sched: fix improper load balance across sched domain

We recently discovered a nasty performance bug in the kernel CPU load
balancer where we were hit by 50% performance regression.

When tasks are assigned to a subset of CPUs that span across
sched_domains (either ccNUMA node or the new multi-core domain) via
cpu affinity, kernel fails to perform proper load balance at
these domains, due to several logic in find_busiest_group() miss
identified busiest sched group within a given domain. This leads to
inadequate load balance and causes 50% performance hit.

To give you a concrete example, on a dual-core, 2 socket numa system,
there are 4 logical cpu, organized as:

CPU0 attaching sched-domain:
 domain 0: span 0003  groups: 0001 0002
 domain 1: span 000f  groups: 0003 000c
CPU1 attaching sched-domain:
 domain 0: span 0003  groups: 0002 0001
 domain 1: span 000f  groups: 0003 000c
CPU2 attaching sched-domain:
 domain 0: span 000c  groups: 0004 0008
 domain 1: span 000f  groups: 000c 0003
CPU3 attaching sched-domain:
 domain 0: span 000c  groups: 0008 0004
 domain 1: span 000f  groups: 000c 0003

If I run 2 tasks with CPU affinity set to 0x5.  There are situation
where cpu0 has run queue length of 2, and cpu2 will be idle.  The
kernel load balancer is unable to balance out these two tasks over
cpu0 and cpu2 due to at least three logics in find_busiest_group()
that heavily bias load balance towards power saving mode. e.g. while
determining "busiest" variable, kernel only set it when
"sum_nr_running > group_capacity".  This test is flawed that
"sum_nr_running" is not necessary same as
sum-tasks-allowed-to-run-within-the sched-group.  The end result is
that kernel "think" everything is balanced, but in reality we have an
imbalance and thus causing one CPU to be over-subscribed and leaving
other idle.  There are two other logic in the same function will also
causing similar effect.  The nastiness of this bug is that kernel not
be able to get unstuck in this unfortunate broken state.  From what
we've seen in our environment, kernel will stuck in imbalanced state
for extended period of time and it is also very easy for the kernel to
stuck into that state (it's pretty much 100% reproducible for us).

So proposing the following fix: add addition logic in
find_busiest_group to detect intrinsic imbalance within the busiest
group.  When such condition is detected, load balance goes into spread
mode instead of default grouping mode.

Signed-off-by: Ken Chen <kenchen@google.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>