ls-files -o now traverses nested repo when given multiple pathspecs

89a1f4aaf7 (dir: if our pathspec might match files under a dir, recurse into it, 2019-09-17) introduced a change in behavior in terms of traversing untracked nested repositories. Say we have a repository that contains a single untracked repository with untracked content:

    $ git init && git init a && touch a/x

Calling ls-files with the nested repository as the sole pathspec does not recurse into that repository:

    $ git ls-files --other a
    a/

However, as of 89a1f4aaf7, adding an additional pathspec results in the nested repository being traversed:

    $ git ls-files --other a foo
    a/
    a/x

Reading 89a1f4aaf7 and skimming the patch series and related thread [*], I haven't found anything that makes me think this change in behavior was intentional.

[*]: https://lore.kernel.org/git/[email protected]/

Hi Kyle,

Thanks for the clear report and pointing out relevant commit(s) and discussion. Apologies in advance if I come off a bit ranty; it's directed to dir.c and its API and not you. It seems to be quite the mess...

Oh man, I guess I shouldn't be surprised by another area of the code that depends on dir.c but lacks any meaningful tests of its behavior, violated the existing contract[1], and depends on the side effects of other bugs -- bugs which don't cover all cases and thus causes it to get different behavior depending on things that otherwise shouldn't matter. Behold, *before* my changes to dir.c:

$ git --version
2.23.0
$ find . -type f | grep -v .git
./empty
./untracked_repo/empty
./untracked_dir/empty
./world
$ git ls-files
world
$ git ls-files -o
empty
untracked_dir/empty
untracked_repo/

So, as you say, it wouldn't traverse into untracked_repo/. Now we start adding pathspecs:

$ git ls-files -o untracked_repo
untracked_repo/

As you mentioned, it won't traverse into it even when specified...

$ git ls-files -o untracked_repo/
untracked_repo/empty

...except that it does traverse into this directory if the user tab completes the name or otherwise manually adds a trailing slash. Weird, let's try multiple pathspecs:

$ git ls-files -o untracked_dir untracked_repo
untracked_dir/empty
untracked_repo/

$ git ls-files -o untracked_dir untracked_repo/
untracked_dir/empty
untracked_repo/

So it will traverse into the untracked_repo when specified as 'untracked_repo/' but not if there are more than one pathspec given?!?
And it traverses into an untracked directory regardless of the
trailing slash? <sarcasm>What a paragon of consistency...</sarcasm>

At least my changes in git-2.24.0 made the behavior consistent; it'll always traverse into a directory that matches a given pathspec. As for whether that's desirable or not when the pathspec is a submodule, I'm not certain. My fixes to dir.c stalled out for over a year and a half despite a few reports that they had fixed issues people were continuing to report in the wild because the whole traversal logic was such a mess and there were so many dependencies on existing behavior built up with that mess that it was really hard to determine what "correct" behavior was and others seemed to be unwilling to wade through the muck to figure out where sanity might lie so they could help give me pointers or opinions on what "correct" was[2].

But here are some possibilities that at least sound sane:

A) ls-files -o should traverse into untracked submodules. This case is easy; the code already does that.

B) ls-files -o should NOT traverse into untracked submodules AND should not even report them. If so, fix looks like this:

index f069a028ce..f144d44d8b 100644
@@ -301,6 +301,9 @@ static void show_files(struct repository *repo,

struct dir_struct *dir)

        int i;
        struct strbuf fullname = STRBUF_INIT;

+ if (!recurse_submodules)
+ dir->flags |= DIR_SKIP_NESTED_GIT;
+

        /* For cached/deleted files we don't need to even do the readdir */
        if (show_others || show_killed) {
                if (!show_others)

C) ls-files -o should NOT traverse into untracked submodules, but should at least report their directory name. If so, the fix is probably to move the DIR_NO_GITLINKS if-block within dir.c:treat_directory() to put it right next to the DIR_SKIP_NESTED_GIT if-block (and maybe even partially combine the two) so that it comes before some of the other code in that function. Might have to be careful about checking for the presence of trailing slashes.

It seems like it should be clear which one of these is "correct", but I seem to be short a few brain cycles right now...either that, or maybe my very rare use of submodules and nested repositories means I just don't have enough context to answer. Maybe it's obvious to someone else...

Elijah

[1] https://lore.kernel.org/git/[email protected]/
[2] https://lore.kernel.org/git/[email protected]/

I think this probably is the most sensible.

The top-level directory of a working tree of a repository other than the current one may exist in the working tree. It is very tempting to declare that, unless we know it is a submodule that has the current repository as its superproject, we should just treat it as a normal subdirectory without *any* files tracked by the current repository, which would mean that we pretend that the ".git/" in that subdirectory is not any special---but that would obviously make things quite messy (e.g. our "ls-files -o" would descend into the other project's working tree and even in its .git/ directory), so we need to special case a directory that has ".git/" in it, whether it is a submodule for our current repository or not.

Thanks for working on this. I agree that dir.c traversal has become messier and messier, especially with its interaction with submodules.

Hi Elijah,

Thanks for the detailed and helpful reply.

Elijah Newren <[email protected]> writes:

[...]

Ah yes, I recall encountering what I think is the same underlying issue when working on a previous series [0,1]. In the context of 'git add untracked_repo/', there's been some discussion related to this trailing slash discrepancy at

https://lore.kernel.org/git/[email protected]/T/#u

Eh, indeed.

I might be getting mixed up, but the changes in 2.24.0 did introduce some inconsistent behavior (in the no trailing slash case) with respect to giving a single pathspec and giving multiple pathspecs, no? Using your example:

    $ git --version
    git version 2.24.0
    $ git ls-files -o untracked_repo
    untracked_repo/
    $ git ls-files -o untracked_repo empty
    empty
    untracked_repo/
    untracked_repo/empty

Hmm, but as shown in the last example, ls-files -o doesn't traverse into untracked submodules for the single pathspec case.

This behavior---which matches the no-slash behavior when no patchspec or a single pathspec is given (on both v2.24.0 and previous version) as well as when multiple pathspecs are given (before v2.24.0)---is the one I prefer. My biased reason for this preference is that in the DataLad project we identify untracked nested repositories based on ls-files -o <untracked directory>... reporting only the directory name for repositories. (Looking into one of our tests that fails with Git v2.24.0 is how I ran into the reported change in behavior [2].)

That some external project relies on unintended ls-files output of course doesn't mean that Git should keep reporting things that way, but it does mean that I _hope_ that not traversing into untracked repositories is the intended behavior and that traversing (either because a slash is appended or as of 89a1f4aaf7 because multiple pathspecs are given) is not intended :>

[0]: https://lore.kernel.org/git/[email protected]
[1]: https://lore.kernel.org/git/[email protected]/
[2]: https://github.com/datalad/datalad/issues/3890#issuecomment-561722194

It looks like the "multiple pathspecs trigger traversal" change isn't limited to nested repositories. It can also be observed with --directory and plain untracked directories. Assume the tree layout from your example again. With a single pathspec (and no slash), 'ls-files -o --directory' will not expand the untracked directory's contents:

    $ git ls-files -o --directory untracked_dir
    untracked_dir/

But, as of 89a1f4aaf7, tacking on an additional pathspec will cause ls-files to traverse into the untracked directory:

    $ git ls-files -o --directory untracked_dir empty
    empty
    untracked_dir/
    untracked_dir/empty

In contrast, on 89a1f4aaf7^ the same command shows

    $ git ls-files -o --directory untracked_dir empty
    empty
    untracked_dir/

Hi Kyle,

Yeah, I spotted that too. You left out a case, a single pathspec with the trailing slash:

   git ls-files -o --directory untracked_dir/

That will traverse into the directory before or after my changes. I also spotted a few other bugs, e.g. try out 'git ls-files -o .git/' (with either git-2.23 or git-2.24). Whoops. We do correctly avoid traversing into the .git directory if multiple pathspecs are provided. Anyway, this whole area seems to be a bug factory. Every time I think I'm close to having some patches to send to the list to fix up the issues I've found, I find the fix isn't where I thought it was and/or find yet another bug. Quite aggravating.

I'm thinking of just sending the patches I have, since they fix up all the issues we've discussed so far (including the .git/ case I just mentioned), and ignoring the 2-3 other bugs I found that are still broken other than providing testcases documenting their breakage.

If you want to take an early look, I've got some patches up at
https://github.com/git/git/pull/676. I plan to write a proper cover letter and submit to the list on Monday.

I can confirm that your patches resolve the cases reported here. Thank you for working on this!