[RFC] Does PHY UTMI data width belong to DWC2 or PHY binding?

[Matt Porter]

On Wed, Oct 23, 2013 at 01:11:15PM -0500, Felipe Balbi wrote:
> Hi,
> 
> On Wed, Oct 23, 2013 at 10:42:42AM -0400, Matt Porter wrote:
> > On Tue, Oct 22, 2013 at 04:38:52PM -0500, Rob Herring wrote:
> > > On 10/22/2013 06:25 AM, Matt Porter wrote:
> > > > On Tue, Oct 22, 2013 at 12:48:29PM +0200, Matthijs Kooijman wrote:
> > > >> Hi Kishon,
> > > >>
> > > >> On Mon, Oct 21, 2013 at 02:57:26PM +0530, Kishon Vijay Abraham I wrote:
> > > >>> I think it makes sense to keep the data width property in the dwc2 node itself.
> > > >>> I mean it describes how the dwc2 IP is configured in that particular SoC (given
> > > >>> that it can be either <8> or <16>).
> > > >> If I'm reading the RT3052 datasheet correctly (GHWCFG4 register), the IP
> > > >> can be configured for 8, 16 or 8 _and_ 16. In the latter case, the "8
> > > >> and 16 supported" would make sense as a property of dwc2 (though this
> > > >> value should be autodetectable through GHWCFG4), while the actual 8 or
> > > >> 16 supported by the PHY would make sense as property of a phy.
> > > > 
> > > > There would be no value in adding a property for an already detectable
> > > > value to dwc2's binding. To be honest, it's pretty much useless
> > > > information due to the existence of the "8 and 16" option.
> > > > 
> > > >> Note sure if this is really useful in practice as well, or if just
> > > >> setting the actual width to use on dwc2 makes more sense...
> > > > 
> > > > The GHWCFG4 information itself is not useful in practice, as described
> > > > in the original thread: https://lkml.org/lkml/2013/10/10/477
> > > > 
> > > > It's certainly useful in practice to have this width property in either
> > > > the dwc2 or the phy binding. One can make a case for either. As I
> > > > mentioned in the original post, if we put it in the phy binding we'll be
> > > > updating the generic phy binding. We'll then need an api added into the
> > > > generic phy framework to fetch the width of a phy.
> > > > 
> > > > Both cases are doable and trivial, we just need the canonical decision
> > > > from a DT maintainer as to where the property belongs. Given that they
> > > > are in ARM ksummit, I'm not expecting to hear anything right this
> > > > moment. :)
> > > 
> > > The host can support both, so it is not a property of the host and is a
> > > property of the phy. It is no different than what mode a SPI slave
> > > requires or whether an i2c slave supports 8 or 10-bit addressing. Those
> > > examples are all 1 to many rather than 1 to 1 where it doesn't really
> > > matter, but the same logic applies.
> > 
> > Makes good sense, thanks.
> > 
> > In this case, given the PHY ownership of width, we can completely avoid
> > any DT properties. The generic phy compliant BCM Kona phy driver can
> > report via the generic phy framework that it is 8-bit wide. There's no
> > support for this type of thing now but it's pretty trivial to add.
> > 
> > I went ahead and did a quick proof-of-concept that adds a free-form
> > phy attributes struct for the generic phy. Given that generic phys can
> > be for any transmission technology this could be filled with a jumble
> > unrelated and often unpopulated attributes over time. In any case, the
> > below patch allows the phy provider to choose to specify utmi_width and
> > a controller driver that cares can use phy_get_attrs() to fetch the
> > optional phy attributes and use the utmi_width field if applicable.
> > 
> > Kishon: I'll start a separate thread to discuss what approach you'd like
> > to see in the generic phy framework to manage this.
> > 
> > -Matt
> > 
> > diff --git a/include/linux/phy/phy.h b/include/linux/phy/phy.h
> > index 6d72269..b763d7b 100644
> > --- a/include/linux/phy/phy.h
> > +++ b/include/linux/phy/phy.h
> > @@ -38,6 +38,14 @@ struct phy_ops {
> >  };
> > 
> >  /**
> > + * struct phy_attrs - represents phy attributes
> > + * @utmi_width: Data path width implemented by UTMI PHY
> > + */
> > +struct phy_attrs {
> > +	int			utmi_width;
> 
> this is supposed to be a generic PHY layer and as such, it shouldn't
> know about USB details such as the UTMI bus. How about calling bus_width
> just to make it more generic ? Then it would work for UTMI, PIPE3, ULPI,
> SLPI (did that even fly ?) or any other PHY <-> link interconnect.

That sounds much better. Yeah, I was also thinking about embedding a
per-phy-class attribute struct and that just looked ugly. I like the simple
generic bus_width.

I'll update and post a real patch.

Thanks,
Matt