Re: [PATCH 2/3] spi: meson: Add support for Amlogic Meson SPIFC

[Beniamino Galvani <b.galvani@gmail.com>]

On Mon, Nov 10, 2014 at 03:11:40PM +0000, Mark Brown wrote:
> On Sun, Nov 09, 2014 at 11:56:50PM +0100, Beniamino Galvani wrote:
> > On Sun, Nov 09, 2014 at 10:17:12AM +0000, Mark Brown wrote:
>
> > > I noticed that the handling of /CS was done in the spifc_txrx() function
> > > - will this do the right thing if the transfer needs to be split for the
> > > buffer size?
> 
> > It should. When the transfer gets split, CS is kept active for all the
> > chunks and the value of CS after that depends on the value of
> > cs_change.
> 
> Can you be more specific about how that works?  I'm just not seeing the
> code that handles this.

It's this:

static int meson_spifc_txrx(struct meson_spifc *spifc,
                            struct spi_transfer *xfer,
                            int offset, int len, bool last_xfer,
                            bool last_chunk)
{
	bool keep_cs = true;

	[...]

        if (last_chunk) {
                if (last_xfer)
                        keep_cs = xfer->cs_change;
                else
                        keep_cs = !xfer->cs_change;
        }

        regmap_update_bits(spifc->regmap, REG_USER4, USER4_CS_ACT,
                           keep_cs ? USER4_CS_ACT : 0);

	/* start transfer */
	[...]
}

The USER4_CS_ACT bit specifies if CS must be kept active after the
transfer.

> > > > +	if (!ret && xfer->delay_usecs)
> > > > +		udelay(xfer->delay_usecs);
> 
> > > The core will do this for you if you implement this as transfer_one().
> 
> > Please correct me if I'm wrong, but I think that transfer_one() can't
> > be used in this case. The hardware doesn't support direct manipulation
> > of CS and allows only to specify if CS must be kept active after the
> > current transfer. So I need to know for each transfer if it's the last
> > and this can be achieved only implementing transfer_one_message().
> 
> This is already in a function that's operating at the transfer_one()
> level, the function is even called transfer_one() and besides it's
> clearly not something specific to this hardware so should be factored
> out into the core instead of open coded.

A way to simplify this at core level could be to add a 'last' flag to
the spi_transfer structure and populate it before calling
transfer_one_message(); in this way, drivers that need to know the
position of the transfer in the message in order to properly handle CS
can use the generic version of transfer_one_message() instead of
reimplementing it. It seems that other existing drivers probably can
benefit from this.

Beniamino