[PATCH] timers fixes/improvements
[linux-2.6.git] / kernel / timer.c
index 207aa4f..8aadc62 100644 (file)
@@ -57,6 +57,11 @@ static void time_interpolator_update(long delta_nsec);
 #define TVN_MASK (TVN_SIZE - 1)
 #define TVR_MASK (TVR_SIZE - 1)
 
+struct timer_base_s {
+       spinlock_t lock;
+       struct timer_list *running_timer;
+};
+
 typedef struct tvec_s {
        struct list_head vec[TVN_SIZE];
 } tvec_t;
@@ -66,9 +71,8 @@ typedef struct tvec_root_s {
 } tvec_root_t;
 
 struct tvec_t_base_s {
-       spinlock_t lock;
+       struct timer_base_s t_base;
        unsigned long timer_jiffies;
-       struct timer_list *running_timer;
        tvec_root_t tv1;
        tvec_t tv2;
        tvec_t tv3;
@@ -77,18 +81,16 @@ struct tvec_t_base_s {
 } ____cacheline_aligned_in_smp;
 
 typedef struct tvec_t_base_s tvec_base_t;
+static DEFINE_PER_CPU(tvec_base_t, tvec_bases);
 
 static inline void set_running_timer(tvec_base_t *base,
                                        struct timer_list *timer)
 {
 #ifdef CONFIG_SMP
-       base->running_timer = timer;
+       base->t_base.running_timer = timer;
 #endif
 }
 
-/* Fake initialization */
-static DEFINE_PER_CPU(tvec_base_t, tvec_bases) = { SPIN_LOCK_UNLOCKED };
-
 static void check_timer_failed(struct timer_list *timer)
 {
        static int whine_count;
@@ -103,7 +105,6 @@ static void check_timer_failed(struct timer_list *timer)
        /*
         * Now fix it up
         */
-       spin_lock_init(&timer->lock);
        timer->magic = TIMER_MAGIC;
 }
 
@@ -156,65 +157,113 @@ static void internal_add_timer(tvec_base_t *base, struct timer_list *timer)
        list_add_tail(&timer->entry, vec);
 }
 
+typedef struct timer_base_s timer_base_t;
+/*
+ * Used by TIMER_INITIALIZER, we can't use per_cpu(tvec_bases)
+ * at compile time, and we need timer->base to lock the timer.
+ */
+timer_base_t __init_timer_base
+       ____cacheline_aligned_in_smp = { .lock = SPIN_LOCK_UNLOCKED };
+EXPORT_SYMBOL(__init_timer_base);
+
+/***
+ * init_timer - initialize a timer.
+ * @timer: the timer to be initialized
+ *
+ * init_timer() must be done to a timer prior calling *any* of the
+ * other timer functions.
+ */
+void fastcall init_timer(struct timer_list *timer)
+{
+       timer->entry.next = NULL;
+       timer->base = &per_cpu(tvec_bases, raw_smp_processor_id()).t_base;
+       timer->magic = TIMER_MAGIC;
+}
+EXPORT_SYMBOL(init_timer);
+
+static inline void detach_timer(struct timer_list *timer,
+                                       int clear_pending)
+{
+       struct list_head *entry = &timer->entry;
+
+       __list_del(entry->prev, entry->next);
+       if (clear_pending)
+               entry->next = NULL;
+       entry->prev = LIST_POISON2;
+}
+
+/*
+ * We are using hashed locking: holding per_cpu(tvec_bases).t_base.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 ->tvX lists.
+ *
+ * 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 timer_base_t *lock_timer_base(struct timer_list *timer,
+                                       unsigned long *flags)
+{
+       timer_base_t *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();
+       }
+}
+
 int __mod_timer(struct timer_list *timer, unsigned long expires)
 {
-       tvec_base_t *old_base, *new_base;
+       timer_base_t *base;
+       tvec_base_t *new_base;
        unsigned long flags;
        int ret = 0;
 
        BUG_ON(!timer->function);
-
        check_timer(timer);
 
-       spin_lock_irqsave(&timer->lock, flags);
+       base = lock_timer_base(timer, &flags);
+
+       if (timer_pending(timer)) {
+               detach_timer(timer, 0);
+               ret = 1;
+       }
+
        new_base = &__get_cpu_var(tvec_bases);
-repeat:
-       old_base = timer->base;
 
-       /*
-        * Prevent deadlocks via ordering by old_base < new_base.
-        */
-       if (old_base && (new_base != old_base)) {
-               if (old_base < new_base) {
-                       spin_lock(&new_base->lock);
-                       spin_lock(&old_base->lock);
-               } else {
-                       spin_lock(&old_base->lock);
-                       spin_lock(&new_base->lock);
-               }
+       if (base != &new_base->t_base) {
                /*
-                * The timer base might have been cancelled while we were
-                * trying to take the lock(s):
+                * We are trying to schedule the timer on the local CPU.
+                * However we can't change timer's base while it is running,
+                * otherwise del_timer_sync() can't detect that the timer's
+                * handler yet has not finished. This also guarantees that
+                * the timer is serialized wrt itself.
                 */
-               if (timer->base != old_base) {
-                       spin_unlock(&new_base->lock);
-                       spin_unlock(&old_base->lock);
-                       goto repeat;
-               }
-       } else {
-               spin_lock(&new_base->lock);
-               if (timer->base != old_base) {
-                       spin_unlock(&new_base->lock);
-                       goto repeat;
+               if (unlikely(base->running_timer == timer)) {
+                       /* The timer remains on a former base */
+                       new_base = container_of(base, tvec_base_t, t_base);
+               } else {
+                       /* See the comment in lock_timer_base() */
+                       timer->base = NULL;
+                       spin_unlock(&base->lock);
+                       spin_lock(&new_base->t_base.lock);
+                       timer->base = &new_base->t_base;
                }
        }
 
-       /*
-        * Delete the previous timeout (if there was any), and install
-        * the new one:
-        */
-       if (old_base) {
-               list_del(&timer->entry);
-               ret = 1;
-       }
        timer->expires = expires;
        internal_add_timer(new_base, timer);
-       timer->base = new_base;
-
-       if (old_base && (new_base != old_base))
-               spin_unlock(&old_base->lock);
-       spin_unlock(&new_base->lock);
-       spin_unlock_irqrestore(&timer->lock, flags);
+       spin_unlock_irqrestore(&new_base->t_base.lock, flags);
 
        return ret;
 }
@@ -232,15 +281,15 @@ void add_timer_on(struct timer_list *timer, int cpu)
 {
        tvec_base_t *base = &per_cpu(tvec_bases, cpu);
        unsigned long flags;
-  
+
        BUG_ON(timer_pending(timer) || !timer->function);
 
        check_timer(timer);
 
-       spin_lock_irqsave(&base->lock, flags);
+       spin_lock_irqsave(&base->t_base.lock, flags);
+       timer->base = &base->t_base;
        internal_add_timer(base, timer);
-       timer->base = base;
-       spin_unlock_irqrestore(&base->lock, flags);
+       spin_unlock_irqrestore(&base->t_base.lock, flags);
 }
 
 
@@ -295,27 +344,22 @@ EXPORT_SYMBOL(mod_timer);
  */
 int del_timer(struct timer_list *timer)
 {
+       timer_base_t *base;
        unsigned long flags;
-       tvec_base_t *base;
+       int ret = 0;
 
        check_timer(timer);
 
-repeat:
-       base = timer->base;
-       if (!base)
-               return 0;
-       spin_lock_irqsave(&base->lock, flags);
-       if (base != timer->base) {
+       if (timer_pending(timer)) {
+               base = lock_timer_base(timer, &flags);
+               if (timer_pending(timer)) {
+                       detach_timer(timer, 1);
+                       ret = 1;
+               }
                spin_unlock_irqrestore(&base->lock, flags);
-               goto repeat;
        }
-       list_del(&timer->entry);
-       /* Need to make sure that anybody who sees a NULL base also sees the list ops */
-       smp_wmb();
-       timer->base = NULL;
-       spin_unlock_irqrestore(&base->lock, flags);
 
-       return 1;
+       return ret;
 }
 
 EXPORT_SYMBOL(del_timer);
@@ -332,72 +376,39 @@ EXPORT_SYMBOL(del_timer);
  * Synchronization rules: callers must prevent restarting of the timer,
  * otherwise this function is meaningless. It must not be called from
  * interrupt contexts. The caller must not hold locks which would prevent
- * completion of the timer's handler.  Upon exit the timer is not queued and
- * the handler is not running on any CPU.
+ * completion of the timer's handler. The timer's handler must not call
+ * add_timer_on(). Upon exit the timer is not queued and the handler is
+ * not running on any CPU.
  *
  * The function returns whether it has deactivated a pending timer or not.
- *
- * del_timer_sync() is slow and complicated because it copes with timer
- * handlers which re-arm the timer (periodic timers).  If the timer handler
- * is known to not do this (a single shot timer) then use
- * del_singleshot_timer_sync() instead.
  */
 int del_timer_sync(struct timer_list *timer)
 {
-       tvec_base_t *base;
-       int i, ret = 0;
+       timer_base_t *base;
+       unsigned long flags;
+       int ret = -1;
 
        check_timer(timer);
 
-del_again:
-       ret += del_timer(timer);
+       do {
+               base = lock_timer_base(timer, &flags);
 
-       for_each_online_cpu(i) {
-               base = &per_cpu(tvec_bases, i);
-               if (base->running_timer == timer) {
-                       while (base->running_timer == timer) {
-                               cpu_relax();
-                               preempt_check_resched();
-                       }
-                       break;
+               if (base->running_timer == timer)
+                       goto unlock;
+
+               ret = 0;
+               if (timer_pending(timer)) {
+                       detach_timer(timer, 1);
+                       ret = 1;
                }
-       }
-       smp_rmb();
-       if (timer_pending(timer))
-               goto del_again;
+unlock:
+               spin_unlock_irqrestore(&base->lock, flags);
+       } while (ret < 0);
 
        return ret;
 }
-EXPORT_SYMBOL(del_timer_sync);
 
-/***
- * del_singleshot_timer_sync - deactivate a non-recursive timer
- * @timer: the timer to be deactivated
- *
- * This function is an optimization of del_timer_sync for the case where the
- * caller can guarantee the timer does not reschedule itself in its timer
- * function.
- *
- * Synchronization rules: callers must prevent restarting of the timer,
- * otherwise this function is meaningless. It must not be called from
- * interrupt contexts. The caller must not hold locks which wold prevent
- * completion of the timer's handler.  Upon exit the timer is not queued and
- * the handler is not running on any CPU.
- *
- * The function returns whether it has deactivated a pending timer or not.
- */
-int del_singleshot_timer_sync(struct timer_list *timer)
-{
-       int ret = del_timer(timer);
-
-       if (!ret) {
-               ret = del_timer_sync(timer);
-               BUG_ON(ret);
-       }
-
-       return ret;
-}
-EXPORT_SYMBOL(del_singleshot_timer_sync);
+EXPORT_SYMBOL(del_timer_sync);
 #endif
 
 static int cascade(tvec_base_t *base, tvec_t *tv, int index)
@@ -415,7 +426,7 @@ static int cascade(tvec_base_t *base, tvec_t *tv, int index)
                struct timer_list *tmp;
 
                tmp = list_entry(curr, struct timer_list, entry);
-               BUG_ON(tmp->base != base);
+               BUG_ON(tmp->base != &base->t_base);
                curr = curr->next;
                internal_add_timer(base, tmp);
        }
@@ -437,7 +448,7 @@ static inline void __run_timers(tvec_base_t *base)
 {
        struct timer_list *timer;
 
-       spin_lock_irq(&base->lock);
+       spin_lock_irq(&base->t_base.lock);
        while (time_after_eq(jiffies, base->timer_jiffies)) {
                struct list_head work_list = LIST_HEAD_INIT(work_list);
                struct list_head *head = &work_list;
@@ -453,8 +464,7 @@ static inline void __run_timers(tvec_base_t *base)
                        cascade(base, &base->tv5, INDEX(3));
                ++base->timer_jiffies; 
                list_splice_init(base->tv1.vec + index, &work_list);
-repeat:
-               if (!list_empty(head)) {
+               while (!list_empty(head)) {
                        void (*fn)(unsigned long);
                        unsigned long data;
 
@@ -462,11 +472,9 @@ repeat:
                        fn = timer->function;
                        data = timer->data;
 
-                       list_del(&timer->entry);
                        set_running_timer(base, timer);
-                       smp_wmb();
-                       timer->base = NULL;
-                       spin_unlock_irq(&base->lock);
+                       detach_timer(timer, 1);
+                       spin_unlock_irq(&base->t_base.lock);
                        {
                                u32 preempt_count = preempt_count();
                                fn(data);
@@ -475,12 +483,11 @@ repeat:
                                        BUG();
                                }
                        }
-                       spin_lock_irq(&base->lock);
-                       goto repeat;
+                       spin_lock_irq(&base->t_base.lock);
                }
        }
        set_running_timer(base, NULL);
-       spin_unlock_irq(&base->lock);
+       spin_unlock_irq(&base->t_base.lock);
 }
 
 #ifdef CONFIG_NO_IDLE_HZ
@@ -499,7 +506,7 @@ unsigned long next_timer_interrupt(void)
        int i, j;
 
        base = &__get_cpu_var(tvec_bases);
-       spin_lock(&base->lock);
+       spin_lock(&base->t_base.lock);
        expires = base->timer_jiffies + (LONG_MAX >> 1);
        list = 0;
 
@@ -547,7 +554,7 @@ found:
                                expires = nte->expires;
                }
        }
-       spin_unlock(&base->lock);
+       spin_unlock(&base->t_base.lock);
        return expires;
 }
 #endif
@@ -1286,9 +1293,9 @@ static void __devinit init_timers_cpu(int cpu)
 {
        int j;
        tvec_base_t *base;
-       
+
        base = &per_cpu(tvec_bases, cpu);
-       spin_lock_init(&base->lock);
+       spin_lock_init(&base->t_base.lock);
        for (j = 0; j < TVN_SIZE; j++) {
                INIT_LIST_HEAD(base->tv5.vec + j);
                INIT_LIST_HEAD(base->tv4.vec + j);
@@ -1302,22 +1309,16 @@ static void __devinit init_timers_cpu(int cpu)
 }
 
 #ifdef CONFIG_HOTPLUG_CPU
-static int migrate_timer_list(tvec_base_t *new_base, struct list_head *head)
+static void migrate_timer_list(tvec_base_t *new_base, struct list_head *head)
 {
        struct timer_list *timer;
 
        while (!list_empty(head)) {
                timer = list_entry(head->next, struct timer_list, entry);
-               /* We're locking backwards from __mod_timer order here,
-                  beware deadlock. */
-               if (!spin_trylock(&timer->lock))
-                       return 0;
-               list_del(&timer->entry);
+               detach_timer(timer, 0);
+               timer->base = &new_base->t_base;
                internal_add_timer(new_base, timer);
-               timer->base = new_base;
-               spin_unlock(&timer->lock);
        }
-       return 1;
 }
 
 static void __devinit migrate_timers(int cpu)
@@ -1331,39 +1332,24 @@ static void __devinit migrate_timers(int cpu)
        new_base = &get_cpu_var(tvec_bases);
 
        local_irq_disable();
-again:
-       /* Prevent deadlocks via ordering by old_base < new_base. */
-       if (old_base < new_base) {
-               spin_lock(&new_base->lock);
-               spin_lock(&old_base->lock);
-       } else {
-               spin_lock(&old_base->lock);
-               spin_lock(&new_base->lock);
-       }
+       spin_lock(&new_base->t_base.lock);
+       spin_lock(&old_base->t_base.lock);
 
-       if (old_base->running_timer)
+       if (old_base->t_base.running_timer)
                BUG();
        for (i = 0; i < TVR_SIZE; i++)
-               if (!migrate_timer_list(new_base, old_base->tv1.vec + i))
-                       goto unlock_again;
-       for (i = 0; i < TVN_SIZE; i++)
-               if (!migrate_timer_list(new_base, old_base->tv2.vec + i)
-                   || !migrate_timer_list(new_base, old_base->tv3.vec + i)
-                   || !migrate_timer_list(new_base, old_base->tv4.vec + i)
-                   || !migrate_timer_list(new_base, old_base->tv5.vec + i))
-                       goto unlock_again;
-       spin_unlock(&old_base->lock);
-       spin_unlock(&new_base->lock);
+               migrate_timer_list(new_base, old_base->tv1.vec + i);
+       for (i = 0; i < TVN_SIZE; i++) {
+               migrate_timer_list(new_base, old_base->tv2.vec + i);
+               migrate_timer_list(new_base, old_base->tv3.vec + i);
+               migrate_timer_list(new_base, old_base->tv4.vec + i);
+               migrate_timer_list(new_base, old_base->tv5.vec + i);
+       }
+
+       spin_unlock(&old_base->t_base.lock);
+       spin_unlock(&new_base->t_base.lock);
        local_irq_enable();
        put_cpu_var(tvec_bases);
-       return;
-
-unlock_again:
-       /* Avoid deadlock with __mod_timer, by backing off. */
-       spin_unlock(&old_base->lock);
-       spin_unlock(&new_base->lock);
-       cpu_relax();
-       goto again;
 }
 #endif /* CONFIG_HOTPLUG_CPU */