7:19 a.m.
Hello, Kazu
2022년 3월 3일 (목) 오후 4:50, HAGIO KAZUHITO(萩尾 一仁) <k-hagio-ab(a)nec.xn--com>-4f21ay07k 작성:
Hi Austin,
-----Original Message-----
> > > This is because output of "ps -l|-m" depends on
task_struct.sched_info.last_arrival.
> > >
> > > Without CONFIG_SCHEDSTATS or CONFIG_SCHED_INFO, 'sched_info' field
is not included
> > > in task_struct.
> > >
> > > So we make "ps -l|-m" option to access 'exec_start'
field of sched_entity
> > > where 'exec_start' is task_struct.se.exec_start.
> >
> > Could you describe what the exec_start means? When is it updated?
> >
>
> The 'task_struct.se.exec_start' contains the most recently-executed
> timestamp when
> process is running in the below cases;
>
> - enqueued to runqueue
> - dequeued from ruqueue
> - scheduler tick is invoked
> - etc
>
> So I guess 'task_struct.se.exec_start' could be one of statistics
> which indicates
> the most recently run timestamp of process activity.
>
> From CFS scheduler's point of view, 'task_struct.se.exec_start' is
> updated within update_curr()
> where its call path is various as below.
>
> - enqueue_task_fair, -dequeue_task_fair, task_tick_fair,
> check_preempt_wakeup, ...
Thank you for looking into this.
As for when the se.exec_start is updated, I think you are right.
(with my understanding, probably it's the last time or tick when a task
is in a runqueue regardless of getting a CPU.)
But I found a problem, is that the se.exec_start is from rq->clock_task,
not from rq->clock like last_arrival. The rq->clock_task may not contain
irq/steal time, please see update_rq_clock_task().
I am understood after looking at below code.
static void sched_info_arrive(struct rq *rq, struct task_struct *t)
{
...
now = rq_clock(rq); //<<--
...
t->sched_info.last_arrival = now;
}
static void update_curr(struct cfs_rq *cfs_rq)
{
u64 now = rq_clock_task(rq_of(cfs_rq)); //<<--
...
curr->exec_start = now;
...
}
This causes the following issues with a vmcore generated on a
machine,
which had run for a while (273 days):
Let me ask a question at this point.
The vmcore was generated from machine which has been running 273 days.
I wonder the kernel image of machine(running 273 days) is without
CONFIG_SCHEDSTATS/CONFIG_SCHED_INFO.
If output is based on vmcore without CONFIG_SCHEDSTATS or CONFIG_SCHED_INFO,
you can ignore above question.
I would like to remind that as commit description said, the patch is
**only** valid
without CONFIG_SCHEDSTATS or CONFIG_SCHED_INFO.
When kernel image is with CONFIG_SCHEDSTATS or CONFIG_SCHED_INFO and then
"ps -l|-m" command is used, task_last_run() is called from its subroutines.
ulonglong
task_last_run(ulong task)
{
ulong last_run;
...
else if (VALID_MEMBER(sched_info_last_arrival)) // With
CONFIG_SCHEDSTATS or CONFIG_SCHED_INFO
timestamp = tt->last_task_read ? ULONGLONG(tt->task_struct +
OFFSET(task_struct_sched_info) +
OFFSET(sched_info_last_arrival)) : 0;
+ else if (VALID_MEMBER(sched_entity_exec_start)) // Without
CONFIG_SCHEDSTATS or CONFIG_SCHED_INFO
+ timestamp = tt->last_task_read ? ULONGLONG(tt->task_struct +
+ OFFSET(task_struct_se) +
+ OFFSET(sched_entity_exec_start)) : 0;
return timestamp;
}
With CONFIG_SCHEDSTATS or CONFIG_SCHED_INFO enabled,
the statement under 'VALID_MEMBER(sched_entity_exec_start)' is not called.
crash> ps -m | head
[ 0 00:59:36.582] [RU] PID: 4023608 TASK: ffff916f7c6b1840 CPU: 15 COMMAND:
"makedumpfile"
^^^^^^^^^(1)
[ 0 00:59:37.831] [ID] PID: 413 TASK: ffff916f772d3080 CPU: 15 COMMAND:
"kworker/15:1H"
[ 0 00:59:39.765] [IN] PID: 3929504 TASK: ffff916f5f0748c0 CPU: 15 COMMAND:
"respawn_actlog"
[ 0 00:59:40.650] [IN] PID: 1974 TASK: ffff91647dc53080 CPU: 15 COMMAND:
"CPU 2/KVM"
[ 0 00:59:41.925] [IN] PID: 1297 TASK: ffff916f63c46100 CPU: 15 COMMAND:
"NetworkManager"
[ 0 00:59:42.944] [ID] PID: 3763057 TASK: ffff9160c4519840 CPU: 15 COMMAND:
"kworker/15:0"
[ 0 00:59:42.944] [IN] PID: 101 TASK: ffff916040c91840 CPU: 15 COMMAND:
"migration/15"
[ 0 00:59:43.078] [IN] PID: 100 TASK: ffff916040c5b080 CPU: 15 COMMAND:
"watchdog/15"
[ 0 00:59:47.533] [IN] PID: 1292 TASK: ffff916f63c43080 CPU: 15 COMMAND:
"lsmd"
[ 0 00:59:49.089] [IN] PID: 113105 TASK: ffff9160412248c0 CPU: 15 COMMAND:
"kvm-nx-lpage-re"
^^^^^^^(2)
(1) large difference from zero
(2) large differences among CPUs (probably due to the differences of irq time)
(1) might be solved with rq->clock_task, but (2) will be misleading and confusing.
So currently I'm thinking that the "ps -l|-m" options should not use the
se.exec_start.
What do you think?
Agreed. (As you indicated) Since rq->clock cannot hold accurate
timestamp we can rely on,
"ps -l|-m" options should not access the se.exec_start in case of kernel image
without CONFIG_SCHEDSTATS or CONFIG_SCHED_INFO.
Thanks for your code review.
If there is something to improve, let me propose a patch again.
BR,
Austin Kim
Thanks,
Kazu