[PATCH] sched/fair: don't use waker's cpu if the waker of sync wake-up is interrupt

From: Barry Song
Date: Mon Apr 26 2021 - 22:45:24 EST

a severe qperf performance decrease was reported in the below use case:
For a hardware with 2 NUMA nodes, node0 has cpu0-31, node1 has cpu32-63.
Ethernet is located in node1.

Run the below commands:
$ taskset -c 32-63 stress -c 32 &
$ qperf tcp_lat
latency = 2.95ms.
Normally the latency should be less than 20us. But in the above test,
latency increased dramatically to 2.95ms.

This is caused by ping-pong of qperf between node0 and node1. Since it
is a sync wake-up and waker's nr_running == 1, WAKE_AFFINE will pull
qperf to node1, but LB will soon migrate qperf back to node0.
Not like a normal sync wake-up coming from a task, the waker in the above
test is an interrupt and nr_running happens to be 1 since stress starts
32 threads on node1 with 32 cpus.

Testing also shows the performance of qperf won't drop if the number
of threads are increased to 64, 96 or larger values:
$ taskset -c 32-63 stress -c 96 &
$ qperf tcp_lat
latency = 14.7us.

Obviously "-c 96" makes "cpu_rq(this_cpu)->nr_running == 1" false in
wake_affine_idle() so WAKE_AFFINE won't pull qperf to node1.

To fix this issue, this patch checks the waker of sync wake-up is a task
but not an interrupt. In this case, the waker will schedule out and give
CPU to wakee.

Reported-by: Yongjia Xie <xieyongjia1@xxxxxxxxxx>
Signed-off-by: Barry Song <song.bao.hua@xxxxxxxxxxxxx>
kernel/sched/fair.c | 7 ++++++-
1 file changed, 6 insertions(+), 1 deletion(-)

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 6d73bdbb2d40..8ad2d732033d 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -5829,7 +5829,12 @@ wake_affine_idle(int this_cpu, int prev_cpu, int sync)
if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;

- if (sync && cpu_rq(this_cpu)->nr_running == 1)
+ /*
+ * If this is a sync wake-up and the only running thread is just
+ * waker, thus, waker is not interrupt, we assume wakee will get
+ * the cpu of waker soon
+ */
+ if (sync && cpu_rq(this_cpu)->nr_running == 1 && in_task())
return this_cpu;

if (available_idle_cpu(prev_cpu))