Re: Possible bug with git restore

From: Jeff King <peff@peff.net>
Date: Aug 20, 2020, 18:27 UTC
Message-ID: <20200820182720.GA2537643@coredump.intra.peff.net>
In-Reply-To: <c3f0d51a-d0e3-ed0a-c9ed-da092704da5c@web.de>

On Thu, Aug 20, 2020 at 07:48:48PM +0200, René Scharfe wrote:

Show 7 quoted lines

> >   - shouldn't that wildcard pathspec match those files? I've confirmed
> >     that the glob characters make it into Git's pathspec machinery, and
> >     since it doesn't have slashes, I think we'd match a basename (and
> >     certainly "git ls-files *test_file.*" does what I expect).
> 
> No, because restore doesn't interpret pathspecs recursively.  I don't
> know why that causes files to disappear, though.  But here's a fix.

I think it's because of this comment from bc96cc87dbb:

  When pathspec.recursive == 0, the behavior depends on match functions:
  non-recursive for tree_entry_interesting() and recursive for
  match_pathspec{,_depth}

So when we read the tree, we don't match recursively, and those entries don't appear. But then we correlate that with the index:

          /*
           * Make sure all pathspecs participated in locating the paths
           * to be checked out.
           */
          for (pos = 0; pos < active_nr; pos++)
                  if (opts->overlay_mode)
                          mark_ce_for_checkout_overlay(active_cache[pos],
                                                       ps_matched,
                                                       opts);
                  else
                          mark_ce_for_checkout_no_overlay(active_cache[pos],
                                                          ps_matched,
                                                          opts);

And in no-overlay mode (the default for restore), we do:

  
  static void mark_ce_for_checkout_no_overlay(struct cache_entry *ce,
                                              char *ps_matched,
                                              const struct checkout_opts *opts)
  {
          ce->ce_flags &= ~CE_MATCHED;
          if (!opts->ignore_skipworktree && ce_skip_worktree(ce))
                  return;
          if (ce_path_match(&the_index, ce, &opts->pathspec, ps_matched)) {
                  ce->ce_flags |= CE_MATCHED;
                  if (opts->source_tree && !(ce->ce_flags & CE_UPDATE))
                          /*
                           * In overlay mode, but the path is not in
                           * tree-ish, which means we should remove it
                           * from the index and the working tree.
                           */
                          ce->ce_flags |= CE_REMOVE | CE_WT_REMOVE;
          }
  }

And that ce_path_match() _does_ treat the pathspec recursively. So we say "yes, it matches in the index but wasn't in the tree, and therefore we must delete it".

So the fundamental issue is treating the pathspec in two different ways, and then correlating the results. We need to either do a recursive match for the tree match (as your patch does), or do non-recursive for this index match (which I don't think is trivial, because of the way the recursive flag works).

> No sign-off because I don't understand why pathspec recursiveness is a
> thing that can be turned off -- I'd expect pathspec syntax to be
> consistent for all commands.  So there might be a good reason why it was
> not enabled for restore (and switch and checkout).

I think it was originally done this way for compatibility of some commands as we unified the pathspec code. But I'm having trouble digging up the exact details.

However, it seems particularly egregious in checkout/restore, because we may also be using the index as a source, in which case the pathspecs _would_ be recursive by default. E.g., in the test repo we've been discussing:

  [make the index and working tree differ]
  $ git reset HEAD^
  Unstaged changes after reset:
  M	incl/test_file.hpp
  M	src/test_file.cpp

  [restore using a wildcard, but out of the index rather than a tree]
  $ git restore -- '*.hpp'

  [and check that we did indeed match]
  $ git status
  On branch master
  Changes not staged for commit:
	modified:   src/test_file.cpp

So I think this inconsistency in pathspec matching between trees and the index has probably existed in git-checkout for ages (and I guess people don't do wildcards with trees often enough for anybody to have noticed). But it didn't cause the index-deletion problem, because that only appeared more recently with the --no-overlay mode. That's the default for restore, but you can trigger the problem with checkout, too:

  $ git reset --hard
  $ git checkout --no-overlay HEAD^ '*.hpp'
  Updated 0 paths from 2668463
  $ git status
  On branch master
  Changes to be committed:
	deleted:    incl/test_file.hpp

-Peff