[ovs-dev] [PATCH v3 3/6] makefile: create destdir when generating files

[Ben Pfaff]

On Wed, Aug 02, 2017 at 05:25:36PM -0400, Aaron Conole wrote:
> Ben Pfaff <blp at ovn.org> writes:
> 
> > On Tue, Aug 01, 2017 at 06:05:40PM -0400, Aaron Conole wrote:
> >> When intermediary files are generated, the destination directory is assumed
> >> to exist.  This has worked so far because most files are built prior to
> >> the dist-packaging step.  However, any files which require rebuild after
> >> the packaging step may end up in failure if the output directory is not
> >> available.  This commit adds a 'mkdir -p' before generating the output files.
> >> 
> >> Signed-off-by: Aaron Conole <aconole at redhat.com>
> >
> > Thanks for working on the makefiles.
> >
> > What's an example of a file for which this makes a difference?  I'm
> > having trouble understanding when this helps.
> 
> Thanks for reviewing, Ben!
> 
> The next patch in the series adds an example of this.  I should have put
> that into the commit message as well.
> 
> I'll go over the effect I observed, and maybe there's a better way to do
> it:
> 
> 1. I added a 'preprocessor' to be able to have dpdk vs. non-dpdk
>    functionality in generated files (ex: the vswitchd service file)
> 
> 2. I execute `./configure --with-dpdk` and observe that the correct
>    lines are included.  So far so good.
> 
> 3. I completely clean the tree, and do a `./configure` without
>    specifying dpdk.  The lines are not included, and still so far so
>    good.
> 
> 4. I use the resulting distribution (made with make dist) to build,
>    doing `./configure --with-dpdk`.  Now, I would think the file should
>    have the correct lines included, but it does not.
> 
> To resolve this, I tried adding the generated file as BUILT_SOURCES, but
> experienced errors when doing a `make distcheck`, since the $(srcdir)
> directory is read only.  So, I implemented the approach where I added
> $(builddir)/config.status which would behave as I expect, but for the
> fact that the sed will do a redirect into a subdirectory that isn't
> created.
> 
> I'm sure a better way could exist, but I don't know it.
> 
> I guess that's a long-winded (apologies) way of saying there are
> currently no examples in tree - this commit adds one.

I think that the root of the problem here is that we're distributing a
generated file, and it's made worse because that generated file depends
on the build-time configuration.  That's a recipe for Autoconf/Automake
hell.

Ideally, we'd stop distributing a generated file entirely.  It's a
little complicated for packaging because we want to fill in part of the
packaging based on what's in OVS, but the straightforward way to do that
is to generate the packaging as part of the OVS build, but the OVS build
is what the packaging is supposed to do anyway.  (Frankly I wish that
OVS didn't come with its own packaging--that would solve lots of
problems.)

What if the spec files themselves did the editing of the service file,
instead of doing it from OVS?  That would solve this particular problem
because we wouldn't need to have a generated file in the OVS tree (just
the source file).