Re: [PATCH v5 1/2] dt-bindings: brcm: clocks: add binding for brcmstb-cpu-clk-div

From: Florian Fainelli
Date: Fri Feb 03 2017 - 15:35:51 EST


On 02/03/2017 12:06 PM, Stephen Boyd wrote:
> On 02/01, Markus Mayer wrote:
>> On 20 January 2017 at 16:52, Stephen Boyd <sboyd@xxxxxxxxxxxxxx> wrote:
>>>
>>> Are these properties used? Please don't put these types of
>>> details in DT.
>>
>> Yeah, unfortunately they are. Luckily, I think the issue can be
>> resolved quite easily, because the user of those properties isn't
>> involved in this series.
>>
>> They are currently being used by a clock driver
>> ("drivers/clk/clk-brcmstb.c") that hasn't been upstreamed yet. I
>> performed some code archeology. While I wasn't 100% successful in
>> tracking down the origins of this interface, it looks like it was
>> designed this way a while back (4+ years or so), probably before
>> device tree best practices were developed or, at least, before they
>> were widely known.
>>
>> So, what I can do is to remove the properties from the official
>> binding. (I'll send an update to that effect shortly.) Once the
>> binding is accepted upstream, we can work on fixing up the design of
>> clk-brcmstb.c, so it doesn't rely on these properties anymore (and
>> derives them from the compatible string instead), and then proceed to
>> upstream that, as well.
>
> Ok. Sounds like some cleanup needs to be done on the way
> upstream.
>
>>> This register really looks like some offset in something larger.
>>> Is there some clock controller? What's the hw block at
>>> 0xf03e2000? Maybe I already asked this.
>>
>> It looks this way, but in this case, looks are deceiving. The address
>> and the length are really correct the way they are.
>>
>> This memory area holds a range of only loosely related configuration
>> registers. It's called the Bus Interface Unit Register Set and deals
>> with configuring the CPU in general. At address 0xf03e257c, there
>> happens to be the clock divider register we need, and it's really just
>> one register, i.e. 4 bytes.
>
> We've seen this style of hardware design before. I'd prefer we
> make the "Bus Interface Unit Register Set" into one device node
> and have a driver probe for it that registers this clock. If
> other things need to be controlled in there then the driver will
> do more than just register one clock, possibly hooking into
> multiple frameworks. The compatible string can indicate which SoC
> it is if the divider register offset changes or if the register
> layout is a total free for all.

We already have another piece of drive code that manipulates registers
in the Bus Interface Unit located in drivers/soc/bcm/brcmstb/biuctrl.c
and which has little to nothing to do with the CPU's clock ratio. And
actually another one being submitted that deals with the CPU's
read-ahead cache. I would very much prefer we keep all of them separate
and dealing with just the register offset they need to do, but that does
not mean the Device Tree binding has to look that way though.

The binding for the BIUCTRL register made it here:

Documentation/devicetree/bindings/arm/bcm/brcm,brcmstb.txt

so we should re-use that, and have a small piece of clock provided that
just uses the relevant register range within that larger register space
and provide the CLOCK_RATIO. Does that work?

>
> Either way, having reg properties end in non-zero values is
> suspect on ARM systems because a device is usually aligned to at
> least a 1k boundary for proper CPU addressing and mapping of the
> device. We can't possibly make a smaller mapping than a page
> anyway, so the registers around this one register will need to be
> mapped with the same attributes, hence the desire to make one
> device for the hardware block.

Yes that's absolutely valid, but this is not really a problem, since
ioremap() is smart enough to re-use existing mappings (or AFAIR).
--
Florian