Re: [PATCH v3 3/5] arm64: dts: ti: Add support for AM642 SoC

[Tony Lindgren <tony@atomide.com>]

* Arnd Bergmann <arnd@kernel.org> [210122 11:24]:
> On Thu, Jan 21, 2021 at 8:57 PM Suman Anna <s-anna@ti.com> wrote:
> > On 1/21/21 12:39 PM, Nishanth Menon wrote:
> > > On 12:13-20210121, Suman Anna wrote:
> > >>
> > >> Hmm, this is kinda counter-intuitive. When I see a dts node, I am expecting the
> > >
> > > What is counter intutive about a -next branch be tested against
> > > linux-next tree?
> >
> > The -next process is well understood. FWIW, you are not sending your PR against
> > -next branch, but against primarily a -rc1 or -rc2 baseline.
> >
> > As a developer, when I am submitting patches, I am making sure that things are
> > functional against the baseline you use. For example, when I split functionality
> > into a driver portions and dts portions, I need to make sure both those
> > individual pieces boot fine and do not cause regressions, even though for the
> > final functionality, you need both.
> > >
> > >
> > > Now, if you want to launch a product with my -next branch - go ahead, I
> > > don't intent it for current kernel version - you are on your own.
> > >
> > > If there is a real risk of upstream next-breaking - speakup with an
> > > real example - All I care about is keeping upstream functional and
> > > useable.
> >
> > This is all moot when your own tree doesn't boot properly. In this case, you are
> > adding MMC nodes, but yet for a boot test, you are saying use linux-next for the
> > nodes that were added or you need additional driver patches (which is not how
> > maintainer-level trees are verified).
> >
> > Arnd,
> > Can you please guide us here as to what is expected in general, given that the
> > pull-request from Nishanth goes through you, and if there is some pre-existing
> > norms around this?
> 
> There are two very different cases to consider, and I'm not sure which one
> we have here:
> 
> - When submitting any changes to a working platform, each patch on
>   a branch that gets merged needs to work incrementally, e.g. a device
>   tree change merged through the soc tree must never stop a platform
>   from booting without a patch that gets merged through another branch
>   in the same merge window, or vice versa.
>   As an extension of this, I would actually appreciate if we never do
>   incompatible binding changes at all. If a driver patch enables a new
>   binding for already supported hardware, a second patch changes
>   the dts file to use the new binding, and a third patch removes the
>   original binding, this could still be done without regressions over
>   multiple merge windows, but it breaks the assumption that a new
>   kernel can boot with an old dtb (or vice versa). This second one
>   is a softer requirement, and we can make exceptions for particularly
>   good reasons, but please explain those in the patch description and
>   discuss with upstream maintainers before submitting patches that do
>   this.
> 
> - For a newly added hardware support, having a runtime dependency
>   on another branch is not a problem, we do that all the time: Adding
>   a device node for an existing board (or a new board) and the driver
>   code in another branch is not a regression because each branch
>   only has incremental changes that improve hardware support, and
>   it will work as soon as both are merged.
>   You raised the point about device bindings, which is best addressed
>   by having one commit that adds the (reviewed) binding document
>   first, and then have the driver branch and the dts branch based on
>   the same commit.
> 
> I hope that clarifies the case you are interested in, let me know if I
> missed something for the specific case at hand.

Hmm and additionally few more mostly obvious things that have helped
quite a bit:

- Each commit in each topic branch should compile and boot so git
  bisect works

- Each topic branch should be ideally based on -rc1 to leave out
  dependencies to other branches

- Aiming for a working and usable -rc1 is worth the effort in case
  git bisect is needed for any top branches based on it :) Otherwise
  you'll be wasting the -rc cycle chasing regressions..

Regards,

Tony