[PATCH v3 00/10] sched/fair: Avoid unnecessary migrations within SMT domains

From: Ricardo Neri
Date: Mon Feb 06 2023 - 23:50:51 EST


Hi,

This is v3 of this series. Previous versions can be found here [1] and
here [2]. To avoid duplication, I do not include the cover letter of the
original submission. You can read it in [1].

Changes since v2:

Vincent correctly indicated that I was abusing asym_packing to force load
balances unrelated to CPU priority. The underlying issue is that the
scheduler cannot not handle load balances between SMT and non-SMT cores
correctly. I added several prework patches to fix it... and I removed the
abuse of asym_packing.

Dietmar helped me to realize that there is a better way to check the idle
state of SMT cores. Now I give the task to the scheduler instead of
architecture-specific overrides. I unconditionally obey CPU priorities
at the SMT level. This keeps Power7 happy. At upper levels (i.e., when
balancing load between cores) the scheduler also considers the idle state
of the core in addition to CPU priority. This satisfies x86.

Ionela spotted a violation of the scheduler topology sanity checks. We did
not find a check that suits both Power7 and x86. For now, I removed the
NEEDS_CHILD flag of SD_ASYM_PACKING.

Hopefully, these patches are in sufficiently good shape to be merged.

Thank you for your feedback and I look forward to getting more of it!

New patches 2, 3, 4, 5, 6, 7, 8
Updated patches: 1
Unchanged patches: 9, 10

BR,
Ricardo

[1]. https://lore.kernel.org/lkml/20220825225529.26465-1-ricardo.neri-calderon@xxxxxxxxxxxxxxx/
[2]. https://lore.kernel.org/lkml/20221122203532.15013-1-ricardo.neri-calderon@xxxxxxxxxxxxxxx/


Ricardo Neri (10):
sched/fair: Generalize asym_packing logic for SMT cores
sched/fair: Move is_core_idle() out of CONFIG_NUMA
sched/fair: Only do asym_packing load balancing from fully idle SMT
cores
sched/fair: Let low-priority cores help high-priority busy SMT cores
sched/fair: Keep a fully_busy SMT sched group as busiest
sched/fair: Use the prefer_sibling flag of the current sched domain
sched/fair: Do not even the number of busy CPUs via asym_packing
sched/topology: Remove SHARED_CHILD from ASYM_PACKING
x86/sched: Remove SD_ASYM_PACKING from the SMT domain flags
x86/sched/itmt: Give all SMT siblings of a core the same priority

arch/x86/kernel/itmt.c | 23 +----
arch/x86/kernel/smpboot.c | 2 +-
include/linux/sched/sd_flags.h | 5 +-
kernel/sched/fair.c | 175 +++++++++++++++++----------------
4 files changed, 99 insertions(+), 106 deletions(-)

--
2.25.1