[linux-sunxi] [PATCH] spi: dts: sun4i: Add support for inter-word wait cycles using the SPI Wait Clock Register

[Marcus Weseloh]

2015-11-22 20:45 GMT+01:00 Maxime Ripard <maxime.ripard at free-electrons.com>:
>> Julien, Rob: thanks for your comments! Ok, I will make the following changes:
>>
>> - remove "sun4i,spi-wdelay" from the sun4i binding and add the
>> property to the spi-bus.txt binding instead
>> - remove the comment about the additional 3 cycles from the documentation
>> - modfy the spi-sun4i driver to take care of the minimum 3 cycle period
>>
>> Does that sound right?
>>
>> And maybe I could also use a more descriptive name for the property,
>> maybe "spi-word-wait-cycles"?
>
> I don't think it should be in a clock-rate dependant unit. Using micro
> or nano-seconds would be more appropriate I guess.

Thanks Maxime, using a time based value instead of cycles sounds like
a much better approach.

However... I'm starting to think that the proposed inter-word wait
time DT property is just an ugly workaround. Let me explain my
use-case:

I'm developing a driver for a sensor that requires a minimum wait time
between words. The wait time depends on the mode the sensor is set to:
37.5us in slow mode, 12.5us in fast mode. I initially used spidev to
test the sensor from userspace. And for that use case, the
"spi-wdelay" property that I proposed works well. But now I am writing
the proper protocol driver and suddenly the explicit wait time setting
seems just wrong. Ideally, the protocol driver would just expose a DT
property that allows to choose between "slow" and "fast" mode.

I think that the correct approach would be to extend the SPI
controller API to allow protocol drivers to set an inter-word delay.
That would keep the magic numbers inside my protocol driver and out of
the devicetree. And an additional ioctl call could set that inter-word
delay from spidev, allowing userspace to set this value as well if
needed.

Mark: would you be open to such a change to the SPI controller API?

I could use the already available spi_transfer.delay_usecs for this,
but I would require that I wrap each word in a single transfer, which
adds significant processing overhead to the communication with the
sensor.

Cheers,

    Marcus