Re: Tracking a merge bug

From: Elijah Newren <newren@gmail.com>
Date: Oct 26, 2020, 15:50 UTC
Message-ID: <CABPp-BFju3vJYMRH8fXBcHu1sDkthWsdO9rJzSNKB7E3pfGiHQ@mail.gmail.com>
In-Reply-To: <120922f1-67a9-9ae2-2e9c-56e20935e0f9@gmail.com>

On Mon, Oct 26, 2020 at 8:11 AM Ephrim Khong <dr.khong@gmail.com> wrote:

Show 16 quoted lines

>
> Dear All,
>
> I am trying to find the root cause for what I believe might be a strange
> bug in git merge.  I have a feature branch A which branched off master
> not too long ago, and want to bring it up to date with master:
>
>     git checkout A
>     git merge master
>
> which yields
>
>     Removing somefile
>     Removing anotherfile
>     error: add_cacheinfo failed to refresh for path 'c/d/e.sh'; merge
> aborting.

"add_cacheinfo failed to refresh"? Wow, that's a new one. Some years back we had a "add_cacheinfo failed for path" corresponding to the other error site within that function, but we fixed that one up long ago. I've never seen anything hit the refresh failure.

The fact that you're seeing this error message means you're using git

>= 2.18.0.  Could you verify the exact version?

Also, I suspect "somefile" and "anotherfile" are placeholders. Any chance we could see the real names?

> the offending file, c/d/e.sh, does not exist in my feature branch but
> was added to master since branching off. After aborting, the working
> directory is in an inconsistent state and c/d/e.sh exists with the
> correct content.

Is c/d a file or submodule or anything in the feature branch? Do you have any untracked files at the beginning of the merge named c/d or c/d/e.sh? I'm trying to guess at what might be weird about the paths here.

Show 6 quoted lines

> Below is a stacktrace - the merger handles the file as a rename
> (apparently there is a similar / identical file 'c/f/g.sh' that is
> renamed to 'c/d/e.sh'), but that fails because the file has MODE_CHANGED
> set. (Which appears strange - at the time where the merge is aborting,
> the file apparently was already written to the working directory. Is it
> renaming two different files to the same target file?).

What's the output of
  git ls-tree -r master | grep c/d/e.sh
  git ls-tree -r feature | grep c/d/f.sh
  git ls-tree -r $(git merge-base master feature) | grep c/d/f.sh
?

And, before the merge, what is the output of
  ls -l c/d/e.sh c/d/f.sh
?

> Any hint is appreciated, especially where to look: Is the root cause
> more likely to be at the filesystem level (the stat returns something
> off), or in the merge logic? What else could be wrong here?

Not sure; I've never hit this before. I'm curious if you've got a weird mode recorded for one of the files in your tree or something. We might have to get someone who understands the index data structure and read-cache.c to take a look.

Show 25 quoted lines

> The stacktrace looks roughly as follows:
>
> -> read-cache.c, ie_modified(): ie_match_stat returned 63, which is
>                 MTIME_CHANGED   | CTIME_CHANGED | OWNER_CHANGED |
>                 MODE_CHANGED | INODE_CHANGED | DATA_CHANGED
>         and is_modified() returns 63 because MODE_CHANGED is set.
>
> -> read-cache.c, refresh_cache_ent(): at the call to ie_modified
>
> -> read-cache.c, refresh_cache_entry()
>
> -> merge-recursive.c, add_cacheinfo(), is in the refresh-path (i.e.
> make_cache_entry() worked, but refresh_cache_entry() will fail)
>
> -> merge-recursive.c, update_file_flags(), after the update_index: label
>
> -> merge-recursive.c, update_file()
>
> -> merge-recursive.c, handle_content_merge() is in the very last
> update_file() call, close to the end of the function
>
> -> merge-recursive.c, handle_rename_normal()
>
> -> merge-recursive.c, process_entry()
>         is in the RENAME_NORMAL / RENAME_ONE_FILE_TO_ONE block

RENAME_NORMAL / RENAME_ONE_FILE_TO_ONE means this is NOT a case of renaming two files to one. From just this information, it could be that both sides renamed c/d/f.sh -> c/d/e.sh -- but you ruled that out when you said c/d/e.sh didn't exist in your branch. So, that basically leaves us with a normal rename; i.e. c/d/f.sh existed in the merge base and in your feature branch, but it was renamed to c/d/e.sh in master.

Any chance this repository is available for others to access to try to
reproduce the problem?  If the repository has sensitive contents, are
the files c/d/f.sh and c/d/e.sh sensitive?  If the repository overall
is sensitive but those two files aren't, we might be able to get a
reproduction by cloning the repository and running
   git filter-repo --path c/d/e.sh --path c/d/f.sh
to filter the repository down to just these two files (after first
installing git-filter-repo; see
https://github.com/newren/git-filter-repo).