RE: [PATCH v1 0/4] cpuidle: governors: teo: Assorted improvements

["Doug Smythies" <dsmythies@telus.net>]

On 2025.11.13 07:22 Christian Loehle wrote:
> On 11/12/25 16:21, Rafael J. Wysocki wrote:
>> Hi,
>> 
>> This is a bunch of teo cpuidle governor improvements, some of which are related
>> to a bug report discussed recently:
>> 
>> https://lore.kernel.org/linux-pm/CAEmPcwsNMNnNXuxgvHTQ93Mx-q3Oz9U57THQsU_qdcCx1m4w5g@mail.gmail.com/
>> 
>> The first patch fixes a bug that may cause an overly deep idle state
>> to be selected when the scheduler tick has been already stopped.
>> 
>> Patch [2/4] removes an unnecessary function argument.
>> 
>> Patch [3/4] makes teo_update() to use s64 as the data type for its local
>> variables more consistently.
>> 
>> The last patch reworks the governor's decay implementation to also decay
>> metric values lower than 8.
>> 
>
> Tested-by: Christian Loehle <christian.loehle@arm.com>
>
> Test results below, although there really isn't anything interesting in there.
> teo-1 to teo-4 (patches 1 to 4 respectively are essentially indistinguishable from
> teo-m = mainline)

I tested the 4 patch set also, and also found no differences in results above
repeatability noise levels.

Additionally, I added another patch (patch 5 of 4):
"cpuidle: governors: teo: Rework the handling of tick wakeups" [1]
Similar findings.

Additionally, I added another patch (patch 6 of 4):
"sched/idle: disable tick in idle=poll idle entry" [2]
And found only one significant improvement, for only one test,
but only for the TEO idle governor:

Kernel 6.18-rc4:
For a 6 pair fast ping-pong test (meaning no work per token stop):
teo: 5.53 uSec per loop, reference test
4 of 4 patches: 5.53 uSec per loop, 0% 
5 of 4 patches: 5.54 uSec per loop, 0.2% (noise)
6 of 4 patches: 4.77 uSec per loop, 13% better
6 of 4 patches (again): 4.81 uSec per loop, 13% better
menu: 5.29 uSec per loop, 4.4% better
menu + patch 6 of 4: 5.28 uSec per loop, 4.5% better

Idle state 0 usage:
18% with patch 6, teo
11% with menu
~1% with mainline and not patch 6, teo.

Idle state 1 usage:
almost 0 with patch 6, teo
~6% with menu
27% with mainline and not patch 6, teo.

Power: About 100 watts. Patch 6 and teo does increase power use by about a watt or 2.

Processor: Intel(R) Core(TM) i5-10600K CPU @ 4.10GHz, 6 cores 12 CPUs.

For clarity my branch log:
3993913d7f81 (HEAD -> rjw-teo) sched/idle: disable tick in idle=poll idle entry
d9b12b8d62bf cpuidle: governors: teo: Rework the handling of tick wakeups
e47178c87272 cpuidle: governors: teo: Decay metrics below DECAY_SHIFT threshold
7fe32e411c2b cpuidle: governors: teo: Use s64 consistently in teo_update()
490e6118e45d cpuidle: governors: teo: Drop redundant function parameter
8f627f86062e cpuidle: governors: teo: Drop incorrect target residency check
6146a0f1dfae (tag: v6.18-rc4, origin/master, origin/HEAD, master) Linux 6.18-rc4

[1] https://lore.kernel.org/linux-pm/6228387.lOV4Wx5bFT@rafael.j.wysocki/
[2] https://lore.kernel.org/linux-pm/aQiWfnnSzxsnwa2o@tpad/