On Tue, Aug 8, 2023 at 3:15 PM <tboegi@xxxxxx> wrote:
> The following sequence leads to loss of work:
> git init
> mkdir README
> touch README/README
> git add .
> git commit -m "Init project"
> echo "Test" > README/README
> mv README/README README2
> rmdir README
> mv README2 README
> git stash
> git stash pop
>
> The problem is, that `git stash` needs to create the directory README/
> and to be able to do this, the file README needs to be removed.
> And this is, where the work was lost.
> There are different possibilities preventing this loss of work:
> a)
> `git stash` does refuse the removel of the untracked file,
s/removel/removal/
> when a directory with the same name needs to be created
s/$/./
> There is a small problem here:
> In the ideal world, the stash would do nothing at all,
> and not do anything but complain.
> The current code makes this hard to achieve
s/$/./
> An other solution could be to do as much stash work as possible,
s/An other/Another/
> but stop when the file/directory conflict is detected.
> This would create some inconsistent state.
>
> b) Create the directory as needed, but rename the file before doing that.
> This would let the `git stash` proceed as usual and create a "new" file,
> which may be surprising for some worlflows.
s/worlflows/workflows/
> This change goes for b), as it seems the most intuitive solution for
> Git users.
>
> Introdue a new function rename_to_untracked_or_warn() and use it
s/Introdue/Introduce/
> in create_directories() in entry.c
>
> Reported-by: Till Friebe <friebetill@xxxxxxxxx>
> Signed-off-by: Torsten Bögershausen <tboegi@xxxxxx>
> ---
> diff --git a/entry.c b/entry.c
> @@ -15,6 +15,28 @@
> +static int rename_to_untracked_or_warn(const char *file)
> +{
> + const size_t file_name_len = strlen(file);
> + const static char *dot_untracked = ".untracked";
> + const size_t dot_un_len = strlen(dot_untracked);
> + struct strbuf sb;
> + int ret;
> +
> + strbuf_init(&sb, file_name_len + dot_un_len);
> + strbuf_add(&sb, file, file_name_len);
> + strbuf_add(&sb, dot_untracked, dot_un_len);
> + ret = rename(file, sb.buf);
This could probably all be simplified to:
char *to = xstrfmt("%s.untracked", file);
ret = rename(...);
...
free(to);
If there is already a file named "foo.untracked", then this will
overwrite it, thus potentially losing work, right? I wonder if it
makes sense to be a bit more careful.
> + if (ret) {
> + int saved_errno = errno;
> + warning_errno(_("unable rename '%s' into '%s'"), file, sb.buf);
> + errno = saved_errno;
> + }
> + strbuf_release(&sb);
> + return ret;
> +}
Do we want to give the user some warning/notification that their file,
as a safety precaution, got renamed to "foo.untracked"?
> diff --git a/t/t3903-stash.sh b/t/t3903-stash.sh
> @@ -1512,4 +1512,27 @@ test_expect_success 'restore untracked files even when we hit conflicts' '
> +test_expect_success 'stash mkdir README needed - README.untracked created' '
> + git init mkdir_needed_file_untracked &&
> + (
> + cd mkdir_needed_file_untracked &&
> + mkdir README &&
> + touch README/README &&
s/touch/>/
> + git add . &&
> + git commit -m "Add README/README" &&
> + echo Version2 > README/README &&
s/> R/>R/
> + mv README/README README2 &&
> + rmdir README &&
> + mv README2 README &&
> + git stash &&
> + test_path_is_file README.untracked &&
> + echo Version2 >expect &&
> + test_cmp expect README.untracked &&
> + rm expect &&
> + git stash pop &&
> + test_path_is_file README.untracked &&
> + echo Version2 >expect &&
> + test_cmp expect README.untracked
> + )
> +'
