Merge conflicts in .gitattributes can cause trouble

Hi,

If there is a conflict in the .gitattributes during a merge then it looks like as if the attributes are not applied (which kind of makes sense as Git would not know what to do). As a result Git can treat e.g. binary files as text and they can end up with changed line endings in the working tree. After resolving the conflict in .gitattributes all files would be marked as binary, again, and the user can easily commit the wrongly changed line endings.

Consider this script on Windows:

$ git init . $ touch first.commit $ git add . $ git commit -m "first commit"

$ git checkout -b branch $ printf "*.bin binary\n" >> .gitattributes $ git add . $ git commit -m "tracking *.bin files"

$ git checkout master
$ printf "binary\ndata\n" > file.dat # <-- Unix line ending!
$ printf "*.dat binary\n" >> .gitattributes # <-- Tell Git to keep Unix line ending!
$ git add .
$ git commit -m "tracking *.dat files"
$ git cat-file -p :file.dat | od -c
0000000   b   i   n   a   r   y  \n   d   a   t   a  \n 
                                ^^^^                ^^^^  <-- Correct!
$ git checkout branch
$ git merge master # <-- Causes merge conflict!
$ printf "*.bin binary\n*.dat binary\n" > .gitattributes # <-- Fix merge conflict!
$ git add .
$ git commit -m "merged"
$ git cat-file -p :file.dat | od -c
0000000   b   i   n   a   r   y  \r  \n   d   a   t   a  \r  \n
                                ^^^^^^^^                ^^^^^^^^  <-- Wrong!

Possible solutions:

1. We could print an appropriate warning if we detect a merge conflict 
   in .gitattributes

2. We could disable all line ending conversions in case of a merge conflict
   (I am not exactly sure about all the implications, though)

3. We could salvage what we could of the .gitattributes file, 
   perhaps by using the version from HEAD (or more likely, the ours stage of
   the index) -- suggested by Peff on the related GitHub issue mentioned below

Thoughts?

Thanks, Lars

PS: I noticed that behavior while working with Git LFS and started a discussion
about it here: https://github.com/github/git-lfs/issues/1544 

On Tue, Oct 4, 2016 at 5:19 PM, Lars Schneider <larsxschneider@gmail.com> wrote:

This is good regardless, to encourage people to resolve conflicts in .gitattributes first. A good place for this warning may be "git status"?

We already have code to fall back to index version in some cases, adding "fall back on merge conflicts" (and probably updating the index lookup code too because it looks for stage 0 now) sounds reasonable (especially with the warning in #1).

BTW whoever fixes this probably should do the same for .gitignore files.

Hi Lars,

On Tue, 4 Oct 2016, Lars Schneider wrote:

I tried to replicate this behavior, to the point where I wrote a patch that demonstrates the breakage so I could single-step in a debugger to find out where things go wrong, and fix them.

Alas, I found out that the .gitattributes are read *before* any merge conflict arises in the case you demonstrated. Which kind of makes sense, because the gitattributes decide over which merge driver to use, among other things.

So in your example:

At this point, the .gitattributes list only .bin files as binary. That is the revision of the .gitattributes used by this command:

And as a consequence, the .gitattributes do not tell Git that it should handle .dat files as binary. Which means that...

... this is actually expected! Why? Because the .gitattributes that were in effect when the user asked to perform a merge said so.

If you adjust .gitattributes *before* merging `master`, it works as you would expect: the line endings are not changed.

The reason to do it this way: we want to respect the .gitattributes as per the current worktree. We go even so far that we respect uncommitted changes to said file...

I would vote for:

4. We keep letting Git read in the *current* version of .gitattributes
   *before* the merge, and apply those attributes while performing the
   merge.

Ciao, Dscho

Johannes Schindelin <Johannes.Schindelin@gmx.de> writes:

Even though this needs a major surgery to the way the attr subsystem reads from these files, I think it is conceptually the cleanest.

Hi Junio,

On Mon, 17 Oct 2016, Junio C Hamano wrote:

To the contrary. As far as I can see, when calling `git merge`, Git currently *does* read .gitattributes from the file, and if that fails, falls back to reading that file from the index.

In other words, option 4. is the current behavior no change required.

Ciao, Dscho

Johannes Schindelin <Johannes.Schindelin@gmx.de> writes:

Hmph.

Assuming that the merge always goes in the index order, I think you are right. When we need to merge path/to/dir/.gitattributes, we would need to know all the .gitattributes files that may affect that path, so .gitattributes, path/.gitattributes, path/to/.gitattributes and the file being merged are all read into core before anything happens and these are kept in attr_stack while merging anything underneath path/to/dir/ hierarchy without being re-read from the filesystem. The original contents of path/to/dir/.gitattributes cached in the attr_stack will be discarded when we start to merge things outside path/to/dir (e.g. merging path/to/another/file), but at that point the contents of path/to/dir/.gitattributes no longer matters to the path being merged, so unless the merge somehow jumps around it is OK.

It may be a fragile assumption in the longer term that the merge always goes in the index order, but I think the assumption holds in the current codebase, and the update planned immediately in the future.