Re: i2c-designware: not possible to write to different i2c addresses in one transfer?

[Hans Verkuil <hverkuil+cisco@kernel.org>]

On 18/10/2025 20:17, Hans Verkuil wrote:
> On 15/10/2025 15:40, Andy Shevchenko wrote:
>> On Wed, Oct 15, 2025 at 01:51:20PM +0200, Wolfram Sang wrote:
>>> Hi Hans,
>>>
>>> lucky you, I happen to have a board with that controller on the table
>>> currently :)
>>>
>>>> This worked fine for the Raspberry Pi 4B using the broadcom i2c driver, but for
>>>> the Raspberry Pi 5 using the designware driver it fails with -EINVAL and these
>>>> kernel messages:
>>>>
>>>> [  272.284689] i2c_designware 1f00074000.i2c: i2c_dw_xfer_msg: invalid target address
>>>
>>> Confirmed.
>>>
>>>> Looking in i2c-designware-master.c it seems it cannot handle consecutive messages for
>>>> different addresses.
>>>
>>> I agree. I leave the final judgement to the experts (Andy, Mika), but I
>>> already anticipate that I need to extend the existing
>>> I2C_AQ_COMB_SAME_ADDR quirk into a more generic one. And set the quirk
>>> in this driver.
>>
>> May I ask a dumb question? Why do we need such an awkward transaction
>> to begin with?
>>
> 
> That's how the datasheet defines the transfer. And in turn it is based on the VESA
> E-DDC standard for reading display EDIDs. By doing it in a single transaction you
> ensure that the bytes are written to (or read from) the correct EEPROM segment. If
> it was split in two commands, then someone could change the segment inadvertently.
> I.e., you set the segment to 1, then someone writes to it as well and sets it to 2,
> then you write the EDID data to the EEPROM, which has now selected the wrong segment.
> 
> For now I have split up the transaction into two independent transfers in my driver,
> one to the segment address, one writing to the EEPROM. This works, but it is not really
> how it is supposed to be done. And I was actually surprised that it worked, since a
> write to the EEPROM without writing to the segment pointer first in the same transaction
> should write to segment 0, according to the cat24c208 datasheet.

Actually, my hack didn't work. It behaved exactly as the datasheet describes: if it is
done as two separate transactions, then when writing to the EEPROM will actually write
to segment 0, not 1.

My test was buggy, and after fixing it I could clearly see that the segment write and
EEPROM write must be part of the same transaction.

Regards,

	Hans

> 
> I couldn't tell from the designware driver code whether this was 1) impossible (HW limitation),
> 2) possible, but not implemented, or 3) a bug. The comment in the code before the
> "invalid target address" error was rather vague.
> 
> Regards,
> 
> 	Hans