> Am 07.10.2019 um 17:11 schrieb Adam Ford <aford173@xxxxxxxxx>:
>
> On Sat, Sep 14, 2019 at 11:12 AM Adam Ford <aford173@xxxxxxxxx> wrote:
>>
>> On Sat, Sep 14, 2019 at 9:38 AM H. Nikolaus Schaller <hns@xxxxxxxxxxxxx> wrote:
>>>
>>>
>>>> Am 14.09.2019 um 15:42 schrieb Adam Ford <aford173@xxxxxxxxx>:
>>>>
>>>> On Sat, Sep 14, 2019 at 4:20 AM H. Nikolaus Schaller <hns@xxxxxxxxxxxxx> wrote:
>>>>>
>>>>>
>>>>>> Am 13.09.2019 um 17:37 schrieb Adam Ford <aford173@xxxxxxxxx>:
>>>>>>
>>>>>> The OMAP3530, AM3517 and DM3730 all show thresholds of 90C and 105C
>>>>>> depending on commercial or industrial temperature ratings. This
>>>>>> patch expands the thermal information to the limits of 90 and 105
>>>>>> for alert and critical.
>>>>>>
>
> Tom / anyone from TI,
>
> I am going to rebase this patch from the current 5.4-RC branch, remove
> the AM3517 references, and leave the throttling only applicable to
> omap34xx and 36xx (like it is now), and remove the RFC. Before I do
> that, I was hoping for some feedback on whether or not there is a
> reason to not do this while acknowledging the thermal sensor isn't
> very accurate.
I wonder if there is a more precise definition what "isn't very accurate"
means?
Is it just because the TI_BANDGAP_FEATURE_UNRELIABLE bit is set in
the driver and we assume that it is right?
Of course the "junction temperature" (TJ) is not well defined (at which
edge? in which area?) and the bandgap sensor can only report a single point
of the die. So e.g. the GPU or the NEON unit may be hotter or cooler.
And, the bandgap sensor + ADC is unlikely to be well calibrated to
0.1°C precision.
But in my experiments there seems to be not much noise and values rise
or fall monotonic according to expectations of processor load.
So a report of 90°C may not be exactly 90°C and some parts of the SoC
may be hotter.
I would also assume that the TJ limits of 90°C have some safety margin
but there seems to be no information in the data sheet.
So, IMHO an "unreliable" bandgap sensor is better than no sensor and
no trips / cooling maps.
One more thing is with the omap3 bandgap sensor (driver?). It appears to
report the value of the previous measurement. So unless it is regularily
polled (like cpufreq seems to do) it will report outdated values. The
first read hours after boot may report the value during probe while booting.
This is also a source of missing accuracy of course. But I haven't
investigated this (can only be tested if thermal management is turned
off) because I think it has no practical influence if cpufreq is polling.
>
> Does anyone have any objections to this?
>
> Other than the omap mailing list, are there other lists that should be CC'd?
>
> adam
>
>>>>>> For boards who never use industrial temperatures, these can be
>>>>>> changed on their respective device trees with something like:
>>>>>>
>>>>>> &cpu_alert0 {
>>>>>> temperature = <85000>; /* millicelsius */
>>>>>> };
>>>>>>
>>>>>> &cpu_crit {
>>>>>> temperature = <90000>; /* millicelsius */
>>>>>> };
>>>>>>
>>>>>> Signed-off-by: Adam Ford <aford173@xxxxxxxxx>
>>>>>> ---
>>>>>> V2: Change the CPU reference to &cpu instead of &cpu0
>>>>>>
>>>>>> diff --git a/arch/arm/boot/dts/omap3-cpu-thermal.dtsi b/arch/arm/boot/dts/omap3-cpu-thermal.dtsi
>>>>>> index 235ecfd61e2d..dfbd0cb0b00b 100644
>>>>>> --- a/arch/arm/boot/dts/omap3-cpu-thermal.dtsi
>>>>>> +++ b/arch/arm/boot/dts/omap3-cpu-thermal.dtsi
>>>>>> @@ -17,4 +17,25 @@ cpu_thermal: cpu_thermal {
>>>>>>
>>>>>> /* sensor ID */
>>>>>> thermal-sensors = <&bandgap 0>;
>>>>>> +
>>>>>> + cpu_trips: trips {
>>>>>> + cpu_alert0: cpu_alert {
>>>>>> + temperature = <90000>; /* millicelsius */
>>>>>> + hysteresis = <2000>; /* millicelsius */
>>>>>> + type = "passive";
>>>>>> + };
>>>>>> + cpu_crit: cpu_crit {
>>>>>> + temperature = <105000>; /* millicelsius */
>>>>>> + hysteresis = <2000>; /* millicelsius */
>>>>>> + type = "critical";
>>>>>> + };
>>>>>> + };
>>>>>> +
>>>>>> + cpu_cooling_maps: cooling-maps {
>>>>>> + map0 {
>>>>>> + trip = <&cpu_alert0>;
>>>>>> + cooling-device =
>>>>>> + <&cpu THERMAL_NO_LIMIT THERMAL_NO_LIMIT>;
>>>>>> + };
>>>>>> + };
>>>>>> };
>>>>>> --
>>>>>> 2.17.1
>>>>>>
>>>>>
>>>>> Here is my test log (GTA04A5 with DM3730CBP100).
>>>>> "high-load" script is driving the NEON to full power
>>>>> and would report calculation errors.
>>>>>
>>>>> There is no noise visible in the bandgap sensor data
>>>>> induced by power supply fluctuations (log shows system
>>>>> voltage while charging).
>>>>>
>>>>
>>>> Great data!
>>>>
>>>>> root@letux:~# ./high-load -n2
>>>>> 100% load stress test for 1 cores running ./neon_loop2
>>>>> Sat Sep 14 09:05:50 UTC 2019 65° 4111mV 1000MHz
>>>>> Sat Sep 14 09:05:50 UTC 2019 67° 4005mV 1000MHz
>>>>> Sat Sep 14 09:05:52 UTC 2019 68° 4000mV 1000MHz
>>>>> Sat Sep 14 09:05:53 UTC 2019 68° 4000mV 1000MHz
>>>>> Sat Sep 14 09:05:55 UTC 2019 72° 3976mV 1000MHz
>>>>> Sat Sep 14 09:05:56 UTC 2019 72° 4023mV 1000MHz
>>>>> Sat Sep 14 09:05:57 UTC 2019 72° 3900mV 1000MHz
>>>>> Sat Sep 14 09:05:59 UTC 2019 73° 4029mV 1000MHz
>>>>> Sat Sep 14 09:06:00 UTC 2019 73° 3988mV 1000MHz
>>>>> Sat Sep 14 09:06:01 UTC 2019 73° 4005mV 1000MHz
>>>>> Sat Sep 14 09:06:03 UTC 2019 73° 4011mV 1000MHz
>>>>> Sat Sep 14 09:06:04 UTC 2019 73° 4117mV 1000MHz
>>>>> Sat Sep 14 09:06:06 UTC 2019 73° 4005mV 1000MHz
>>>>> Sat Sep 14 09:06:07 UTC 2019 75° 3994mV 1000MHz
>>>>> Sat Sep 14 09:06:08 UTC 2019 75° 3970mV 1000MHz
>>>>> Sat Sep 14 09:06:09 UTC 2019 75° 4046mV 1000MHz
>>>>> Sat Sep 14 09:06:11 UTC 2019 75° 4005mV 1000MHz
>>>>> Sat Sep 14 09:06:12 UTC 2019 75° 4023mV 1000MHz
>>>>> Sat Sep 14 09:06:14 UTC 2019 75° 3970mV 1000MHz
>>>>> Sat Sep 14 09:06:15 UTC 2019 75° 4011mV 1000MHz
>>>>> Sat Sep 14 09:06:16 UTC 2019 77° 4017mV 1000MHz
>>>>> Sat Sep 14 09:06:18 UTC 2019 77° 3994mV 1000MHz
>>>>> Sat Sep 14 09:06:19 UTC 2019 77° 3994mV 1000MHz
>>>>> Sat Sep 14 09:06:20 UTC 2019 77° 3988mV 1000MHz
>>>>> Sat Sep 14 09:06:22 UTC 2019 77° 4023mV 1000MHz
>>>>> Sat Sep 14 09:06:23 UTC 2019 77° 4023mV 1000MHz
>>>>> Sat Sep 14 09:06:24 UTC 2019 78° 4005mV 1000MHz
>>>>> Sat Sep 14 09:06:26 UTC 2019 78° 4105mV 1000MHz
>>>>> Sat Sep 14 09:06:27 UTC 2019 78° 4011mV 1000MHz
>>>>> Sat Sep 14 09:06:28 UTC 2019 78° 3994mV 1000MHz
>>>>> Sat Sep 14 09:06:30 UTC 2019 78° 4123mV 1000MHz
>>>>> ...
>>>>> Sat Sep 14 09:09:57 UTC 2019 88° 4082mV 1000MHz
>>>>> Sat Sep 14 09:09:59 UTC 2019 88° 4164mV 1000MHz
>>>>> Sat Sep 14 09:10:00 UTC 2019 88° 4058mV 1000MHz
>>>>> Sat Sep 14 09:10:01 UTC 2019 88° 4058mV 1000MHz
>>>>> Sat Sep 14 09:10:03 UTC 2019 88° 4082mV 1000MHz
>>>>> Sat Sep 14 09:10:04 UTC 2019 88° 4058mV 1000MHz
>>>>> Sat Sep 14 09:10:06 UTC 2019 88° 4146mV 1000MHz
>>>>> Sat Sep 14 09:10:07 UTC 2019 88° 4041mV 1000MHz
>>>>> Sat Sep 14 09:10:08 UTC 2019 88° 4035mV 1000MHz
>>>>> Sat Sep 14 09:10:10 UTC 2019 88° 4052mV 1000MHz
>>>>> Sat Sep 14 09:10:11 UTC 2019 88° 4087mV 1000MHz
>>>>> Sat Sep 14 09:10:12 UTC 2019 88° 4152mV 1000MHz
>>>>> Sat Sep 14 09:10:14 UTC 2019 88° 4070mV 1000MHz
>>>>> Sat Sep 14 09:10:15 UTC 2019 88° 4064mV 1000MHz
>>>>> Sat Sep 14 09:10:17 UTC 2019 88° 4170mV 1000MHz
>>>>> Sat Sep 14 09:10:18 UTC 2019 88° 4058mV 1000MHz
>>>>> Sat Sep 14 09:10:19 UTC 2019 88° 4187mV 1000MHz
>>>>> Sat Sep 14 09:10:21 UTC 2019 88° 4093mV 1000MHz
>>>>> Sat Sep 14 09:10:22 UTC 2019 88° 4087mV 1000MHz
>>>>> Sat Sep 14 09:10:23 UTC 2019 90° 4070mV 1000MHz
>>>>
>>>> Should we be a little more conservative? Without knowing the
>>>> accuracy, i believe we do not want to run at 800 or 1GHz at 90C, so if
>>>> we made this value 89 instead of 90, we would throttle a little more
>>>> conservatively.
>>>
>>> Well, the OMAP5 also defines exactly 100°C in the device tree.
>>>
>>> I would assume that the badgap sensor accuracy is so that it
>>> never reports less than the real temperature. So if we
>>> throttle at reported 90° TJ is likely lower.
>>>
>>>>> Sat Sep 14 09:10:25 UTC 2019 88° 4123mV 800MHz
>>>>> Sat Sep 14 09:10:26 UTC 2019 88° 4064mV 1000MHz
>>>>> Sat Sep 14 09:10:28 UTC 2019 90° 4058mV 1000MHz
>>>>
>>>> Again here, I interpret the data sheet correctly, we're technically out of spec
>>>
>>> I read the data sheet as if 90°C at OPP1G is still within spec.
>>> 91 would be obviously outside (if bandgap sensor is precise).
>>>
>>>>
>>>>> Sat Sep 14 09:10:29 UTC 2019 88° 4076mV 1000MHz
>>>>> Sat Sep 14 09:10:30 UTC 2019 88° 4064mV 1000MHz
>>>>> Sat Sep 14 09:10:32 UTC 2019 88° 4117mV 1000MHz
>>>>> Sat Sep 14 09:10:33 UTC 2019 88° 4105mV 800MHz
>>>>> Sat Sep 14 09:10:34 UTC 2019 88° 4070mV 1000MHz
>>>>> Sat Sep 14 09:10:36 UTC 2019 88° 4076mV 1000MHz
>>>>> Sat Sep 14 09:10:37 UTC 2019 88° 4087mV 1000MHz
>>>>> Sat Sep 14 09:10:39 UTC 2019 88° 4017mV 1000MHz
>>>>> Sat Sep 14 09:10:40 UTC 2019 88° 4093mV 1000MHz
>>>>> Sat Sep 14 09:10:41 UTC 2019 88° 4058mV 800MHz
>>>>> Sat Sep 14 09:10:42 UTC 2019 88° 4035mV 1000MHz
>>>>> Sat Sep 14 09:10:44 UTC 2019 90° 4058mV 1000MHz
>>>>> Sat Sep 14 09:10:45 UTC 2019 88° 4064mV 1000MHz
>>>>> Sat Sep 14 09:10:47 UTC 2019 88° 4064mV 1000MHz
>>>>> Sat Sep 14 09:10:48 UTC 2019 88° 4029mV 1000MHz
>>>>> Sat Sep 14 09:10:50 UTC 2019 90° 4046mV 1000MHz
>>>>> ^Ckill 4680
>>>>> root@letux:~# cpufreq-info
>>>>> cpufrequtils 008: cpufreq-info (C) Dominik Brodowski 2004-2009
>>>>> Report errors and bugs to cpufreq@xxxxxxxxxxxxxxx, please.
>>>>> analyzing CPU 0:
>>>>> driver: cpufreq-dt
>>>>> CPUs which run at the same hardware frequency: 0
>>>>> CPUs which need to have their frequency coordinated by software: 0
>>>>> maximum transition latency: 300 us.
>>>>> hardware limits: 300 MHz - 1000 MHz
>>>>> available frequency steps: 300 MHz, 600 MHz, 800 MHz, 1000 MHz
>>>>> available cpufreq governors: conservative, userspace, powersave, ondemand, performance
>>>>> current policy: frequency should be within 300 MHz and 1000 MHz.
>>>>> The governor "ondemand" may decide which speed to use
>>>>> within this range.
>>>>> current CPU frequency is 600 MHz (asserted by call to hardware).
>>>>> cpufreq stats: 300 MHz:22.81%, 600 MHz:2.50%, 800 MHz:2.10%, 1000 MHz:72.59% (1563)
>>>>> root@letux:~#
>>>>>
>>>>> So OPP is reduced if bandgap sensor reports >= 90°C
>>>>> which almost immediately makes the temperature
>>>>> go down.
>>>>>
>>>>> No operational hickups were observed.
>>>>>
>>>>> Surface temperature of the PoP chip did rise to
>>>>> approx. 53°C during this test.
>>>>>
>>>>> Tested-by: H. Nikolaus Schaller <hns@xxxxxxxxxxxxx> # on GTA04A5 with dm3730cbp100
>>>>>
>>>
>>> BTW: this patch (set) is even independent of my 1GHz OPP patches.
>>> Should also work with OPP-v1 definitions so that maintainers can
>>> decide which one to apply first.
>>
>> If I am going integrate the cooling references into &cpu node, I'll
>> probably base it on your work since the cooling isn't really that
>> important until we exceed 800MHz. If I do it on the current linux
>> master or omap for-next branch, it may not apply cleanly.
>>
>>>
>>> It is just more difficult to reach TJ of 90°C without 1GHz.
>>
>> If it even does at all without external influences.
>>
>> adam
>>>
>>> BR,
>>> Nikolaus
>>>
