On Tue, Nov 05, 2019 at 10:58:39AM +0000, John Garry wrote:
> On 04/11/2019 19:24, Mark Brown wrote:
> > On Tue, Nov 05, 2019 at 12:51:36AM +0800, John Garry wrote:

> > > +		if (len) {
> > > +			u32 val;
> > > +
> > > +			val = __raw_readl(host->regbase + CMD_DATABUF(words));
> > > +
> > > +			to += words * 4;
> > > +			for (i = 0; i < len; i++, val >>= 8, to++)
> > > +				*to = (u8)val;
> > > +		}
> > > +	} else {
> > > +		for (i = 0; i < DIV_ROUND_UP(len, 4); i++) {
> > > +			u32 val = __raw_readl(host->regbase + CMD_DATABUF(i));
> > > +			int j;

> > The more usual pattern for these would be to do some unaligned accesses
> > for the start/end of the buffer to get to alignment and then transfer
> > the rest as aligned data.

> Yeah, I understand you, but for that I would need to generate multiple
> transactions in the driver, and I wanted to keep 1x transaction per
> spi_controller_mem_ops.exec_op call.

> So maybe I can do some trickery in my adjust_op_size method to generate
> these multiple transactions: a. any unaligned start data b. the 32b-aligned
> data b. unaligned end. I think that the HW should be able to handle that.

Right, that's what I was expecting.