0

Yesterday the arch community announced the successful migration to git. Congrats at this point.

Reading the news article and the steps to use the new repositories I stumbled over the lines "merge the pacman pacnew /etc/pacman.conf.pacnew file".

I compared the two files (my existing and the new one) with nvim -d (or vimdiff) and thought is it possible to merge the files like I am used to within a git repository?

MaKaNu
  • 762
  • 8
  • 25

2 Answers2

1

My experience is from Fedora/Centos where rpm might produce *.rpmsave or *.rpmnew files, but I am assuming *.pacnew files behave in the exact same way.

and thought is it possible to merge the files like I am used to within a git repository?

Yes it is. It requires a little bit of setup but it is absolutely doable.

Step 1, Install etckeeper

Etckeeper is a program that

hooks into package managers like apt, yum, dnf or pacman to automatically commit changes made to /etc during package upgrades.

THIS IS PROGRAM THAT EVERYONE SHOULD INSTALL, independently of plans of using git merge.

Etckeeper has had full support for Arch since 2021.

Step 2, create an additional worktree for /etc

Worktree in general

When you clone or initializes a repository with git you end up having the repo object storage where git stores commits in the .git directory, the index/staging/cache which also is stored in .git and a working directory which contains the files corresponding to the current commit. And there is a one to one relation between them.

However git supports connecting multiple working directories to the same repo object storage through the worktree command. The additional worktrees shares the original .git directory but gets their own indexes, and the files are checked out in different directories. Only one single worktree can check out a given branch (although you can of course check out any specific commit, or use a mirroring branch).

Typical use cases for creating additional worktrees are for instance

  • running git test
  • check out an old version where you want to be able to compare the old version towards the current version for multiple files at once (difftool only lets you compare one file at the time)
  • you have a rebase with conflicts ongoing that you do not want to abandon while you need to create a fix on a different branch before that work is finished.

Worktree for storing the contents of /etc

If you use /etc as your worktree when doing interactive rebase, then notice that git checks out older commits, and thus /etc suddenly contains old versions of various config files. This not unlikely to represent a problem for some services. And when there is a conflicts where git inserts the "<<<<<"/">>>>>" markers, you will have config files that definitely have invalid syntax/content which is unlikely to not represent a problem.

So for these reasons, you should consider /etc as the production system that you normally do not touch and rather create an additional worktree as the staging system where you are free to break and make modifications to.

cd /etc
git worktree add /root/etc-worktree worktree-main main
cd /root/etc-worktree

Step 3, taking advantage of git for merging

In order to make use of git's merge power you need to keep your own changes separate from the upstream changes on separate branches (this sounds more complicated than it actually is). By having a branch which contains the upstream changes, you can then merge this branch and git is able to use its merge algorithm properly.

You only need to do this for files that end up producing .pacnew or .rpmnew files and you can do this retroactively, so it is fine to just start with one main branch and one worktree-main branch, and add upstream branches as needed over time.

You need one branch per file with upstream changes you want to merge (e.g. this only applies to files that you have made local changes to which then prevents the package manager to update).

Example

One of my machines started out being installed with Fedora 26 in 2017. Why do I know that? Because etckeeper is the second program to install on my list of what to do on a new machine (only installing a full vim version has higher priority), and the first commit which /etc/hosts is present in has commit message "Fresh install of Fedora 26". Did I mention how awesome etckeeper is?

Over time I have added a few entries to /etc/hosts (and thanks to etckeeper I know exactly when!). The initial hosts from Fedora 26 only contained two lines, but then in September 2022 (how do I know when?...) an update to the setup package came with a new hosts file which contained 5 additional comment lines. Since I had modified hosts this new version ended up as /etc/hosts.rpmnew.

So this is how I handled that update:

I ran git log -p hosts to figure out what the newest commit before I started modifying it was. It was actually commit 1d49be6b Fresh install of Fedora 26. So that is the starting point for the upstream update branch.

My branch naming policy is to prefix all such upstream branches with rpmnew since *.rpmnew files will be the source for the content of those. You can decide to use the full path in the branch or just the file name. For some cases like /etc/mock/default.cfg using rpmnew/mock/default.cfg is obviously a much better choice than rpmnew/default.cfg while using rpmnew/sshd_config for /etc/ssh/sshd_config might be ok, just make a choice (where a hybrid approach is a possibility).

So with that in mind, I created a rpmnew/hosts branch and updated it with the new hosts.rpmnew content.

cd /root/etc-worktree
git branch rpmnew/hosts 1d49be6b
git switch rpmnew/hosts
cp /etc/hosts.rpmnew hosts
git add hosts
git commit -m /etc/hosts.rpmnew
rm /etc/hosts.rpmnew

Notice that this resets the worktree's content back to 2017, but that's fine because is not used for anything else. If I had worked directly in /etc that would have been a huge problem.

So the upstream changes branch is updated and I can then merge it into the main branch (via worktree-main first).

cd /root/etc-worktree
git switch worktree-main
git merge --ff main  # `worktree-main` should strictly follow `main` and only
                     # occasionally be ahead when doing a merge.
git merge rpmnew/hosts

The last merge command might result in conflicts, and if so you need to resolve those. I highly recommend using KDiff3 for conflict resolution.

After the merge is done then worktree-main is one merge commit ahead of main. This one we want to update /etc to use.

cd /etc
git switch main
git merge --ff worktree-main

and that's it. /etc was moved forward from whatever today's content was to today's content plus an update to hosts and git made use of its merge functionality along the way.

hlovdal
  • 26,565
  • 10
  • 94
  • 165
  • Wow, this does not answer my question in the first place, but sounds probably like the solution for the other problem (manually creating temporary baks to update /etc). Most of what you talked about I am familiar with, but `worktree` and `etckeeper` are new for me. I always wanted too use git for system configs. I definitly try this workflow. – MaKaNu May 29 '23 at 10:10
  • I used KDiff3 in the past, but never understood the 3-Way merge it always confused me, what which file is. The other problem: Is KDiff3 also remote capable? This is the point why I started the question with vimdiff, because it is the tool I have available on all my remote machines. – MaKaNu May 29 '23 at 10:14
  • KDiff3 is possible to run remotely, just run `ssh` with [`-X` or `-Y` option](https://askubuntu.com/questions/35512/what-is-the-difference-between-ssh-y-trusted-x11-forwarding-and-ssh-x-u). – hlovdal May 30 '23 at 11:00
  • Now I am testing this approach I struggled with the creation of the worktree: If I use your command I prompted with the usage info of `git worktree`. If I use `git worktree add /root/etc-worktree main` or `worktree-main` I got invalid reference. – MaKaNu May 30 '23 at 13:37
  • @MaKaNu Maybe you have not configured the default branch to be `main` in git. – hlovdal May 30 '23 at 13:46
  • yes this was the case it was the old "master" but still the usage page. It worked if I only used the path, which probably used inherently the HEAD branch and named the worktree "etc-worktree" – MaKaNu May 30 '23 at 14:00
-1

While writing this question I realized, I could do what everybody is doing these days and ask the dump AI. Since I didn't found an answer to my problem, I thought: Why not test was the AI is saying an provide the answer for future readers.

So this is what the AI told me to do if I want to merge two files. I describe it with the example for the pacman.conf:

open the two files in diff mode:

sudo cp /etc/pacman.conf /etc/pacman.conf.bak

sudo vimdiff /etc/pacman.conf /etc/pacman.conf.pacnew

or

sudo nvim -d /etc/pacman.conf /etc/pacman.conf.pacnew

To navigate between the various diffs use the following commands:

]c   # move the cursor down to the next diff
[c   # move the cursor up to the previous diff

To merge the entries at a diff use the following commands:

do   # merge the content of right file to left file (diff obtain)
dp   # merge the content of left file to right file (diff put)

I must mention that at the end something happened, which I didn't expected:

If the last diff create changes which makes the last lines equal (or the complete file equal) the diff will collapse the last lines. This might be a bit irritating at the first try.

MaKaNu
  • 762
  • 8
  • 25