Re: [PATCH] sched: pull tasks when CPU is about to run SCHED_IDLE tasks

[chin <ultrachin@163.com>]

At 2020-12-23 19:30:26, "Vincent Guittot" <vincent.guittot@linaro.org> wrote:
>On Wed, 23 Dec 2020 at 09:32, <ultrachin@163.com> wrote:
>>
>> From: Chen Xiaoguang <xiaoggchen@tencent.com>
>>
>> Before a CPU switches from running SCHED_NORMAL task to
>> SCHED_IDLE task, trying to pull SCHED_NORMAL tasks from other
>
>Could you explain more in detail why you only care about this use case

>in particular and not the general case?


We want to run online tasks using SCHED_NORMAL policy and offline tasks
using SCHED_IDLE policy. The online tasks and the offline tasks run in
the same computer in order to use the computer efficiently.
The online tasks are in sleep in most times but should responce soon once
wake up. The offline tasks are in low priority and will run only when no online
tasks.

The online tasks are more important than the offline tasks and are latency
sensitive we should make sure the online tasks preempt the offline tasks
as soon as possilbe while there are online tasks waiting to run.
So in our situation we hope the SCHED_NORMAL to run if has any.

Let's assume we have 2 CPUs,
In CPU1 we got 2 SCHED_NORMAL tasks.
in CPU2 we got 1 SCHED_NORMAL task and 2 SCHED_IDLE tasks.

             CPU1                      CPU2
        curr       rq1            curr          rq2
      +------+ | +------+       +------+ | +----+ +----+
t0    |NORMAL| | |NORMAL|       |NORMAL| | |IDLE| |IDLE|
      +------+ | +------+       +------+ | +----+ +----+

                                 NORMAL exits or blocked
      +------+ | +------+                | +----+ +----+
t1    |NORMAL| | |NORMAL|                | |IDLE| |IDLE|
      +------+ | +------+                | +----+ +----+

                                 pick_next_task_fair
      +------+ | +------+         +----+ | +----+
t2    |NORMAL| | |NORMAL|         |IDLE| | |IDLE|
      +------+ | +------+         +----+ | +----+

                                 SCHED_IDLE running
t3    +------+ | +------+        +----+  | +----+
      |NORMAL| | |NORMAL|        |IDLE|  | |IDLE|
      +------+ | +------+        +----+  | +----+
                 
                                 run_rebalance_domains
      +------+ |                +------+ | +----+ +----+
t4    |NORMAL| |                |NORMAL| | |IDLE| |IDLE|
      +------+ |                +------+ | +----+ +----+

As we can see
t1: NORMAL task in CPU2 exits or blocked
t2: CPU2 pick_next_task_fair would pick a SCHED_IDLE to run while
another SCHED_NORMAL in rq1 is waiting. 
t3: SCHED_IDLE run in CPU2 while a SCHED_NORMAL wait in CPU1.
t4: after a short time, periodic load_balance triggerd and pull
SCHED_NORMAL in rq1 to rq2, and SCHED_NORMAL likely preempts SCHED_IDLE.

In this scenario, SCHED_IDLE is running while SCHED_NORMAL is waiting to run.
The latency of this SCHED_NORMAL will be high which is not acceptble.

Do a load_balance before running the SCHED_IDLE may fix this problem.

This patch works as below:

             CPU1                      CPU2
        curr       rq1            curr          rq2
      +------+ | +------+       +------+ | +----+ +----+
t0    |NORMAL| | |NORMAL|       |NORMAL| | |IDLE| |IDLE|
      +------+ | +------+       +------+ | +----+ +----+

                                 NORMAL exits or blocked
      +------+ | +------+                | +----+ +----+
t1    |NORMAL| | |NORMAL|                | |IDLE| |IDLE|
      +------+ | +------+                | +----+ +----+

t2                            pick_next_task_fair (all se are SCHED_IDLE)

                                 newidle_balance
      +------+ |                 +------+ | +----+ +----+
t3    |NORMAL| |                 |NORMAL| | |IDLE| |IDLE|
      +------+ |                 +------+ | +----+ +----+


t1: NORMAL task in CPU2 exits or blocked
t2: pick_next_task_fair check all se in rbtree are SCHED_IDLE and calls
newidle_balance who tries to pull a SCHED_NORMAL(if has).
t3: pick_next_task_fair would pick a SCHED_NORMAL to run instead of
SCHED_IDLE(likely).

>
>> CPU by doing load_balance first.
>>
>> Signed-off-by: Chen Xiaoguang <xiaoggchen@tencent.com>
>> Signed-off-by: Chen He <heddchen@tencent.com>
>> ---
>>  kernel/sched/fair.c | 5 +++++
>>  1 file changed, 5 insertions(+)
>>
>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
>> index ae7ceba..0a26132 100644
>> --- a/kernel/sched/fair.c
>> +++ b/kernel/sched/fair.c
>> @@ -7004,6 +7004,11 @@ struct task_struct *
>>         struct task_struct *p;
>>         int new_tasks;
>>
>> +       if (prev &&
>> +           fair_policy(prev->policy) &&
>
>Why do you need a prev and fair task  ? You seem to target the special
>case of pick_next_task  but in this case why not only testing rf!=null
> to make sure to not return immediately after jumping to the idle

>label?
We just want to do load_balance only when CPU switches from SCHED_NORMAL
to SCHED_IDLE.
If not check prev, when the running tasks are all SCHED_IDLE, we would
do newidle_balance everytime in pick_next_task_fair, it makes no sense
and kind of wasting.

>

>Also why not doing that for default case too ? i.e. balance_fair() ?
You are right, if you think this scenario makes sense, we will send a
refined patch soon :-)

>
>> +           sched_idle_cpu(rq->cpu))
>> +               goto idle;
>> +
>>  again:
>>         if (!sched_fair_runnable(rq))
>>                 goto idle;
>> --
>> 1.8.3.1
>>
>>