[PATCH] fix scheduler regression from "sched/fair: Rework load_balance()"

["Chris Mason" <clm@fb.com>]

Hi everyone,

We’re validating a new kernel in the fleet, and compared with v5.2, 
performance is ~2-3% lower for some of our workloads.  After some 
digging, Johannes found that our involuntary context switch rate was ~2x 
higher, and we were leaving a CPU idle a higher percentage of the time, 
even though the workload was trying to saturate the system.

We were able to reproduce the problem with schbench, and Johannes 
bisected down to:

commit 0b0695f2b34a4afa3f6e9aa1ff0e5336d8dad912
Author: Vincent Guittot <vincent.guittot@linaro.org>
Date:   Fri Oct 18 15:26:31 2019 +0200

     sched/fair: Rework load_balance()

Our working theory is the load balancing changes are leaving processes 
behind busy CPUs instead of moving them onto idle ones.  I made a few 
schbench modifications to make this easier to demonstrate:

https://git.kernel.org/pub/scm/linux/kernel/git/mason/schbench.git/

My VM has 40 cpus (20 cores, 2 threads per core), and my schbench 
command line is:

schbench -t 20 -r 0 -c 1000000 -s 1000 -i 30 -z 120

This has two message threads, and 20 workers per message thread.  Once 
woken up, the workers think for a full second, which means you’ll have 
some long latencies if you’re stuck behind one of these workers in the 
runqueue.  The message thread does a little bit of work and then sleeps, 
so we end up with 40 threads hammering full blast on the CPU and 2 
threads popping in and out of idle.

schbench times the delay from when a message thread wakes a worker to 
when the worker runs.  On a good kernel, the output looks like this:

Latency percentiles (usec) runtime 1290 (s) (3280 total samples)
         50.0th: 155 (1653 samples)
         75.0th: 189 (808 samples)
         90.0th: 216 (501 samples)
         95.0th: 227 (163 samples)
         *99.0th: 256 (123 samples)
         99.5th: 1510 (16 samples)
         99.9th: 3132 (13 samples)
         min=21, max=3286

With 0b0695f2b34a, we get this:

Latency percentiles (usec) runtime 1440 (s) (4480 total samples)
         50.0th: 147 (2261 samples)
         75.0th: 182 (1116 samples)
         90.0th: 205 (671 samples)
         95.0th: 224 (215 samples)
         *99.0th: 12240 (173 samples) <—— much higher p99 and up
         99.5th: 12752 (22 samples)
         99.9th: 13104 (18 samples)
         min=21, max=13172

Since the idea is to fully load the machine with schbench, use schbench 
-t <your_num_cpus/2>, and make sure the box doesn’t have other stuff 
running in the background.  I used a VM because it ended up giving more 
consistent results on our kernel test machines, which have some periodic 
noise running in the background.

We’ve tried a few different approaches, but don’t quite have a solid 
fix yet.  I thought I’d kick off the discussion with my most useful 
hunks so far:

diff a/kernel/sched/fair.c b/kernel/sched/fair.c
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c

-chris