[RFC PATCH] watchdog: da9062: Correct the timeout values [Klartext]

[Christoph Niedermaier]

From: linux-arm-kernel [mailto:linux-arm-kernel-bounces at lists.infradead.org]
Sent: Monday, December 13, 2021 11:45 PM
>>>>>> Thanks anyway, so now I know it must be
>>>>>> problem with my DA9061 chip.
>>>>>>
>>>>>> @Adam
>>>>>> Where can it come from?
>>>>>> Can you give we a hint what to check?
>>>>>
>>>>> I've spoken internally and have been informed that this is down to the fact that
>>>>> DA9061 runs only from an internal oscillator which may be slower. The indication
>>>>> is that the values for TWDSCALE describe the window where if a kick/ping occurs
>>>>> within that period then the watchdog is guaranteed *not* to timeout. The actual
>>>>> timeout would be at some point after the selected timeout period, assuming no
>>>>> ping/kick occurred.
>>>>>
>>>>> Table 8 in the datasheet specifies a minimum watchdog timeout of 2.5s (tWDMAX)
>>>>> under specific operating conditions, so if the minimum 2s window was chosen
>>>>> (TWDSCALE = 1) then earliest the watchdog would actually timeout, following a
>>>>> ping, is 2.5s, assuming the conditions matched those described.
>>>>>
>>>>> If you have further questions it probably makes sense to contact Dialog/Renesas
>>>>> support as they will be able to provide more detailed info on this.
>>>>
>>>> So a DA9061 runs only from an internal oscillator, whereas a DA9062
>>>> can run on either an internal or an external oscillator. So this
>>>> means that the DA9061 timeout values are differ from the DA9062
>>>> with an external oscillator not only on my device but on all DA9061
>>>> devices.
>>>>
>>>> This are the values (in seconds) in comparison:
>>>> DA9062 (from driver): 0  2  4   8  16  32  65 131
>>>> DA9061 (measured):    0  3  6  12  25  51 102 204
>>>> =================================================
>>>> Difference:           0 +1 +2  +4  +9 +19 +37 +73
>>>>
>>>> In my opinion, the differences in the higher values are very huge.
>>>> If I expect that the watchdog triggers and I have to wait more than
>>>> a minute for that to happen I ask myself is there something wrong.
>>>>
>>>> @Andrej
>>>> I guess, you are using an external oscillator, aren't you?
>>>>
>>>> @Adam
>>>> Is there a way to check in the driver which oscillator is in use?
>>>>
>>>> @Maintainers
>>>> Is in the driver a need to distinguish between an external and an
>>>> internal oscillator to get the timeout values more accurate?
>>>>
>>>
>>> It would be very desirable to get timeout values more accurate.
>>> I would not want to dictate how to implement it, though.
>>> It could be automatically detected if that is possible, there
>>> could be a devicetree clock property providing the clock
>>> frequency, or maybe there is some other solution.
>>>
>>> Guenter
>>
>> I am open for a good solution.
>> Meanwhile I measured the timeout values of my 8 available DA9061
>> watchdogs. I derived the following formula from the given formula
>> at the data sheet and the clock divider of 2^16:
>>
>> f = 2^(15+TWDSCALE) / t
>>
>> Formula check with the external oscillator (32kHz) TWDSCALE=7 @ 131s:
>> f = 2^(15+7) / 131 = 32017Hz (=> should be OK)
>>
>> The timeouts of my 8 watchdogs (9061-AA) with TWDSCALE=7:
>> t7 = 211s => 19878Hz
>> t7 = 197s => 21291Hz
>> t7 = 203s => 20662Hz
>> t7 = 204s => 20560Hz
>> t7 = 206s => 20361Hz
>> t7 = 198s => 21662Hz
>> t7 = 200s => 20972Hz
>>
>> According to the data sheet the internal oscillator should run at 25kHz.
>> The average frequency of my 8 devices is 20.6kHz. Maybe the data sheet
>> Clock value is a max value. The timeout difference on my 8 devices are
>> 14s. So the values vary from device to device, and maybe there is also a
>> temperature component.
>>
>> @Adam
>> Is there a way to check which oscillator is in use?
>> Is there a way to find the current oscillator frequency?
>> Are there any other ideas/solutions to get the timeout values more accurate?
> 
> As I sent my E-Mail I didn't see your answer, so here it is:
> 
>> A quick scan of the DA9062 datasheet shows that reg/field EN_32K/CRYSTAL will
>> indicate the presence of a 32KHz crystal oscillator. Obviously on DA9061 that
>> option isn't available.
>>
>> I think the problem seems to lie around determining the internal oscillator's
>> frequency. Datasheet references 25Mhz in Table 9 (Watchdog Electrical
>> Characteristics), but that doesn't seem to tally with your timings and I don't
>> see an obvious way in the regmap to calculate this at run time. *If* the
>> oscillator frequency varies from part to part, or under different environmental
>> conditions, then it's going to be tough to tie this down, and you wouldn't want
>> to state a timeout value that's longer than reality.
> 
> @Adam
> What do you think if I add also the timeout values for the internal
> oscillator?
> Than the values would be:
> 
> t = 2^(15+TWDSCALE) / f
> 
> Internal 25kHz oscillator: 0  2  5  10  20  41  83 167
> 
> In comparison to the
> external 32kHz oscillator: 0  2  4   8  16  32  65 131
> 
> So the difference to my measured values will decrease and it will improve
> the
> current driver towards more accurate values.
> 
> Btw if I calculate the values with 32768Hz (the typical 32kHz oscillator), I
> get the
> following values:
> 32768Hz oscillator:        0  2  4   8  16  32  64 128

I come to the conclusion that under normal condition (room temperature) the
timeout values could be improved, but due to the used internal ring oscillator
on the DA9061 chip under extreme conditions like automotive (125 degrees) the
values come close to the DA9062 timeout values, if an external oscillator
is used. In order not to trigger too early in extreme situations, the values
must be left as they are. But I told Dialog to add a note in the data sheet
explaining the deviation when the internal oscillator is used (DA9061).

Thanks for helping me and best regards
Christoph