Re: [PATCH 08/16] update submodules: add depopulate_submodule

[Stefan Beller <sbeller@google.com>]

On Thu, Nov 17, 2016 at 2:29 PM, Brandon Williams <bmwill@google.com> wrote:
> On 11/17, Stefan Beller wrote:
>> Well first you get the warning:
>>
>>     "cannot remove submodule '%s' because it (or one of "
>>     "its nested submodules) uses a .git directory"),
>>
>> and in case a d/f/ conflict arises in a later stage (e.g. when the submodule
>> is replaced by a file or symlink), you get another related error with
>> less helpful description how to debug it.
>
> Maybe a warning isn't the right thing?  Shouldn't the checkout fail if
> there are any issues?  This would force the user to stash/commit their
> changes and then retry.

Well if the path is not reused, e.g. you just delete a submodule in a commit
without anything else, you could proceed and have the submodule laying
around dirty?

>
>> >  All other submodules will actually be removed? Couldn't
>> > you end up in an undesirable state with a checkout effecting one
>> > submodule but not another?
>>
>> Yes you could. Maybe it's time to add
>> "git submodule intern-git-dir", which can be given as a helpful hint
>> or even run here first.
>
> That would be a good idea, does that functionality already exist in one
> form or another?  I'm assuming it must since git update does just that
> when cloning a submodule.

No it doesn't (it is roughly these three steps):

    mv ${SUBMODULE_PATH}/.git ${GIT_DIR}/modules/${SUBMODULE_NAME}
    git config -f ${GIT_DIR}/modules/${SUBMODULE_NAME}/config
core.worktree ${SUBMODULE_PATH}
    echo "gitdir: ${GIT_DIR}/modules/${SUBMODULE_NAME}" >
${SUBMODULE_PATH}/.git

The last 2 steps are done via

    void connect_work_tree_and_git_dir(const char *work_tree, const
char *git_dir);

in submodule.{c,h}

However we'd need to make sure the first step is performed correctly. (and make
damn sure we don't loose that git dir), so I think rename(2) does the
correct thing
for directories, except when these two locations are on a different mount point.

I think I'll just write this functionality in C and optionally expose
it via the submodule--helper,
such that the user facing git-submodule.sh only has to call that helper.