Re: [PATCH 1/2] rm: changes in the '.gitmodules' are staged after using '--cached'

From: Junio C Hamano <gitster@pobox.com>
Date: Feb 18, 2021, 22:03 UTC
Message-ID: <xmqqblchdoej.fsf@gitster.g>
In-Reply-To: <20210218184931.83613-2-periperidip@gmail.com>

Shourya Shukla <periperidip@gmail.com> writes:

Show 21 quoted lines

> +		if (list.entry[i].is_submodule) {
> +			/*
> +			 * Then, unless we used "--cached", remove the filenames from
> +			 * the workspace. If we fail to remove the first one, we
> +			 * abort the "git rm" (but once we've successfully removed
> +			 * any file at all, we'll go ahead and commit to it all:
> +			 * by then we've already committed ourselves and can't fail
> +			 * in the middle)
> +			 */
> +			if (!index_only) {
> +				struct strbuf buf = STRBUF_INIT;
>  				strbuf_reset(&buf);
>  				strbuf_addstr(&buf, path);
>  				if (remove_dir_recursively(&buf, 0))
>  					die(_("could not remove '%s'"), path);
>  
>  				removed = 1;
> -				if (!remove_path_from_gitmodules(path))
> -					gitmodules_modified = 1;
> -				continue;
> +				strbuf_release(&buf);

Since we won't come to this block when doing index_only, we are allowed to touch the working tree contents and files. We indeed do "rm -rf" of the submodule working tree and touch .gitmodules file that is in the working tree.

>  			}
> +			if (!remove_path_from_gitmodules(path))
> +				gitmodules_modified = 1;
> +			continue;

But this looks wrong. It might be OK to remove from the .gitmodules stored in the index, but I fail to see why it is justified to touch the working tree file when "--cached" is given.

Show 12 quoted lines

> +		}
> +		if (!index_only) {
>  			if (!remove_path(path)) {
>  				removed = 1;
>  				continue;
> @@ -396,11 +398,15 @@ int cmd_rm(int argc, const char **argv, const char *prefix)
>  			if (!removed)
>  				die_errno("git rm: '%s'", path);
>  		}
> -		strbuf_release(&buf);
> -		if (gitmodules_modified)
> -			stage_updated_gitmodules(&the_index);

Assuming that it is somehow justifiable that removing the entry from the .gitmodules in the index (again, I do not think it is justifiable to remove from the working tree file), we no longer can use stage_updated_gitmodules() helper as-is.

I think you'd need to

 - Add a variant of remove_path_from_gitmodules() that can remove
   the given submodule from the .gitmodules in the index entry
   without touching the working tree.  The change could be to update
   the function to take an extra "index_only" parameter and a
   pointer to an index_state instance, and

   (1) if !index_only, then edit the .gitmodules file in the working
       tree to remove the entry for path;

   (2) in both !index_only and index_only cases, read .gitmodules
       file from the index, edit to remove the entry for path, and
       add the result to the index.

   and return 0 for success (e.g. if path is not a submoudle or no
   entry for it is found in .gitmodules, it may return -1).

 - Since the previous point will maintain the correct contents in
   the index in all cases, get rid of gitmodules_modified and calls
   to stage_updated_gitmodules().  The call to write_locked_index()
   at the end will take care of the actual writing out of the index.

if we want to teach "rm --cached" to update only the index, and "rm" to update both the index and the working tree, of ".gitmodules".

Having said that, I still do not think it is a good direction to go to teach low level "rm", "mv" etc. to know about ".gitmodules" (yes, yes, I know that some changes that I consider to be mistakes have already happened---that does not mean we cannot correct our course and it does not mean it is OK to make things even worse). Such a "how does a submodule get managed" policy decision belongs to the "git submodule" subcommand, I would think.

Thanks.

Show 10 quoted lines

> +	/*
> +	 * Remove the entry of the submodule from the ".gitmodules" irrespective
> +	 * whether "--cached" was passed or not.
> +	 */
> +	if (gitmodules_modified)
> +		stage_updated_gitmodules(&the_index);
> +
>  	if (write_locked_index(&the_index, &lock_file,
>  			       COMMIT_LOCK | SKIP_IF_UNCHANGED))
>  		die(_("Unable to write new index file"));