Em Fri, 12 Sep 2025 13:16:12 +0300
Jani Nikula [off-list ref] escreveu:

On Fri, 12 Sep 2025, Mauro Carvalho Chehab [off-list ref] wrote:

quoted

Em Thu, 11 Sep 2025 13:47:54 -0600
Jonathan Corbet [off-list ref] escreveu:
 

quoted

Jani Nikula [off-list ref] writes:
  

quoted

On Thu, 11 Sep 2025, Jonathan Corbet [off-list ref] wrote:    

quoted

A couple of times I have looked into using intersphinx, making each book
into an actually separate book.  The thing I always run into is that
doing a complete docs build, with working references, would require
building everything twice.  This is probably worth another attempt one
of these years...    

There are a couple of different usecase scenarios for building docs.

1) The first and most important one is to produce book(s) for people
   to use. This is usually done by some automation, and the result is
   placed on places like:
	- https://docs.kernel.org/

   and on subsystem-specific places like:
	- https://linuxtv.org/downloads/v4l-dvb-apis-new/

for scenario (1), taking twice the time to build is not an issue, as
nobody will be sitting on a chair waiting for the build to finish.

On such scenario, SPHINXDIRS is important on subsystem-specific docs.
For instance, on media, we use SPHINXDIRS to pick parts of 3 different
books:

	- Documentation/admin-guide/media/
	- Documentation/driver-api/media/
	- Documentation/userspace-api/media/

What media automation does, once per day, is:

	# Non-essencial parts of index.rst dropped
	cat <<END >Documentation/media/index.rst
	================================
	Linux Kernel Media Documentation
	================================

	.. toctree::

	        admin-guide/index
        	driver-api/index
	        userspace-api/index
	END

	rsync -uAXEHlaSx -W --inplace --delete Documentation/admin-guide/media/ Documentation/media/admin-guide
	rsync -uAXEHlaSx -W --inplace --delete Documentation/driver-api/media/ Documentation/media/driver-api
	rsync -uAXEHlaSx -W --inplace --delete Documentation/userspace-api/media/ Documentation/media/userspace-api

	make SPHINXDIRS='media' CSS='$CSS' DOCS_THEME='$DOCS_THEME' htmldocs
	make SPHINXDIRS='media' pdfdocs
	make SPHINXDIRS='media' epubdocs  

I was actually wondering how [1] was built. So it's not a complete build
of anything upstream, but rather something cobbled together downstream.

It used to be a direct build from upstream. I had to do this hack when
we decided to split subsystem docs on 3 separate books.

So your scenario (1) above is actually *two* wildly different scenarios.

And if upstream needs to cater for pretty much random subsets of
documentation being built, cherry-picking documentation from here and
there, I don't know what hope there is in radically refactoring how
documentation gets built upstream.

I presume you have one or more of a) get bunch of broken link warnings
at build, b) get broken links in the output, c) avoid links outside of
your subset altogether.

There aren't many broken links, and this is not due to (c): almost all
cross-references we have are between media kAPI and media uAPI. Those
were solved when we artificially joined two books and used SPHINXDIRS
feature to produce the docs.

If we had intersphinx support, I would be building the docs in
separate using the standard SPHINXDIRS logic to create such
cross references, pointing to linuxtv.org for media docs and to
docs.kernel.org for other ones.

[1] https://linuxtv.org/downloads/v4l-dvb-apis-new/

quoted

2) CI tests. Here, taking more time usually is not a problem, except
   when CI is used before pushing stuff, and the developer has to wait
   it to finish before pushing.

For scenario (2), a build time increase is problematic, as, if it now
takes twice the time, a change like that will require twice the
resources for the build with may increase costs.

3) developers who touched docs. They want a way to quickly build and
   verify the output for their changes.

Here, any time increase is problematic, and SPHINXDIRS play an important 
hole by allowing them to build only the touched documents.  

This is actually problematic, because the SPHINXDIRS partial builds will
give you warnings for unresolved references that are just fine if the
entire documentation gets built.

True, but if you pick them before/after a chanseset, you can notice
if the warning was introduced or not by the changeset. Only if it was
introduced by the patchset you need to wait 3 minutes for the full build.

quoted

For instance, when I was developing Netlink yaml plugin, I had to use
dozens of times:

	make SPINXDRS=Documentation/netlink/specs/ htmldocs

If I had to build the entire documentation every time, the development
time would increase from days to weeks.

Looking on those three scenarios, the only one where intersphinx is
useful is (1).  

It's also helpful for 3, and it could be helpful for 2 if CI only checks
some parts of the documentation.

I'm not arguing against intersphinx. I do think having it is something
we need to aim for.

The question is: does it replace SPHINXDIRS by providing quick builds
if only some of the books were changed?

quoted

From my PoV, we should support intersphinx, but this should be optional.  

Per my understanding making this somehow optional is not easily
achieved. And you end up with a bunch of extra complexity.

True, but I guess extra complexity is unavoidable: intersphinx
requires a list of books with reference locations, with is not the
same for everyone.

This is what expect once we have intersphinx in place:

Use linuxtv.org URLs for all references from:
	- Documentation/admin-guide/media/
	- Documentation/driver-api/media/
	- Documentation/userspace-api/media/

everything else: from kernel.org.

As they were generated from media next branch.

If implement it for DRM, in a way to track what DRM next branches
have, and if you have kapi, uapi and per-driver apis on different
books, you will probably want to solve intersphinx dependencies with 
a FDO specific "search" order, like:

	- xe and i915 books: from intel next branches;
	- amd books: from amd next branches;
	- drm core: from drm-next;
	- everything else: from kernel.org.

So, it is not just making it optional: you also need to provide a
way to allow it to be adjusted were it is needed. IMO, the easiest
way would be to have a separate .py file with intersphinx specifics:

	make SPHINXREFMAP=intersphinx_mapping.py htmldocs

This way, I could create a media_mapping.py file that would include
intersphinx_mapping.py and replace some defaults to do my own mapping.

quoted

Also, one has to point from where intersphinx will point unsolved
symbols. So, we would need something like:

	make SPHINXREFMAP=intersphinx_mapping.py htmldocs

where intersphinx_mapping.py would be a file containing intersphinx
configuration. We would add a default map at Documentation/, while
letting it to be overridden if some subsystem has different requirements
or is using a different CSS tamplate or not using alabaster.
 

quoted

quoted

I think the main factor in that should be whether it makes sense from
overall documentation standpoint, not the technical details.  

Agreed.
 

quoted

quoted

Having several books might make sense. It might even be helpful in
organizing the documentation by audiences. But having the granularity of
SPHINXDIRS with that would be overkill.   

On the contrary. SPHINXDIRS granuarity is very important for scenario (3).  

Sphinx does support incremental builds, and it's only the very first
build that's slow. IMO a handful of books that you can actually build
without warnings (unlike SPHINXDIRS) with incremental builds is a good
compromise.

That's not quite true: when Sphinx detects a missing file, it expires
the caches related to it and don't do incremental builds anymore. I had
to write a patch during the last development cycle due to that, as -rc1
came up with a broken reference because of a file rename. This was only
solved 3 months after the fact.

quoted

quoted

quoted

And there needs to be a book to
bring them together, and link to the other books, acting as the landing
page.    

Well, I think that the number of existing directories needs to be
reduced rather further.  I made progress in that direction by coalescing
all the arch docs under Documentation/arch/.  I would like to do
something similar with all the device-specific docs, creating
Documentation/devices/.  Then we start to get to a reasonable number of
books.  

I don't think reducing the number of books should be the goal, but,
instead, to have them with a clear and coherent organization with focus
on the audience that will be actually using them.

After reorg, we may have less books. That's fine. But it is also fine
if we end with more books.

I lost the battle years ago, but I still believe that, at least for
some subsystems like media, i2c, DRM, security and others, a 
subsystem-specific book could be better. After all, the audience for
such subsystems is very specialized.
 

quoted

quoted

I believe it should be possible to generate the intersphinx inventory
without generating the full html or pdf documentation. So I don't think
it's actually two complete docs builds. It might speed things up to have
a number of independent documentation builds.    

That's a good point, I hadn't looked into that part.  The builder phase
takes a lot of the time, if that could be cut out things would go
faster.   

Indeed, but we need to double check if .doctree cache expiration will
happen the right way for all books affected by a partial build.

During this merge window, I sent a RFC patch in the middle of a comment
with a conf.py logic to detect Sphinx cache expiration. I remember I
added a comment asking if we should upstream it or not, but, as nobody
answered, I ended forgetting about it.

If we're willing to experiment with that, I recommend looking on such
patch and add a variant of it, enabled via V=1 or via some debug
parameter.

The goal would be to check if a change on a file will ensure that all
books using it will have cache expiration and be rebuilt.
 

quoted

quoted

As to the working references, IIUC partial builds with SPHINXDIRS
doesn't get that part right if there are references outside of the
designated dirs, leading to warnings.    

That is true.  My point though is that, to get the references right with
a *full* build, a two-pass approach is needed though, as you suggest,
perhaps the first pass could be faster.  

How fast? during development time, SPHINXDIRS means a couple of seconds:

	$ make clean; time make SPHINXDIRS="peci" htmldocs
	...
	real    0m1,373s
	user    0m1,348s

Even more complex builds, even when picking more than one book, like this:

	$ make clean; time make SPHINXDIRS="driver-api/media/ userspace-api/media/" htmldocs
	...
	real    0m11,801s
	user    0m31,381s
	sys     0m6,880s

it still fits at the seconds range. Can interphinx first pass have a
similar build time?  

Probably not. Can you add links from media to non-media documentation
without warnings? Probably not also.

No, but I can count on my fingers the number of times I made such
changes: 99.9% of the time, doc changes aren't on the few docs that
have subsystem interdependencies. Even the number of dependencies
between media kapi and uapi are not high.

Thanks,
Mauro