The try_to_wake_up function has an optimization where it can queue
a task for wakeup on its previous CPU, if the task is still in the
middle of going to sleep inside schedule().
Once schedule() re-enables IRQs, the task will be woken up with an
IPI, and placed back on the runqueue.
If we have such a wakeup pending, there is no need to search other
CPUs for runnable tasks. Just skip (or bail out early from) newidle
balancing, and run the just woken up task.
For a memcache like workload test, this reduces total CPU use by
about 2%, proportionally split between user and system time,
and p99 and p95 application response time by 2-3% on average.
The schedstats run_delay number shows a similar improvement.
Signed-off-by: Rik van Riel <[email protected]>
---
v2:
- fix !SMP build error and prev-not-CFS case by moving check into newidle_balance
- fix formatting of if condition
- audit newidle_balance return value use to make sure we get that right
- reset idle_stamp when breaking out of the loop due to ->ttwu_pending
kernel/sched/fair.c | 13 +++++++++++--
1 file changed, 11 insertions(+), 2 deletions(-)
diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 69680158963f..5e26f013e182 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -10594,6 +10594,14 @@ static int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
u64 curr_cost = 0;
update_misfit_status(NULL, this_rq);
+
+ /*
+ * There is a task waiting to run. No need to search for one.
+ * Return 0; the task will be enqueued when switching to idle.
+ */
+ if (this_rq->ttwu_pending)
+ return 0;
+
/*
* We must set idle_stamp _before_ calling idle_balance(), such that we
* measure the duration of idle_balance() as idle time.
@@ -10661,7 +10669,8 @@ static int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
* Stop searching for tasks to pull if there are
* now runnable tasks on this rq.
*/
- if (pulled_task || this_rq->nr_running > 0)
+ if (pulled_task || this_rq->nr_running > 0 ||
+ this_rq->ttwu_pending)
break;
}
rcu_read_unlock();
@@ -10688,7 +10697,7 @@ static int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
if (this_rq->nr_running != this_rq->cfs.h_nr_running)
pulled_task = -1;
- if (pulled_task)
+ if (pulled_task || this_rq->ttwu_pending)
this_rq->idle_stamp = 0;
rq_repin_lock(this_rq, rf);
--
2.25.4
On Mon, 19 Apr 2021 at 18:51, Rik van Riel <[email protected]> wrote:
>
> The try_to_wake_up function has an optimization where it can queue
> a task for wakeup on its previous CPU, if the task is still in the
> middle of going to sleep inside schedule().
>
> Once schedule() re-enables IRQs, the task will be woken up with an
> IPI, and placed back on the runqueue.
>
> If we have such a wakeup pending, there is no need to search other
> CPUs for runnable tasks. Just skip (or bail out early from) newidle
> balancing, and run the just woken up task.
>
> For a memcache like workload test, this reduces total CPU use by
> about 2%, proportionally split between user and system time,
> and p99 and p95 application response time by 2-3% on average.
> The schedstats run_delay number shows a similar improvement.
>
> Signed-off-by: Rik van Riel <[email protected]>
> ---
> v2:
> - fix !SMP build error and prev-not-CFS case by moving check into newidle_balance
> - fix formatting of if condition
> - audit newidle_balance return value use to make sure we get that right
> - reset idle_stamp when breaking out of the loop due to ->ttwu_pending
>
> kernel/sched/fair.c | 13 +++++++++++--
> 1 file changed, 11 insertions(+), 2 deletions(-)
>
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index 69680158963f..5e26f013e182 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -10594,6 +10594,14 @@ static int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
> u64 curr_cost = 0;
>
> update_misfit_status(NULL, this_rq);
> +
> + /*
> + * There is a task waiting to run. No need to search for one.
> + * Return 0; the task will be enqueued when switching to idle.
> + */
> + if (this_rq->ttwu_pending)
> + return 0;
> +
> /*
> * We must set idle_stamp _before_ calling idle_balance(), such that we
> * measure the duration of idle_balance() as idle time.
> @@ -10661,7 +10669,8 @@ static int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
> * Stop searching for tasks to pull if there are
> * now runnable tasks on this rq.
> */
> - if (pulled_task || this_rq->nr_running > 0)
> + if (pulled_task || this_rq->nr_running > 0 ||
> + this_rq->ttwu_pending)
> break;
> }
> rcu_read_unlock();
> @@ -10688,7 +10697,7 @@ static int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
> if (this_rq->nr_running != this_rq->cfs.h_nr_running)
> pulled_task = -1;
>
> - if (pulled_task)
> + if (pulled_task || this_rq->ttwu_pending)
This needs at least a comment to explain why we must clear
this_rq->idle_stamp when this_rq->ttwu_pending is set whereas it is
also done during sched_ttwu_pending()
> this_rq->idle_stamp = 0;
>
> rq_repin_lock(this_rq, rf);
> --
> 2.25.4
>
>
On Tue, 2021-04-20 at 11:04 +0200, Vincent Guittot wrote:
> On Mon, 19 Apr 2021 at 18:51, Rik van Riel <[email protected]> wrote:
> >
> > @@ -10688,7 +10697,7 @@ static int newidle_balance(struct rq
> > *this_rq, struct rq_flags *rf)
> > if (this_rq->nr_running != this_rq->cfs.h_nr_running)
> > pulled_task = -1;
> >
> > - if (pulled_task)
> > + if (pulled_task || this_rq->ttwu_pending)
>
> This needs at least a comment to explain why we must clear
> this_rq->idle_stamp when this_rq->ttwu_pending is set whereas it is
> also done during sched_ttwu_pending()
>
> > this_rq->idle_stamp = 0;
I spent some time staring at sched_ttwu_pending and
the functions it calls, but I can't seem to spot
where it clears rq->idle_stamp, except inside
ttwu_do_wakeup where it will end up adding a
non-idle period into the rq->avg_idle, which seems
wrong.
If we are actually idle, and get woken up with a
ttwu_queue task, we do not come through newidle_balance,
and we end up counting the idle time into the avg_idle
number.
However, if a task is woken up while the CPU is
in newidle_balance, because prev != idle, we should
not count that period towards rq->avg_idle, for
the same reason we do so when we pulled a task.
I'll add a comment in v3 explaining why idle_stamp
needs to be 0.
--
All Rights Reversed.
On Tue, 20 Apr 2021 at 17:20, Rik van Riel <[email protected]> wrote:
>
> On Tue, 2021-04-20 at 11:04 +0200, Vincent Guittot wrote:
> > On Mon, 19 Apr 2021 at 18:51, Rik van Riel <[email protected]> wrote:
> > >
> > > @@ -10688,7 +10697,7 @@ static int newidle_balance(struct rq
> > > *this_rq, struct rq_flags *rf)
> > > if (this_rq->nr_running != this_rq->cfs.h_nr_running)
> > > pulled_task = -1;
> > >
> > > - if (pulled_task)
> > > + if (pulled_task || this_rq->ttwu_pending)
> >
> > This needs at least a comment to explain why we must clear
> > this_rq->idle_stamp when this_rq->ttwu_pending is set whereas it is
> > also done during sched_ttwu_pending()
> >
> > > this_rq->idle_stamp = 0;
>
> I spent some time staring at sched_ttwu_pending and
> the functions it calls, but I can't seem to spot
> where it clears rq->idle_stamp, except inside
> ttwu_do_wakeup where it will end up adding a
> non-idle period into the rq->avg_idle, which seems
> wrong.
Not sure that this is really wrong because it ends up scheduling the
idle task which is immediately preempted. But the preemption happened
in the idle task, isn't it ?
>
> If we are actually idle, and get woken up with a
> ttwu_queue task, we do not come through newidle_balance,
> and we end up counting the idle time into the avg_idle
> number.
>
> However, if a task is woken up while the CPU is
> in newidle_balance, because prev != idle, we should
> not count that period towards rq->avg_idle, for
> the same reason we do so when we pulled a task.
As mentioned above, we have effectively schedule the idle task in your
case whereas we don't in the other cases
IIUC, your problem comes from rq->avg_idle decreasing a lot in such
cases. And because rq->avg_idle is used to decide if you have time to
run newlyidle_balance,you skip it more often.
>
> I'll add a comment in v3 explaining why idle_stamp
> needs to be 0.
Yes please.
>
> --
> All Rights Reversed.