Re: [PATCH v2 9/9] sched/topology: Define and use shortcut pointers for wakeup sd_flag scan

From: Dietmar Eggemann
Date: Thu Mar 19 2020 - 06:46:26 EST


On 11.03.20 19:16, Valentin Schneider wrote:
> Reworking select_task_rq_fair()'s domain walk exposed that !want_affine
> wakeups only look for highest sched_domain with the required sd_flag
> set. This is something we can cache at sched domain build time to slightly
> optimize select_task_rq_fair(). Note that this isn't a "free" optimization:
> it costs us 3 pointers per CPU.
>
> Add cached per-CPU pointers for the highest domains with SD_BALANCE_WAKE,
> SD_BALANCE_EXEC and SD_BALANCE_FORK. Use them in select_task_rq_fair().
>
> Signed-off-by: Valentin Schneider <valentin.schneider@xxxxxxx>
> ---
> kernel/sched/fair.c | 25 +++++++++++++------------
> kernel/sched/sched.h | 3 +++
> kernel/sched/topology.c | 12 ++++++++++++
> 3 files changed, 28 insertions(+), 12 deletions(-)
>
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index a6fca6817e92..40fb97062157 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -6595,17 +6595,6 @@ select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags)
> int want_affine = 0;
> int sd_flag;
>
> - switch (wake_flags & (WF_TTWU | WF_FORK | WF_EXEC)) {
> - case WF_TTWU:
> - sd_flag = SD_BALANCE_WAKE;
> - break;
> - case WF_FORK:
> - sd_flag = SD_BALANCE_FORK;
> - break;
> - default:
> - sd_flag = SD_BALANCE_EXEC;
> - }
> -
> if (wake_flags & WF_TTWU) {
> record_wakee(p);
>
> @@ -6621,7 +6610,19 @@ select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags)
>
> rcu_read_lock();
>
> - sd = highest_flag_domain(cpu, sd_flag);
> + switch (wake_flags & (WF_TTWU | WF_FORK | WF_EXEC)) {
> + case WF_TTWU:
> + sd_flag = SD_BALANCE_WAKE;
> + sd = rcu_dereference(per_cpu(sd_balance_wake, cpu));

IMHO, since we hard-code 0*SD_BALANCE_WAKE in sd_init(), sd would always
be NULL, so !want_affine (i.e. wake_wide()) would still go sis().

SD_BALANCE_WAKE is no a topology related sd_flag so it can't be set from
outside. Since the sd->flags sysctl is now read-only, wouldn't this case
be redundant?

[...]