On 2/10/25 17:09, Juri Lelli wrote:
> Hi Christian,
>
> Thanks for taking a look as well.
>
> On 07/02/25 15:55, Christian Loehle wrote:
>> On 2/7/25 14:04, Jon Hunter wrote:
>>>
>>>
>>> On 07/02/2025 13:38, Dietmar Eggemann wrote:
>>>> On 07/02/2025 11:38, Jon Hunter wrote:
>>>>>
>>>>> On 06/02/2025 09:29, Juri Lelli wrote:
>>>>>> On 05/02/25 16:56, Jon Hunter wrote:
>>>>>>
>>>>>> ...
>>>>>>
>>>>>>> Thanks! That did make it easier :-)
>>>>>>>
>>>>>>> Here is what I see ...
>>>>>>
>>>>>> Thanks!
>>>>>>
>>>>>> Still different from what I can repro over here, so, unfortunately, I
>>>>>> had to add additional debug printks. Pushed to the same branch/repo.
>>>>>>
>>>>>> Could I ask for another run with it? Please also share the complete
>>>>>> dmesg from boot, as I would need to check debug output when CPUs are
>>>>>> first onlined.
>>>>
>>>> So you have a system with 2 big and 4 LITTLE CPUs (Denver0 Denver1 A57_0
>>>> A57_1 A57_2 A57_3) in one MC sched domain and (Denver1 and A57_0) are
>>>> isol CPUs?
>>>
>>> I believe that 1-2 are the denvers (even thought they are listed as 0-1 in device-tree).
>>
>> Interesting, I have yet to reproduce this with equal capacities in isolcpus.
>> Maybe I didn't try hard enough yet.
>>
>>>
>>>> This should be easy to set up for me on my Juno-r0 [A53 A57 A57 A53 A53 A53]
>>>
>>> Yes I think it is similar to this.
>>>
>>> Thanks!
>>> Jon
>>>
>>
>> I could reproduce that on a different LLLLbb with isolcpus=3,4 (Lb) and
>> the offlining order:
>> echo 0 > /sys/devices/system/cpu/cpu5/online
>> echo 0 > /sys/devices/system/cpu/cpu1/online
>> echo 0 > /sys/devices/system/cpu/cpu3/online
>> echo 0 > /sys/devices/system/cpu/cpu2/online
>> echo 0 > /sys/devices/system/cpu/cpu4/online
>>
>> while the following offlining order succeeds:
>> echo 0 > /sys/devices/system/cpu/cpu5/online
>> echo 0 > /sys/devices/system/cpu/cpu4/online
>> echo 0 > /sys/devices/system/cpu/cpu1/online
>> echo 0 > /sys/devices/system/cpu/cpu2/online
>> echo 0 > /sys/devices/system/cpu/cpu3/online
>> (Both offline an isolcpus last, both have CPU0 online)
>>
>> The issue only triggers with sugov DL threads (I guess that's obvious, but
>> just to mention it).
>
> It wasn't obvious to me at first :). So thanks for confirming.
>
>> I'll investigate some more later but wanted to share for now.
>
> So, problem actually is that I am not yet sure what we should do with
> sugovs' bandwidth wrt root domain accounting. W/o isolation it's all
> good, as it gets accounted for correctly on the dynamic domains sugov
> tasks can run on. But with isolation and sugov affected_cpus that cross
> isolation domains (e.g., one BIG one little), we can get into troubles
> not knowing if sugov contribution should fall on the DEF or DYN domain.
>
> Hummm, need to think more about it.
That is indeed tricky.
I would've found it super appealing to always just have sugov DL tasks activate
on this_cpu and not have to worry about all this, but then you have contention
amongst CPUs of a cluster and there are energy improvements from always
having little cores handle all sugov DL tasks, even for the big CPUs,
that's why I introduced
commit 93940fbdc468 ("cpufreq/schedutil: Only bind threads if needed")
but that really doesn't make this any easier.
