timers: Forward the wheel clock whenever possible
authorThomas Gleixner <tglx@linutronix.de>
Mon, 4 Jul 2016 09:50:36 +0000 (09:50 +0000)
committerIngo Molnar <mingo@kernel.org>
Thu, 7 Jul 2016 08:35:11 +0000 (10:35 +0200)
The wheel clock is stale when a CPU goes into a long idle sleep. This has the
side effect that timers which are queued end up in the outer wheel levels.
That results in coarser granularity.

To solve this, we keep track of the idle state and forward the wheel clock
whenever possible.

Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: George Spelvin <linux@sciencehorizons.net>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Len Brown <lenb@kernel.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160704094342.512039360@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
kernel/time/tick-internal.h
kernel/time/tick-sched.c
kernel/time/timer.c

index 966a5a6..f738251 100644 (file)
@@ -164,3 +164,4 @@ static inline void timers_update_migration(bool update_nohz) { }
 DECLARE_PER_CPU(struct hrtimer_cpu_base, hrtimer_bases);
 
 extern u64 get_next_timer_interrupt(unsigned long basej, u64 basem);
+void timer_clear_idle(void);
index 69abc7b..5d81f9a 100644 (file)
@@ -700,6 +700,12 @@ static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts,
        delta = next_tick - basemono;
        if (delta <= (u64)TICK_NSEC) {
                tick.tv64 = 0;
+
+               /*
+                * Tell the timer code that the base is not idle, i.e. undo
+                * the effect of get_next_timer_interrupt():
+                */
+               timer_clear_idle();
                /*
                 * We've not stopped the tick yet, and there's a timer in the
                 * next period, so no point in stopping it either, bail.
@@ -809,6 +815,12 @@ static void tick_nohz_restart_sched_tick(struct tick_sched *ts, ktime_t now)
        tick_do_update_jiffies64(now);
        cpu_load_update_nohz_stop();
 
+       /*
+        * Clear the timer idle flag, so we avoid IPIs on remote queueing and
+        * the clock forward checks in the enqueue path:
+        */
+       timer_clear_idle();
+
        calc_load_exit_idle();
        touch_softlockup_watchdog_sched();
        /*
index 658051c..9339d71 100644 (file)
@@ -196,9 +196,11 @@ struct timer_base {
        spinlock_t              lock;
        struct timer_list       *running_timer;
        unsigned long           clk;
+       unsigned long           next_expiry;
        unsigned int            cpu;
        bool                    migration_enabled;
        bool                    nohz_active;
+       bool                    is_idle;
        DECLARE_BITMAP(pending_map, WHEEL_SIZE);
        struct hlist_head       vectors[WHEEL_SIZE];
 } ____cacheline_aligned;
@@ -519,24 +521,37 @@ static void internal_add_timer(struct timer_base *base, struct timer_list *timer
 {
        __internal_add_timer(base, timer);
 
+       if (!IS_ENABLED(CONFIG_NO_HZ_COMMON) || !base->nohz_active)
+               return;
+
        /*
-        * Check whether the other CPU is in dynticks mode and needs
-        * to be triggered to reevaluate the timer wheel.  We are
-        * protected against the other CPU fiddling with the timer by
-        * holding the timer base lock. This also makes sure that a
-        * CPU on the way to stop its tick can not evaluate the timer
-        * wheel.
-        *
-        * Spare the IPI for deferrable timers on idle targets though.
-        * The next busy ticks will take care of it. Except full dynticks
-        * require special care against races with idle_cpu(), lets deal
-        * with that later.
+        * TODO: This wants some optimizing similar to the code below, but we
+        * will do that when we switch from push to pull for deferrable timers.
         */
-       if (IS_ENABLED(CONFIG_NO_HZ_COMMON) && base->nohz_active) {
-               if (!(timer->flags & TIMER_DEFERRABLE) ||
-                   tick_nohz_full_cpu(base->cpu))
+       if (timer->flags & TIMER_DEFERRABLE) {
+               if (tick_nohz_full_cpu(base->cpu))
                        wake_up_nohz_cpu(base->cpu);
+               return;
        }
+
+       /*
+        * We might have to IPI the remote CPU if the base is idle and the
+        * timer is not deferrable. If the other CPU is on the way to idle
+        * then it can't set base->is_idle as we hold the base lock:
+        */
+       if (!base->is_idle)
+               return;
+
+       /* Check whether this is the new first expiring timer: */
+       if (time_after_eq(timer->expires, base->next_expiry))
+               return;
+
+       /*
+        * Set the next expiry time and kick the CPU so it can reevaluate the
+        * wheel:
+        */
+       base->next_expiry = timer->expires;
+       wake_up_nohz_cpu(base->cpu);
 }
 
 #ifdef CONFIG_TIMER_STATS
@@ -844,10 +859,11 @@ static inline struct timer_base *get_timer_base(u32 tflags)
        return get_timer_cpu_base(tflags, tflags & TIMER_CPUMASK);
 }
 
-static inline struct timer_base *get_target_base(struct timer_base *base,
-                                                unsigned tflags)
+#ifdef CONFIG_NO_HZ_COMMON
+static inline struct timer_base *
+__get_target_base(struct timer_base *base, unsigned tflags)
 {
-#if defined(CONFIG_NO_HZ_COMMON) && defined(CONFIG_SMP)
+#ifdef CONFIG_SMP
        if ((tflags & TIMER_PINNED) || !base->migration_enabled)
                return get_timer_this_cpu_base(tflags);
        return get_timer_cpu_base(tflags, get_nohz_timer_target());
@@ -856,6 +872,43 @@ static inline struct timer_base *get_target_base(struct timer_base *base,
 #endif
 }
 
+static inline void forward_timer_base(struct timer_base *base)
+{
+       /*
+        * We only forward the base when it's idle and we have a delta between
+        * base clock and jiffies.
+        */
+       if (!base->is_idle || (long) (jiffies - base->clk) < 2)
+               return;
+
+       /*
+        * If the next expiry value is > jiffies, then we fast forward to
+        * jiffies otherwise we forward to the next expiry value.
+        */
+       if (time_after(base->next_expiry, jiffies))
+               base->clk = jiffies;
+       else
+               base->clk = base->next_expiry;
+}
+#else
+static inline struct timer_base *
+__get_target_base(struct timer_base *base, unsigned tflags)
+{
+       return get_timer_this_cpu_base(tflags);
+}
+
+static inline void forward_timer_base(struct timer_base *base) { }
+#endif
+
+static inline struct timer_base *
+get_target_base(struct timer_base *base, unsigned tflags)
+{
+       struct timer_base *target = __get_target_base(base, tflags);
+
+       forward_timer_base(target);
+       return target;
+}
+
 /*
  * We are using hashed locking: Holding per_cpu(timer_bases[x]).lock means
  * that all timers which are tied to this base are locked, and the base itself
@@ -1417,16 +1470,49 @@ u64 get_next_timer_interrupt(unsigned long basej, u64 basem)
 
        spin_lock(&base->lock);
        nextevt = __next_timer_interrupt(base);
-       spin_unlock(&base->lock);
+       base->next_expiry = nextevt;
+       /*
+        * We have a fresh next event. Check whether we can forward the base:
+        */
+       if (time_after(nextevt, jiffies))
+               base->clk = jiffies;
+       else if (time_after(nextevt, base->clk))
+               base->clk = nextevt;
 
-       if (time_before_eq(nextevt, basej))
+       if (time_before_eq(nextevt, basej)) {
                expires = basem;
-       else
+               base->is_idle = false;
+       } else {
                expires = basem + (nextevt - basej) * TICK_NSEC;
+               /*
+                * If we expect to sleep more than a tick, mark the base idle:
+                */
+               if ((expires - basem) > TICK_NSEC)
+                       base->is_idle = true;
+       }
+       spin_unlock(&base->lock);
 
        return cmp_next_hrtimer_event(basem, expires);
 }
 
+/**
+ * timer_clear_idle - Clear the idle state of the timer base
+ *
+ * Called with interrupts disabled
+ */
+void timer_clear_idle(void)
+{
+       struct timer_base *base = this_cpu_ptr(&timer_bases[BASE_STD]);
+
+       /*
+        * We do this unlocked. The worst outcome is a remote enqueue sending
+        * a pointless IPI, but taking the lock would just make the window for
+        * sending the IPI a few instructions smaller for the cost of taking
+        * the lock in the exit from idle path.
+        */
+       base->is_idle = false;
+}
+
 static int collect_expired_timers(struct timer_base *base,
                                  struct hlist_head *heads)
 {
@@ -1440,7 +1526,7 @@ static int collect_expired_timers(struct timer_base *base,
 
                /*
                 * If the next timer is ahead of time forward to current
-                * jiffies, otherwise forward to the next expiry time.
+                * jiffies, otherwise forward to the next expiry time:
                 */
                if (time_after(next, jiffies)) {
                        /* The call site will increment clock! */