[PATCH 8/9] buildman: Propose a format for extra boards

[Tom Rini]

On Sun, Nov 17, 2024 at 12:48:14PM -0700, Simon Glass wrote:
> Hi Tom,
> 
> On Fri, 15 Nov 2024 at 07:44, Tom Rini <trini at konsulko.com> wrote:
> >
> > On Fri, Nov 15, 2024 at 06:37:18AM -0700, Simon Glass wrote:
> > > Hi Tom,
> > >
> > > On Wed, 13 Nov 2024 at 15:20, Tom Rini <trini at konsulko.com> wrote:
> > > >
> > > > On Wed, Nov 13, 2024 at 09:03:35AM -0700, Simon Glass wrote:
> > > > > Hi Tom,
> > > > >
> > > > > On Tue, 12 Nov 2024 at 19:40, Tom Rini <trini at konsulko.com> wrote:
> > > > > >
> > > > > > On Fri, Nov 08, 2024 at 08:23:49AM -0700, Simon Glass wrote:
> > > > > >
> > > > > > > It has become more common to use config fragments to extend or adjust
> > > > > > > the functionality of boards in U-Boot.
> > > > > > >
> > > > > > > Propose a format for how to deal with this. It is not implemented as
> > > > > > > yet.
> > > > > > >
> > > > > > > Signed-off-by: Simon Glass <sjg at chromium.org>
> > > > > >
> > > > > > I think that the first problem is that this patch series is an
> > > > > > inappropriate method and place to start the discussion.
> > > > >
> > > > > We had a discussion a year ago but it trailed off.
> > > >
> > > > OK. Still an inappropriate place to resurrect it.
> > >
> > > What do you suggest? I am trying to solve the problem, not just talk
> > > about it, so patches are often best for that.
> >
> > Well, in that there's a problem to be solved, it's that buildman makes
> > #include more painful that it should be to use.
> >
> > Or, we validate config fragments by building them. A requires-buildman
> > solution is not appropriate. That:
> > https://patchwork.ozlabs.org/project/uboot/patch/20241114115049.2070780-2-patrick.rudolph@9elements.com/
> > doesn't use the acpi.config fragment is because of the parsing issue
> > buildman has.
> >
> > > > > > I also think this gets things backwards as the common case is "make",
> > > > > > not "buildman". We need more defconfig's that are just base +
> > > > > > fragment(s) if they're important enough to be a combination that needs
> > > > > > to be tested and work. A board is not a time-expensive part of CI. A
> > > > > > pytest run is, a new job itself is.
> > > > >
> > > > > OK, that would work too. It would also avoid the problem of combinatorial
> > > > > explosion. But I am not seeing people actually doing that, with rare
> > > > > exceptions.
> > > >
> > > > I think it's the exception, not the rule, where config fragments are not
> > > > being put to use in a defconfig. And that's possibly because not a lot
> > > > of people seem to know about the #include option, and then when I
> > > > explain it exists to people the next problem is "Oh, I have to do what
> > > > so that buildman also works?".
> > >
> > > Are you looking for any feature in Buildman for this? This series is
> > > intended to fix [1]
> >
> > Yes, the part of this series that, if I read it right, causes buildman
> > to read the preprocessed config file, rather than the raw defconfig, is
> > what should fix that, as the final patch in the series confirmed.
> 
> Are you saying you want a validation solution that doesn't use
> buildman. If so, why? CI relies entirely on buildman for building
> things.

Because (a) fragments should be used as a general rule and (b) it will
introduce a new failure mode for new platforms, platform added with
fragments, whatever-table not updated.

> The first 7 patches of this series resolve an issue that you filed.
> Above you say 'because of the parsing issue that buildman has'.
> So...should we fix it?

Yes, patches 1-7 are good and I'm leaning towards picking them up for
-next soon. They haven't quite had the general 2 week waiting period for
others to comment on.

> I don't like having fragments in the tree where we don't (at least
> programmatically) know what board they relate to. At least I think it
> should be an error to add a fragment with no in-tree defconfig users.

Fragments don't have specific relationships to boards. That's what I
keep saying. And we *intentionall* have fragments without in-tree
defconfig users too, as I spelled out in another part of this thread.

> In any case, I think the file format I have come up with is somewhat
> reasonable. My main concern with it is that it might greatly increase
> the CI load.

Yes, but I don't want the thing you're proposing and don't think it will
have value and that it will make life harder not easier because of
introducing a new CI-only failure mode for submissions.

-- 
Tom
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 659 bytes
Desc: not available
URL: <https://lists.denx.de/pipermail/u-boot/attachments/20241117/de0ec901/attachment.sig>