Re: [PATCH/RFC] Hacky version of a glob() driven config include

["Ævar Arnfjörð Bjarmason" <avarab@gmail.com>]

On Fri, May 7, 2010 at 20:46, Jakub Narebski <jnareb@gmail.com> wrote:
> Ævar Arnfjörð Bjarmason  <avarab@gmail.com> writes:
>
>> This is not ready for inclusion in anything. Commiting for RFC on
>> whether this way of doing it is sane in theory.
>
> I think this is a good idea at least in theory.

Thanks, and thanks for all your comments below. They were very useful.

>> Known bugs:
>>
>>   * Breaks the model of being able to *set* config values. That
>>     doesn't work for the included files. Maybe not a bug.
>
> Errr... do I understand correctly that it simply means that you are
> not able to set config values that came from included files, in
> included files?
>
> This is quite serious limitation.

It is. And recap, you can now you can set Git's config in either
places .git/config, ~/.gitconfig and $prefix/etc/gitconfig.

With inclusion this is a bit more complex. If my ~/.gitconfig includes
a seekrt.key=foobar via an include in ~/.gitconfig/seekrt, what should
`git config --global seekrt.key newkey` do? How about `git config
--global seekrt.some_new_value blah`?

I think it's best to not try to get into that mess and just let the
user manage included files manually, or with `git config --file`.

>>   * Errors in the git_config_from_file() call in glob_include_config()
>>     aren't passed upwards.
>
> Hmmm...
>
>>
>>   * It relies on the GNU GLOB_TILDE extension with no
>>     alternative. That can be done by calling getenv("HOME") and
>>     s/~/$home/.
>
> "git config --path <variable>" expands leading '~' to $HOME, and ~user
> to home directory of given user.  Why not use this?

I didn't spot it. expand_user_path() does do everything I need,
yippie! I'll use that then.

>>   * The whole bit with saving/restoring global state for config
>>     inclusion is evil, but then again so is the global state.
>
> Why not encapsulate those global variables in a struct, passed to
> appropriate functions, with a global variable holding an instance of
> such struct (IIRC similarly to what is done for "the_index").

That's indeed the sane way to go. I'll do that (and look at
the_index).

>>   * We don't check for recursion. But Git gives up eventually after
>>     after spewing a *lot* of duplicate entry errors. Not sure how to
>>     do this sanely w/symlinks.
>
> The alternates mechanism has some depth limit; why not use it also for
> config file inclusion?  The machanism is quite similar...

Yet another example of prior art in git itself I have to check out,
thanks.

>> Not-signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
>
> You can simply do not add Signed-off-by for an RFC patch...
>
>> ---
>>
>> > On Sun, Apr 4, 2010 at 07:50, Eli Barzilay <eli@barzilay.org> wrote:
>> > > Isn't it better to have a way to include files instead?
>> >
>> > Probably yes. Programs like Apache HTTPD, rsyslog and others just use
>> > ${foo}conf.d by convention by supporting config inclusion.
>>
>> Here's an evil implementation of this. I know the code is horrid &
>> buggy (see above). But is the general idea sane. I thought it would be
>> better to submit this for comments before I went further with it.
>>
>>  config.c               |   55 +++++++++++++++++++++++++++++++++++++++++++++++-
>>  t/t1300-repo-config.sh |   43 +++++++++++++++++++++++++++++++++++++
>>  2 files changed, 97 insertions(+), 1 deletions(-)
>
> No documentation.

A final patch will have that. Since some of the semantics are "munges
your .gitconfig" (a bug). I'll write dosc for it once it's behaving.

> [...]
>> diff --git a/t/t1300-repo-config.sh b/t/t1300-repo-config.sh
>> index f11f98c..4df6658 100755
>> --- a/t/t1300-repo-config.sh
>> +++ b/t/t1300-repo-config.sh
>> @@ -824,4 +824,47 @@ test_expect_success 'check split_cmdline return' "
>>       test_must_fail git merge master
>>       "
>>
>> +cat > .git/config << EOF
>> +[some]
>> +     variable = blah
>> +[voodoo]
>> +     include = .git/more_config_*
>> +EOF
>
> I don't like this syntax.

Me neither.

> First, it forces git-config to hide all 'include' keys.  I think there
> might be some legitimate <section>.include config variables (perhaps
> outside git-core); with this patch they are impossible.

It's only hiding the full 'voodoo.include' key currently, you can
still have e.g. 'bleh.include'.

> Second, I guess that the section name has absolutely no meaning here.
> If included config file has section.key config variable, i.e.:
>
>  [section]
>        key = value
>
> the variable in master config file (visible by git-config) would not
> be voodoo.section.key.

No. All includes are done at the top level. The end result should be
equivalent to having used cat(1) to stitch a bunch of config files
together.

> Third, what happens with the sections in master config file?  If I
> have the following in .git/config
>
>  [voodoo]
>        var1 = val1
>        include = .git/more_config
>        var2 = val2
>
> and the .git/more_config has
>
>  [foo]
>        bar = baz
>
> would "git config --list" see 'voodoo.var2' (i.e. sections in included
> file does not change parsing of master file), or would it see
> 'foo.var2'?

Only voodoo.include should be hidden, nothing else. But let's move on
to..

> I would propose
>
>  include .git/more_config_*
>
> if not for the fast that it would trip older git.  Perhaps

>  ## include ".git/more_config_*"

Probably not a good idea to mix up comments & configuration like
that. Some (semi-broken) parsers of .gitconfig also use INI parsers to
parse it, which breaks on # comments. Those are already broken, but it
would be nice if a feature didn't require them.

>  [include .git/more_config_*]

Syntax error on older Gits.

>  [include ".git/more_config_*"]

I like this one the best. It's also easy to modify the parser (so it
doesn't think it's a section) to handle it. And it doesn't incur the
confusion of looking like a normal configuration variable.