From: Brendan Jackman
Date: Thu Oct 05 2017 - 07:45:29 EST

This patchset optimises away an unused comparison, and fixes some corner cases in
the find_idlest_group path of select_task_rq_fair.

Changes v2 -> v3:

- 1/5 through 4/5 unchanged.

- 5/5: Essentially undid [1]. Vincent pointed out that this isn't really a
bugfix but more of a change in policy. Now find_idlest_group still returns
NULL when the local group is idlest.

I previously said some extra code could be removed in this case, but after
more careful inspection I changed my mind.

Changes v1 -> v2:

- Reworked task affinity checks to avoid repeating them, as per Vincent's
suggestion. To avoid excessive indentation this required moving code into its
own function, as per PeterZ's suggestion.

- Split up the patches.

- Altered the caller of find_idlest_group so that it now unconditionally uses
find_idlest_group_cpu (formerly find_idlest_cpu). This means that we more
often use the maligned "perspective-switching" logic at the bottom of the
while(sd) loop, but it also means the task placement algorithm is more
consistent between whether the idlest group is local or remote.

As mentioned in 5/5 an alternative would be to just initialise @new_cpu to
@cpu instead of @prev_cpu (which is what PeterZ suggested in v1 review). In
that case, some extra code could be removed in & around

Brendan Jackman (5):
sched/fair: Move select_task_rq_fair slow-path into its own function
sched/fair: Remove unnecessary comparison with -1
sched/fair: Fix find_idlest_group when local group is not allowed
sched/fair: Fix use of find_idlest_group when no groups are allowed
sched/fair: Fix use of find_idlest_group when local group is idlest

kernel/sched/fair.c | 93 ++++++++++++++++++++++++++++++++---------------------
1 file changed, 56 insertions(+), 37 deletions(-)