Re: [PATCH 12/14] doc-rst: add ABI documentation to the admin-guide book

[Mauro Carvalho Chehab <mchehab+samsung@kernel.org>]

Em Fri, 14 Jun 2019 16:06:03 +0200
Greg Kroah-Hartman <gregkh@linuxfoundation.org> escreveu:

> On Fri, Jun 14, 2019 at 04:42:20PM +0300, Jani Nikula wrote:
> > On Thu, 13 Jun 2019, Mauro Carvalho Chehab <mchehab+samsung@kernel.org> wrote:  
> > > From: Mauro Carvalho Chehab <mchehab@s-opensource.com>
> > >
> > > As we don't want a generic Sphinx extension to execute commands,
> > > change the one proposed to Markus to call the abi_book.pl
> > > script.
> > >
> > > Use a script to parse the Documentation/ABI directory and output
> > > it at the admin-guide.  
> > 
> > We had a legacy kernel-doc perl script so we used that instead of
> > rewriting it in python. Just to keep it bug-for-bug compatible with the
> > past. That was the only reason.
> > 
> > I see absolutely zero reason to add a new perl monstrosity with a python
> > extension to call it. All of this could be better done in python,
> > directly.
> > 
> > Please don't complicate the documentation build. I know you know we all
> > worked hard to take apart the old DocBook Rube Goldberg machine to
> > replace it with Sphinx. Please don't turn the Sphinx build to another
> > complicated mess.
> > 
> > My strong preferences are, in this order:
> > 
> > 1) Convert the ABI documentation to reStructuredText  
> 
> What would that exactly look like?  What would it require for new
> developers for when they write new entries?  Why not rely on a helper
> script, that allows us to validate things better?

Funny enough, this e-mail arrived here after Greg's reply, and my
reply over his one :-)

-

With regards to the script, my idea is to have it run on a new
"validate" mode, when the Kernel is built with COMPILE_TEST:

	https://git.linuxtv.org/mchehab/experimental.git/log/?h=abi_patches_v4

NB: the last patch is not yet... somehow, the building system is not
passing CONFIG_WARN_ABI_ERRORS to Documentation/Makefile. I'm
debugging it.

Personally, I would prefer to keep it the way it is, with two
additions:

1) I would add a SPDX header at the fist line of each file there;

2) It would make sense to have a new field - or indicator - to let
add ReST markups at the description. 

The advantage of using a parseable ABI file is that it is possible
to parse it, for example, to search for a symbol:

	$ ./scripts/get_abi.pl voltage_max

	/sys/class/power_supply/<supply_name>/voltage_max
	-------------------------------------------------

	Date:			January 2008
	Contact:		linux-pm@vger.kernel.org
	Defined on file:	Documentation/ABI/testing/sysfs-class-power

	Description:

	Reports the maximum VBUS voltage the supply can support.

	Access: Read
	Valid values: Represented in microvolts

	...

> 
> > 2) Have the python extension read the ABI files directly, without an
> >    extra pipeline.  
> 
> He who writes the script, get's to dictate the language of the script :)

No idea about how much time it would take if written in python,
but this perl script is really fast:

	$ time ./scripts/get_abi.pl search voltage_max >/dev/null
	real	0m0,139s
	user	0m0,132s
	sys	0m0,006s

That's the time it takes here (SSD disks) to read all files under
Documentation/ABI, parse them and seek for a string.

That's about half of the time a python script takes to just import the
the sphinx modules and print its version, running at the same machine:

	$ time sphinx-build --version >/dev/null

	real	0m0,224s
	user	0m0,199s
	sys	0m0,024s

> Personally, this looks sane to me, I'm going to apply the ABI fixups to
> my tree at least, and then see how the script works out.  The script can
> always be replaced with a different one in a different language at a
> later point in time of people think it really mattes.

Thanks,
Mauro