[PATCH] fix next_timer_interrupt() for hrtimer
[pandora-kernel.git] / kernel / hrtimer.c
index 690efd9..14bc9cf 100644 (file)
  *  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
  */
 
@@ -66,6 +72,12 @@ 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
@@ -108,6 +120,7 @@ void ktime_get_ts(struct timespec *ts)
        set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec,
                                ts->tv_nsec + tomono.tv_nsec);
 }
+EXPORT_SYMBOL_GPL(ktime_get_ts);
 
 /*
  * Functions and macros which are different for UP/SMP systems are kept in a
@@ -271,10 +284,10 @@ void unlock_hrtimer_base(const struct hrtimer *timer, unsigned long *flags)
  * @interval:  the interval to forward
  *
  * Forward the timer expiry so it will expire in the future.
- * The number of overruns is added to the overrun field.
+ * Returns the number of overruns.
  */
 unsigned long
-hrtimer_forward(struct hrtimer *timer, const ktime_t interval)
+hrtimer_forward(struct hrtimer *timer, ktime_t interval)
 {
        unsigned long orun = 1;
        ktime_t delta, now;
@@ -286,6 +299,9 @@ hrtimer_forward(struct hrtimer *timer, const ktime_t interval)
        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)) {
                nsec_t incr = ktime_to_ns(interval);
 
@@ -313,7 +329,6 @@ hrtimer_forward(struct hrtimer *timer, const ktime_t interval)
 static void enqueue_hrtimer(struct hrtimer *timer, struct hrtimer_base *base)
 {
        struct rb_node **link = &base->active.rb_node;
-       struct list_head *prev = &base->pending;
        struct rb_node *parent = NULL;
        struct hrtimer *entry;
 
@@ -329,22 +344,23 @@ static void enqueue_hrtimer(struct hrtimer *timer, struct hrtimer_base *base)
                 */
                if (timer->expires.tv64 < entry->expires.tv64)
                        link = &(*link)->rb_left;
-               else {
+               else
                        link = &(*link)->rb_right;
-                       prev = &entry->list;
-               }
        }
 
        /*
-        * Insert the timer to the rbtree and to the sorted list:
+        * 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);
-       list_add(&timer->list, prev);
 
        timer->state = HRTIMER_PENDING;
-}
 
+       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
@@ -354,9 +370,11 @@ static void enqueue_hrtimer(struct hrtimer *timer, struct hrtimer_base *base)
 static void __remove_hrtimer(struct hrtimer *timer, struct hrtimer_base *base)
 {
        /*
-        * Remove the timer from the sorted list and from the rbtree:
+        * Remove the timer from the rbtree and replace the
+        * first entry pointer if necessary.
         */
-       list_del(&timer->list);
+       if (base->first == &timer->node)
+               base->first = rb_next(&timer->node);
        rb_erase(&timer->node, &base->active);
 }
 
@@ -400,8 +418,19 @@ hrtimer_start(struct hrtimer *timer, ktime_t tim, const enum hrtimer_mode mode)
        /* Switch the timer base, if necessary: */
        new_base = switch_hrtimer_base(timer, base);
 
-       if (mode == HRTIMER_REL)
+       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);
@@ -476,30 +505,61 @@ ktime_t hrtimer_get_remaining(const struct hrtimer *timer)
        return rem;
 }
 
+#ifdef CONFIG_NO_IDLE_HZ
 /**
- * hrtimer_rebase - rebase an initialized hrtimer to a different base
+ * hrtimer_get_next_event - get the time until next expiry event
  *
- * @timer:     the timer to be rebased
- * @clock_id:  the clock to be used
+ * Returns the delta to the next expiry event or KTIME_MAX if no timer
+ * is pending.
  */
-void hrtimer_rebase(struct hrtimer *timer, const clockid_t clock_id)
+ktime_t hrtimer_get_next_event(void)
 {
-       struct hrtimer_base *bases;
+       struct hrtimer_base *base = __get_cpu_var(hrtimer_bases);
+       ktime_t delta, mindelta = { .tv64 = KTIME_MAX };
+       unsigned long flags;
+       int i;
 
-       bases = per_cpu(hrtimer_bases, raw_smp_processor_id());
-       timer->base = &bases[clock_id];
+       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, const clockid_t clock_id)
+void hrtimer_init(struct hrtimer *timer, clockid_t clock_id,
+                 enum hrtimer_mode mode)
 {
+       struct hrtimer_base *bases;
+
        memset(timer, 0, sizeof(struct hrtimer));
-       hrtimer_rebase(timer, clock_id);
+
+       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];
 }
 
 /**
@@ -515,9 +575,8 @@ int hrtimer_get_res(const clockid_t which_clock, struct timespec *tp)
 {
        struct hrtimer_base *bases;
 
-       tp->tv_sec = 0;
        bases = per_cpu(hrtimer_bases, raw_smp_processor_id());
-       tp->tv_nsec = bases[which_clock].resolution;
+       *tp = ktime_to_timespec(bases[which_clock].resolution);
 
        return 0;
 }
@@ -528,22 +587,24 @@ int hrtimer_get_res(const clockid_t which_clock, struct timespec *tp)
 static inline void run_hrtimer_queue(struct hrtimer_base *base)
 {
        ktime_t now = base->get_time();
+       struct rb_node *node;
 
        spin_lock_irq(&base->lock);
 
-       while (!list_empty(&base->pending)) {
+       while ((node = base->first)) {
                struct hrtimer *timer;
                int (*fn)(void *);
                int restart;
                void *data;
 
-               timer = list_entry(base->pending.next, struct hrtimer, list);
+               timer = rb_entry(node, struct hrtimer, node);
                if (now.tv64 <= timer->expires.tv64)
                        break;
 
                fn = timer->function;
                data = timer->data;
                set_curr_timer(base, timer);
+               timer->state = HRTIMER_RUNNING;
                __remove_hrtimer(timer, base);
                spin_unlock_irq(&base->lock);
 
@@ -559,6 +620,10 @@ static inline void run_hrtimer_queue(struct hrtimer_base *base)
 
                spin_lock_irq(&base->lock);
 
+               /* Another CPU has added back the timer */
+               if (timer->state != HRTIMER_RUNNING)
+                       continue;
+
                if (restart == HRTIMER_RESTART)
                        enqueue_hrtimer(timer, base);
                else
@@ -580,6 +645,137 @@ void hrtimer_run_queues(void)
                run_hrtimer_queue(&base[i]);
 }
 
+/*
+ * Sleep related functions:
+ */
+
+/**
+ * schedule_hrtimer - sleep until timeout
+ *
+ * @timer:     hrtimer variable initialized with the correct clock base
+ * @mode:      timeout value is abs/rel
+ *
+ * Make the current task sleep until @timeout is
+ * elapsed.
+ *
+ * You can set the task state as follows -
+ *
+ * %TASK_UNINTERRUPTIBLE - at least @timeout is guaranteed to
+ * pass before the routine returns. The routine will return 0
+ *
+ * %TASK_INTERRUPTIBLE - the routine may return early if a signal is
+ * delivered to the current task. In this case the remaining time
+ * will be returned
+ *
+ * The current task state is guaranteed to be TASK_RUNNING when this
+ * routine returns.
+ */
+static ktime_t __sched
+schedule_hrtimer(struct hrtimer *timer, const enum hrtimer_mode mode)
+{
+       /* fn stays NULL, meaning single-shot wakeup: */
+       timer->data = current;
+
+       hrtimer_start(timer, timer->expires, mode);
+
+       schedule();
+       hrtimer_cancel(timer);
+
+       /* Return the remaining time: */
+       if (timer->state != HRTIMER_EXPIRED)
+               return ktime_sub(timer->expires, timer->base->get_time());
+       else
+               return (ktime_t) {.tv64 = 0 };
+}
+
+static inline ktime_t __sched
+schedule_hrtimer_interruptible(struct hrtimer *timer,
+                              const enum hrtimer_mode mode)
+{
+       set_current_state(TASK_INTERRUPTIBLE);
+
+       return schedule_hrtimer(timer, mode);
+}
+
+static long __sched nanosleep_restart(struct restart_block *restart)
+{
+       struct timespec __user *rmtp;
+       struct timespec tu;
+       void *rfn_save = restart->fn;
+       struct hrtimer timer;
+       ktime_t rem;
+
+       restart->fn = do_no_restart_syscall;
+
+       hrtimer_init(&timer, (clockid_t) restart->arg3, HRTIMER_ABS);
+
+       timer.expires.tv64 = ((u64)restart->arg1 << 32) | (u64) restart->arg0;
+
+       rem = schedule_hrtimer_interruptible(&timer, HRTIMER_ABS);
+
+       if (rem.tv64 <= 0)
+               return 0;
+
+       rmtp = (struct timespec __user *) restart->arg2;
+       tu = ktime_to_timespec(rem);
+       if (rmtp && copy_to_user(rmtp, &tu, sizeof(tu)))
+               return -EFAULT;
+
+       restart->fn = rfn_save;
+
+       /* 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 timer;
+       struct timespec tu;
+       ktime_t rem;
+
+       hrtimer_init(&timer, clockid, mode);
+
+       timer.expires = timespec_to_ktime(*rqtp);
+
+       rem = schedule_hrtimer_interruptible(&timer, mode);
+       if (rem.tv64 <= 0)
+               return 0;
+
+       /* Absolute timers do not update the rmtp value and restart: */
+       if (mode == HRTIMER_ABS)
+               return -ERESTARTNOHAND;
+
+       tu = ktime_to_timespec(rem);
+
+       if (rmtp && copy_to_user(rmtp, &tu, sizeof(tu)))
+               return -EFAULT;
+
+       restart = &current_thread_info()->restart_block;
+       restart->fn = nanosleep_restart;
+       restart->arg0 = timer.expires.tv64 & 0xFFFFFFFF;
+       restart->arg1 = timer.expires.tv64 >> 32;
+       restart->arg2 = (unsigned long) rmtp;
+       restart->arg3 = (unsigned long) 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:
  */
@@ -588,11 +784,8 @@ 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++) {
+       for (i = 0; i < MAX_HRTIMER_BASES; i++, base++)
                spin_lock_init(&base->lock);
-               INIT_LIST_HEAD(&base->pending);
-               base++;
-       }
 }
 
 #ifdef CONFIG_HOTPLUG_CPU