Re: [PATCH v2 6/6 RFT] sched/fair: change sched asym checking condition

From: kuiliang Shi
Date: Thu Feb 01 2024 - 06:40:42 EST




On 2/1/24 9:10 AM, Ricardo Neri wrote:
> On Tue, Jan 30, 2024 at 09:17:08PM +0800, alexs@xxxxxxxxxx wrote:
>> From: Alex Shi <alexs@xxxxxxxxxx>
>>
>> Asym only used on SMT sd, or core sd with ITMT and core idled.
>> !sched_smt_active isn't necessary.
>
> sched_smt_active() is implemented as a static key. Thus, if SMT is not
> enabled, we can quickly return without having to check the rest of the
> conditions, as we should.

Hi Ricardo,

Thanks a lot for comments! I will drop this patch in this series.

But forgive my stupidity, asym feature is possible when SMT enabled instead of SMT disable. Why no SMT is a condition for asm feature? For this asym feature, I only see the SMT and MC domain use this, correct me if I'm wrong.

>
>>
>> Signed-off-by: Alex Shi <alexs@xxxxxxxxxx>
>> To: Ricardo Neri <ricardo.neri-calderon@xxxxxxxxxxxxxxx>
>> To: linux-kernel@xxxxxxxxxxxxxxx
>> To: Valentin Schneider <vschneid@xxxxxxxxxx>
>> To: Daniel Bristot de Oliveira <bristot@xxxxxxxxxx>
>> To: Ben Segall <bsegall@xxxxxxxxxx>
>> To: Steven Rostedt <rostedt@xxxxxxxxxxx>
>> To: Dietmar Eggemann <dietmar.eggemann@xxxxxxx>
>> To: Vincent Guittot <vincent.guittot@xxxxxxxxxx>
>> To: Juri Lelli <juri.lelli@xxxxxxxxxx>
>> To: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
>> To: Ingo Molnar <mingo@xxxxxxxxxx>
>> ---
>> kernel/sched/fair.c | 4 ++--
>> 1 file changed, 2 insertions(+), 2 deletions(-)
>>
>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
>> index 6680cb39c787..0b7530b93429 100644
>> --- a/kernel/sched/fair.c
>> +++ b/kernel/sched/fair.c
>> @@ -9744,8 +9744,8 @@ static bool sched_use_asym_prio(struct sched_domain *sd, int cpu)
>> if (!(sd->flags & SD_ASYM_PACKING))
>> return false;
>>
>> - return (!sched_smt_active()) ||
>> - (sd->flags & SD_SHARE_CPUCAPACITY) || is_core_idle(cpu);
>> + return (sd->flags & SD_SHARE_CPUCAPACITY) ||
>> + (is_core_idle(cpu) && test_bit(cpu_core_flags(), (void *)&sd->flags));
>
> cpu_core_flags() can contain more than one flag, AFAICS. Which bit should
> it check? Moreover, it is implemented differently for each architecture.

It seems only x86 using the function. But there is still a error which SMT/CLUSTER domain also has this flags bit.
$ git grep 'cpu_core_flags('
arch/x86/kernel/smpboot.c: return cpu_core_flags() | x86_sched_itmt_flags();
include/linux/sched/topology.h:static inline int cpu_core_flags(void)

> Also, as stated, in x86 asym_packing is also used in domains other than MC.

For the feature SD_ASYM_PACKING and SD_ASYM_CPUCAPACITY, for guts of 2 features, is it possible to combine them into one, if we give a little bit more capacity to priority cpus, like 5%?

Thanks
Alex