Re: [PATCH v2 05/11] spi: cadence-qspi: add FIFO depth detection quirk

["Théo Lebrun" <theo.lebrun@bootlin.com>]

Hello,

On Mon Apr 8, 2024 at 4:51 PM CEST, Mark Brown wrote:
> On Mon, Apr 08, 2024 at 04:38:56PM +0200, Théo Lebrun wrote:
> > On Mon Apr 8, 2024 at 4:10 PM CEST, Mark Brown wrote:
> > > On Fri, Apr 05, 2024 at 05:02:15PM +0200, Théo Lebrun wrote:
>
> > > > +	if (ddata && ddata->quirks & CQSPI_DETECT_FIFO_DEPTH) {
> > > > +		cqspi->fifo_depth = fifo_depth;
> > > > +		dev_dbg(dev, "using FIFO depth of %u\n", fifo_depth);
> > > > +	} else if (fifo_depth != cqspi->fifo_depth) {
> > > > +		dev_warn(dev, "detected FIFO depth (%u) different from config (%u)\n",
> > > > +			 fifo_depth, cqspi->fifo_depth);
> > > > +	}
>
> > > It's not obvious to me that we should ignore an explicitly specified
> > > property if the quirk is present
>
> > DT value isn't expected for compatibles with CQSPI_DETECT_FIFO_DEPTH
> > quirk, therefore we do not ignore a specified property. Bindings agree:
> > prop is false with EyeQ5 compatible.
>
> Sure, but it's not obvious that that is the most helpful or constructive
> way to handle things.

Agreed, a simpler solution can be found.

> > > - if anything I'd more expect to see
> > > the new warning in that case, possibly with a higher severity if we're
> > > saying that the quirk means we're more confident that the data reported
> > > by the hardware is reliable.  I think what I'd expect is that we always
> > > use an explicitly specified depth (hopefully the user was specifying it
> > > for a reason?).
>
> > The goal was a simpler devicetree on Mobileye platform. This is why we
> > add this behavior flag. You prefer the property to be always present?
> > This is a only a nice-to-have, you tell me what you prefer.
>
> I would prefer that the property is always optional, or only required on
> platforms where we know that the depth isn't probeable.
>
> > I wasn't sure all HW behaved in the same way wrt read-only bits in
> > SRAMPARTITION, and I do not have access to other platforms exploiting
> > this driver. This is why I kept behavior reserved for EyeQ5-integrated
> > IP block.
>
> Well, if there's such little confidence that the depth is reported then
> we shouldn't be logging an error.
>
> > > Pulling all the above together can we just drop the quirk and always do
> > > the detection, or leave the quirk as just controlling the severity with
> > > which we log any difference between detected and explicitly configured
> > > depths?
>
> > If we do not simplify devicetree, then I'd vote for dropping this patch
> > entirely. Adding code for detecting such an edge-case doesn't sound
> > useful. Especially since this kind of error should only occur to people
> > adding new hardware support; those probably do not need a nice
> > user-facing error message. What do you think?
>
> I'm confused why you think dropping the patch is a good idea?

Sorry I was unclear. I'll recap here options I see possible.

 - (1) Require DT property for all compatibles. That would be my
   preferred option *if* you think we should keep the DT property
   mandatory. I do not think requiring property AND detecting at
   runtime is useful.

 - (2) Require DT property for all but EyeQ5 compatible. On this
   platform, runtime detection is done.
    - (2a) On others, warn if value is different from DT property.
    - (2b) On others, do not detect+warn.

 - (3) Make DT property optional for all compatibles.
    - (3a) If provided, warn if runtime detect value is different.
    - (3b) If provided, do not detect+warn.

My preference would go to (3a):
 - we avoid a new quirk,
 - we avoid dt-bindings conditionals based on compatible,
 - we add a warning for a potentially buggy behavior and,
 - we do not modify FIFO depth used for existing devicetrees.

To make a choice, it'd be useful to know other platform behaviors. I
have no reason to think this SRAMPARTITION behavior isn't reproducable
on other platforms but I cannot guarantee anything. I just tested on TI
J7200 EVM with the quad SPI-NOR instance (spi@47040000) and it works as
expected.

Thanks,

--
Théo Lebrun, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com