Re: [PATCH] psi: fix aggregation idle shut-off

From: Andrew Morton
Date: Mon Jan 28 2019 - 18:06:40 EST


On Wed, 16 Jan 2019 14:35:01 -0500 Johannes Weiner <hannes@xxxxxxxxxxx> wrote:

> psi has provisions to shut off the periodic aggregation worker when
> there is a period of no task activity - and thus no data that needs
> aggregating. However, while developing psi monitoring, Suren noticed
> that the aggregation clock currently won't stay shut off for good.
>
> Debugging this revealed a flaw in the idle design: an aggregation run
> will see no task activity and decide to go to sleep; shortly
> thereafter, the kworker thread that executed the aggregation will go
> idle and cause a scheduling change, during which the psi callback will
> kick the !pending worker again. This will ping-pong forever, and is
> equivalent to having no shut-off logic at all (but with more code!)
>
> Fix this by exempting aggregation workers from psi's clock waking
> logic when the state change is them going to sleep. To do this, tag
> workers with the last work function they executed, and if in psi we
> see a worker going to sleep after aggregating psi data, we will not
> reschedule the aggregation work item.
>
> What if the worker is also executing other items before or after?
>
> Any psi state times that were incurred by work items preceding the
> aggregation work will have been collected from the per-cpu buckets
> during the aggregation itself. If there are work items following the
> aggregation work, the worker's last_func tag will be overwritten and
> the aggregator will be kept alive to process this genuine new activity.
>
> If the aggregation work is the last thing the worker does, and we
> decide to go idle, the brief period of non-idle time incurred between
> the aggregation run and the kworker's dequeue will be stranded in the
> per-cpu buckets until the clock is woken by later activity. But that
> should not be a problem. The buckets can hold 4s worth of time, and
> future activity will wake the clock with a 2s delay, giving us 2s
> worth of data we can leave behind when disabling aggregation. If it
> takes a worker more than two seconds to go idle after it finishes its
> last work item, we likely have bigger problems in the system, and
> won't notice one sample that was averaged with a bogus per-CPU weight.
>
> --- a/kernel/sched/psi.c
> +++ b/kernel/sched/psi.c
> @@ -480,9 +481,6 @@ static void psi_group_change(struct psi_group *group, int cpu,
> groupc->tasks[t]++;
>
> write_seqcount_end(&groupc->seq);
> -
> - if (!delayed_work_pending(&group->clock_work))
> - schedule_delayed_work(&group->clock_work, PSI_FREQ);
> }
>
> static struct psi_group *iterate_groups(struct task_struct *task, void **iter)

This breaks Suren's "psi: introduce psi monitor":

--- kernel/sched/psi.c~psi-introduce-psi-monitor
+++ kernel/sched/psi.c
@@ -752,8 +1012,25 @@ static void psi_group_change(struct psi_

write_seqcount_end(&groupc->seq);

- if (!delayed_work_pending(&group->clock_work))
- schedule_delayed_work(&group->clock_work, PSI_FREQ);
+ /*
+ * Polling flag resets to 0 at the max rate of once per update window
+ * (at least 500ms interval). smp_wmb is required after group->polling
+ * 0-to-1 transition to order groupc->times and group->polling writes
+ * because stall detection logic in the slowpath relies on groupc->times
+ * changing before group->polling. Explicit smp_wmb is missing because
+ * cmpxchg() implies smp_mb.
+ */
+ if ((state_mask & group->trigger_mask) &&
+ atomic_cmpxchg(&group->polling, 0, 1) == 0) {
+ /*
+ * Start polling immediately even if the work is already
+ * scheduled
+ */
+ mod_delayed_work(system_wq, &group->clock_work, 1);
+ } else {
+ if (!delayed_work_pending(&group->clock_work))
+ schedule_delayed_work(&group->clock_work, PSI_FREQ);
+ }
}

and I'm too lazy to go in and figure out how to fix it.

If we're sure about "psi: fix aggregation idle shut-off" (and I am not)
then can I ask for a redo of "psi: introduce psi monitor"?