Merge branch 'perf/nmi' into perf/core
[pandora-kernel.git] / kernel / timer.c
index e8de5eb..c29e2d4 100644 (file)
@@ -319,6 +319,24 @@ unsigned long round_jiffies_up_relative(unsigned long j)
 }
 EXPORT_SYMBOL_GPL(round_jiffies_up_relative);
 
+/**
+ * set_timer_slack - set the allowed slack for a timer
+ * @slack_hz: the amount of time (in jiffies) allowed for rounding
+ *
+ * Set the amount of time, in jiffies, that a certain timer has
+ * in terms of slack. By setting this value, the timer subsystem
+ * will schedule the actual timer somewhere between
+ * the time mod_timer() asks for, and that time plus the slack.
+ *
+ * By setting the slack to -1, a percentage of the delay is used
+ * instead.
+ */
+void set_timer_slack(struct timer_list *timer, int slack_hz)
+{
+       timer->slack = slack_hz;
+}
+EXPORT_SYMBOL_GPL(set_timer_slack);
+
 
 static inline void set_running_timer(struct tvec_base *base,
                                        struct timer_list *timer)
@@ -550,6 +568,7 @@ static void __init_timer(struct timer_list *timer,
 {
        timer->entry.next = NULL;
        timer->base = __raw_get_cpu_var(tvec_bases);
+       timer->slack = -1;
 #ifdef CONFIG_TIMER_STATS
        timer->start_site = NULL;
        timer->start_pid = -1;
@@ -715,6 +734,46 @@ int mod_timer_pending(struct timer_list *timer, unsigned long expires)
 }
 EXPORT_SYMBOL(mod_timer_pending);
 
+/*
+ * Decide where to put the timer while taking the slack into account
+ *
+ * Algorithm:
+ *   1) calculate the maximum (absolute) time
+ *   2) calculate the highest bit where the expires and new max are different
+ *   3) use this bit to make a mask
+ *   4) use the bitmask to round down the maximum time, so that all last
+ *      bits are zeros
+ */
+static inline
+unsigned long apply_slack(struct timer_list *timer, unsigned long expires)
+{
+       unsigned long expires_limit, mask;
+       int bit;
+
+       expires_limit = expires;
+
+       if (timer->slack >= 0) {
+               expires_limit = expires + timer->slack;
+       } else {
+               unsigned long now = jiffies;
+
+               /* No slack, if already expired else auto slack 0.4% */
+               if (time_after(expires, now))
+                       expires_limit = expires + (expires - now)/256;
+       }
+       mask = expires ^ expires_limit;
+       if (mask == 0)
+               return expires;
+
+       bit = find_last_bit(&mask, BITS_PER_LONG);
+
+       mask = (1 << bit) - 1;
+
+       expires_limit = expires_limit & ~(mask);
+
+       return expires_limit;
+}
+
 /**
  * mod_timer - modify a timer's timeout
  * @timer: the timer to be modified
@@ -745,6 +804,8 @@ int mod_timer(struct timer_list *timer, unsigned long expires)
        if (timer_pending(timer) && timer->expires == expires)
                return 1;
 
+       expires = apply_slack(timer, expires);
+
        return __mod_timer(timer, expires, false, TIMER_NOT_PINNED);
 }
 EXPORT_SYMBOL(mod_timer);
@@ -955,6 +1016,47 @@ static int cascade(struct tvec_base *base, struct tvec *tv, int index)
        return index;
 }
 
+static void call_timer_fn(struct timer_list *timer, void (*fn)(unsigned long),
+                         unsigned long data)
+{
+       int preempt_count = preempt_count();
+
+#ifdef CONFIG_LOCKDEP
+       /*
+        * It is permissible to free the timer from inside the
+        * function that is called from it, this we need to take into
+        * account for lockdep too. To avoid bogus "held lock freed"
+        * warnings as well as problems when looking into
+        * timer->lockdep_map, make a copy and use that here.
+        */
+       struct lockdep_map lockdep_map = timer->lockdep_map;
+#endif
+       /*
+        * Couple the lock chain with the lock chain at
+        * del_timer_sync() by acquiring the lock_map around the fn()
+        * call here and in del_timer_sync().
+        */
+       lock_map_acquire(&lockdep_map);
+
+       trace_timer_expire_entry(timer);
+       fn(data);
+       trace_timer_expire_exit(timer);
+
+       lock_map_release(&lockdep_map);
+
+       if (preempt_count != preempt_count()) {
+               WARN_ONCE(1, "timer: %pF preempt leak: %08x -> %08x\n",
+                         fn, preempt_count, preempt_count());
+               /*
+                * Restore the preempt count. That gives us a decent
+                * chance to survive and extract information. If the
+                * callback kept a lock held, bad luck, but not worse
+                * than the BUG() we had.
+                */
+               preempt_count() = preempt_count;
+       }
+}
+
 #define INDEX(N) ((base->timer_jiffies >> (TVR_BITS + (N) * TVN_BITS)) & TVN_MASK)
 
 /**
@@ -998,45 +1100,7 @@ static inline void __run_timers(struct tvec_base *base)
                        detach_timer(timer, 1);
 
                        spin_unlock_irq(&base->lock);
-                       {
-                               int preempt_count = preempt_count();
-
-#ifdef CONFIG_LOCKDEP
-                               /*
-                                * It is permissible to free the timer from
-                                * inside the function that is called from
-                                * it, this we need to take into account for
-                                * lockdep too. To avoid bogus "held lock
-                                * freed" warnings as well as problems when
-                                * looking into timer->lockdep_map, make a
-                                * copy and use that here.
-                                */
-                               struct lockdep_map lockdep_map =
-                                       timer->lockdep_map;
-#endif
-                               /*
-                                * Couple the lock chain with the lock chain at
-                                * del_timer_sync() by acquiring the lock_map
-                                * around the fn() call here and in
-                                * del_timer_sync().
-                                */
-                               lock_map_acquire(&lockdep_map);
-
-                               trace_timer_expire_entry(timer);
-                               fn(data);
-                               trace_timer_expire_exit(timer);
-
-                               lock_map_release(&lockdep_map);
-
-                               if (preempt_count != preempt_count()) {
-                                       printk(KERN_ERR "huh, entered %p "
-                                              "with preempt_count %08x, exited"
-                                              " with %08x?\n",
-                                              fn, preempt_count,
-                                              preempt_count());
-                                       BUG();
-                               }
-                       }
+                       call_timer_fn(timer, fn, data);
                        spin_lock_irq(&base->lock);
                }
        }
@@ -1619,11 +1683,14 @@ static int __cpuinit timer_cpu_notify(struct notifier_block *self,
                                unsigned long action, void *hcpu)
 {
        long cpu = (long)hcpu;
+       int err;
+
        switch(action) {
        case CPU_UP_PREPARE:
        case CPU_UP_PREPARE_FROZEN:
-               if (init_timers_cpu(cpu) < 0)
-                       return NOTIFY_BAD;
+               err = init_timers_cpu(cpu);
+               if (err < 0)
+                       return notifier_from_errno(err);
                break;
 #ifdef CONFIG_HOTPLUG_CPU
        case CPU_DEAD:
@@ -1649,7 +1716,7 @@ void __init init_timers(void)
 
        init_timer_stats();
 
-       BUG_ON(err == NOTIFY_BAD);
+       BUG_ON(err != NOTIFY_OK);
        register_cpu_notifier(&timers_nb);
        open_softirq(TIMER_SOFTIRQ, run_timer_softirq);
 }