Re: [PATCH v2 3/2] sched/deadline: Check bandwidth overflow earlier for hotplug

[Jon Hunter <jonathanh@nvidia.com>]

On 11/02/2025 10:42, Juri Lelli wrote:
> On 11/02/25 10:15, Christian Loehle wrote:
>> 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.
> 
> What about we actually ignore them consistently? We already do that for
> admission control, so maybe we can do that when rebuilding domains as
> well (until we find maybe a better way to deal with them).
> 
> Does the following make any difference?
> 
> ---
>   kernel/sched/deadline.c | 2 +-
>   1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/kernel/sched/deadline.c b/kernel/sched/deadline.c
> index b254d878789d..8f7420e0c9d6 100644
> --- a/kernel/sched/deadline.c
> +++ b/kernel/sched/deadline.c
> @@ -2995,7 +2995,7 @@ void dl_add_task_root_domain(struct task_struct *p)
>   	struct dl_bw *dl_b;
>   
>   	raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
> -	if (!dl_task(p)) {
> +	if (!dl_task(p) || dl_entity_is_special(&p->dl)) {
>   		raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
>   		return;
>   	}
> 

I have tested this on top of v6.14-rc2, but this is still not resolving 
the issue for me :-(

Jon

-- 
nvpublic