changelog: Document import Dgit field parsing fix

[dgit.git] / dgit-user.7.pod

=head1 NAME

dgit-user - making and sharing changes to Debian packages, with git

=head1 INTRODUCTION

dgit lets you fetch the source code to every package on your
system
as if your distro used git to maintain all of it.

You can then edit it,
build updated binary packages (.debs)
and install and run them.
You can also share your work with others.

This tutorial gives some recipes and hints for this.
It assumes you have basic familiarity with git.
It does not assume any initial familiarity with
Debian's packaging processes.

If you are a package maintainer within Debian; a DM or DD;
and/or a sponsee:
this tutorial is not for you.
Try L<dgit-nmu-simple(7)>, L<dgit-maint-*(7)>,
or L<dgit(1)> and L<dgit(7)>.

=head1 SUMMARY

(These runes will be discussed later.)

=over 4

    % dgit clone glibc jessie,-security
    % cd glibc
    % wget 'https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=28250;mbox=yes;msg=89' | patch -p1 -u
    % git commit -a -m 'Fix libc lost output bug'
    % gbp dch -S --since=dgit/dgit/sid --ignore-branch --commit
    % sudo apt-get build-dep glibc
    % dpkg-buildpackage -uc -b
    % sudo dpkg -i ../libc6_*.deb

=back

Occasionally:

=over 4

    % git clean -xdf
    % git reset --hard

=back

Later:

=over 4

    % cd glibc
    % dgit pull jessie,-security
    % gbp dch -S --since=dgit/dgit/sid --ignore-branch --commit
    % dpkg-buildpackage -uc -b
    % sudo dpkg -i ../libc6_*.deb

=back

=head1 FINDING THE RIGHT SOURCE CODE - DGIT CLONE

=over 4

    % dgit clone glibc jessie,-security
    % cd glibc

=back

dgit clone needs to be told the source package name
(which might be different to the binary package name,
which was the name you passed to "apt-get install")
and the codename or alias of the Debian release
(this is called the "suite").

=head2 Finding the source package name

For many packages, the source package name is obvious.
Otherwise, if you know a file that's in the package,
you can look it up with dpkg:

=over 4

    % dpkg -S /lib/i386-linux-gnu/libc.so.6 
    libc6:i386: /lib/i386-linux-gnu/libc.so.6
    % dpkg -s libc6:i386
    Package: libc6
    Status: install ok installed
    ...
    Source: glibc

=back

(In this example,
libc6 is a "multi-arch: allowed" package,
 which means that it exists in several different builds
 for different architectures.
That's where C<:i386> comes from.)

=head2 Finding the Debian release (the "suite")

Internally,
Debian (and derived) distros normally refer to their releases by codenames.
Debian also has aliases which refer to the current stable release etc.
So for example, at the time of writing
Debian C<jessie> (Debian 8) is Debian C<stable>; and
the current version of Ubuntu is C<yakkety> (Yakkety Yak, 16.10).
You can specify either
the codename C<jessie> or the alias C<stable>.
If you don't say, you get C<sid>,
which is Debian C<unstable> - the main work-in progress branch.

If you don't know what you're running, try this:

=over 4

    % grep '^deb' /etc/apt/sources.list
    deb http://the.earth.li/debian/ jessie main non-free contrib
    ...
    %

=back

For Debian, you should add C<,-security>
to the end of the suite name,
unless you're on unstable or testing.
Hence, in our example
C<jessie> becomes C<jessie,-security>.
(Yes, with a comma.)

=head1 WHAT DGIT CLONE PRODUCES

=head2 What branches are there

dgit clone will give you a new working tree,
and arrange for you to be on a branch named like
C<dgit/jessie,-security> (yes, with a comma in the branch name).

For each release (like C<jessie>)
there is a tracking branch for the contents of the archive, called
C<remotes/dgit/dgit/jessie>
(and similarly for other suites).  This can be updated with
C<dgit fetch jessie>.
This, the I<remote suite branch>,
is synthesized by your local copy of dgit.
It is fast forwarding.

Debian separates out the security updates, into C<*-security>.
Telling dgit C<jessie,-security> means that it should include 
any updates available in C<jessie-security>.
The comma notation is a request to dgit to track jessie,
or jessie-security if there is an update for the package there.

(You can also dgit fetch in a tree that wasn't made by dgit clone.
If there's no C<debian/changelog>
you'll have to supply a C<-p>I<package> option to dgit fetch.)

=head2 What kind of source tree do you get

If the Debian package is based on some upstream release,
the code layout should be like the upstream version.
You should find C<git grep> helpful to find where to edit.

The package's Debian metadata and the scripts for building binary
packages are under C<debian/>.
C<debian/control>, C<debian/changelog> and C<debian/rules> are the
starting points.
The Debian Policy Manual has most of the in-depth
technical details.

For many Debian packages,
there will also be some things in C<debian/patches/>.
It is best to ignore these.
Insofar as they are relevant
the changes there will have been applied to the actual files,
probably by means of actual comments in the git history.
The contents of debian/patches are ignored
when building binaries
from dgitish git branches.

(For Debian afficionados:
the git trees that come out of dgit are
"patches-applied packaging branches
without a .pc directory".)

=head2 What kind of history you get

If you're lucky, the history will be a version of,
or based on,
the Debian maintainer's own git history,
or upstream's git history.

But for many packages the real git history
does not exist,
or has not been published in a dgitish form.
So yuu may find that the history is a rather short
history invented by dgit.

dgit histories often contain automatically-generated commits,
including commits which make no changes but just serve
to make a rebasing branch fast-forward.
This is particularly true of
combining branches like
C<jessie,-security>.

If the package maintainer is using git then
after dgit clone
you may find that there is a useful C<vcs-git> remote
referring to the Debian package maintainer's repository
for the package.
You can see what's there with C<git fetch vcs-git>.
But use what you find there with care:
Debian maintainers' git repositories often have
contents which are very confusing and idiosyncratic.
In particular, you may need to manually apply the patches
that are in debian/patches before you do anything else!

=head1 BUILDING

=head2 Always commit before building

=over 4

    % wget 'https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=28250;mbox=yes;msg=89' | patch -p1 -u
    % git commit -a -m 'Fix libc lost output bug'

=back

Debian package builds are often quite messy:
they may modify files which are also committed to git,
or leave outputs and teporary files not covered by C<.gitignore>.

Kf you always commit,
you can use

=over 4

    % git clean -xdf
    % git reset --hard

=back

to tidy up after a build.
(If you forgot to commit, don't use those commands;
instead, you may find that you can use C<git add -p>
to help commit what you actually wanted to keep.)

These are destructive commands which delete all new files
(so you B<must> remember to say C<git add>)
and throw away edits to every file
(so you B<must> remember to commit).

=head2 Update the changelog (at least once) before building

=over 4

    % gbp dch -S --since=dgit/dgit/sid --ignore-branch --commit

=back

The binaries you build will have a version number which ultimately
comes from the C<debian/changelog>.
You want to be able to tell your
binaries apart from your distro's.

So you should update C<debian/changelog>
to add a new stanza at the top,
for your build.

This rune provides an easy way to do this.
It adds a new changelog
entry with an uninformative message and a plausible version number
(containing a bit of your git commit id).

If you want to be more sophisticated,
the package C<dpkg-dev-el> has a good Emacs mode
for editing changelogs.
Alternatively, you could edit the changelog with another text editor,
or run C<dch> or C<gbp dch> with different options.
Choosing a good version number is slightly tricky and
a complete treatment is beyond the scope of this tutorial.

=head2 Actually building

=over 4

    % sudo apt-get build-dep glibc
    % dpkg-buildpackage -uc -b

=back

apt-get build-dep installs the build dependencies according to the
official package, not your modified one.  So if you've changed the
build dependencies you might have to install some of them by hand.

dpkg-buildpackage is the primary tool for building a Debian source
package.
C<-uc> means not to pgp-sign the results.
C<-b> means build all binary packages,
but not to build a source package.

=head1 INSTALLING

=head2 Debian Jessie or older

=over 4

    % sudo dpkg -i ../libc6_*.deb

=back

You can use C<dpkg -i> to install the
.debs that came out of your package.

If the dependencies aren't installed,
you will get an error, which can usually be fixed with
C<apt-get -f install>.

=head2 Debian Stretch or newer

=over 4

    % sudo apt install ../libc6_*.deb

=back

=head1 Multiarch

If you're working on a library package and your system has multiple
architectures enabled,
you may see something like this:

=over 4

    dpkg: error processing package libpcre3-dev:amd64 (--configure):
     package libpcre3-dev:amd64 2:8.39-3~3.gbp8f25f5 cannot be configured because libpcre3-dev:i386 is at a different version (2:8.39-2)

=back

The multiarch system used by Debian requires each package which is
present for multiple architectures to be exactly the same across
all the architectures for which it is installed.

The proper solution
is to build the package for all the architectures you
have enabled.
You'll need a chroot for each of the secondary architectures.
This iw somewhat tiresome,
even though Debian has excellent tools for managing chroots.
C<sbuild-createchroot> from the sbuild package is a
good starting point.

Otherwise you could deinstall the packages of interest
for those other architectures
with something like C<dpkg --remove libpcre3:i386>.

If neither of those are an option,
your desperate last resort is to try
using the same version number
as the official package for your own package.
(The verseion is controlled by C<debian/changelog> - see above,)
This is not ideal because it makes it hard to tell what is installed,
because it will mislead and confuse apt.

With the "same number" approach you may still get errors like

=over 4

trying to overwrite shared '/usr/include/pcreposix.h', which is different from other instances of package libpcre3-dev

=back

but passing C<--force-overwrite> to dpkg will help
- assuming you know what you're doing.

=head1 SHARING YOUR WORK

The C<dgit/jessie,-security> branch (or whatever) is a normal git branch.
You can use C<git push> to publish it on any suitable git server.

Anyone who gets that git branch from you
will be able to build binary packages (.deb)
just as you did.

If you want to contribute your changes back to Debian,
you should probably send them as attachments to 
an email to the
L<Debian Bug System|https://bugs.debian.org/>
(either a followup to an existing bug, or a new bug).
Patches in C<git-format-patch> format are usually very welcome.

=head2 Source packages

The
git branch is not sufficient to build a source package
the way Debian does.
Source packages are somewhat awkward to work with.
Indeed many plausible git histories or git trees
cannot be converted into a suitable source package.
So I recommend you share your git branch instead.

If a git branch is not enough, and
you need to provide a source package
but don't care about its format/layout
(for example because some software you have consumes source packages,
not git histories)
you can use this recipe to generate a C<3.0 (native)>
source package, which is just a tarball
with accompanying .dsc metadata file:

=over 4

    % echo '3.0 (native)' >debian/source/format
    % git commit -m 'switch to native source format' debian/source/format
    % dgit -wgf build-source

=back

If you need to provide a good-looking source package,
be prepared for a lot more work.
You will need to read much more, perhaps starting with
L<dgit-nmu-simple(7)>,
L<dgit-sponsorship(7)> or
L<dgit-maint-*(7)>

=head1 SEE ALSO

dgit(1), dgit(7)