dgit-nmu-simple - tutorial for DDs wanting to NMU with git
This tutorial describes how a Debian Developer can do a straightforward NMU of a
package in Debian, using dgit.
This document won't help you decide whether an NMU is a good idea or whether it
be well received. The Debian Developers' Reference has some (sometimes
questionable) guidance on this.
Conversely, you do not need to know anything about the usual maintainer's git
workflow. If appropriate, you can work on many different packages, making
similar changes, without worrying about the individual maintainers' git
This tutorial only covers changes which can sensibly be expressed as a
reasonably small number of linear commits (whether to Debian packaging or to
upstream files or both).
If you want to do a new upstream version, you probably want to do as the
maintainer would have done. You'll need to find out what the maintainer's git
practices are and consult the appropriate "dgit-maint-*(7)" workflow
% dgit clone glibc jessie
% cd glibc
% git am ~/glibc-security-fix.diff
% dch --nmu "Apply upstream's fix for foo bug."
% git add debian/changelog && git commit -m"NMU changelog entry"
% dpkg-buildpackage -uc -b
[ run your tests ]
% dch -r && git add debian/changelog && git commit -m"Finalise NMU"
% dgit -wgf sbuild -A -c jessie
[ final tests on generated .debs ]
% dgit -wgf [--delayed=5] push jessie
[ enter your gnupg passphrase as prompted ]
[ see that push and upload are successful ]
[ prepare and email NMU diff (git-diff, git-format-patch) ]
When preparing an NMU, the git commits you make on the dgit branch should be
simple linear series of commits with good commit messages. The commit messages
will be published in various ways, including perhaps being used as the cover
messages for generated quilt patches.
Do not make merge commits. Do not try to rebase to drop patches - if you need to
revert a change which is actually a Debian patch, use git-revert.
If you need to modify a Debian patch, make a new commit which fixes what needs
fixing, and explain in the commit message which patch it should be squashed
with (perhaps by use of a commit message in "git rebase --autosquash
(Of course if you have specific instructions from the maintainer, you can follow
those instead. But the procedure in this tutorial is legitimate for any
maintainer, in the sense that it should generate an upload to which the
maintainer cannot reasonably object.)
dgit clone will put you on a branch like "dgit/sid". There is a
pseudo-remote called "dgit" which also contains a branch like
"dgit/sid", so you do things like "git diff dgit/dgit/sid"
to see what changes you have made.
Don't forget to "git add" any new files you create. Otherwise git
clean (which is requested with the "-wgf" option in the recipe
above) will delete them.
Many package builds leave dirty git trees. So, commit before building. That way
you can use "git reset --hard".
If you follow this approach you don't need to care about the build dirtying the
tree. It also means you don't care about the package clean target, which is
just as well because many package clean targets are broken.
The dgit git history (visible in gitk and git log) is not necessarily related to
the maintainer's or upstream's git history (if any).
If the maintainer has advertised a git repo with Vcs-Git dgit will set up a
remote for it, so you can do
% git fetch vcs-git
You can cherry pick changes from there, for example. Note that the maintainer's
git history may not be suitable for use with dgit. For example, it might be a
patches-unapplied branch or even contain only a debian/ directory.
You can use dgit's --delayed
option to upload to the DELAYED queue.
However, you should read the warning about this option in dgit