Re: [PATCH v2 6/8] sched/idle: Move busy_cpu accounting to idle callback

[Aubrey Li <aubrey.li@linux.intel.com>]

Hi Valentin,

On 5/11/21 7:51 PM, Valentin Schneider wrote:
> On 06/05/21 22:15, Srikar Dronamraju wrote:
>> diff --git a/kernel/sched/topology.c b/kernel/sched/topology.c
>> index 8db40c8a6ad0..00e4669bb241 100644
>> --- a/kernel/sched/topology.c
>> +++ b/kernel/sched/topology.c
>> @@ -647,6 +647,7 @@ DEFINE_PER_CPU(int, sd_llc_id);
>>  #ifdef CONFIG_SCHED_SMT
>>  DEFINE_PER_CPU(int, smt_id);
>>  #endif
>> +DEFINE_PER_CPU(int, is_idle);
> 
> This + patch 8 immediately reminds me of Aubrey's patch:
> 
>   http://lore.kernel.org/r/1615872606-56087-1-git-send-email-aubrey.li@intel.com
> 
> last I looked it seemed OK, even the test bot seems happy. Aubrey, did you
> have any more work to do on that one (other than rebasing)?
> 

Thanks to mention this patch, in terms of the patch itself, I don't have any more
work, other than rebasing it to Rik's bring-back-select_idle_smt() patch.

But I have some other ideas based on this patch need to verify, for example, if we
have an idle_cpu mask when we scan the idle CPU, we should be able to remove SIS_PROP
feature at all. This removes scan cost computation and should reduce the task wakeup
latency. But I need a while to understand some benchmark regressions.

Thanks,
-Aubrey