Hi,

On Thu, 29 Jun 2023, Jeff King wrote:

On Tue, Jun 27, 2023 at 05:02:30PM +0100, Tiago d'Almeida wrote:

quoted

Attached to this email follow the `git bugreport` and global `config`
files, and the git_bug repo.

Thanks for providing your config; it was very important to reproducing.
The bug comes from the combination of "core.splitIndex" and
"index.skipHash" (the latter is triggered in your config by
"feature.manyFiles").

Here's a quick reproduction:

  git init repo
  cd repo
  touch file
  git -c core.splitIndex=true -c index.skipHash=true add file

I ran into this issue while debugging the `commit -am` issue I worked on
in https://github.com/gitgitgadget/git/pull/1554.

The reason is that `write_shared_index()` calls `do_write_index()` without
any additional flags (see
https://github.com/git/git/blob/v2.41.0/read-cache.c#L3300) and
`do_write_index()` heeds the `index.skipHash` setting always (see
https://github.com/git/git/blob/v2.41.0/read-cache.c#L2900).

I briefly experimented with this diff, which is ugly and should not be
used as is, but it seemed to fix the issue for me:

-- snip --

diff --git a/read-cache.c b/read-cache.c
index ee6bcf40351..92a4aa2f25a 100644
--- a/read-cache.c
+++ b/read-cache.c

@@ -3292,14 +3294,17 @@ static int write_shared_index(struct index_state *istate,
 			      struct tempfile **temp, unsigned flags)
 {
 	struct split_index *si = istate->split_index;
-	int ret, was_full = !istate->sparse_index;
+	int ret, was_full = !istate->sparse_index, saved_skip_hash;

 	move_cache_to_base_index(istate);
 	convert_to_sparse(istate, 0);

 	trace2_region_enter_printf("index", "shared/do_write_index",
 				   the_repository, "%s", get_tempfile_path(*temp));
+	saved_skip_hash = si->base->repo->settings.index_skip_hash;
+	si->base->repo->settings.index_skip_hash = 0;
 	ret = do_write_index(si->base, *temp, WRITE_NO_EXTENSION, flags);
+	si->base->repo->settings.index_skip_hash = saved_skip_hash;
 	trace2_region_leave_printf("index", "shared/do_write_index",
 				   the_repository, "%s", get_tempfile_path(*temp));

-- snap --

The reason why this is needed is that the shared index _must_ have an
identifer that the split index can use, and that's that index hash.
Skipping it breaks that pattern.

Probably a much better idea than above-mentioned diff would be to add a
new flag as a sibling to `COMMIT_LOCK` (i.e. here:
https://github.com/git/git/blob/v2.41.0/cache.h#L346-L348) and use that
only in `write_shared_index()` to force the index hash to be computed and
written.

I won't have time to work on this, though.

Ciao,
Johannes

That should add "file" to the index but doesn't. Removing either the
splitIndex option or the skipHash option makes it work. I didn't dig
further than that.

Adding the author of skipHash to the cc.

-Peff