On 01/27/2016 01:18 PM, Stéphan Kochen wrote:

On Fri, Jan 22, 2016 at 06:53:50PM -0800, Guenter Roeck wrote:

quoted

The thermal subsystem supports setting trip points, which I would think
is what you are looking for here. Question is if an how you can use the
information from the thermal subsystem (and thus the thermal zone configuration)
to set the various limits in the lm90 driver. This should hopefully be sufficient
to fix your immediate problem. For handling alerts, I guess we'll have to wait for
thermal subsystem improvements (unless of course you volunteer to do that work.

I may take a shot at this. So in short, the goal is to have a device
tree thermal zone communicate trip points to the sensor driver, and use
interrupts to act on trips.

(This indirectly solves my problem of my sensor having weird initial
values. Perhaps we also want a solution for this case if the thermal
subsystem is disabled, but for me there'd be no need.)

Here's what I see:

  - The thermal core layer already supports interrupt driven systems.
    Support is missing from thermal OF, the layer between thermal core
    and the sensor driver implementing device tree support.

  - Thermal OF registers a device in thermal core for each zone defined
    in the device tree.

  - In theory, a thermal zone in the device tree can have multiple
    sensors, and multiple zones can refer to the same sensor, but the
    current implementation only supports 1-on-1 relations.

  - There are already exports thermal_zone_device_update and
    thermal_notify_framework in thermal core, which allow external code
    to trigger an update.

  - Updates happen by explicit calls to such exports, or by an optional
    and configurable interval in thermal core.

What I think we want:

  - Any additions should be optional extensions implemented by sensor
    drivers. Polling should keep on working for all sensor drivers
    already supporting thermal OF, with no interface changes.

  - For interrupt-capable sensor drivers, the thermal OF device should
    keep the sensor driver updated with the current nearest trip
    temperature and hysteresis. (Don't burden drivers with a full list of
    trip points.)

  - In the case of LM90, this would set the low and high alert
    temperatures. LM90 can have additional alerts (critical, emergency),
    but a sensor driver registered with thermal OF should disable any
    additional alerts.

... and thus disable any external chip signals associated with those limits.

  - Similarly, a sensor driver should disable alerts when there is no
    current trip temperature or hysteresis. (E.g., we're below the lowest
    trip point.)

The idea with most if not all temperature sensors is that multiple trip points
would be configured as multiple limits in the chip. Often those limits,
when reached, are tied to pins to cause a specific action (eg to cause
an interrupt, turn on a fan, or shut down the hardware). In effect, you
suggest to re-define this mechanism and, for all practical purposes,
use just one of the limits provided by the chip(s).

Personally I don't think that would be a good idea, because it would have
impact on hardware design. It would effectively limit the use of the thermal
subsystem with temperature sensor chips to hardware designs which take
the thermal subsystem's expectations and assumptions into account.
At the same time, it would for all practical purposes mandate the use
of the thermal subsystem on such systems, because the hardware would depend
on the thermal subsystem's implementation to control the temperature
in the system.

While it may be feasible in some situations to have the thermal subsystem
dynamically set free-moving limits, there are many other situations where
those limits are tied to hardware responses, and the limits need to be static.

Maybe this is just a different world view. The thermal subsystem may see the
world assuming that it always has an unlimited number of trip points available,
and the chip it supports only support lower and upper boundaries (or trip points),
which can be set and changed as needed. This is somewhat different to the
traditional world view, implemented not only in many temperature sensors,
but also in fan controller or Super-IO chips, where a set of temperatures
is programmed into the chip only once. I hope that it is possible to support
both mechanisms.

Thanks,
Guenter

Implementation-wise:

  - The struct thermal_zone_of_device_ops needs an additional function to
    set the current sensor trip temperature and hysteresis. Presence of
    this function indicates the sensor driver has interrupt support.

  - The thermal OF device will call this function every time the
    temperature slides across trip points. (Or when trip points are
    altered.)

  - The thermal OF device should ignore the polling delay (set it to 0)
    if its sensor has interrupt support. (In this case, we can also make
    polling-delay optional. It's currently required in the device tree.)

  - On interrupt, the sensor driver should call
    thermal_zone_device_update with its thermal_zone_device, as returned
    by thermal_zone_of_sensor_register.

My only concern is that I don't understand kernel contexts and interrupt
handling well enough. It looks like at least its up to the sensor driver
to ensure calls into the thermal subsystem have long left the hardware
interrupt context, which I think should be sufficient.

Hoping all of this sounds about right!