Re: [PATCH] sched: Make ASYM_PACKING compile-time arch config

[Shrikanth Hegde]

On 10/18/24 03:16, Christian Loehle wrote:
> Only x86 and Power7 set SD_ASYM_PACKING at boot-time depending on the

Power10 on Shared Processor LPAR also enables that.

> system. All other platforms don't make use of asym-packing anyway,
> so introduce auxiliary ARCH_HAS_SCHED_ASYM_PACKING and guard all the
> related logic behind that so it isn't compiled when not needed.
>
> On arm64 this reduces
> size kernel/sched/fair.o
>     text	   data	    bss	    dec	    hex	filename
>    74120	   4097	     88	  78305	  131e1	kernel/sched/fair.o
> to
> size kernel/sched/fair.o
>     text	   data	    bss	    dec	    hex	filename
>    72896	   4065	     88	  77049	  12cf9	kernel/sched/fair.o
>
> Most of that is on the load-balance hot-path, in particular
> need_active_balance() reduces from 141 to 84 instructions.

This patch sprinkes #ifdef across the code. IMHO it makes the code 
difficult to read. Its possible to do slight such reductions for other 
group types as well, such as group_misfit_task(which allows right load 
balancing when SD_ASYM_CPUCAPACITY is on, except for a case when 
capacity is reduced due to side activity).

sd->flags should be cache hot and its a percpu variable. Hence the 
access to it shouldn't be costly. the branch predictors should do a good
job since it doesn't change often. Correct me if i am wrong.

> hackbench -pTl 20000 on a rk3399 goes from
> 58.4664 to 57.6056 (-1.5%), mean over 20 iterations.
>
> Signed-off-by: Christian Loehle <christian.loehle@xxxxxxx>
> ---
>   arch/Kconfig         |  3 ++
>   arch/powerpc/Kconfig |  1 +
>   arch/x86/Kconfig     |  1 +
>   kernel/sched/fair.c  | 86 ++++++++++++++++++++++++++++++--------------
>   kernel/sched/sched.h |  8 ++++-
>   5 files changed, 71 insertions(+), 28 deletions(-)
>
> diff --git a/arch/Kconfig b/arch/Kconfig
> index 8af374ea1adc..9f6ec10723a3 100644
> --- a/arch/Kconfig
> +++ b/arch/Kconfig
> @@ -1684,4 +1684,7 @@ config CC_HAS_SANE_FUNCTION_ALIGNMENT
>   config ARCH_NEED_CMPXCHG_1_EMU
>   	bool
...