breaking DT compatibility

[Andre Przywara]

Hi,

On 10/02/16 14:37, Maxime Ripard wrote:
> Hi Rob,
> 
> On Wed, Feb 10, 2016 at 07:42:02AM -0600, Rob Herring wrote:
>> On Wed, Feb 10, 2016 at 6:30 AM, Andre Przywara <andre.przywara at arm.com> wrote:
>>> Hi,
>>>
>>> just a ping:
>>>
>>> Are we really OK with breaking existing DTs in 4.6? (per the code in
>>> -next: f7d372ba54ea04d528a291b8dbe34716507bb60b, which is this patch).
>>
>> I only warn and make sure people are aware of the issue. I leave that
>> up to platform maintainers to decide. It depends on the maturity of
>> the platform and users.
> 
> The impacted SoCs support is really partial. For the most supported
> one, big things like display or sound are totally missing, and we
> still update them on a regular basis to add support for new
> devices. As such, users are very likely to upgrade the DT from one
> version to another just because there's new devices available to
> them. And the newest SoC impacted just got introduced in 4.5, and only
> has the UART and MMC devices available.

But I still don't see why we have to break things - can't we just
improve support _without_ breaking compatibility? For instance keeping
the old behaviour around? I can dig out my old x86 hat and find a
_compatible_ solution for this, if you prefer ;-)
Actually doesn't Jean-Francois' patch fixes the PLL8 issue without
breaking anything?

Also were do you want to draw the line for "partial support"? The
sun6i-a31.dtsi is around since 3.12, which was released more than 2
years ago.
If we want to wait for "stabilization", we will probably wait forever. I
find discussions about DT stabilization going back to the very beginning
of DT for ARM - as well as the request for moving the DT bindings out of
the kernel (which I think we should really eventually do now).

Frankly I don't care so much about these particular boards, but just
want avoid to set a bad example for the future - in particular sneaking
this behaviour into the arm64 world, where DTs _really_ come with the
board or are even generated by the (UEFI-)firmware.
U-Boot already can embed a device tree, sounds like a perfect place to
have it stored for me - as long as there is _one_ best version of it.

>> If people complain about it then it's their mess. For platforms
>> supported in distros such as debian or fedora, I would strongly
>> recommend against breaking compatibility.
> 
> None of them are officially supported:
> https://www.debian.org/releases/stable/armhf/ch02s01.html.en
> https://fedoraproject.org/wiki/Architectures/ARM#Fedora_23
> 
> Only the older one are, and they are not affected by this patch.
> 
>> They do ship dtbs, but it's a chicken and egg problem. If dtbs were
>> stable and provided by firmware, then they wouldn't have to provide
>> them. If dtbs are unstable, then they have no choice.
> 
> And while it might work great on platforms that have all the needed
> documentation, or a perfect one, which is our case. Almost each
> release, we discover that something is not working as it was
> documented, when it was documented in the first place.

I understand that it's not an ideal situation for those Allwinner chips
- but nevertheless I would like us to at least _try_ to comply with this
idea. As said before: this patch is a nice rework/cleanup - but
definitely not a good reason to break compatibility.

> It also seems that even on well documented platforms, supported by the
> vendors, the stable ABI dream is not going to happen:
> https://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/tree/Documentation/arm/Atmel/README#n105
> https://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/tree/Documentation/devicetree/bindings/arm/marvell,berlin.txt#n4

I find those two statements really unfortunate ...

> If it makes things clearer, I can also add such a statement.

and would rather see us referring solely to this document instead:
	Documentation/devicetree/bindings/ABI.txt

Cheers,
Andre.

>>> Also I think one needs ACKs from DT maintainers before merging something
>>> in the respective directories, which I don't see here.
>>
>> It can go in with subsystem maintainers ack, but there are problems
>> with this one regardless of compatibility.
>>
>>> As I am somewhat blocked on that patch, I'd like to have some discussion
>>> on the list.
>>>
>>> Thanks,
>>> Andre.
>>>
>>> On 05/02/16 17:59, Andre Przywara wrote:
>>>> Hi Maxime,
>>>>
>>>> just found this while looking at your current git branch, so sorry for
>>>> the late reply.
>>>>
>>>> CC:ing DT people, since you touch both existing DTs(!) and the binding doc.
>>>>
>>>> On 01/02/16 20:20, Maxime Ripard wrote:
>>>>> Remove the fixed dividers from the PLL6 driver to be able to have a
>>>>> reusable driver that can be used across several SoCs that share the same
>>>>> controller, but don't have the same set of dividers for this clock, and to
>>>>> also be reused multiple times in the same SoC, since we're droping the
>>>>> clock name.
>>
>> Removing a compatible name or not has nothing to do with sharing code.
> 
> This was not about the compatible name, but the hardcoded name in our
> structure associated to that compatible. And since we can't register
> two clocks with the same name, we couldn't use the same compatible
> several times.
> 
> Maxime
>