Commit 258cafe4 authored Apr 17, 2019 by Abhijeet Dharmapurikar

sched/fair: correct fastpath



idle_get_state_idx() returns the idle state of the cpu - 0 being the
shallowest and higher values indicate deeper idle state.

The code currently chooses prev_cpu for a idle state of less than 1. 1
on our platforms corresponds to a deep idle state,and for mid and high
capacity cpus exiting from it incurs additional latency.

So amend the code such that for mid and high cluster cpus we limit
prev_cpu fastpath to the shallowest idle state. For min capacity cpus
using idle state of 1 is fine, as it doesn't incur huge idle exit
latencies.

Change-Id: Idb544015c8b533a2b1ad9c1355db3e975162f074
Signed-off-by: Abhijeet Dharmapurikar <adharmap@codeaurora.org>

parent c0835c6a

kernel/sched/fair.c

+3 −1

Original line number	Original line	Diff line number	Diff line
	@@ -6766,7 +6766,9 @@ static void find_best_target(struct sched_domain sd, cpumask_t cpus,
	!cpu_isolated(prev_cpu) && cpu_online(prev_cpu) &&		!cpu_isolated(prev_cpu) && cpu_online(prev_cpu) &&
	idle_cpu(prev_cpu)) {		idle_cpu(prev_cpu)) {

	if (idle_get_state_idx(cpu_rq(prev_cpu)) <= 1) {		if (idle_get_state_idx(cpu_rq(prev_cpu)) <=
			(is_min_capacity_cpu(prev_cpu) ? 1 : 0)) {

	target_cpu = prev_cpu;		target_cpu = prev_cpu;

	fbt_env->fastpath = PREV_CPU_FASTPATH;		fbt_env->fastpath = PREV_CPU_FASTPATH;