Re: Using a temperature sensor with 1-bit output for CPU throttling

[Javi Merino <javi.merino@arm.com>]

On Wed, Apr 29, 2015 at 02:47:13PM +0100, Mason wrote:
> On 28/04/2015 13:27, Mason wrote:
> 
> > The SoC I'm working on provides a temperature sensor (NXP) in the CPU block.
> > The sensor seems to be very primitive, so I wanted to ask experienced people
> > what would be the best way to use it from Linux.
> > 
> > General Description
> > "The sensor generates an output signal that indicates if the die temperature
> > exceeds a programmable threshold. This makes it particularly suitable for
> > detecting overheating."
> > 
> > So it seems that the original purpose of this sensor was to periodically
> > check that the temperature has not exceeded a given threshold.
> > 
> > - Is the CPU temp higher than 100°C ?
> > - No.
> > - OK. Business as usual.
> > 
> > (1 second later)
> > - Is the CPU temp higher than 100°C ?
> > - Yes.
> > - Uh-oh! I need to do something about it.
> > 
> > 
> > Basic Functions
> > "The temp sensor uses a bandgap type of circuit to compare a voltage which
> > has a negative temperature coefficient with a voltage that is proportional
> > to absolute temperature. A resistor bank allows 40 different temperature
> > thresholds to be selected and the logic output 'out_temperature' will then
> > indicate whether the actual die temperature lies above or below the selected
> > threshold."
> > 
> > The available thresholds seem to be chosen somewhat arbitrarily:
> > 
> >   -45.1, -39.7, -33.7, -29.4, -24.4, -20.4, -15.4, -10.1,
> >   -6.4, -1.4, 3.6, 7.6, 12.9, 16.6, 20.6, 25.6, 30.9,
> >   34.9, 38.6, 43.9, 48.9, 52.9, 57.9, 61.9, 66.9, 70.9,
> >   76.3, 81.3, 85.3, 90.3, 95.3, 98.9, 102.9, 108.3, 111.9,
> >   117.3, 122.3, 126.3, 131.3, 135.3, 139.3
> > 
> > The spacing between values seems arbitrary also.
> > (Is there an underlying physical explanation?)
> > 
> > I'm not sure that there is much point in testing for temperatures lower
> > than 50°C ? (I'm told that the SoC can reliably function up to 125°C.)
> > 
> > Do higher temperatures shorten the lifespan of a component?
> > In other words, would a CPU running 24/7 at 100°C "break" sooner
> > than one running 24/7 at 50°C ?
> > 
> > 
> > Characteristics
> > 
> > Symbol      Parameter             Min  Typ  Max  Unit
> > 
> > (Operating conditions)
> > Tjunc      Junction temperature   -40   25   125  °C
> > Vdd        Supply voltage         1.0  1.1  1.26   V
> > 
> > (Normal operating mode)
> > Idd         Supply current              50    60  μA
> > Vbandgapref Ref output voltage   0.72  0.8  0.88   V
> > ∆outtemp    Absolute Temp               ±2   ±10  °C
> >             threshold error
> > T_res       Temp resolution        3    4.5    7  °C
> > 
> > 
> > Given the semantics of the temperature sensor hardware block, I was
> > tempted to implement something along these lines:
> > 
> > Create a kernel thread that runs periodically (e.g. every second)
> > to check if the temperature is above 100°C.
> > - If not, do nothing
> > - If yes, somehow prevent the CPU from using the highest frequencies
> > defined in cpufreq's freq table
> > (They are 1000, 500, 333, 200, 100 MHz)
> > 
> > Is that a sensible approach?
> > Is there a way to implement this using the thermal framework?
> > 
> > Or am I looking at this wrong, and things should be done a
> > different way? (I'm using 3.14 by the way.)
> > 
> > I suppose I could perform some kind of binary search to zoom in
> > on the current threshold (although it might change during the
> > measurements, so I'd rather not go there.)
> 
> I'm aware that I posted many questions. I'd be grateful if someone
> would answer even a tiny subset. That would get the ball rolling.
> 
> If I understand correctly, if I want to use the CPU throttling
> framework, I need to define a "thermal zone device" and a
> "cooling device". AFAIU, the cooling device is taken care of
> by cpu_cooling.c
> 
>   cpufreq_cooling_register(cpu_present_mask);

Correct

> My temperature sensor would be the thermal zone device?
> How do I tie the two devices together?

Your temperature sensor would be the input to the thermal zone
device.  You register it with thermal_zone_device_register().  See
Documentation/thermal/sysfs-api.txt .  Your thermal sensor
doesn't actually report temperature but the thermal framework expects
a temperature, so your thermal zone's get_temp() function should
report a fake temperature.  For example, if you've configure your
sensor for 50C, then you could report 45C if the sensor reads as 0 and
50C if the sensor reads as 1.  It's a hack, but it should work.  Bear
in mind that get_temp() should report in millicelsius.

> Is that where a thermal governor comes in play?

Because you want on/off behavior, the bangbang governor is the
simplest to use and should do the work.  You can choose as the default
in your kernel configuration or you can choose it by passing it as
part of the tzp that you pass to thermal_zone_device_register()

Put a trip point at the temperature you've set up your sensor to and
bind the cpu cooling device to it.

Hope this helps,
Javi