Message-ID: <CAKv+Gu-m-KYnsgyFBUuez3gRhwPF=_sD4fGpE6BDugzW9NUnsA@mail.gmail.com>
Date:   Sat, 4 Nov 2017 20:39:06 +0000
From:   Ard Biesheuvel <ard.biesheuvel@...aro.org>
To:     Andreas Färber <afaerber@...e.de>
Cc:     Leif Lindholm <leif.lindholm@...aro.org>,
        Grant Likely <Grant.Likely@....com>,
        Masahiro Yamada <yamada.masahiro@...ionext.com>,
        Satoru OKAMOTO <okamoto.satoru@...ionext.com>,
        Arnd Bergmann <arnd@...db.de>, Olof Johansson <olof@...om.net>,
        Mark Rutland <mark.rutland@....com>,
        Rob Herring <robh@...nel.org>,
        "devicetree@...r.kernel.org" <devicetree@...r.kernel.org>,
        Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
        Amit Kucheria <amit.kucheria@...aro.org>,
        linux-arm-kernel <linux-arm-kernel@...ts.infradead.org>,
        Yang Zhang <yang.zhang@...oards.org>
Subject: Re: [PATCH 3/5] dt-bindings: arm: Document Socionext MB86S71 and
 Fujitsu F-Cue

On 4 November 2017 at 20:06, Andreas Färber <afaerber@...e.de> wrote:
> Am 04.11.2017 um 23:39 schrieb Ard Biesheuvel:
>> On 4 November 2017 at 15:30, Andreas Färber <afaerber@...e.de> wrote:
>>> Am 04.11.2017 um 22:57 schrieb Ard Biesheuvel:
>>>> On 4 November 2017 at 13:44, Andreas Färber <afaerber@...e.de> wrote:
>>>>> Hi everyone,
>>>>>
>>>>> The non-building clk driver has been removed for 4.14, but this patchset
>>>>> seems stuck on matters of naming and maintenance...
>>>>>
>>>>> Am 30.06.2017 um 01:18 schrieb Masahiro Yamada:
>>>>>> Hi Andreas,
>>>>>>
>>>>>> 2017-06-29 21:53 GMT+09:00 Andreas Färber <afaerber@...e.de>:
>>>>>>> Hi Masahiro-san,
>>>>>>>
>>>>>>> Am 29.06.2017 um 14:18 schrieb Masahiro Yamada:
>>>>>>>> 2017-06-29 1:46 GMT+09:00 Rob Herring <robh@...nel.org>:
>>>>>>>>> On Sun, Jun 25, 2017 at 07:00:18PM +0200, Andreas Färber wrote:
>>>>>>>>>> For consistency with existing SoC bindings, use "fujitsu,mb86s71" but
>>>>>>>>>> socionext.txt.
>>>>>>>>>>
>>>>>>>>>> Signed-off-by: Andreas Färber <afaerber@...e.de>
>>>>>>>>>> ---
>>>>>>>>>>  Documentation/devicetree/bindings/arm/socionext.txt | 17 +++++++++++++++++
>>>>>>>>>>  1 file changed, 17 insertions(+)
>>>>>>>>>>  create mode 100644 Documentation/devicetree/bindings/arm/socionext.txt
>>>>>>>>>
>>>>>>>>> Acked-by: Rob Herring <robh@...nel.org>
>>>>>>>>> --
>>>>>>>>
>>>>>>>> I do not mind this, but
>>>>>>>> please note there are multiple product lines in Socionext
>>>>>>>> because Socionext merged LSI divisions from Panasonic and Fujitsu.
>>>>>>>>
>>>>>>>> I maintain documents for Socionext UniPhier SoC family
>>>>>>>> (which inherits SoC architecture of Panasonic)
>>>>>>>> in Documentation/devicetree/bindings/arm/uniphier/.
>>>>>>>
>>>>>>> Actually you seemed to be lacking bindings beyond the cache controller
>>>>>>> for Uniphier:
>>>>>>>
>>>>>>> https://git.kernel.org/pub/scm/linux/kernel/git/next/linux-next.git/tree/Documentation/devicetree/bindings/arm/uniphier
>>>>>>>
>>>>>>> The SoC compatible, e.g. "socionext,uniphier-ld11", needs to be defined
>>>>>>> somewhere too, as done here. A git-grep for that particular compatible
>>>>>>> only finds derived clk and reset bindings.
>>>>>>
>>>>>> I can care to send a patch later, but it is off-topic here.
>>>>>
>>>>> [The relevance was that had there been any bindings precedence from
>>>>> UniPhier, it would've influenced my naming choices.]
>>>>>
>>>>>>> Using socionext.txt allows you to add those bindings to a shared file;
>>>>>>> if you prefer to host them separately below uniphier/ or as uniphier.txt
>>>>>>
>>>>>> I was thinking of this way.
>>>>>>
>>>>>> For example, TI has omap/, keystone/, davinci.txt, etc.
>>>>>> in this directory level.
>>>>>>
>>>>>>> do you have a better name suggestion for this one? I was trying to keep
>>>>>>> our options open to later add SC2A11 in socionext.txt, and I also saw
>>>>>>> some mb8ac300 or so (MB86S7x predecessor?) in downstream sources, so I
>>>>>>> don't know a good common name for the non-Panasonic parts. And if we
>>>>>>> take fujitsu.txt for MB86S7x to match the vendor prefix then we will
>>>>>>> need a separate file for the new SC2A11 IIUC.
>>>>>>
>>>>>> I have no idea.
>>>>>> Actually, I am not familiar with those SoCs.
>>>>>>
>>>>>> I am not sure if there exists a common name for those Fujitsu-derived SoCs.
>>>>>> I think a SoC family name will be helpful to avoid proliferating
>>>>>> arch/arm/mach-{mb86s7x,mb8ac300, ...}.
>>>>>>
>>>>>> I see some Socionext guys CC'ed in this mail,
>>>>>> somebody might have information about this.
>>>>>>
>>>>>> As I said before, I do not mind adding socionext.txt
>>>>>> and it seems reasonable enough
>>>>>> if there is no common name for those SoCs.
>>>>>>
>>>>>>> Also if you can tell us where the cut between Fujitsu and Socionext
>>>>>>> should be done, we can certainly adapt. NXP is still adding all their
>>>>>>> new SoCs in fsl.txt, it seems.
>>>>>>> (A similar naming issue exists for my not-yet-submitted FM4 patches,
>>>>>>> where it changed owners from Fujitsu to Spansion and then to Cypress.)
>>>>>>
>>>>>> Right, vendor names are not future-proof in some cases.
>>>>>>
>>>>>> We use "uniphier" because it is convenient to
>>>>>> make a group of SoCs with similar architecture,
>>>>>> and it will work even if UniPhier product lines are sold again in the
>>>>>> future.  :-)
>>>>>
>>>>> Summarizing: Masahiro-san only wants to maintain the UniPhier family of
>>>>> Socionext SoCs, not this MB86S71. No one from Socionext or Linaro has
>>>>> volunteered as maintainer for these F-Cue MB86S71 patches - that seems
>>>>> to indicate I'll again have to set up a new repository and start
>>>>> maintaining it myself.
>>>>>
>>>>> Naming it linux-socionext.git wouldn't quite be right due to UniPhier
>>>>> also being Socionext.
>>>>>
>>>>> It's also unclear whether and by whom there may be SC2A11 patches - I
>>>>> hear for now Linaro are maintaining a SynQuacer DT in EDK2, rebelling
>>>>> against linux.git.
>>>>
>>>> Eh, wait, what? "Rebelling against linux.git"? What is that supposed
>>>> to mean exactly?
>>>
>>> Bindings must be submitted to the devicetree mailing list, acked by Rob
>>> and merged into the Linux tree before compatible strings may be used in
>>> Linux drivers or in-tree .dts[i] files. checkpatch.pl checks for that.
>>
>> OK, so where did we deviate from this in your opinion?
>
> I was very clear which new Socionext 96board platform my remark was
> about. There not being a socionext.txt file before this patch here hints
> at there being no official SoC binding defined for any.
>
> Adding one would require coordinating how to go about these platforms,
> as unsuccessfully attempted here.
>
> I was also referring to off-list remarks wrt EDK2 DT vs. Linux, which
> could've well come from you, given your rant.

I honestly have no idea what you are talking about here. You see
rebellion and argument where there is none.

>  Note that both the
> bindings maintainer and one arm-soc maintainer are from Linaro afaik, so
> you're essentially fighting against your colleagues here...
>

I am not fighting anyone, but you seem to think we are rebelling.

The reality is that the platform topology is not finalized yet, and
submitting and merging what are essentially draft version of DT
descriptions and then having to update them is something we are trying
very hard to avoid. There is enough churn in upstream DTs as it is.

> Last but not least pretty much every shipping 96Board to date had a
> pre-installed DT that did not have official bindings submitted (I don't
> count the Cello as shipping) and thus later changed incompatibly.
>

We are working very hard to make sure the upstream bindings and
drivers in the kernel are in shape to support the SoC we are
developing firmware support for. When the time comes, we may or may
not decide to propose the SoC DT for inclusion in the kernel. But
those are two different things.

This has nothing to do with 'rebelling against linux.git' (your words)

>>> U-Boot only allows import of new .dts files from Linux, too.
>>>
>>> So Linus' linux.git is the location to merge any vendor prefixes and
>>> device tree bindings,
>>
>> Agreed.
>>
>>> as well as the central location for device trees.
>>
>> Why should there be a central location for device trees?
>
> Because with a .dtsi for SoC and SoM we can share DT code across boards
> and bootloaders. And at least one .dts is needed for compile-testing it.
> From a central location files can be either individually copied or
> potentially aggregated as submodules.
>
> DT patch reviews have resulted in GIC nodes growing virtualization
> support for instance, and there have been in-tree refactorings fixing
> the arch timer's interrupt type iirc. Comparing modeling across vendors
> is also useful to not live in silos.
>
> Also considering that drivers are spread over subsystem rather than SoC
> directories these days, the central location being Linux facilitates
> checking which platforms have been enabled already and to which degree.
>

I understand all that. But being able to review SoC DTs is one thing,
and deciding that the kernel repository shall be the authoritative
source is another.

In my view, this argues for having a shared repository, mailing list
etc between the Linux, u-boot, freebsd and EDK2 communities, because
it would increase review coverage, and make it easier to push back
against frivolous DT changes.

>>> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/Documentation/devicetree/bindings
>>> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/arch/arm/boot/dts
>>> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/arch/arm64/boot/dts
>>>
>>> If you're unfamiliar with that, talk to your Linaro colleagues please!
>>
>> Please don't lecture me on this.
>
> If you ask questions, you may get answers. Blame yourself if you
> intended them as rhetorical questions.
>
>> The bindings are the contract between the OS and the platform, and I
>> agree that those need to be reviewed and maintained in a central
>> location, and the kernel tree, while not optimal, is a suitable place
>> for that.
>>
>> However, the existence of such contracts means that there is no need
>> to have a central location for device trees, nor should there be a
>> need for distros to ever ship device trees with the kernel. That
>> practice defeats the entire purpose of having contracts in the first
>> place, especially with the pathetic churn we are seeing in device
>> trees that are kept in the kernel.
>
> Please don't argue with me about these fundamentals. Not my decisions!
>
> This is not about distros shipping things; there is no openSUSE on these
> platforms precisely because code is not in central repositories.
>

By 'these platforms' you mean 96boards? Fair point. But please don't
extrapolate from Hikey (and FYI, Cello has been running mainline Linux
and EDK2 from the beginning)

> It's rather about three platforms by one vendor all going into different
> directions: SC2A11 into some seemingly non-mainline EDK2, MB86S71 as
> vendor kernel / U-Boot tarballs (and my kernel patches here), and
> UniPhier as only one in mainline kernel and mainline U-Boot.
>
> Have you ever tried to boot e.g. a HiKey with a DT not from the mainline
> kernel you are trying to boot? Good luck... Even server vendors
> occasionally change their device trees with firmware updates.
>

You say that like it's a bad thing. Do you honestly think it is better
for the OS to pick its own DT from some tarball that shipped with it?

If the kernel driver developers wouldn't keep making backward
incompatible changes to the bindings, they could actually stabilize to
the point where it makes even more sense for the platform vendor to
fix bugs in their DT descriptions.

> Fact is, as platforms get enabled, device trees grow additional nodes
> and properties. DTs don't fall from skies in some final form for new SoCs.
>
> Especially not if they keep getting written by volunteers like me - I'm
> already enabling the fourth 96board because Linaro and its partners keep
> throwing ~3.10 kernels at users again and again and again, which as I've
> pointed out at BUD17 is highly annoying! Even more annoying every single
> time Linaro marketing asks people to join as paying members because
> Linaro supposedly does so awesome Open Source work and such great
> 96Boards. Just this Friday again.

Who's ranting now?

Again, please don't extrapolate. For the SynQuacer based platform we
are involved with, there will be no vendor trees whatsoever. And I
share your frustration, especially when it comes to HiKey, because
they claim to implement UEFI (but building your own bootloader from
some random collection of EDK2 source files != UEFI), and keep trying
to upstream it.

> Nobody needs Reference Platform kernels if everything just timely gets
> merged into linux-next and mainline, where code belongs.

Agreed.

> Same for device trees, until the powers that be decide on a different
> location - which would not change the problem at all: Every review will
> be regarded as churn by someone else, and some people will always be
> lazy in submitting patches to whatever widely agreed-upon location, be
> it kernel drivers or DTs or U-Boot or EDK2 or ATF or something else.
>
> We wouldn't be having this argument if everything were great with the
> various non-central 96boards trees and downstream DTs.
>

Everything we need in the kernel and EDK2 will be upstream before the
product ships, and we are trying to do the same for ARM Trusted
Firmware, but this is a bit problematic due to the sorry state the
code is in.

Again, I am not the one who is ranting here. You hit a nerve by
accusing me of 'rebelling against linux.git' while this is quite the
opposite of what I am doing.

Powered by blists - more mailing lists