Re: [PATCH v5 0/3] Thermal ACPI APIs for generic trip points

From: Rafael J. Wysocki
Date: Thu Jan 19 2023 - 07:18:15 EST


On Thu, Jan 19, 2023 at 12:04 AM srinivas pandruvada
<srinivas.pandruvada@xxxxxxxxxxxxxxx> wrote:
>
> On Wed, 2023-01-18 at 23:14 +0100, Daniel Lezcano wrote:
> > On 18/01/2023 22:16, srinivas pandruvada wrote:
> > > On Wed, 2023-01-18 at 22:01 +0100, Daniel Lezcano wrote:
> > > > On 18/01/2023 21:53, srinivas pandruvada wrote:
> > > > > On Wed, 2023-01-18 at 21:00 +0100, Daniel Lezcano wrote:
> > > > > > On 18/01/2023 20:16, srinivas pandruvada wrote:
> > > > > >
> > > > > > [ ... ]
> > > > > >
> > > > > > > > > But we'd better wait for the thermald test result from
> > > > > > > > > Srinvias.
> > > > > > > >
> > > > > > > > A quick test show that things still work with thermald
> > > > > > > > and
> > > > > > > > these
> > > > > > > > changes.
> > > > > > >
> > > > > > > But I have a question. In some devices trip point
> > > > > > > temperature
> > > > > > > is
> > > > > > > not
> > > > > > > static. When hardware changes, we get notification. For
> > > > > > > example
> > > > > > > INT3403_PERF_TRIP_POINT_CHANGED for INT3403 drivers.
> > > > > > > Currently get_trip can get the latest changed value. But if
> > > > > > > we
> > > > > > > preregister, we need some mechanism to update them.
> > > > > >
> > > > > > When the notification INT3403_PERF_TRIP_POINT_CHANGED
> > > > > > happens, we
> > > > > > call
> > > > > > int340x_thermal_read_trips() which in turn updates the trip
> > > > > > points.
> > > > > >
> > > > >
> > > > > Not sure how we handle concurrency here when driver can freely
> > > > > update
> > > > > trips while thermal core is using trips.
> > > >
> > > > Don't we have the same race without this patch ? The thermal core
> > > > can
> > > > call get_trip_temp() while there is an update, no ?
> > > Yes it is. But I can add a mutex locally here to solve.
> > > But not any longer.
> > >
> > > I think you need some thermal_zone_read_lock/unlock() in core,
> > > which
> > > can use rcu. Even mutex is fine as there will be no contention as
> > > updates to trips will be rare.
> >
> > I was planning to provide a thermal_trips_update(tz, trips) and from
> > there handle the locking.
> >
> > As the race was already existing, can we postpone this change after
> > the
> > generic trip points changes?
> I think so.

Well, what if this bug is reported by a user and a fix needs to be
backported to "stable"?

Are we going to backport the whole framework redesign along with it?

Or is this extremely unlikely?