Re: [PATCH v2 0/2] iio: humidity: si7020: Check device property for skipping reset in probe

[Eddie James <eajames@linux.ibm.com>]

On 5/14/22 10:02, Peter Rosin wrote:
> 2022-05-14 at 15:43, Jonathan Cameron wrote:
>> On Sat, 14 May 2022 00:48:51 +0200
>> Peter Rosin <peda@axentia.se> wrote:
>>
>>> Hi!
>>>
>>> 2022-05-13 at 18:45, Jonathan Cameron wrote:
>>>> On Thu, 12 May 2022 12:08:07 -0500
>>>> Eddie James <eajames@linux.ibm.com> wrote:
>>>>    
>>>>> On 5/12/22 11:48, Jonathan Cameron wrote:
>>>>>> On Thu, 12 May 2022 11:20:18 -0500
>>>>>> Eddie James <eajames@linux.ibm.com> wrote:
>>>>>>     
>>>>>>> I2C commands issued after the SI7020 is starting up or after reset
>>>>>>> can potentially upset the startup sequence. Therefore, the host
>>>>>>> needs to wait for the startup sequence to finish before issuing
>>>>>>> further i2c commands. This is impractical in cases where the SI7020
>>>>>>> is on a shared bus or behind a mux, which may switch channels at
>>>>>>> any time (generating I2C traffic). Therefore, check for a device
>>>>>>> property that indicates that the driver should skip resetting the
>>>>>>> device when probing.
>>>>>> Why not lock the bus?  It's not ideal, but then not resetting and hence
>>>>>> potentially ending up in an unknown state isn't great either.
>>>>>
>>>>> Agreed, but locking the bus doesn't work in the case where the chip is
>>>>> behind a mux. The mux core driver deselects the mux immediately after
>>>>> the transfer to reset the si7020, causing some i2c traffic, breaking the
>>>>> si7020. So it would also be a requirement to configure the mux to idle
>>>>> as-is... That's why I went with the optional skipping of the reset.
>>>>> Maybe I should add the bus lock too?
>>>>>   
>>>> +Cc Peter and linux-i2c for advice as we should resolve any potential
>>>> issue with the mux side of things rather than hiding it in the driver
>>>> (if possible!)
>>> IIUC, the chip in question cannot handle *any* action on the I2C bus
>>> for 15ms (or so) after a "soft reset", or something bad<tm> happens
>>> (or at least may happen).
>>>
>>> If that's the case, then providing a means of skipping the reset is
>>> insufficient. If you don't lock the bus, you would need to *always*
>>> skip the reset, because you don't know for certain if something else
>>> does I2C xfers.
>>>
>>> So, in order to make the soft reset not be totally dangerous even in
>>> a normal non-muxed environment, the bus must be locked for the 15ms.
>>>
>>> However, Eddie is correct in that the I2C mux code may indeed do its
>>> muxing xfer right after the soft reset command. There is currently
>>> no way to avoid that muxing xfer. However, it should be noted that
>>> there are ways to mux an I2C bus without using xfers on the bus
>>> itself, so it's not problematic for *all* mux variants.
>>>
>>> It can be debated if the problem should be worked around with extra
>>> dt properties like this, or if a capability should be added to delay
>>> a trailing muxing xfer.
>>>
>>> I bet there are other broken chips that have drivers that do in fact
>>> lock the bus to give the chip a break, but then it all stumbles
>>> because of the unexpected noise if there's a (wrong kind of) mux in
>>> the mix.
>> Ok, so for now I think we need the bus lock for the reset + either
>> a work around similar to this series, or additions to the i2c mux code
>> to stop it doing a muxing xfer if the bus is locked?
> I think there might be cases where it might be valid to restore the mux
> directly after an xfer even if the mux is externally locked prior to the
> muxed xfer. But I'm not sure? In any case, it will be a bit convoluted
> for the mux code to remember that it might need to restore the mux
> later. And it will get even hairier when multiple levels of muxing is
> considered...
>
> Maybe some kind of hook/callback that could be installed temporarily on
> the I2C adapter that is called right after the "real" xfer, where the
> driver could then make the needed mdelay call?
>
> I.e.
> 1. lock the bus
> 2. install this new hook/callback
> 3. do an unlocked xfer, get notified and call mdelay
> 5. uninstall the hook/callback
> 6. unlock the bus
>
> The hook/callback could be uninstalled automatically on unlock, then
> you would not need to handle multiple notifications. But then again,
> there is probably some existing framework that should be used that
> handles all than neatly and efficiently.


Hm, interesting. Sounds a bit complicated, though very flexible. For a 
less flexible, but less complex, approch, we could add a i2c_msg flag 
that says to do a delay in the core? And then si7020 could just submit a 
couple of raw messages rather than smbus... What do you think?


Thanks,

Eddie



>
> Me waves hand a bit...
>
> Cheers,
> Peter