Re: [PATCH] spi: Add spi-bits-per-word binding.

[Adrian Fiergolski]

Hi Geert,

On 13.03.2017 at 20:57, Geert Uytterhoeven wrote:
> On Mon, Mar 13, 2017 at 7:12 PM, Adrian Fiergolski
> <Adrian.Fiergolski@xxxxxxx> wrote:
>> On 13.03.2017 at 18:55, Mark Brown wrote:
>>>> In my case, xilinx_spi_probe function (of spi-xilinx controller) sets
>>>> bits_per_word_mask of spi_master struct only to 16 bits support. Later,
>>>> xilinx_spi_probe calls of_register_spi_devices, which calls
>>>> of_register_spi_devices. The last one allocates an empty spi_device
>>>> struct and configures different options of the spi_device according to a
>>>> device tree. bits_per_word are not covered here (why?), thus it is left
>>>> 0 (value after allocation), which, by convention, means 8 bits support.
>>>> At the end, the same function (of_register_spi_device) calls
>>>> spi_add_device which finally calls spi_setup. The last call, according
>>>> to convention, changes bits_per_word to 8 and calls
>>>> __spi_validate_bits_per_word which fails, as master doesn't support 8
>>>> bit transmission. This fails registration sequence of a device driver.
>>>> As you see, the device driver doesn't have possibility to modify
>>>> bits_per_word during the registration process, thus it can't provide
>>>> support for such limited controllers.
>>> I can't see any way in which it follows from the above that it's a good
>>> idea to try to override bits per word settings in the device tree, that
>>> just wastes user time and is an abstraction failure.  We need better
>>> handling of defaults done purely in the kernel.
>> If enforcing by device tree specific for a given device driver SPI_CPHA,
>> SPIC_CPOL, SPI_CS_HIGH, max_speed_hz, etc. if fine form the abstraction
>> point of view, why it doesn't apply to bits_per_word ?
> Because unlike polarity, phase, and speed, bits_per_word is a property
> of the communication protocol.
>
> E.g. you can talk to the same EEPROM using different polarities, phase, or
> speed, but bits_per_word is fixed.
In this case, currently, what is the proper way to handle SPI
controllers (spi-xilinx) without 8-bit transmission support ?