[PATCH] nohz: Fix spurious warning when hrtimer and clockevent get out of sync

[Frederic Weisbecker <fweisbec@gmail.com>]

On Wed, Jun 07, 2017 at 09:36:45PM +0000, Levin, Alexander (Sasha Levin) wrote:
> On Wed, Jun 07, 2017 at 04:14:03PM +0200, Frederic Weisbecker wrote:
> > On Wed, Jun 07, 2017 at 04:17:41AM +0000, Levin, Alexander (Sasha Levin) wrote:
> > > > Thanks Sasha!
> > > > 
> > > > I couldn't reproduce it, that config boots fine on my kvm.
> > > > Would you have the time to dump some traces for me?
> > > > 
> > > > I'd need you to add this boot option: trace_event=hrtimer_cancel,hrtimer_start,hrtimer_expire_entry
> > > > And boot your kernel with the below patch. This will dump the timer traces to your console.
> > > > I would be very interested in the resulting console dump file.
> > > 
> > > Attached. Let me know if you need anything else.
> > 
> > Great! So now I can deduce that the problem doesn't come from the nohz code as
> > ts->next_tick matches the hrtimer deadline. But the dev->next_event from the
> > clockevent seems to be out of line.
> > 
> > Sorry to bother you again, but I'm chasing this bug for several weeks now and
> > you're one of the rare person who can reproduce it. So I may need some more
> > tracing details.
> 
> I take payment in beers ;)

Duly noted ;-)

> 
> But really, not a problem.
> 
> > Here is another version of the debugging patch (not a delta), I added more trace_printk,
> > namely the places where we set this dev->next_event. Can you please apply the below and do
> > the dump again?
> 
> Attached.

Awesome, these traces have been very helpful! So now I think I get what's going on.
Can you please test the following fix?

Thanks a lot!

---
>From 604a46c3e821c61411cee2205fa1cb65e5b04174 Mon Sep 17 00:00:00 2001
From: Frederic Weisbecker <fweisbec@gmail.com>
Date: Thu, 8 Jun 2017 16:32:58 +0200
Subject: [PATCH] nohz: Fix spurious warning when hrtimer and clockevent get
 out of sync

The sanity check ensuring that the tick expiry cache (ts->next_tick)
is actually in sync with the hardware clock (dev->next_event) makes the
wrong assumption that the clock can't be programmed later than the
hrtimer deadline.

In fact the clock hardware can be programmed later on some conditions
such as:

    * The hrtimer deadline is already in the past.
    * The hrtimer deadline is earlier than the minimum delay supported
      by the hardware.

Such conditions can be met when we program the tick, for example if the
last jiffies update hasn't been seen by the current CPU yet, we may
program the hrtimer to a deadline that is earlier than ktime_get()
because last_jiffies_update is our timestamp base to compute the next
tick.

As a result, we can randomly observe such warning:

	WARNING: CPU: 5 PID: 0 at kernel/time/tick-sched.c:794 tick_nohz_stop_sched_tick kernel/time/tick-sched.c:791 [inline]
	Call Trace:
	 tick_nohz_irq_exit
	 tick_irq_exit
	 irq_exit
	 exiting_irq
	 smp_call_function_interrupt
	 smp_call_function_single_interrupt
	 call_function_single_interrupt

Therefore, let's rather make sure that the tick expiry cache is sync'ed
with the tick hrtimer deadline, against which it is not supposed to
drift away. The clock hardware instead has its own will and can't be
used as a reliable comparison point.

Reported-by: Sasha Levin <alexander.levin@verizon.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: James Hartsock <hartsjc@redhat.com>
Cc: Tim Wright <tim@binbash.co.uk>
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
---
 kernel/time/tick-sched.c | 7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)

diff --git a/kernel/time/tick-sched.c b/kernel/time/tick-sched.c
index 9d31f1e..b55547f 100644
--- a/kernel/time/tick-sched.c
+++ b/kernel/time/tick-sched.c
@@ -768,7 +768,8 @@ static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts,
 	/* Skip reprogram of event if its not changed */
 	if (ts->tick_stopped && (expires == ts->next_tick)) {
 		/* Sanity check: make sure clockevent is actually programmed */
-		if (likely(dev->next_event <= ts->next_tick))
+		if (tick != KTIME_MAX &&
+		    ts->next_tick == hrtimer_get_expires(&ts->sched_timer))
 			goto out;
 
 		WARN_ON_ONCE(1);
@@ -806,8 +807,10 @@ static ktime_t tick_nohz_stop_sched_tick(struct tick_sched *ts,
 		goto out;
 	}
 
+	hrtimer_set_expires(&ts->sched_timer, tick);
+
 	if (ts->nohz_mode == NOHZ_MODE_HIGHRES)
-		hrtimer_start(&ts->sched_timer, tick, HRTIMER_MODE_ABS_PINNED);
+		hrtimer_start_expires(&ts->sched_timer, HRTIMER_MODE_ABS_PINNED);
 	else
 		tick_program_event(tick, 1);
 out:
-- 
2.7.4