Re: [RFC PATCH 2/2] sched/fair: skip the cache hot CPU in select_idle_cpu()

[K Prateek Nayak <kprateek.nayak@amd.com>]

Hello Chenyu,

On 9/14/2023 4:13 PM, Chen Yu wrote:
> Hi Prateek,
> 
> On 2023-09-14 at 11:00:02 +0530, K Prateek Nayak wrote:
>> Hello Chenyu,
>>
>> One question ...
>>
>> On 9/11/2023 8:20 AM, Chen Yu wrote:
>>> [..snip..]
>>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
>>> index e20f50726ab8..fe3b760c9654 100644
>>> --- a/kernel/sched/fair.c
>>> +++ b/kernel/sched/fair.c
>>> [..more snip..]
>>> @@ -7052,10 +7072,14 @@ static int select_idle_core(struct task_struct *p, int core, struct cpumask *cpu
>>>  	int cpu;
>>>  
>>>  	for_each_cpu(cpu, cpu_smt_mask(core)) {
>>> -		if (!available_idle_cpu(cpu)) {
>>> +		bool cache_hot = sched_feat(SIS_CACHE) ?
>>> +			sched_clock_cpu(cpu) < cpu_rq(cpu)->cache_hot_timeout : false;
>>> +
>>> +		if (!available_idle_cpu(cpu) || cache_hot) {
>>>  			idle = false;
>>>  			if (*idle_cpu == -1) {
>>> -				if (sched_idle_cpu(cpu) && cpumask_test_cpu(cpu, p->cpus_ptr)) {
>>> +				if (sched_idle_cpu(cpu) && cpumask_test_cpu(cpu, p->cpus_ptr) &&
>>> +				    !cache_hot) {
>>
>> Here, the CPU is running a SCHED_IDLE task ...
>>
>>>  					*idle_cpu = cpu;
>>>  					break;
>>>  				}
>>
>> ... but just below this, there are following lines to cache the idle_cpu:
>>
>> 		}
>> 		if (*idle_cpu == -1 && cpumask_test_cpu(cpu, p->cpus_ptr))
>> 			*idle_cpu = cpu;
>>
>> Would it make sense to also add the same "cache_hot" check here when we
>> come across an idle CPU during the search for an idle core? Something
>> like:
>>
>> -		if (*idle_cpu == -1 && cpumask_test_cpu(cpu, p->cpus_ptr))
> 
> When we reached above code, the following condition should be true:
>  (available_idle_cpu(cpu) && !cache_hot)
> because the previous 'if' statement is false. So I guess we already
> has !cache_hot ?

Ah! You are right. I missed the break at end of the if block. Thank you
for pointing it out to me :)

> 
>> +		if (*idle_cpu == -1 && !cache_hot && cpumask_test_cpu(cpu, p->cpus_ptr))
>> 			*idle_cpu = cpu;
>>
>> Implications with the above change:
>>
>> If the entire core is idle, "select_idle_core()" will return the core
>> and the search will bail out in "select_idle_cpu()". Otherwise, the
>> cache-hot idle CPUs encountered during the search for idle core will be
>> ignored now and if "idle_cpu" is not -1, it contains an idle CPU that is
>> not cache-hot.
>>
>> Thoughts?
>>
> 
> Yes, agree, we want to skip the cache-hot idle CPU if that entire core is not idle
> in your case.
> 
> thanks,
> Chenyu
 
--
Thanks and Regards,
Prateek