[PATCH v3] clk: shmobile: Add R8A7740-specific clock support

[Magnus Damm]

HI Laurent,

On Fri, May 23, 2014 at 8:18 AM, Laurent Pinchart
<laurent.pinchart at ideasonboard.com> wrote:
> Hi Magnus,
>
> On Thursday 22 May 2014 20:16:06 Magnus Damm wrote:
>> On Thu, May 22, 2014 at 7:22 PM, Laurent Pinchart wrote:
>> > On Thursday 22 May 2014 09:37:40 Magnus Damm wrote:
>> >> On Thu, May 22, 2014 at 12:41 AM, Laurent Pinchart wrote:
>> >> > On Wednesday 21 May 2014 16:21:26 Ulrich Hecht wrote:
>> >> >> Driver for the R8A7740's clocks that are too specific to be supported
>> >> >> by a generic driver.
>> >> >>
>> >> >> Signed-off-by: Ulrich Hecht <ulrich.hecht+renesas at gmail.com>
>> >> >
>> >> > The implementation looks globally sane to me. There's still quite a few
>> >> > missing clocks, but there's no hurry in adding support for them at the
>> >> > moment. I'd like to get the bindings reviewed by someone outside of our
>> >> > team, but that would require making the CPG documentation (or at least
>> >> > the block diagram) available. Magnus, is there a chance for that to
>> >> > happen ?
>> >>
>> >> Regarding documentation, as much as I'd like to see this, in practice it
>> >> feels highly unlikely since I'm not in control of the actual data sheet
>> >> distribution policy. I believe Emma Mobile series data sheet are
>> >> available for public download, but the rest of the SoCs are not
>> >> unfortunately.
>> >>
>> >> This issue with closed documentation is not specific to r8a7740 though,
>> >> so earlier developed CCF implementations included in upstream like
>> >> r8a7790, r8a7791, r7s72100 and r8a7779 are in the same state as r8a7740.
>> >
>> > Yes they are, although the hardware is simpler in those cases, so there's
>> > less potential issues. I don't have a specific concern here, just a
>> > feeling of uneasiness coming from publishing DT bindings that can't be
>> > properly reviewed by someone out of our team. We're too familiar with the
>> > hardware to take a step back and see the bindings from an external point
>> > of view, which is why I'd like an external review before committing to
>> > any DT binding stability.
>>
>> I understand that you cannot vouch for the stability of this binding. =)
>>
>> And I agree that external review would help in making it stabler quicker.
>>
>> > There's of course no reason to delay this patch (well, except for all the
>> > other small comments I've made :-)), but if it were me I would mark the
>> > corresponding bindings with a big "EXPERIMENTAL" warning.
>>
>> We can chose to treat it as relatively experimental if we want to. So
>> your warning is fine!
>>
>> How long stabilization time would you recommend for this kind of
>> thing? I would guess half a year?
>
> I see several main reasons why DT bindings need time to stabilize.

Thanks for listing these. I agree that we need time to make the bindings stable.

> - When implementing support for currently unsupported features of the device
> we could realize that the bindings were designed in a way that makes support
> of new features difficult in a backward-compatible way.

Right, I can see how it is possible to end up in that situation by not
supporting all device hardware features in the driver from the
beginning and just "enabling" DT support. It does however seem to me
that in such case the DT binding must have been written to match the
driver and not the hardware. We should try not to. =)

So perhaps we would need more time to develop the initial DT binding
for the particular device version? Of course if this happens then we
need to fix it - no doubt about that - but exactly how to fix it needs
to be discussed and handled case-by-case IMO. Having plenty of
stabilization time (and users) will help here I think.

> - When a new SoC comes out we could realize that device features or SoC
> integration parameters that we thought were set in stone are actually not, and
> don't play well with the current bindings.

So if the documentation gets a surprise update then we should of
course update the software too. Depending on how long time the DT
binding has been in use I believe we have different options how to
handle backwards compatibility. We can ignore the DT stability
altogether or make sure to keep compatibility for some time. Exactly
what to do depends on the timing, number of users and difficulty of
maintaining compatibility and our users in the particular product
segment. Stabilization time should help here too.

> - When other vendors implement bindings for similar devices we could realize
> that common needs would benefit from common bindings.

This one is a bit different IMO.

> At least the last reason advocates for experimental DT bindings being
> committed to the mainline kernel, to make it possible for developers to
> identify common needs between vendors.

While I agree that vendors should submit their DT bindings upstream as
first step I would not necessarily use the word "experimental" for
such DT bindings. Instead I would argue that it is natural for the
vendors to cover only their own hardware to begin with. I would also
say that it is pretty normal for the kernel to over time come up with
better ways to describe the hardware.

My opinion is that it is seldom right to break DT compatibility just
because a slightly nicer common binding has been designed. Instead I
think it makes more sense to use the new-and-improved DT style for
newer SoCs and keep compatibility for older SoCs. Basically I believe
we should handle different SoC generations and their devices with
different style of bindings and keep on moving forward to better and
better ways to describe the hardware.

I don't think stabilization time would help in this case.

> In this specific case I'm mostly concerned about the first reason, as I don't
> expect our new SoCs to come out with a CPG nearly-but-not-quite-identical to
> the r8a7740's. The stabilization time would thus depend entirely on our effort
> to implement support for the currently missing features. Integration with MSTP
> is the main one (the overlapping register ranges problem I've mentioned before
> is a good example, even if it doesn't cause a real problem today), and support
> for at least some of the missing clocks would be good as well. If we actively
> maintain this driver I expect that one or two kernel releases should be enough
> for the bindings to stabilize.

I agree about your suggestion about r8a7740 stabilization time.
Regarding the particular case with MSTP my opinion is that we must
have discussed this nesting option when we designed the MSTP DT
bindings in the first place. Of course, new information may have come
up so we need to adjust. I do however think that it must make sense to
handle r8a7740 and newer SoCs in the new DT binding format but refrain
from breaking existing MSTP DT bindings for already supported SoCs.

> Please feel free to also discuss the other reasons I've mentioned above and
> how you would like to see them being handled.

As you can tell, I don't think there is any one-way-fits-all for DT
stability. In the end I think this all is about keeping a good balance
between making a stable enough kernel for our users while also not
collecting too much compatibility cruft. One way to handle this is to
separate the timing for introduction and removal of DT bindings and
give users time to move over.

Cheers,

/ magnus