Re: [PATCH 0/3] spi-nor: Add Octal SPI support

[Boris Brezillon]

On Thu, 4 Oct 2018 16:05:36 +0530
Vignesh R <vigneshr@xxxxxx> wrote:

> >>
> >>  .../devicetree/bindings/mtd/cadence-quadspi.txt       |  1 +
> >>  drivers/mtd/spi-nor/cadence-quadspi.c                 |  9 +++++++++  
> > 
> > On a slightly different topic, do you plan to convert the Cadence
> > driver to spi-mem? And if you don't, is it because you don't have time
> > or because some features are missing in spi-mem (I remember you
> > mentioned a few things back when you were reviewing the spi-mem series)?
> >   
> 
> I do not have plans to convert cadence QSPI driver to spi-mem yet,
> mainly due to lack of time. Also, not sure if original author Marek and
> other altera people are okay with that.
> 
> I see couple of issues in the way of conversion:
> 1. I would wait to know what direction would direct mapping APIs[1] go
> before starting spi-mem conversion for Cadence QSPI driver. Else, we
> have may to re write again if direct mapping APIs are merged.

I'd suggest reviewing the proposal I posted so that you can influence
the design of this new API ;-).

> 2. New Cadence OSPI IP has an integrated PHY to support high throughput
> OSPI flashes operating up 200MHz in Octal DDR mode. In order to work
> with such flashes, PHY DLLs need to be calibrated. Highly simplified
> calibration sequence is as below(See [2] for actual sequence):
> -Read flash ID at low speed and store it.
> -Enable PHY and set DLLs to a defined initial value
> -Increment RX DLL value
> -Read flash ID and check for correctness of data read
> -repeat above two steps until a band of passing values is obtained for
> RX DLL where flash ID is correctly read.
> -DLL needs to set to middle of the passing band.

Is the Read ID operation hardcoded or do you just use it as a way to
trigger predictable transfers on the IO bus?

> 
> I am trying to figure out how to fit this into the spi-mem framework as
> controller would to need to store READ ID opcode and expected JEDEC ID
> before starting calibration sequence.

I think this should be split in 2:

- the SPI NOR framework passing the operation to use to do the
  calibration (here a READ ID)
- the SPI controller framework replaying the same operation with
  different DLL configs until it finds the best match

So, it would basically be added as a new hook:

	int (*calibrate)(struct spi_mem *mem,
			 const struct spi_mem_op *tmpl);

and a new function provided by the spi-mem API

int spi_mem_calibrate(struct spi_mem *mem,
		      const struct spi_mem_op *tmpl);

and calibration outcome would be somehow attached to the spi_mem
object.

This way we stay memory agnostic but still provide the necessary blocks
at the spi-mem level to do such callibrations.

Would that work?