node);
+               delta.tv64 = timer->expires.tv64;
+               spin_unlock_irqrestore(&base->lock, flags);
+               delta = ktime_sub(delta, base->get_time());
+               if (delta.tv64 < mindelta.tv64)
+                       mindelta.tv64 = delta.tv64;
+       }
+       if (mindelta.tv64 < 0)
+               mindelta.tv64 = 0;
+       return mindelta;
+}
+#endif
+
+/**
+ * hrtimer_init - initialize a timer to the given clock
+ *
+ * @timer:     the timer to be initialized
+ * @clock_id:  the clock to be used
+ * @mode:      timer mode abs/rel
+ */
+void hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
+                 enum hrtimer_mode mode)
+{
+       struct hrtimer_base *bases;
+
+       memset(timer, 0, sizeof(struct hrtimer));
+
+       bases = per_cpu(hrtimer_bases, raw_smp_processor_id());
+
+       if (clock_id == CLOCK_REALTIME && mode != HRTIMER_ABS)
+               clock_id = CLOCK_MONOTONIC;
+
+       timer->base = &bases[clock_id];
+       timer->node.rb_parent = HRTIMER_INACTIVE;
+}
+EXPORT_SYMBOL_GPL(hrtimer_init);
+
+/**
+ * hrtimer_get_res - get the timer resolution for a clock
+ *
+ * @which_clock: which clock to query
+ * @tp:                 pointer to timespec variable to store the resolution
+ *
+ * Store the resolution of the clock selected by which_clock in the
+ * variable pointed to by tp.
+ */
+int hrtimer_get_res(const clockid_t which_clock, struct timespec *tp)
+{
+       struct hrtimer_base *bases;
+
+       bases = per_cpu(hrtimer_bases, raw_smp_processor_id());
+       *tp = ktime_to_timespec(bases[which_clock].resolution);
+
+       return 0;
+}
+EXPORT_SYMBOL_GPL(hrtimer_get_res);
+
+/*
+ * Expire the per base hrtimer-queue:
+ */
+static inline void run_hrtimer_queue(struct hrtimer_base *base)
+{
+       struct rb_node *node;
+
+       if (!base->first)
+               return;
+
+       if (base->get_softirq_time)
+               base->softirq_time = base->get_softirq_time();
+
+       spin_lock_irq(&base->lock);
+
+       while ((node = base->first)) {
+               struct hrtimer *timer;
+               int (*fn)(struct hrtimer *);
+               int restart;
+
+               timer = rb_entry(node, struct hrtimer, node);
+               if (base->softirq_time.tv64 <= timer->expires.tv64)
+                       break;
+
+               fn = timer->function;
+               set_curr_timer(base, timer);
+               __remove_hrtimer(timer, base);
+               spin_unlock_irq(&base->lock);
+
+               restart = fn(timer);
+
+               spin_lock_irq(&base->lock);
+
+               if (restart != HRTIMER_NORESTART) {
+                       BUG_ON(hrtimer_active(timer));
+                       enqueue_hrtimer(timer, base);
+               }
+       }
+       set_curr_timer(base, NULL);
+       spin_unlock_irq(&base->lock);
+}
+
+/*
+ * Called from timer softirq every jiffy, expire hrtimers:
+ */
+void hrtimer_run_queues(void)
+{
+       struct hrtimer_base *base = __get_cpu_var(hrtimer_bases);
+       int i;
+
+       hrtimer_get_softirq_time(base);
+
+       for (i = 0; i < MAX_HRTIMER_BASES; i++)
+               run_hrtimer_queue(&base[i]);
+}
+
+/*
+ * Sleep related functions:
+ */
+static int hrtimer_wakeup(struct hrtimer *timer)
+{
+       struct hrtimer_sleeper *t =
+               container_of(timer, struct hrtimer_sleeper, timer);
+       struct task_struct *task = t->task;
+
+       t->task = NULL;
+       if (task)
+               wake_up_process(task);
+
+       return HRTIMER_NORESTART;
+}
+
+void hrtimer_init_sleeper(struct hrtimer_sleeper *sl, task_t *task)
+{
+       sl->timer.function = hrtimer_wakeup;
+       sl->task = task;
+}
+
+static int __sched do_nanosleep(struct hrtimer_sleeper *t, enum hrtimer_mode mode)
+{
+       hrtimer_init_sleeper(t, current);
+
+       do {
+               set_current_state(TASK_INTERRUPTIBLE);
+               hrtimer_start(&t->timer, t->timer.expires, mode);
+
+               schedule();
+
+               hrtimer_cancel(&t->timer);
+               mode = HRTIMER_ABS;
+
+       } while (t->task && !signal_pending(current));
+
+       return t->task == NULL;
+}
+
+static long __sched nanosleep_restart(struct restart_block *restart)
+{
+       struct hrtimer_sleeper t;
+       struct timespec __user *rmtp;
+       struct timespec tu;
+       ktime_t time;
+
+       restart->fn = do_no_restart_syscall;
+
+       hrtimer_init(&t.timer, restart->arg3, HRTIMER_ABS);
+       t.timer.expires.tv64 = ((u64)restart->arg1 << 32) | (u64) restart->arg0;
+
+       if (do_nanosleep(&t, HRTIMER_ABS))
+               return 0;
+
+       rmtp = (struct timespec __user *) restart->arg2;
+       if (rmtp) {
+               time = ktime_sub(t.timer.expires, t.timer.base->get_time());
+               if (time.tv64 <= 0)
+                       return 0;
+               tu = ktime_to_timespec(time);
+               if (copy_to_user(rmtp, &tu, sizeof(tu)))
+                       return -EFAULT;
+       }
+
+       restart->fn = nanosleep_restart;
+
+       /* The other values in restart are already filled in */
+       return -ERESTART_RESTARTBLOCK;
+}
+
+long hrtimer_nanosleep(struct timespec *rqtp, struct timespec __user *rmtp,
+                      const enum hrtimer_mode mode, const clockid_t clockid)
+{
+       struct restart_block *restart;
+       struct hrtimer_sleeper t;
+       struct timespec tu;
+       ktime_t rem;
+
+       hrtimer_init(&t.timer, clockid, mode);
+       t.timer.expires = timespec_to_ktime(*rqtp);
+       if (do_nanosleep(&t, mode))
+               return 0;
+
+       /* Absolute timers do not update the rmtp value and restart: */
+       if (mode == HRTIMER_ABS)
+               return -ERESTARTNOHAND;
+
+       if (rmtp) {
+               rem = ktime_sub(t.timer.expires, t.timer.base->get_time());
+               if (rem.tv64 <= 0)
+                       return 0;
+               tu = ktime_to_timespec(rem);
+               if (copy_to_user(rmtp, &tu, sizeof(tu)))
+                       return -EFAULT;
+       }
+
+       restart = &current_thread_info()->restart_block;
+       restart->fn = nanosleep_restart;
+       restart->arg0 = t.timer.expires.tv64 & 0xFFFFFFFF;
+       restart->arg1 = t.timer.expires.tv64 >> 32;
+       restart->arg2 = (unsigned long) rmtp;
+       restart->arg3 = (unsigned long) t.timer.base->index;
+
+       return -ERESTART_RESTARTBLOCK;
+}
+
+asmlinkage long
+sys_nanosleep(struct timespec __user *rqtp, struct timespec __user *rmtp)
+{
+       struct timespec tu;
+
+       if (copy_from_user(&tu, rqtp, sizeof(tu)))
+               return -EFAULT;
+
+       if (!timespec_valid(&tu))
+               return -EINVAL;
+
+       return hrtimer_nanosleep(&tu, rmtp, HRTIMER_REL, CLOCK_MONOTONIC);
+}
+
+/*
+ * Functions related to boot-time initialization:
+ */
+static void __devinit init_hrtimers_cpu(int cpu)
+{
+       struct hrtimer_base *base = per_cpu(hrtimer_bases, cpu);
+       int i;
+
+       for (i = 0; i < MAX_HRTIMER_BASES; i++, base++)
+               spin_lock_init(&base->lock);
+}
+
+#ifdef CONFIG_HOTPLUG_CPU
+
+static void migrate_hrtimer_list(struct hrtimer_base *old_base,
+                               struct hrtimer_base *new_base)
+{
+       struct hrtimer *timer;
+       struct rb_node *node;
+
+       while ((node = rb_first(&old_base->active))) {
+               timer = rb_entry(node, struct hrtimer, node);
+               __remove_hrtimer(timer, old_base);
+               timer->base = new_base;
+               enqueue_hrtimer(timer, new_base);
+       }
+}
+
+static void migrate_hrtimers(int cpu)
+{
+       struct hrtimer_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);
+
+       local_irq_disable();
+
+       for (i = 0; i < MAX_HRTIMER_BASES; i++) {
+
+               spin_lock(&new_base->lock);
+               spin_lock(&old_base->lock);
+
+               BUG_ON(old_base->curr_timer);
+
+               migrate_hrtimer_list(old_base, new_base);
+
+               spin_unlock(&old_base->lock);
+               spin_unlock(&new_base->lock);
+               old_base++;
+               new_base++;
+       }
+
+       local_irq_enable();
+       put_cpu_var(hrtimer_bases);
+}
+#endif /* CONFIG_HOTPLUG_CPU */
+
+static int hrtimer_cpu_notify(struct notifier_block *self,
+                                       unsigned long action, void *hcpu)
+{
+       long cpu = (long)hcpu;
+
+       switch (action) {
+
+       case CPU_UP_PREPARE:
+               init_hrtimers_cpu(cpu);
+               break;
+
+#ifdef CONFIG_HOTPLUG_CPU
+       case CPU_DEAD:
+               migrate_hrtimers(cpu);
+               break;
+#endif
+
+       default:
+               break;
+       }
+
+       return NOTIFY_OK;
+}
+
+static struct notifier_block hrtimers_nb = {
+       .notifier_call = hrtimer_cpu_notify,
+};
+
+void __init hrtimers_init(void)
+{
+       hrtimer_cpu_notify(&hrtimers_nb, (unsigned long)CPU_UP_PREPARE,
+                         (void *)(long)smp_processor_id());
+       register_cpu_notifier(&hrtimers_nb);
+}
+