Re: [GSoC][PATCH 06/13] submodule: port submodule subcommand 'deinit' from shell to C

[Stefan Beller <sbeller@google.com>]

On Mon, Jul 31, 2017 at 1:56 PM, Prathamesh Chavan <pc44800@gmail.com> wrote:
> The same mechanism is used even for porting this submodule
> subcommand, as used in the ported subcommands till now.
> The function cmd_deinit in split up after porting into three
> functions: module_deinit(), for_each_submodule_list() and
> deinit_submodule().
>
> Mentored-by: Christian Couder <christian.couder@gmail.com>
> Mentored-by: Stefan Beller <sbeller@google.com>
> Signed-off-by: Prathamesh Chavan <pc44800@gmail.com>
> ---
> In this new version, the following changes have been made:
> * In the function deinit_submodule, since the test is_git_directory()
>   adds an additional condition, instead is_directory() is used to check
>   if "sm_path/.git" is a directory.

Thanks for writing these patches.
I wonder if (some of) these notes are best put into the code
as a comment such as

    /* NEEDSWORK: convert to is_submodule_active */

such that people reading this code later realize that checking
for a directory may not be the "correct" thing, but a thing which
was easy to express using shell.

> +struct deinit_cb {
> +       const char *prefix;
> +       unsigned int quiet: 1;
> +       unsigned int force: 1;
> +       unsigned int all: 1;

The value 'all' seems to be unused, i.e. we assign it but never read it?

> +};
> +#define DEINIT_CB_INIT { NULL, 0, 0, 0 }
> +
> +static void deinit_submodule(const struct cache_entry *list_item,
> +                            void *cb_data)
> +{
> +       struct deinit_cb *info = cb_data;
> +       const struct submodule *sub;
> +       char *displaypath = NULL;
> +       struct child_process cp_config = CHILD_PROCESS_INIT;
> +       struct strbuf sb_config = STRBUF_INIT;
> +       char *sub_git_dir = xstrfmt("%s/.git", list_item->name);
> +       mode_t mode = 0777;
> +
> +       sub = submodule_from_path(null_sha1, list_item->name);
> +
> +       if (!sub || !sub->name)
> +               goto cleanup;
> +
> +       displaypath = get_submodule_displaypath(list_item->name, info->prefix);
> +
> +       /* remove the submodule work tree (unless the user already did it) */
> +       if (is_directory(list_item->name)) {
> +               struct stat st;
> +               /* protect submodules containing a .git directory */

Here may a good place to put:
  /* NEEDSWORK: automatically call absorbgitdirs before warning/die. */
(It was not in the shell version, so feel free to ignore)

> +               if (!info->force) {
> +                       struct child_process cp_rm = CHILD_PROCESS_INIT;
> +                       cp_rm.git_cmd = 1;
> +                       argv_array_pushl(&cp_rm.args, "rm", "-qn",
> +                                        list_item->name, NULL);

A bug that exists in the shell version as well as here:
What if the submodule has the name '--cached', which happens
to be a valid argument for git-rm?

The call to git-rm would die claiming that the <file> is missing,
as the file name was miss-interpreted as another flag.

To solve this problem we would insert a '--' after the options,
before the file name to state that the last argument is a <file>.

Not sure if we want to fix the bug while we're here or if we rather
want to add

    /* NEEDSWORK: add '--' to confirm <file> argument */

> +       if (mkdir(list_item->name, mode))
> +               die(_("could not create empty submodule directory %s"),
> +                     displaypath);

In the shell version we used the shell mkdir, which on failure
would print the error to stderr on its own. The added "|| say"
is just to clarify the bigger picture.

mkdir in C doesn't print the actual error (e.g.
no diskspace, permissions) so we have to do it.
Use 'die_errno' instead as then the reason will be given as well.

> +       cp_config.git_cmd = 1;
> +       argv_array_pushl(&cp_config.args, "config", "--get-regexp", NULL);
> +       argv_array_pushf(&cp_config.args, "submodule.%s\\.", sub->name);

It would be nice if the commit message could give hints as why a literal
translation was chosen here and not e.g. one of the builtin config
calls, as that would omit spawning a process.