Re: [RFC] Introducing yamldt, a yaml to dtb compiler

On Mon, Jul 31, 2017 at 12:15:14PM -0500, Rob Herring wrote:

On Mon, Jul 31, 2017 at 8:11 AM, David Gibson
[off-list ref] wrote:

quoted

On Fri, Jul 28, 2017 at 10:07:10AM -0500, Rob Herring wrote:

quoted

On Fri, Jul 28, 2017 at 7:23 AM, Pantelis Antoniou
[off-list ref] wrote:

quoted

Hi Rob,

On Thu, 2017-07-27 at 21:12 -0500, Rob Herring wrote:

quoted

On Thu, Jul 27, 2017 at 7:51 PM, Tom Rini [off-list ref] wrote:

quoted

On Thu, Jul 27, 2017 at 06:00:00PM -0500, Rob Herring wrote:

quoted

On Thu, Jul 27, 2017 at 4:46 PM, Pantelis Antoniou
[off-list ref] wrote:

quoted

Hi Frank,

On Thu, 2017-07-27 at 13:22 -0700, Frank Rowand wrote:

quoted

Hi Pantelis,

Keep in mind one of the reasons Linus says he is very direct is to
avoid leading a developer on, so that they don't waste a lot of time
trying to resolve the maintainer's issues instead of realizing that
the maintainer is saying "no". Please read my current answer as being
"no, not likely to ever be accepted", not "no, not in the current form".

My first reaction is: no, this is not a good idea for the Linux kernel.

This has nothing to do with the kernel. It spits out valid DTBs that the
kernel (or anything else) may use.

Let me rephrase Frank's statement: this is not a good idea for the
main repository of dts files.

But sure, DTS is already not the only source of DTBs. It comes from
firmware on Power systems.

Yes, but unless they're generated from something other than a (at the
time) normal DTS, that's not a good example, IMHO.

They aren't. I'm talking about IBM systems. The firmware has its own
representation and flattens that to a DTB is how I understand it.

quoted

quoted

If you want to create and maintain your own
source format, then that is perfectly fine. But based on the current
understanding, I'm not seeing a reason we'd convert DTS files to YAML.

Can I propose one?  To borrow a phrase, Validation, Validation,
Validation.  Let me point to fe496e23b748 in the kernel for a moment.  I
found that as part of helping a new engineer come up to speed on doing
device tree work.  What I found was a case where:
- The binding doc gives one value for compatible as the required value.
- The code accepts only a single, different value.
- A few in-kernel dts files have different still values.

If the common dts source file was in yaml, binding docs would be written
so that we could use them as validation and hey, the above wouldn't ever
have happened.  And I'm sure this is not the only example that's in-tree
right now.  These kind of problems create an artificially high barrier
to entry in a rather important area of the kernel (you can't trust the
docs, you have to check around the code too, and of course the code
might have moved since the docs were written).

I'm all for validation, but the binding doc or schema and files that
describe platforms (aka DTS files) are not the same thing. The schema
is what are the constraints for a binding. Maybe some bindings are
fixed where there's only one valid binding implementation, but that's
the easy case (we could use DTS for that). I'll take YAML for binding
docs yesterday. Believe me, I'm tired of reviewing free form binding
docs. If that's where you want to go, reply to my reply that went
unanswered on Matt Porter's YAML proposal from 2 years ago (or maybe 3
now). I had the whole binding doc tree converted over to an initial
YAML schema. We just need to agree on the schema. Or we can keep
waiting for Grant to publish what he started on...

The way I see it there's a validation hierarchy.

There are the bindings that describe the schema of the resulting source
files. The bindings must be validated against a binding schema.

For the source files, at first they must be valid against the core
language (i.e. DTS or DT YAML variant) schema.

Next for each node that a binding exists in a valid format, it must be
validated against it. I.e. if an interrupt property exist it must point
to valid interrupt node etc.

Up next a number of per-platform/configuration validation passes.
I.e. for a complete source file which is using a specific SoC family
i.e. "ti,am33xx" the pass may verify that for the given peripherals
their configuration is correct, i.e. that the interrupt numbers for a
given peripheral are the correct ones for the target board etc.
This may be possible by having a golden master configuration when those
number can be retrieved and compared against.

Finally you could have a per-application/vendor/end-user final rule
check, i.e. the regulators may be configured in a manner that the power
consumption is under some specified threshold, etc. This is something
that is completely out of the kernel scope, but may have have to
vendors.

Why don't you share what you've been working on and see what we can do
using it as a base?

I did. 2 years ago:

https://git.kernel.org/pub/scm/linux/kernel/git/robh/linux.git/log/?h=dt-yaml-v2

It's very rough, but I was at the point of wanting feedback on the
schema format. Only the crickets gave me any.

It doesn't validate anything, but is purely binding docs mass
converted to YAML using DTS files as input.

Efforts on schemas have started and petered out several times :(.

I think the fundamental problem is that there just isn't critical mass
of people with the time to work on it.  Lots of people want better
validation, but not enough to put significant time and effort into
it.  I hope someone proves me wrong about that.

I would say part of the problem is the validation plans are always
just too grand. We need to start small.

Right, exactly.

Just having a machine
parseable documentation format alone would be a win. The validation
can come later IMO. It's going to come later anyway if we do nothing.

Well, maybe.  I fear that a "machine readable" format that doesn't
have some sort of automatic validation attached to it will end up
being not as machine readable as originally hoped.

quoted

Can I suggest (again) that one approach might be to add more pieces to
dtc's "checks" system to at least look for the more common errors.
It's not nearly a complete solution, but it gets you something with
much less difficulty than defining a whole schema system.  Some
rudimentary checking of unit addresses has been added relatively
recently, but not a lot else in the way of semantic checks.

Yes, I obviously agree. And at least for me, it's in a language I
already know. I've thought of a few more checks to add in the course
of this thread. Perhaps we should start a todo list if you have any
specific ideas. I've focused on things I repeatedly catch in binding
reviews (if only I could have a check for needing more specific
compatible strings :) ). Where we hit limits with the checks is when
we need specific compatible(s) to key checks on. This is any binding
that lacks a "class" property like gpio-controller or
interrupt-controller. For example I2C or SPI controllers and buses.
Certainly every common binding with a #*-cells property could have
some level of checks.

Go for it.  Checks are fairly easy to right, and easy to review, so
send 'em through and we should be able to merge them pretty quickly.

-- 
David Gibson			| I'll have my music baroque, and my code
david AT gibson.dropbear.id.au	| minimalist, thank you.  NOT _the_ _other_
				| _way_ _around_!
http://www.ozlabs.org/~dgibson