sched: cpufreq: Allow remote cpufreq callbacks
authorViresh Kumar <viresh.kumar@linaro.org>
Fri, 28 Jul 2017 06:46:38 +0000 (12:16 +0530)
committerRafael J. Wysocki <rafael.j.wysocki@intel.com>
Tue, 1 Aug 2017 12:24:53 +0000 (14:24 +0200)
commit674e75411fc260b0d4532701228cfe12fc090da8
tree78752d7e6e2ec87c6f0bfb3c2b1bf2c6a7ed51dc
parent251accf98591d7f59f7a2bac2e05c66d16bf2811
sched: cpufreq: Allow remote cpufreq callbacks

With Android UI and benchmarks the latency of cpufreq response to
certain scheduling events can become very critical. Currently, callbacks
into cpufreq governors are only made from the scheduler if the target
CPU of the event is the same as the current CPU. This means there are
certain situations where a target CPU may not run the cpufreq governor
for some time.

One testcase to show this behavior is where a task starts running on
CPU0, then a new task is also spawned on CPU0 by a task on CPU1. If the
system is configured such that the new tasks should receive maximum
demand initially, this should result in CPU0 increasing frequency
immediately. But because of the above mentioned limitation though, this
does not occur.

This patch updates the scheduler core to call the cpufreq callbacks for
remote CPUs as well.

The schedutil, ondemand and conservative governors are updated to
process cpufreq utilization update hooks called for remote CPUs where
the remote CPU is managed by the cpufreq policy of the local CPU.

The intel_pstate driver is updated to always reject remote callbacks.

This is tested with couple of usecases (Android: hackbench, recentfling,
galleryfling, vellamo, Ubuntu: hackbench) on ARM hikey board (64 bit
octa-core, single policy). Only galleryfling showed minor improvements,
while others didn't had much deviation.

The reason being that this patch only targets a corner case, where
following are required to be true to improve performance and that
doesn't happen too often with these tests:

- Task is migrated to another CPU.
- The task has high demand, and should take the target CPU to higher
  OPPs.
- And the target CPU doesn't call into the cpufreq governor until the
  next tick.

Based on initial work from Steve Muckle.

Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Acked-by: Saravana Kannan <skannan@codeaurora.org>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
drivers/cpufreq/cpufreq_governor.c
drivers/cpufreq/intel_pstate.c
include/linux/cpufreq.h
kernel/sched/cpufreq_schedutil.c
kernel/sched/deadline.c
kernel/sched/fair.c
kernel/sched/rt.c
kernel/sched/sched.h