Skip to contentSkip to conversation

nohz: Avoid double clock write

PATCH In review Frederic Weisbecker 1 participant 4 messages Jul 31, 2014 – Aug 11, 2014 Nohz, dynticks support

lore
Share & tools
lore
#0 lore raw

cover nohz: Avoid double clock write

Hi Thomas,

The tick reschedules itself unconditionally. That's what we want as long as the CPU is in periodic mode. It's not that relevant when the CPU is in dynticks mode though as the clock write is likely to be overwritten by the nohz code on irq exit.

In low-res mode, that even result in unexpected periodic behaviour when no timer is scheduled in the future.

So here is a set that proposes the removal of these double writes.

git://git.kernel.org/pub/scm/linux/kernel/git/frederic/linux-dynticks.git
origin/nohz/drop-double-write-v2

Thanks,
Frederic
---

Viresh Kumar (2):
nohz: Fix spurious periodic tick behaviour in low-res dynticks mode
nohz: Avoid tick's double reprogramming in highres mode

 kernel/time/tick-sched.c | 8 ++++++++
 1 file changed, 8 insertions(+)
#1 lore raw

PATCH 1/2 nohz: Fix spurious periodic tick behaviour in low-res dynticks mode

From: Viresh Kumar <[email protected]>

When we reach the end of the tick handler, we unconditionally reschedule the next tick to the next jiffy. Then on irq exit, the nohz code overrides that setting if needed and defers the next tick as far away in the future as possible.

Now in the best dynticks case, when we actually don't need any tick in the future (ie: expires == KTIME_MAX), low-res and high-res behave differently. What we want in this case is to cancel the next tick programmed by the previous one. That's what we do in high-res mode. OTOH we lack a low-res mode equivalent of hrtimer_cancel() so we simply don't do anything in this case and the next tick remains scheduled to jiffies + 1.

As a result, in low-res mode, when the dynticks code determines that no tick is needed in the future, we can recursively get a spurious tick every jiffy because then the next tick is always reprogrammed from the tick handler and is never cancelled. And this can happen indefinetly until some subsystem actually needs a precise tick in the future and only then we eventually overwrite the previous tick handler setting to defer the next tick.

We are fixing this by introducing the ONESHOT_STOPPED mode which will let us pause a clockevent when no further interrupt is needed. Meanwhile we can't expect all drivers to support this new mode.

So lets reduce much of the symptoms by skipping the nohz-blind tick rescheduling from the tick-handler when the CPU is in dynticks mode. That tick rescheduling wrongly assumed periodicity and the low-res dynticks code can't cancel such decision. This breaks the recursive (and thus the worst) part of the problem. In the worst case now, we'll get only one extra tick due to uncancelled tick scheduled before we entered dynticks mode.

This also removes a needless clockevent write on idle ticks. Since those clock write are usually considered to be slow, it's a general win.

---
 kernel/time/tick-sched.c | 4 ++++
 1 file changed, 4 insertions(+)
kernel/time/tick-sched.c +4 -0
index 6558b7a..b07ba8e 100644
@@ -956,6 +956,10 @@ static void tick_nohz_handler(struct clock_event_device *dev)
 	tick_sched_do_timer(now);
 	tick_sched_handle(ts, regs);
 
+	/* Skip reprogramming next event if we are running tickless  */
+	if (unlikely(ts->tick_stopped))
+		return;
+
 	while (tick_nohz_reprogram(ts, now)) {
 		now = ktime_get();
 		tick_do_update_jiffies64(now);
Signature · 1 line
1.8.3.1
#2 lore raw

PATCH 2/2 nohz: Avoid tick's double reprogramming in highres mode

In reply to Frederic Weisbecker

From: Viresh Kumar <[email protected]>

In highres mode, the tick reschedules itself unconditionally to the next jiffies.

However while this clock reprogramming is relevant when the tick is in periodic mode, it's not that interesting when we run in dynticks mode because irq exit is likely going to overwrite the next tick to some randomly deferred future.

So lets just get rid of this tick self rescheduling in dynticks mode. This way we can avoid some clockevents double write in favourable scenarios like when we stop the tick completely in idle while no other hrtimer is pending.

---
 kernel/time/tick-sched.c | 4 ++++
 1 file changed, 4 insertions(+)
kernel/time/tick-sched.c +4 -0
index b07ba8e..a7098f7 100644
@@ -1087,6 +1087,10 @@ static enum hrtimer_restart tick_sched_timer(struct hrtimer *timer)
 	if (regs)
 		tick_sched_handle(ts, regs);
 
+	/* Do not restart, when we are in idle or full dynticks mode */
+	if (unlikely(ts->tick_stopped))
+		return HRTIMER_NORESTART;
+
 	hrtimer_forward(timer, now, tick_period);
 
 	return HRTIMER_RESTART;
Signature · 1 line
1.8.3.1
#3 lore raw
In reply to Frederic Weisbecker

Ping?

Quoted 26 lines · Frederic Weisbecker

Hi Thomas,

The tick reschedules itself unconditionally. That's what we want as long as the CPU is in periodic mode. It's not that relevant when the CPU is in dynticks mode though as the clock write is likely to be overwritten by the nohz code on irq exit.

In low-res mode, that even result in unexpected periodic behaviour when no timer is scheduled in the future.

So here is a set that proposes the removal of these double writes.

git://git.kernel.org/pub/scm/linux/kernel/git/frederic/linux-dynticks.git
origin/nohz/drop-double-write-v2

Thanks,
Frederic
---

Viresh Kumar (2):
nohz: Fix spurious periodic tick behaviour in low-res dynticks mode
nohz: Avoid tick's double reprogramming in highres mode

 kernel/time/tick-sched.c | 8 ++++++++
 1 file changed, 8 insertions(+)