Re: [RFC PATCH 2/2] config: add 'config.superproject' file

[Emily Shaffer <emilyshaffer@google.com>]

On Fri, Apr 09, 2021 at 03:29:46PM -0700, Junio C Hamano wrote:
> 
> Matheus Tavares Bernardino <matheus.bernardino@usp.br> writes:
> 
> > Hi, Emily
> >
> > I'm not familiar enough with this code to give a full review and I
> > imagine you probably want comments more focused on the design level,
> > while this is an RFC, but here are some small nitpicks I found while
> > reading the patch. I Hope it helps :)
> >
> > On Thu, Apr 8, 2021 at 8:39 PM Emily Shaffer <emilyshaffer@google.com> wrote:
> >>
> >> diff --git a/Documentation/git-config.txt b/Documentation/git-config.txt
> >> index 4b4cc5c5e8..a33136fb08 100644
> >> --- a/Documentation/git-config.txt
> >> +++ b/Documentation/git-config.txt
> >> @@ -48,7 +48,7 @@ unset an existing `--type` specifier with `--no-type`.
> >>
> >>  When reading, the values are read from the system, global and
> >>  repository local configuration files by default, and options
> >> -`--system`, `--global`, `--local`, `--worktree` and
> >> +`--system`, `--global`, `--superproject`, `--local`, `--worktree` and
> >>  `--file <filename>` can be used to tell the command to read from only
> >>  that location (see <<FILES>>).
> >>
> >> @@ -127,6 +127,17 @@ rather than from all available files.
> >>  +
> >>  See also <<FILES>>.
> >>
> >> +--superproject::
> >> +       For writing options: write to the superproject's
> >> +       `.git/config.superproject` file, even if run from a submodule of that
> >> +       superproject.
> >
> > Hmm, I wonder what happens if a repo is both a submodule and a
> > superproject (i.e. in case of nested submodules).
> 
> Another thing I am not sure about the design is that a repository
> can be shared as a submodule by more than one superprojects.  The
> superprojects may want their submodule checkouts at different
> submodule commits, but that is something doable by having multiple
> worktrees connected to a single submodule repository.

I think the implementation as-written actually handles this
sharing-via-worktree case you describe gracefully, as it discovers the
gitdir belonging to the worktree above the worktree where it is being
run now:

superproject-a
-> sub-a <gitdir at superproject-a/.git/modules/sub-a/>
superproject-b
-> sub-b <gitdir at superproject-a/.git/modules/sub-a/worktrees/sub-b/>

In this case running `git config --list --superproject` in sub-a will
yield superproject-a/.git/config.superproject and running it in sub-b
will yield superproject-b/.git/config.superproject; that seems logical
to me. If I am adding libc as a submodule via worktree to Git as well
as, say, Wireshark, just to save me on disk space, I wouldn't want my
Wireshark hooks to run in Git project or vice versa.

> I think our design principle has been that it is perfectly OK for a
> superproject to be in total control if its submodules, but
> submodules should not even be aware of being used as a submodule by
> a superproject, and that allows a submodule repository to be shared
> by multiple superprojects.  As "write to the superproject's X file"
> requires a submodule to know who THE superproject of itself is, this
> feature itself (not the implementation) feels somewhat iffy.

As for this, I wonder what the reasoning is. I guess to simplify the
code, and to make the behavior more predictable (for example, 'git
commit' doesn't suddenly make a commit in some project that isn't this
one)?

One could imagine some really nice quality-of-life improvements if the
submodule is allowed to know it's a submodule (even by a config, for
example):

 - We could teach 'git status' to indicate the state when the submodule
   index is clean, but the superproject does not contain a commit
   pointing to the submodule's HEAD - which could still be considered a
   dirty state, since the change isn't associated with the larger project
   yet. I could imagine there might be other handy information related
   to submodule/superproject status we may want to display too.
 - We could teach 'git log' to decorate commits which are referred to by
   superproject commits, perhaps?
 - We could teach 'git push' to, by option or config, push the entire
   superproject-and-submodules package at once, to make it easier to
   coordinate changes across the whole superproject
 - One could envision some other niceties like 'git stash
   --whole-superproject' or similar, where a user could operate on the
   entire overall project (that is, the superproject and all its
   submodules) without needing to switch context away

I don't think lacking these things would stop a user from doing
something they want to do, but it does seem like they could make life
more comfortable for a user developing in a project made up of many
submodules.

 - Emily