Message-ID: <CAJZ5v0gwgBmj0XaLhBz0puSOA_OYUy0oFEjoBVCWjT3XW+qZVg@mail.gmail.com>
Date:   Tue, 22 Aug 2023 12:57:59 +0200
From:   "Rafael J. Wysocki" <rafael@...nel.org>
To:     Nícolas F. R. A. Prado 
        <nfraprado@...labora.com>
Cc:     "Rafael J. Wysocki" <rafael@...nel.org>,
        Daniel Lezcano <daniel.lezcano@...aro.org>,
        AngeloGioacchino Del Regno 
        <angelogioacchino.delregno@...labora.com>, kernel@...labora.com,
        Amit Kucheria <amitk@...nel.org>,
        Zhang Rui <rui.zhang@...el.com>, linux-kernel@...r.kernel.org,
        linux-pm@...r.kernel.org
Subject: Re: [PATCH] thermal/core: Don't update trip points inside the
 hysteresis range

On Tue, Aug 22, 2023 at 12:25 AM Nícolas F. R. A. Prado
<nfraprado@...labora.com> wrote:
>
> On Mon, Aug 21, 2023 at 11:10:27PM +0200, Rafael J. Wysocki wrote:
> > On Mon, Jul 3, 2023 at 7:15 PM Nícolas F. R. A. Prado
> > <nfraprado@...labora.com> wrote:
> > >
> > > When searching for the trip points that need to be set, the nearest trip
> > > point's temperature is used for the high trip, while the nearest trip
> > > point's temperature minus the hysteresis is used for the low trip. The
> > > issue with this logic is that when the current temperature is inside a
> > > trip point's hysteresis range, both high and low trips will come from
> > > the same trip point. As a consequence instability can still occur like
> > > this:
> > > * the temperature rises slightly and enters the hysteresis range of a
> > >   trip point
> > > * polling happens and updates the trip points to the hysteresis range
> > > * the temperature falls slightly, exiting the hysteresis range, crossing
> > >   the trip point and triggering an IRQ, the trip points are updated
> > > * repeat
> > >
> > > So even though the current hysteresis implementation prevents
> > > instability from happening due to IRQs triggering on the same
> > > temperature value, both ways, it doesn't prevent it from happening due
> > > to an IRQ on one way and polling on the other.
> > >
> > > To properly implement a hysteresis behavior, when inside the hysteresis
> > > range, don't update the trip points. This way, the previously set trip
> > > points will stay in effect, which will in a way remember the previous
> > > state (if the temperature signal came from above or below the range) and
> > > therefore have the right trip point already set. The exception is if
> > > there was no previous trip point set, in which case a previous state
> > > doesn't exist, and so it's sensible to allow the hysteresis range as
> > > trip points.
> > >
> > > Signed-off-by: Nícolas F. R. A. Prado <nfraprado@...labora.com>
> > >
> > > ---
> > >
> > >  drivers/thermal/thermal_trip.c | 21 +++++++++++++++++++--
> > >  1 file changed, 19 insertions(+), 2 deletions(-)
> > >
> > > diff --git a/drivers/thermal/thermal_trip.c b/drivers/thermal/thermal_trip.c
> > > index 907f3a4d7bc8..c386ac5d8bad 100644
> > > --- a/drivers/thermal/thermal_trip.c
> > > +++ b/drivers/thermal/thermal_trip.c
> > > @@ -57,6 +57,7 @@ void __thermal_zone_set_trips(struct thermal_zone_device *tz)
> > >  {
> > >         struct thermal_trip trip;
> > >         int low = -INT_MAX, high = INT_MAX;
> > > +       int low_trip_id = -1, high_trip_id = -2;
> > >         int i, ret;
> > >
> > >         lockdep_assert_held(&tz->lock);
> > > @@ -73,18 +74,34 @@ void __thermal_zone_set_trips(struct thermal_zone_device *tz)
> > >
> > >                 trip_low = trip.temperature - trip.hysteresis;
> > >
> > > -               if (trip_low < tz->temperature && trip_low > low)
> > > +               if (trip_low < tz->temperature && trip_low > low) {
> > >                         low = trip_low;
> > > +                       low_trip_id = i;
> > > +               }
> > >
> >
> > I think I get the idea, but wouldn't a similar effect be achieved by
> > adding an "else" here?
>
> No. That would only fix the problem in one direction, namely, when the
> temperature entered the hysteresis range from above. But when the temperature
> entered the range from below, we'd need to check the high threshold first to
> achieve the same result.
>
> The way I've implemented here is the simplest I could think of that works for
> both directions.

Well, what about the replacement patch below (untested)?

---
 drivers/thermal/thermal_trip.c |   19 +++++++++++++++++--
 1 file changed, 17 insertions(+), 2 deletions(-)

Index: linux-pm/drivers/thermal/thermal_trip.c
===================================================================
--- linux-pm.orig/drivers/thermal/thermal_trip.c
+++ linux-pm/drivers/thermal/thermal_trip.c
@@ -55,6 +55,7 @@ void __thermal_zone_set_trips(struct the
 {
     struct thermal_trip trip;
     int low = -INT_MAX, high = INT_MAX;
+    bool same_trip = false;
     int i, ret;

     lockdep_assert_held(&tz->lock);
@@ -63,6 +64,7 @@ void __thermal_zone_set_trips(struct the
         return;

     for (i = 0; i < tz->num_trips; i++) {
+        bool low_set = false;
         int trip_low;

         ret = __thermal_zone_get_trip(tz, i , &trip);
@@ -71,18 +73,31 @@ void __thermal_zone_set_trips(struct the

         trip_low = trip.temperature - trip.hysteresis;

-        if (trip_low < tz->temperature && trip_low > low)
+        if (trip_low < tz->temperature && trip_low > low) {
             low = trip_low;
+            low_set = true;
+            same_trip = false;
+        }

         if (trip.temperature > tz->temperature &&
-            trip.temperature < high)
+            trip.temperature < high) {
             high = trip.temperature;
+            same_trip = low_set;
+        }
     }

     /* No need to change trip points */
     if (tz->prev_low_trip == low && tz->prev_high_trip == high)
         return;

+    /*
+     * If "high" and "low" are the same, skip the change unless this is the
+     * first time.
+     */
+    if (same_trip && (tz->prev_low_trip != -INT_MAX ||
+        tz->prev_high_trip != INT_MAX))
+        return;
+
     tz->prev_low_trip = low;
     tz->prev_high_trip = high;

Powered by blists - more mailing lists