Re: [PATCH] psi: fix possible trigger missing in the window

[Zhaoyang Huang <huangzhaoyang@gmail.com>]

On Tue, Dec 21, 2021 at 3:58 AM Suren Baghdasaryan <surenb@google.com> wrote:
>
> On Fri, Dec 17, 2021 at 10:03 PM Zhaoyang Huang <huangzhaoyang@gmail.com> wrote:
> >
> > loop Suren
>
> Thanks.
>
>
> >
> > On Fri, Dec 17, 2021 at 2:08 PM Huangzhaoyang <huangzhaoyang@gmail.com> wrote:
> > >
> > > From: Zhaoyang Huang <zhaoyang.huang@unisoc.com>
> > >
> > > There could be missing wake up if the rest of the window remain the
> > > same stall states as the polling_total updates for every polling_min_period.
>
> Could you please expand on this description? I'm unclear what the
> problem is. I assume "polling_min_period" in this description refers
> to the group->poll_min_period.
>
> From the code, looks like the change results in update_triggers()
> calling window_update() once there was a new stall recorded for the
> trigger state and until the tracking window is complete. I don't see
> the point of calling window_update() if there was no stall change
> since the last call to window_update(). The resulting growth will not
> increase if there is no new stall.
> Maybe what you want to achieve here is more than one trigger per
> window if the stall limit was breached? If so, then this goes against
> the design for psi triggers in which we want to rate-limit the number
> of generated triggers per tracking window (see:
> https://elixir.bootlin.com/linux/latest/source/kernel/sched/psi.c#L545).
> Please clarify the issue and the intentions here.
> Thanks!
Please correct me if I am wrong. Imagine that there is a new stall
during the 1st polling_min_period among 10 of them in the window and
group->polling_total will be updated to total without trigger. If the
rest of 9 polling_min_periods remain the same states, the trigger will
be missed when window timing is reached.
>
> > >
> > > Signed-off-by: Zhaoyang Huang <zhaoyang.huang@unisoc.com>
> > > ---
> > >  include/linux/psi_types.h |  2 ++
> > >  kernel/sched/psi.c        | 30 ++++++++++++++++++------------
> > >  2 files changed, 20 insertions(+), 12 deletions(-)
> > >
> > > diff --git a/include/linux/psi_types.h b/include/linux/psi_types.h
> > > index 0a23300..9533d2e 100644
> > > --- a/include/linux/psi_types.h
> > > +++ b/include/linux/psi_types.h
> > > @@ -132,6 +132,8 @@ struct psi_trigger {
> > >
> > >         /* Refcounting to prevent premature destruction */
> > >         struct kref refcount;
> > > +
> > > +       bool new_stall;
> > >  };
> > >
> > >  struct psi_group {
> > > diff --git a/kernel/sched/psi.c b/kernel/sched/psi.c
> > > index 1652f2b..402718c 100644
> > > --- a/kernel/sched/psi.c
> > > +++ b/kernel/sched/psi.c
> > > @@ -458,9 +458,12 @@ static void psi_avgs_work(struct work_struct *work)
> > >  static void window_reset(struct psi_window *win, u64 now, u64 value,
> > >                          u64 prev_growth)
> > >  {
> > > +       struct psi_trigger *t = container_of(win, struct psi_trigger, win);
> > > +
> > >         win->start_time = now;
> > >         win->start_value = value;
> > >         win->prev_growth = prev_growth;
> > > +       t->new_stall = false;
> > >  }
> > >
> > >  /*
> > > @@ -515,7 +518,6 @@ static void init_triggers(struct psi_group *group, u64 now)
> > >  static u64 update_triggers(struct psi_group *group, u64 now)
> > >  {
> > >         struct psi_trigger *t;
> > > -       bool new_stall = false;
> > >         u64 *total = group->total[PSI_POLL];
> > >
> > >         /*
> > > @@ -523,19 +525,26 @@ static u64 update_triggers(struct psi_group *group, u64 now)
> > >          * watchers know when their specified thresholds are exceeded.
> > >          */
> > >         list_for_each_entry(t, &group->triggers, node) {
> > > -               u64 growth;
> > > -
> > >                 /* Check for stall activity */
> > >                 if (group->polling_total[t->state] == total[t->state])
> > >                         continue;
> > >
> > >                 /*
> > > -                * Multiple triggers might be looking at the same state,
> > > -                * remember to update group->polling_total[] once we've
> > > -                * been through all of them. Also remember to extend the
> > > -                * polling time if we see new stall activity.
> > > +                * update the trigger if there is new stall which will be
> > > +                * reset when run out of the window
> > >                  */
> > > -               new_stall = true;
> > > +               t->new_stall = true;
> > > +
> > > +               memcpy(&group->polling_total[t->state], &total[t->state],
> > > +                               sizeof(group->polling_total[t->state]));
> > > +       }
> > > +
> > > +       list_for_each_entry(t, &group->triggers, node) {
> > > +               u64 growth;
> > > +
> > > +               /* check if new stall happened during this window*/
> > > +               if (!t->new_stall)
> > > +                       continue;
> > >
> > >                 /* Calculate growth since last update */
> > >                 growth = window_update(&t->win, now, total[t->state]);
> > > @@ -552,10 +561,6 @@ static u64 update_triggers(struct psi_group *group, u64 now)
> > >                 t->last_event_time = now;
> > >         }
> > >
> > > -       if (new_stall)
> > > -               memcpy(group->polling_total, total,
> > > -                               sizeof(group->polling_total));
> > > -
> > >         return now + group->poll_min_period;
> > >  }
> > >
> > > @@ -1152,6 +1157,7 @@ struct psi_trigger *psi_trigger_create(struct psi_group *group,
> > >         t->last_event_time = 0;
> > >         init_waitqueue_head(&t->event_wait);
> > >         kref_init(&t->refcount);
> > > +       t->new_stall = false;
> > >
> > >         mutex_lock(&group->trigger_lock);
> > >
> > > --
> > > 1.9.1
> > >