Re: [PATCH] sched: move rr_timeslice sysctls to rt.c

[Peter Zijlstra]

On Thu, Feb 10, 2022 at 02:08:31PM +0800, Zhen Ni wrote:
> move rr_timeslice sysctls to rt.c and use the new
> register_sysctl_init() to register the sysctl interface.
> 
> Signed-off-by: Zhen Ni <nizhen@xxxxxxxxxxxxx>

OK, I've had it with this nonsense. Can you *please* redo all of sched
such that:

 - In the Subject:, the first letter after the subsystem prefix (sched:)
   is capitalized.
 - the lot actually applies to tip/sched/core (so far not a single one
   of these patches applied without needing -- mostly trivial --
   fixups).
 - Do obvious cleanups.. see below.
 - Don't have more than a single *sysctl_init() per .c file.
 - It's a full series that does all instead of random little patches
   that conflict with one another when applied out of turn.

> ---
>  include/linux/sched/sysctl.h |  3 ---
>  kernel/sched/rt.c            | 28 ++++++++++++++++++++++++++--
>  kernel/sysctl.c              |  7 -------
>  3 files changed, 26 insertions(+), 12 deletions(-)
> 
> diff --git a/include/linux/sched/sysctl.h b/include/linux/sched/sysctl.h
> index d416d8f45186..f6466040883c 100644
> --- a/include/linux/sched/sysctl.h
> +++ b/include/linux/sched/sysctl.h
> @@ -45,11 +45,8 @@ extern unsigned int sysctl_sched_uclamp_util_min_rt_default;
>  extern unsigned int sysctl_sched_autogroup_enabled;
>  #endif
>  
> -extern int sysctl_sched_rr_timeslice;
>  extern int sched_rr_timeslice;

Why leave sched_rr_timeslice here? It doesn't belong here.


> +#ifdef CONFIG_SYSCTL
> +static struct ctl_table sched_rr_sysctls[] = {
> +	{
> +		.procname       = "sched_rr_timeslice_ms",
> +		.data           = &sysctl_sched_rr_timeslice,
> +		.maxlen         = sizeof(int),
> +		.mode           = 0644,
> +		.proc_handler   = sched_rr_handler,
> +	},
> +	{}
> +};
> +
> +static void __init sched_rr_sysctl_init(void)
> +{
> +	register_sysctl_init("kernel", sched_rr_sysctls);
> +}
> +#else
> +#define sched_rr_sysctl_init() do { } while (0)
> +#endif
> +
>  static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
>  
>  struct rt_bandwidth def_rt_bandwidth;
> @@ -2471,6 +2494,7 @@ void __init init_sched_rt_class(void)
>  		zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
>  					GFP_KERNEL, cpu_to_node(i));
>  	}
> +	sched_rr_sysctl_init();
>  }

When I combine this with: patches/zhen_ni-sched-move_rt_period_runtime_sysctls_to_rt_c.patch

That ends up as:

@@ -2471,6 +2535,8 @@ void __init init_sched_rt_class(void)
                zalloc_cpumask_var_node(&per_cpu(local_cpu_mask, i),
                                        GFP_KERNEL, cpu_to_node(i));
        }
+       sched_rt_sysctl_init();
+       sched_rr_sysctl_init();
 }
 #endif /* CONFIG_SMP */

Like srsly?


So I've dropped the whole lot I had:

patches/zhen_ni-sched-move_energy_aware_sysctls_to_topology_c.patch
patches/zhen_ni-sched-move_cfs_bandwidth_slice_sysctls_to_fair_c.patch
patches/zhen_ni-sched-move_uclamp_util_sysctls_to_core_c.patch
patches/zhen_ni-sched-move_schedstats_sysctls_to_core_c.patch
patches/zhen_ni-sched-move_deadline_period_sysctls_to_deadline_c.patch
patches/zhen_ni-sched-move_rt_period_runtime_sysctls_to_rt_c.patch
patches/zhen_ni-sched-move_rr_timeslice_sysctls_to_rt_c.patch

And I expect a single coherent series or I'll forgo all this.