[PATCH 0/4] sched/core: fix cfs_prio_less

[Cruz Zhao]

The update of vruntime snapshot will cause unfair sched, especially when
tasks enqueue/dequeue frequently.

Consider the following case: 
 - Task A1 and A2 share a cookie, and task B has another cookie.
 - A1 is a short task, waking up frequently but running short everytime.
 - A2 and B are long tasks.
 - A1 and B runs on ht0 and A2 runs on ht1.

ht0			ht1		fi_before	fi	update
switch to A1		switch to A2	0		0	1
A1 sleeps
switch to B		A2 force idle	0		1	1
A1 wakes up
switch to A1		switch to A1	1		0	1
A1 sleeps
switch to B		A2 force idle	0		1	1

In this case, cfs_rq->min_vruntime_fi will update every schedule, and
prio of B and A2 will be pulled to the same level, no matter how long A2
and B have run before, which is not fair enough. Extramely, we observed
that the latency of a task became several minutes due to this reason,
which should be 100ms.

To fix this problem, a possible approach is to maintain another vruntime
relative to the core, called core_vruntime, and we compare the priority
of ses using core_vruntime directly, instead of vruntime snapshot. To
achieve this goal, we need to introduce cfs_rq->core, similarity to
rq->core, and record core_min_vruntime in cfs_rq->core.

Cruz Zhao (4):
  sched/core: Introduce core_id
  sched: Introduce cfs_rq->core
  sched: introduce core_vruntime and core_min_vruntime
  fix vruntime snapshot

 include/linux/sched.h |  3 ++
 kernel/sched/core.c   | 37 +++++++---------
 kernel/sched/fair.c   | 98 ++++++++++++++++++++++++++-----------------
 kernel/sched/sched.h  |  5 ++-
 4 files changed, 81 insertions(+), 62 deletions(-)

-- 
2.39.3