Date:   Sun, 18 Sep 2022 22:16:53 -0700
From:   Suren Baghdasaryan <surenb@...gle.com>
To:     Hao Lee <haolee.swjtu@...il.com>
Cc:     Johannes Weiner <hannes@...xchg.org>,
        Peter Zijlstra <peterz@...radead.org>,
        Zhaoyang Huang <zhaoyang.huang@...soc.com>,
        LKML <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH] psi: fix possible missing or delayed pending event

On Sun, Sep 18, 2022 at 3:55 AM Hao Lee <haolee.swjtu@...il.com> wrote:
>
> On Sat, Sep 17, 2022 at 09:44:12PM -0700, Suren Baghdasaryan wrote:
> > On Sat, Sep 17, 2022 at 12:31 AM Hao Lee <haolee.swjtu@...il.com> wrote:
> > >
> > > On Fri, Sep 16, 2022 at 11:08:34PM -0700, Suren Baghdasaryan wrote:
> > > > On Wed, Sep 14, 2022 at 2:30 AM Hao Lee <haolee.swjtu@...il.com> wrote:
> > > > >
> > > > > When a pending event exists and growth is less than the threshold, the
> > > > > current logic is to skip this trigger without generating event. However,
> > > > > from e6df4ead85d9 ("psi: fix possible trigger missing in the window"),
> > > > > our purpose is to generate event as long as pending event exists and the
> > > > > rate meets the limit. This patch fixes the possible pending-event
> > > > > missing or delay.
> > > > >
> > > > > Fixes: e6df4ead85d9 ("psi: fix possible trigger missing in the window")
> > > > > Signed-off-by: Hao Lee <haolee.swjtu@...il.com>
> > > > > ---
> > > > >  kernel/sched/psi.c | 2 +-
> > > > >  1 file changed, 1 insertion(+), 1 deletion(-)
> > > > >
> > > > > diff --git a/kernel/sched/psi.c b/kernel/sched/psi.c
> > > > > index 9711827e3..0bae4ee2b 100644
> > > > > --- a/kernel/sched/psi.c
> > > > > +++ b/kernel/sched/psi.c
> > > > > @@ -539,7 +539,7 @@ static u64 update_triggers(struct psi_group *group, u64 now)
> > > > >
> > > > >                         /* Calculate growth since last update */
> > > > >                         growth = window_update(&t->win, now, total[t->state]);
> > > > > -                       if (growth < t->threshold)
> > > > > +                       if (growth < t->threshold && !t->pending_event)
> > > >
> > > > I'm not sure how this additional condition changes things. Current
> > > > logic is to set t->pending_event=true whenever growth exceeds the
> > > > t->threshold. This patch will change this logic into setting
> > > > t->pending_event=true also when t->pending_event=true.
> > >
> > > This is right.
> > >
> > > > But why would
> > > > you want to set t->pending_event=true if it's already true? What am I
> > > > missing?
> > >
> > > If I expand this if-else branch and the pending_event statement
> > > to a more detailed snippet, it will be like this:
> > >
> > > if (growth < t->threshold && !t->pending_event) // under threshold && no pending event. Skip.
> > >         continue;
> > > else if (growth >= t->threshold) // above threshold. Try to generate event.
> > >         t->pending_event = true;
> > > else // under threshold && have pending events. Try to generate event.
> > >         ; // pending_event is already true. do nothing
> > >
> > >
> > > The original code didn't handle the `else` condition properly.
> >
> > The `else` condition in your code does nothing, and that's why the
> > original code does not implement a handler for that case.
> >
> > > It will
> > > skip the trigger when its growth is under the threshold, even though it
> > > has a pending event. This patch handles this condition correctly.
> > >
> > > But I think assigning true to pending_event when it's already true doesn't
> > > have other side effects, so I eliminate the `else if` branch. Maybe we'd
> > > better make it explicit, like the above snippet? Thanks.
> >
> > The new code you posted is functionally the same as the original one
> > while being more verbose and IMO less readable. Unless you can explain
> > the problem with the original code, I don't see any reason to change
> > it.
>
> Hi, for the original code, let's assume t->pending_event is true:
>     * if new_stall is false, we will try to check event ratelimit and
>       generate an event for this psi_trigger. This case is right.
>     * but if new_stall is true, we will skip this psi_trigger if growth
>       growth < t->threshold. I think we shouldn't skip this psi_trigger
>       in this case because it has a pending event.

Ok, I see the issue now. I think the following fix would be the simplest:

                       /* Calculate growth since last update */
                        growth = window_update(&t->win, now, total[t->state]);
-                       if (growth < t->threshold)
-                                continue;
+                       if (!t->pending_event) {
+                              if (growth < t->threshold)
+                                        continue;

-                        t->pending_event = true;
+                               t->pending_event = true;
+                       }


>
> >
> > >
> > > >
> > > > >                                 continue;
> > > > >
> > > > >                         t->pending_event = true;
> > > > > --
> > > > > 2.21.0
> > > > >

Powered by blists - more mailing lists