Re: [PATCH 1/2] usb: doc: dwc3-xilinx: Add devicetree bindings

[Rob Herring]

On Wed, Nov 18, 2015 at 5:02 PM, Felipe Balbi <balbi@xxxxxx> wrote:
>
> Hi,
>
> Rob Herring <robh@xxxxxxxxxx> writes:
>> On Wed, Nov 18, 2015 at 06:20:31PM +0530, Subbaraya Sundeep Bhatta wrote:
>>> This patch adds binding doc for Xilinx DWC3 glue driver.
>>>
>>> Signed-off-by: Subbaraya Sundeep Bhatta <sbhatta@xxxxxxxxxx>
>>
>> I really dislike how the DWC3 binding has been done. The sub-node here
>> is pointless. The only thing the outer node does is add clocks which
>> should simply be part of the DWC3 node. Presumably the IP block needs
>> some clocks...
>>
>> Anyway, it's self-contained within the DWC3, so I won't make you clean
>> up the mess.
>
> heh, it's definitely not pointless. We get a lot of free goodies by
> doing it that way. I'm just not going to repeat it once again. But in
> summary:
>
> a) we force people to write glue layers for *only* their HW specific
> details
>
> b) there's really no way to abuse dwc3 core because there's no symbol
> exported from dwc3 core to glue

Both are doable independent of DT.

> c) PM (both system sleep and runtime) becomes a lot simpler with core
> only caring about what PM details delivered by SNPS (e.g. Hibernation
> mode from DWC3) and glue only has to consider the SoC integration parts
> of PM (for OMAP that would be PRCM details, hwmod, etc).

No doubt OMAP is special.

> I'm definitely not going to change dwc3 because it has made my life a
> lot saner. Definitely a lot saner than MUSB. Besides, DTS is supposed to
> describe the HW and that's, really, how the HW is.

So reading the description, the DWC3 has no clocks?

> There's a piece of HW which is somewhat platform agnostic and delivered
> by SNPS as a black box (SNPS customers don't touch SNPS RTL) and there's
> another piece of HW which bridges this dwc3 to the platform specific
> details and integration context of the platform (for OMAP that would the
> SYSCONFIG/SYSSTATUS registers, Clock autogating, PRCM, etc, etc etc).

It is a black box, but with dozens of configuration options typically.

> So you _do_ in fact, have two separate pieces of HW which are SW visible
> and controllable independently. They each have their own IRQs (though in
> some SoCs they share the same line), they're own address space, yada
> yada yada.

When that is the case, I have no problem with this split, but I don't
see any of these examples in this particular case. So how should the
binding look when there is no vendor specific glue? Are we supposed to
keep the binding structure to appease the needs of the driver?

Rob
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/