home of the madduck/ blog/
Packaging with Git

Introduction

I gave a joint presentation with Manoj at Debconf7 about using distributed version control for Debian packaging, and I volunteered to do an on-line workshop about using Git for the task, so it's about time that I should know how to use Git for Debian packaging, but it turns out that I don't. Or well, didn't.

After I made a pretty good mess out of the mdadm packaging repository (which is not a big problem as it's just ugly history up to the point when I start to get it right), I decided to get down with the topic and figure it out once and for all. I am writing this post as I put the pieces together. It's been cooking for a week, simply so I could gather enough feedback. I am aware that Git is not exactly a showcase of usability, so I took some extra care to not add to the confusion.

It may be the first post in a series, because this time, I am just covering the case of mdadm, for which upstream also uses Git and where I am the only maintainer, and I shall pretend that I am importing mdadm to version control for the first time, so there won't be any history juggling. Future posts could well include tracking Subversion repositories with git-svn, and importing packages previously tracked therewith.

I realise that git-buildpackage exists, but imposes a rather strict branch layout and tagging scheme, which I don't want to adhere to. And gitpkg (Romain blogged about it recently), deserves another look since, according to its author, it does not impose anything on its user. But in any case, before using such tools (and possibly extending them to allow for other layouts), I'd really rather have done it by hand a couple of times to get the hang of it and find out where the culprits lie.

Now, enough of the talking, just one last thing: I expect this blog post to change quite a bit as I get feedback. Changes shall be highlighted in bold typeface.

Setting up the infrastructure

First, we prepare a shared repository on git.debian.org for later use (using collab-maint for illustration purposes), download the Debian source package we want to import (version 2.6.3+200709292116+4450e59-3 at time of writing, but I pretend it's -2 because we shall create -3 further down…), set up a local repository, and link it to the remote repository. Note that there are other ways to set up the infrastructure, but this happens to be the one I prefer, even though it's slightly more complicated:

$ ssh alioth
$ cd /git/collab-maint
$ ./setup-repository pkg-mdadm mdadm Debian packaging
$ exit
$ apt-get source --download-only mdadm
$ mkdir mdadm && cd mdadm
$ git init
$ git remote add origin ssh://git.debian.org/git/collab-maint/pkg-mdadm
$ git config branch.master.remote origin
$ git config branch.master.merge refs/heads/master

Now we can use git-pull and git-push, except the remote repository is empty and we can't pull from there yet. We'll save that for later.

Instead, we tell the repository about upstream's Git repository. I am giving you the git.debian.org URL though, simply because I don't want upstream repository (which lives on an ADSL line) hammered in response to this blog post:

$ git remote add upstream-repo git://git.debian.org/git/pkg-mdadm/mdadm

Since we're using the upstream branch of the pkg-mdadm repository as source (and don't want all the other mess I created in that repository), we'll first limit the set of branches to be fetched (I could have used the -t option in the above git-remote command, but I prefer to make it explicit that we're doing things slightly differently to protect upstream's ADSL line).

$ git config remote.upstream-repo.fetch \
    +refs/heads/upstream:refs/remotes/upstream-repo/upstream

And now we can pull down upstream's history and create a local branch off it. The "no common commits" warning can be safely ignored since we don't have any commits at all at that point (so there can't be any in common between the local and remote repository), but we know what we're doing, even to the point that we can forcefully give birth to a branch, which is because we do not have a HEAD commit yet (our repository is still empty):

$ git fetch upstream-repo
warning: no common commits
[…]
  # in the real world, we'd be branching off upstream-repo/master
$ git checkout -b upstream upstream-repo/upstream
warning: You appear to be on a branch yet to be born.
warning: Forcing checkout of upstream-repo/upstream.
Branch upstream set up to track remote branch
  refs/remotes/upstream-repo/upstream.
$ git branch
* upstream
$ ls | wc -l
77

Importing the Debian package

Now it's time to import Debian's diff.gz — remember how I pretend to use version control for package maintenance for the first time. Oh, and sorry about the messy file names, but I decided it's best to stick with real data in case you are playing along:

Since we're applying the diff against version 2.6.3+200709292116+4450e59, we ought to make sure to have the repository at the same state. Upstream never "released" that version, but I encoded the commit ID of the tip when I snapshotted it: 4450e59, so we branch off there. Since we are actually tracking the git.debian.org pkg-mdadm repository instead of upstream, you can use the tag I made. Otherwise you could consider tagging yourself:

$ #git tag -s mdadm-2.6.3+200709292116+4450e59 4450e59
$ git checkout -b master mdadm-2.6.3+200709292116+4450e59
$ zcat ../mdadm_2.6.3+200709292116+4450e59-2.diff.gz | git apply

The local tree is now "debianised", but Git does not know about the new and changed files, which you can verify with git-status. We will split the changes made by Debian's diff.gz across several branches.

The idea of feature branches

We could just create a debian branch, commit all changes made by the diff.gz there, and be done with it. However, we might want to keep certain aspects of Debianisation separate, and the way to do that is with feature branches (also known as "topic" branches). For the sake of this demonstration, let's create the following four branches in addition to the master branch, which holds the standard Debian files, such as debian/changelog, debian/control, and debian/rules:

If you're importing a Debian package using dpatch, you might want to convert every dpatch into a single branch, or at least collect logical units into separate branches. Up to you. For now, our simple example suffices. Keep in mind that it's easy to merge two branch and less trivial to split one into two.

Why? Well, good question. As you will see further down, the separation between master and deb/initramfs actually makes things more complicated when you are working on an issue spanning across both. However, feature branches also bring a whole lot of flexibility. For instance, with the above separation, I could easily create mdadm packages without initramfs integration (see #434934), a disk-space-conscious distribution like grml might prefer to leave out the extra documentation, and maybe another derivative doesn't like the fact that the configuration file is in a different place from upstream. With feature branches, all these issues could be easily addressed by leaving out unwanted branches from the merge into the integration/build branch (see further down).

Whether you use feature branches, and how many, or whether you'd like to only separate upstream and Debian stuff is entirely up to you. For the purpose of demonstration, I'll go the more complicated way.

Setting up feature branches

So let's commit the individual files to the branches. The output of the git-checkout command shows modified files that have not been committed yet (which I trim after the first example); Git keeps these across checkouts/branch changes. Note that the ./debian/ directory does not show up as Git does not know about it yet (git-status will tell you that it's untracked, or rather: contains untracked files since Git does not track directories at all):

$ git checkout -b upstream-patches mdadm-2.6.3+200709292116+4450e59
M Makefile
M ReadMe.c
M mdadm.8
M mdadm.conf.5
M mdassemble.8
M super1.c
$ git add super1.c     #444682
$ git commit -s

  # i now branch off master, but that's the same as 4450e59 actually
  # i just do it so i can make this point…
$ git checkout -b deb/conffile-location master
$ git add Makefile ReadMe.c mdadm.8 mdadm.conf.5 mdassemble.8
$ git commit -s

$ git checkout -b deb/initramfs master
$ git add debian/initramfs/*
$ git commit -s

$ git checkout -b deb/docs master
$ git add RAID5_versus_RAID10.txt md.txt rootraiddoc.97.html
$ git commit -s

  # and finally, the ./debian/ directory:
$ git checkout master
$ chmod +x debian/rules
$ git add debian
$ git commit -s

$ git branch
  deb/conffile-location
  deb/docs
* master
  upstream
  upstream-patches

At this time, we push our work so it won't get lost if, at this moment, aliens land on the house, or any other completely plausible event of apocalypse descends upon you. We'll push our work to git.debian.org (the origin, which is the default destination and thus needs not be specified) by using git-push --all, which conveniently pushes all local branches, thus including the upstream code; you may not want to push the upstream code, but I prefer it since it makes it easier to work with the repository, and since most of the objects are needed for the other branches anyway — after all, we branched off the upstream branch.

Specifying --tags instead of --all pushes tags instead of heads (branches); you couldn't have guessed that! See this thread if you (rightfully) think that one should be able to do this in a single command (which is not git push refs/heads/* refs/tags/*)…

$ git push --all
$ git push --tags

Done. Well, almost…

Building the package (theory)

Let's build the package. There seem to be two (sensible) ways we could do this, considering that we have to integrate (merge) the branches we just created, before we fire off the building scripts:

  1. by using a temporary (or "throw-away") branch off upstream, where we integrate all the branches we have just created, build the package, tag our master branch (it contains debian/changelog), and remove the temporary branch. When a new package needs to be built, we repeat the process.

  2. by using a long-living integration branch off upstream, into which we merge all our branches, tag the branch, and build the package off the tag. When a new package comes around, we re-merge our branches, tag, and build.

Both approaches have a certain appeal to me, but I settled for the second, for two reasons, the first of which leads to the second:

  1. When I upload a package to the Debian archive, I want to create a tag which captures the exact state of the tree from which the package was built, for posterity (I will return to this point later). Since the throw-away branches are not designed to persist and are not uploaded to the archive, tagging the merging commit makes no sense. Thus, the only way to properly identify a source tree across all involved branches would be to run git-tag $branch/$tagname $branch for each branch, which is purely semantic and will get messy sooner or later.

  2. As a result of the above: when Debian makes a new stable release, I would like to create a branch corresponding to the package in the stable archive at the time, for security and other proposed updates. I could rename my throw-away branch, if it still existed, or I could create a new branch and merge all other branches, using the (semantic) tags, but that seems rather unfavourable.

So instead, I use a long-living integration branch, notoriously tag the merge commits which produced the tree from which I built the package I uploaded, and when a certain version ends up in a stable Debian release, I create a maintenance branch off the one, single tag which corresponds to the very version of the package distributed as part of the Debian release.

So much for the theory. Let's build, already!

Building the package (practise)

So we need a long-living integration branch, and that's easier done than said:

$ git checkout -b build mdadm-2.6.3+200709292116+4450e59

Now we're ready to build, and the following procedure should really be automated. I thus write it like a script, called poor-mans-gitbuild, which takes as optional argument the name of the (upstream) tag to use, defaulting to upstream (the tip):

#!/bin/sh
set -eu
git checkout master
debver=$(dpkg-parsechangelog | sed -ne 's,Version: ,,p')
git checkout build
git merge ${1:-upstream}
git merge upstream-patches
git merge master
for b in $(git for-each-ref --format='%(refname)' refs/heads/deb/*); do
  git merge -- $b
done
git tag -s debian/$debver
debuild -i.git
git checkout master

Kumar Appaiah spotted that -i.git is actually needed in the debuild call to make it exclude the .git directory from the generated diff.gz.

Note how we are merging each branch in turn, instead of using the octopus merge strategy (which would create a commit with more than two parents) for reasons outlined in this post. An octopus-merge would actually work in our situation, but it will not always work, so better safe than sorry (although you could still achieve the same result).

If you discover during the build that you forgot something, or the build script failed to run, just remove the tag, undo the merges, checkout the branch to which you need to commit to fix the issue, and then repeat the above build process:

$ git tag -d debian/$debver
$ git checkout build
$ git reset --hard upstream
$ git checkout master
$ editor debian/rules    # or whatever
$ git add debian/rules
$ git commit -s

$ poor-mans-gitbuild

Before you upload, it's a good idea to invoke gitk --all and verify that all goes according to plan:

screenshot of gitk after the above steps

When you're done and the package has been uploaded, push your work to git.debian.org, as before. Instead of using --all and --tags, I now specify exactly which refs to push. This is probably a good habit to get into to prevent publishing unwanted refs:

$ git push origin build tag debian/2.6.3+200709292116+4450e59-3

Now take your dog for a walk, or play outside, or do something else not involving a computer or entertainment device.

Uploading a new Debian version

If you are as lucky as I am, the package you uploaded still has a bug in the upstream code and someone else fixes it before upstream releases a new version, then you might be in the position to release a new Debian version. Or maybe you just need to make some Debian-specific changes against the same upstream version. I'll let the commands speak for themselves:

$ git checkout upstream-patches
$ git-apply < patch-from-lunar.diff   #444682 again
$ git commit --author 'Jérémy Bobbio <lunar@debian.org>' -s

  # this should also be automated, see below
$ git checkout master
$ dch -i
$ dpkg-parsechangelog | sed -ne 's,Version: ,,p'
2.6.3+200709292116+4450e59-3
$ git commit -s debian/changelog

$ poor-mans-gitbuild

$ git push
$ git push origin tag debian/2.6.3+200709292116+4450e59-3

That first git-push may require a short explanation: without any arguments, git-push updates only the intersection of local and remote branches, so it would never push a new local branch (such as build above), but it updates all existing ones; thus, you cannot inadvertedly publish a local branch. Tags still need to be published explicitly.

Hacking on the software

Imagine: on a rainy Saturday afternoon you get bored and decide to implement a better way to tell mdadm when to start which array. Since you're a genius, it'll take you only a day, but you do make mistakes here and there, so what could be better than to use version control? However, rather than having a branch that will live forever, you are just creating a local branch, which you will not publish. When you are done, you'll feed your work back into the existing branches.

Git makes branching really easy and as you may have spotted, the poor-mans-gitbuild script reserves an entire branch namespace for people like you:

$ git checkout -b tmp/start-arrays-rework master

Unfortunately (or fortunately), fixing this issue will require work on two branches, since the initramfs script and hook are maintained in a separate branch. There are (again) two ways in which we can (sensibly) approach this:

I am undecided on this, but maybe the best would be a combination: merge both into a temporary branch and later cherry-pick the commits into two additional, temporary branches until you got it right, and then fast-forward the official branches to their tips:

$ git merge master deb/initramfs
$ editor debian/mdadm-raid                     # …
$ git commit -s debian/mdadm-raid
$ editor debian/initramfs/script.local-top     # …
$ git commit -s debian/initramfs/script.local-top
[many hours of iteration pass…]

[… until you are done]
$ git checkout -b tmp/start-arrays-rework-init master
  # for each commit $c in tmp/start-arrays-rework
  # applicable to the master branch:
$ git cherry-pick $c
$ git checkout -b tmp/start-arrays-rework-initramfs deb/initramfs
  # for each commit $c in tmp/start-arrays-rework
  # applicable to the deb/initramfs branch:
$ git cherry-pick $c

This is assuming that all your commits are logical units. If you find several commits which would better be bundled together into a single commit, this is the time to do it:

$ git cherry-pick --no-commit <commit7>
$ git cherry-pick --no-commit <commit4>
$ git cherry-pick --no-commit <commit5>
$ git commit -s

Before we now merge this into the official branches, let me briefly intervene and introduce the concept of a fast-forward. Git will "fast-forward" a branch to a new tip if it decides that no merge is needed. In the above example, we branched a temporary branch (T) off the tip of an official branch (O) and then worked on the temporary one. If we now merge the temporary one into the official one, Git determines that it can actually squash the ancestry into a single line and push the official branch tip to the same ref as the temporary branch tip. In cheap (poor man's), ASCII notation:

- - - O             >> merge T >>     - - - = - - OT
       ` - - T      >>  into O >>

This works because no new commits have been made on top of O (if there would be any, we might be able to rebase, but let's not go there quite yet; rebasing is how you shoot yourself in the foot with Git). Thus we can simply do the following:

$ git checkout deb/initramfs
$ git merge tmp/start-arrays-rework-initramfs
$ git checkout master
$ git merge tmp/start-arrays-rework-init

and test/build/push the result. Or well, since you are not an mdadm maintainer (We\^W I have open job positions! Applications welcome!), you'll want to submit your work as patches via email:

$ git format-patch -s -M origin/master

This will create a number of files in the current directory, one corresponding for each commit you made since origin/master. Assuming each commit is a logical unit, you can now submit these to an email address. The --compose option lets you write an introductory message, which is optional:

$ git send-email --compose --to your@email.address <file1> <file2> <…>

Once you've verified that everything is alright, swap your email address for the bug number (or the pkg-mdadm-devel list address).

Thanks (in advance) for your contribution!

Of course, you may also be working on a feature that you want to go upstream, in which case you'd probably branch off upstream-patches (if it depends on a patch not yet in upstream's repository), or upstream (if it does not):

$ git checkout -b tmp/cool-feature upstream
[…]

… when a new upstream version comes around

After a while, upstream may have integrated your patches, in addition to various other changes, to give birth to mdadm-2.6.4. We thus first fetch all the new refs and merge them into our upstream branch:

$ git fetch upstream-repo
$ git checkout upstream
$ git merge upstream-repo/master

we could just as well have executed git-pull, which with the default configuration would have done the same; however, I prefer to separate the process into fetching and merging.

Now comes the point when many Git people think about rebasing. And in fact, rebasing is exactly what you should be doing, iff you're still working on an unpublished branch, such as the previous tmp/cool-feature off upstream. By rebasing your branch onto the updated upstream branch, you are making sure that your patch will apply cleanly when upstream tries it, because potential merge conflicts would be handled by you as part of the rebase, rather than by upstream:

$ git checkout tmp/cool-feature
$ git rebase upstream

What rebasing does is quite simple actually: it takes every commit you made since you branched off the parent branch and records the diff and commit message. Then, for each diff/commit_message pair, it creates a new commit on top of the new parent branch tip, thus rewrites history, and orphans all your original commits. Thus, you should only do this if your branch has never been published or else you would leave people who cloned from your published branch with orphans.

If this still does not make sense, try it out: create a (source) repository, make a commit (with a meaningful commit message), branch B off the tip, make a commit on top of B (with a meaningful message), clone that repository and return to the source repository. There, checkout the master, make a commit (with a …), checkout B, rebase it onto the tip of master, make a commit (with a …), and now git-pull from the clone; use gitk to figure out what's going on.

So you should almost never rebase a published branch, and since all your branches outside of the tmp/* namespace are published on git.debian.org, you should not rebase those.

But then again, Pierre actually rebases a published branch in his workflow, and he does so with reason: his patches branch is just a collection of branches to go upstream, from which upstream cherry-picks or which upstream merges, but which no one tracks (or should be tracking).

But we can't (or at least will not at this point) do this for our feature branches (though we could treat upstream-patches that way), so we have to merge. At first, it suffices to merge the new upstream into the long-living build branch, and to call poor-mans-gitbuild, but if you run into merge conflicts or find that upstream's changes affect the functionality contained in your feature branches, you need to actually fix those.

For instance, let's say that upstream started providing md.txt (which I previously provided in the deb/docs branch), then I need to fix that branch:

$ git checkout deb/docs
$ git rm md.txt
$ git commit -s

That was easy, since I could evade the conflict. But what if upstream made a change to Makefile, which got in the way with my configuration file location change? Then I'd have to merge upstream into deb/conffile-location, resolve the conflicts, and commit the change:

$ git checkout deb/conffile-location
$ git merge upstream
CONFLICT!
$ git-mergetool
$ git commit -s

When all conflicts have been resolved, I can prepare a new release, as before:

$ git checkout master
$ dch -i
$ dpkg-parsechangelog | sed -ne 's,Version: ,,p'
2.6.3+200709292116+4450e59-3
# git commit -s debian/changelog

$ poor-mans-gitbuild

# git push
$ git push origin tag debian/2.6.3+200709292116+4450e59-3

Note that Git often appears smart about commits that percolated upstream: since upstream included the two commits in upstream-patches in his 2.6.4 release, my upstream-patches branch got effectively annihilated, and Git was smart enough to figure that out without a conflict. But before you rejoice, let it be told that this does not always work.

Creating and using a maintenance branch

Let's say Debian "lenny" is released with mdadm 2.7.6-1, then:

$ git checkout -b maint/lenny debian/2.7.6-1

You might do this to celebrate the release, or you may wait until the need arises. We've already left the domain of reality ("lenny" is not yet released), so the following is just theory.

Now, assume that a security bug is found in mdadm 2.7.6 after "lenny" was released. Upstream is already on mdadm 2.7.8 and commits deadbeef and c0ffee fix the security issue, then you'd cherry-pick them into the maint/lenny branch:

$ git checkout upstream
$ git pull
$ git checkout maint/lenny
$ git cherry-pick deadbeef
$ git cherry-pick c0ffee

If there are no merge conflicts (which you'd resolve with git-mergetool), we can just go ahead to prepare the new package:

$ dch -i
$ dpkg-parsechangelog | sed -ne 's,Version: ,,p'
2.7.6-1lenny1
$ git commit -s debian/changelog

$ poor-mans-gitbuild

$ git push origin maint/lenny
$ git push origin tag debian/2.7.6-1lenny1

Future directions

It should be trivial to create the Debian source package directly from the repository, and in fact, in response to a recent blog post of mine on the dispensability of pristine upstream tarballs, two people showed me their scripts to do it.

My post also caused Joey Hess to clarify his position on pristine tarballs, before he went out to implement dpkg-source v3. This looks very promising.

Yet, as Romain argues, there are benefits with simple patch management systems. Exciting times ahead!

In addition to creating source packages from version control, a couple of other ideas have been around for a while:

And I am sure there are more. If you have any, I'd be interested to hear about them!

Wrapping up

I hope this post was useful. Thank you for reading to the end, this was probably my longest blog post ever.

I want to thank Pierre Habouzit, Johannes Schindelin, and all the others on the #git/freenode IRC channel for their tutelage. Thanks also to Manoj Srivastava, whose pioneering work on packaging with GNU arch got me started on most of the concepts I use in the above. And of course, the members of the the vcs-pkg mailing list for the various discussions on this subject, especially those who participated in the thread leading up to this post. Finally, thanks to Linus and Junio for Git and the continuously outstanding high level of support they give.

If you are interested in the topic of using version control for distro packaging, I invite you to join the vcs-pkg mailing list and/or the #vcs-pkg/irc.oftc.net IRC channel.

NP: Aphex Twin: Selected Ambient Works, Volume 2 (at least when I started writing…)