Linux – git: show which files have changed between kernel 3.18.6 and 3.18.7

All of the commits and the files that they reference would be stored as objects in the objects directory. Git creates those as read-only, so they should all still be present.

To recover, I'd advise creating a new, empty repository and copying the contents of your broken repository's objects directory into that of the new one. That should get you to a point where git will at least recognize that it's a repository and it will have all of your objects. Working with a copy will also help to avoid causing even more damage while trying to fix things.

Shell commands to create the temporary repository and copy over the objects:

git init /tmp/recovery
cd /tmp/recovery
cp -r /path/to/broken/repo/.git/objects .git

Once that is done, you could use git fsck to get a list of objects that aren't referenced by anything. This should include all of the branch heads, but it would also include any commits that were made obsolete by git commit --amend or by rebasing.

Since you still have the logs directory that is likely to be an even bigger help. There should be a logs/refs/heads/<branch> file for each branch that you had. The second column of the last line will contain the ID of the commit which was at the head of that branch when the deletion was done. There should also be logs/HEAD with the same information for where the HEAD was, but unless you'd been working with a detached HEAD it's probably better to just recover the branches and then checkout a branch normally.

For each branch that you want to restore you can run:

git branch <name> <commit_id>

Once you've restored the branches you can copy over the config file, and you should be fairly close to where you were as of the latest commit.

This can be done by using a custom dlagent. I do not really understand Arch packaging or how the dlagents work, so I only have a hack answer, but it gets the job done.

The idea is to modify the PKGBUILD to use a custom download agent. I modified the source

"${pkgname%-git}::git+http://github.com/Itseez/opencv.git"

into

"${pkgname%-git}::mygit://opencv.git"

and then defined a new dlagent called mygit which does a shallow clone. I did this by adding to the DLAGENTS array in /etc/makepkg.conf the following dlagent:

'mygit::/usr/bin/git clone --depth 1 http://github.com/Itseez/opencv.git'

My guess is you could probably define this download agent somewhere else, but I do not know how. Also notice that the repository that is being cloned is hard coded into the command. Again, this can probably be avoided. Finally, the download location is not what the PKGBUILD expects. To work around this, I simply move the repository after downloading it. I do this by adding

mv "${srcdir}/../mygit:/opencv.git" "${srcdir}/../${pkgname%-git}"

at the beginning of the pkgver function.

I think the cleaner solution would be to figure out what the git+http dlagent is doing and redfine that temporarily. This should avoid all the hack aspects of the solution.