[PATCH] Fix c-state transitions when !NOHZ (nohz=off kernel parameter)

[PATCH] Fix c-state transitions when !NOHZ (nohz=off kernel parameter)

[Prarit Bhargava]

Some users do still boot with "nohz=off" which re-enables the old "tickfull"
kernel and still want the (diminished) power savings of c-states.

When examining the c-state transitions for !NOHZ it was noticed that the
processors never leaving the polling state.

ie) powertop shows 98-99% residency in POLLING for all cpus

This occurs  because the value of the current tick_sched's sleep_length
is never updated from its initialized value if !NOHZ.  sleep_length is used
by the cpuidle code (via a call to tick_nohz_get_sleep_length()) to determine
how long a cpu should sleep for and in what c-state.

Admittedly I am not overly familiar with this code and there may be some
unknown repercussion from doing this.  I can say that experimentally across
several systems I do see the cpus going into deeper c-states after making
this change, and there does not appear to be anything wrong with the system.

----8<----

Fix c-state transitions when !NOHZ (nohz=off kernel parameter)

The value of sleep_length needs to be updated for the !NOHZ case in
order for processor c-state transitions to work.

Cc: Len Brown <len.brown@intel.com>
---
 kernel/time/tick-sched.c |    4 +++-
 1 files changed, 3 insertions(+), 1 deletions(-)

diff --git a/kernel/time/tick-sched.c b/kernel/time/tick-sched.c
index 7656642..d2e2782 100644
--- a/kernel/time/tick-sched.c
+++ b/kernel/time/tick-sched.c
@@ -300,8 +300,10 @@ static void tick_nohz_stop_sched_tick(struct tick_sched *ts)
 			tick_do_timer_cpu = TICK_DO_TIMER_NONE;
 	}
 
-	if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
+	if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) {
+		ts->sleep_length = ktime_sub(dev->next_event, now);
 		return;
+	}
 
 	if (need_resched())
 		return;
-- 
1.7.1

Re: [PATCH] Fix c-state transitions when !NOHZ (nohz=off kernel parameter)

[Prarit Bhargava]

On 01/17/2012 01:12 PM, Prarit Bhargava wrote:
> Some users do still boot with "nohz=off" which re-enables the old "tickfull"
> kernel and still want the (diminished) power savings of c-states.
> 
> When examining the c-state transitions for !NOHZ it was noticed that the
> processors never leaving the polling state.
> 
> ie) powertop shows 98-99% residency in POLLING for all cpus
> 
> This occurs  because the value of the current tick_sched's sleep_length
> is never updated from its initialized value if !NOHZ.  sleep_length is used
> by the cpuidle code (via a call to tick_nohz_get_sleep_length()) to determine
> how long a cpu should sleep for and in what c-state.
> 
> Admittedly I am not overly familiar with this code and there may be some
> unknown repercussion from doing this.  I can say that experimentally across
> several systems I do see the cpus going into deeper c-states after making
> this change, and there does not appear to be anything wrong with the system.
> 
> ----8<----
> 
> Fix c-state transitions when !NOHZ (nohz=off kernel parameter)
> 
> The value of sleep_length needs to be updated for the !NOHZ case in
> order for processor c-state transitions to work.
> 
> Cc: Len Brown <len.brown@intel.com>

Nuts -- I completely forgot

Signed-off-by: Prarit Bhargava <prarit@redhat.com>

(Thanks Len)

P.

> ---
>  kernel/time/tick-sched.c |    4 +++-
>  1 files changed, 3 insertions(+), 1 deletions(-)
> 
> diff --git a/kernel/time/tick-sched.c b/kernel/time/tick-sched.c
> index 7656642..d2e2782 100644
> --- a/kernel/time/tick-sched.c
> +++ b/kernel/time/tick-sched.c
> @@ -300,8 +300,10 @@ static void tick_nohz_stop_sched_tick(struct tick_sched *ts)
>  			tick_do_timer_cpu = TICK_DO_TIMER_NONE;
>  	}
>  
> -	if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE))
> +	if (unlikely(ts->nohz_mode == NOHZ_MODE_INACTIVE)) {
> +		ts->sleep_length = ktime_sub(dev->next_event, now);
>  		return;
> +	}
>  
>  	if (need_resched())
>  		return;