On Sat, 22 Mar 2008, Gabriel C wrote:
 > Now some time later CPU1 gets woken by an interrupt/IPI and runs the
quoted text
> > timer wheel. At this point the pm_timer which is the reference clock
> > has already wrapped around, so the watchdog thinks that there is a
> > huge time difference and marks the TSC unstable.
> > 
> > Aside of that watchdog issue this also affects the other users of
> > add_timer_on(): e.g. queue_delayed_work_on().
> > 
> > Can you please apply the patch below and verify it with Andi's
> > watchdog patch applied ? 
> 
> 
> Did that , git head , Andi's + your patch but TSC is still marked unstable.

Doh, stupid me. We do not reevaluate the timer wheel, when we just
wake up via the smp_reschedule IPI when the resched flag on the other
CPU is not set. That's a separate vector which is not going through
irq_enter() / irq_exit(). 

Does the patch below solve the problem ?

Thanks,

	tglx

---
 include/linux/tick.h     |    4 +++
 kernel/time/tick-sched.c |   50 +++++++++++++++++++++++++++++++++++++++++++++++
 kernel/timer.c           |   14 ++++++++++++-
 3 files changed, 67 insertions(+), 1 deletion(-)

Index: linux-2.6/include/linux/tick.h
===================================================================
--- linux-2.6.orig/include/linux/tick.h
+++ linux-2.6/include/linux/tick.h
@@ -111,6 +111,8 @@ extern void tick_nohz_update_jiffies(voi
 extern ktime_t tick_nohz_get_sleep_length(void);
 extern void tick_nohz_stop_idle(int cpu);
 extern u64 get_cpu_idle_time_us(int cpu, u64 *last_update_time);
+extern int tick_nohz_cpu_needs_wakeup(int cpu);
+extern void tick_nohz_rescan_timers_on(int cpu);
 # else
 static inline void tick_nohz_stop_sched_tick(void) { }
 static inline void tick_nohz_restart_sched_tick(void) { }
@@ -123,6 +125,8 @@ static inline ktime_t tick_nohz_get_slee
 }
 static inline void tick_nohz_stop_idle(int cpu) { }
 static inline u64 get_cpu_idle_time_us(int cpu, u64 *unused) { return 0; }
+static inline int tick_nohz_cpu_needs_wakeup(int cpu) { return 0; }
+static inline void tick_nohz_rescan_timers_on(int cpu) { }
 # endif /* !NO_HZ */
 
 #endif
Index: linux-2.6/kernel/time/tick-sched.c
===================================================================
--- linux-2.6.orig/kernel/time/tick-sched.c
+++ linux-2.6/kernel/time/tick-sched.c
@@ -183,6 +183,56 @@ u64 get_cpu_idle_time_us(int cpu, u64 *l
 }
 
 /**
+ * tick_nohz_cpu_needs_wakeup - check possible wakeup of cpu in add_timer_on()
+ *
+ * when add_timer_on() happens on a CPU which is in a long idle sleep,
+ * then we need to wake it up so the timer wheel gets reevaluated.
+ *
+ * Note: we use idle_cpu() which checks the idle state lockless, but
+ * we are ordered against the other cpu which might be on the way to
+ * idle by the timer base lock, which we hold.
+ */
+int tick_nohz_cpu_needs_wakeup(int cpu)
+{
+	return tick_nohz_enabled && idle_cpu(cpu) &&
+		(cpu != smp_processor_id());
+}
+
+/*
+ * Rescan the timer wheel, when
+ *
+ * - the CPU is idle
+ * - the CPU is not processing an interupt
+ * - the need_resched flag is off
+ */
+static void tick_nohz_rescan_timers(void *unused)
+{
+	int cpu = smp_processor_id();
+
+	if (!idle_cpu(cpu) || in_interrupt() || need_resched())
+		return;
+
+	tick_nohz_stop_idle(cpu);
+	tick_nohz_update_jiffies();
+	tick_nohz_stop_sched_tick();
+}
+
+/**
+ * tick_nohz_rescan_timers_on - reevaluate the idle sleep time of a CPU
+ *
+ * When a CPU is idle and a timer got added to this CPU timer wheel
+ * via add_timer_on() then we need to make sure that the CPU
+ * reevaluates the timer wheel. Otherwise the timer might be delayed
+ * for a real long time.
+ */
+void tick_nohz_rescan_timers_on(int cpu)
+{
+	if (tick_nohz_enabled && idle_cpu(cpu))
+		smp_call_function_single(cpu, tick_nohz_rescan_timers, NULL,
+					 0, 0);
+}
+
+/**
  * tick_nohz_stop_sched_tick - stop the idle tick from the idle task
  *
  * When the next event is more than a tick into the future, stop the idle tick
Index: linux-2.6/kernel/timer.c
===================================================================
--- linux-2.6.orig/kernel/timer.c
+++ linux-2.6/kernel/timer.c
@@ -445,15 +445,27 @@ void add_timer_on(struct timer_list *tim
 {
 	struct tvec_base *base = per_cpu(tvec_bases, cpu);
 	unsigned long flags;
+	int wakeidle;
 
 	timer_stats_timer_set_start_info(timer);
 	BUG_ON(timer_pending(timer) || !timer->function);
 	spin_lock_irqsave(&base->lock, flags);
 	timer_set_base(timer, base);
 	internal_add_timer(base, timer);
+	/*
+	 * Check whether the other CPU is idle and needs to be
+	 * triggered to reevaluate the timer wheel when nohz is
+	 * active. 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 idle can not evaluate
+	 * the timer wheel.
+	 */
+	wakeidle = tick_nohz_cpu_needs_wakeup(cpu);
 	spin_unlock_irqrestore(&base->lock, flags);
-}
 
+	if (wakeidle)
+		tick_nohz_rescan_timers_on(cpu);
+}
 
 /**
  * mod_timer - modify a timer's timeout
--
unsubscribe notice
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/