NAME
git-debrebase - git data model for Debian packaging
INTRODUCTION
git-debrebase is a tool for representing in git, and manpulating, Debian packages based on upstream source code.
The Debian packaging has a fast forwarding history. The delta queue (changes to upstream files) is represented as a series of individual git commits, which can worked on with rebase, and also shared.
DISCUSSION
git-debrebase is designed to work well with dgit.
git-debrebase can also be used in workflows without source
packages, for example to work on Debian-format packages
outside or alongside Debian.
git-debrebase itself is not very suitable for use by Debian derivatives, to work on packages inherited from Debian, because it assumes that you want to throw away any packaging provided by your upstream. However, use of git-debrebase in Debian does not make anything harder for derivatives, and it can make some things easier.
When using gitk on branches managed by git-debrebase, gitk --date-order, gitk --first-parent and gitk -- :. (or gitk .) produce more useful output than the default.
TERMINOLOGY
Pseudomerge
A merge which does not actually merge the trees; instead, it is constructed by taking the tree from one of the parents (ignoring the contents of the other parents). These are used to make a rewritten history fast forward from a previous tip, so that it can be pushed and pulled normally. Manual construction of pseudomerges can be done with "git merge -s ours" but is not normally needed when using git-debrebase.
Packaging files
Files in the source tree within debian/, excluding anything in debian/patches/.
Upstream
The version of the package without Debian’s packaging. Typically provided by the actual upstream project, and sometimes tracked by Debian contributors in a branch "upstream".
Upstream contains upstream files, but some upstreams also contain packaging files in debian/. Any such non-upstream files found in upstream are thrown away by git-debrebase each time a new upstream version is incorporated.
Upstream files
Files in the source tree outside debian/. These may include unmodified source from upstream, but also files which have been modified or created for Debian.
Delta queue
Debian’s changes to upstream files: a series of git commits.
Quilt patches
Files in debian/patches/ generated for the benefit of dpkg-source’s 3.0 (quilt) .dsc source package format. Not used, often deleted, and regenerated when needed (such as when uploading to Debian), by git-debrebase.
Interchange branch; breakwater; stitched; laundered
See " BRANCHES AND BRANCH STATES - OVERVIEW" .
Anchor; Packaging
See " BRANCH CONTENTS - DETAILED SPECIFICATION" .
ffq-prev; debrebase-last
See " STITCHING, PSEUDO-MERGES, FFQ RECORD" .
DIAGRAM
------/--A!----/--B3!--%--/--> interchange view
/ / / with debian/ directory
% % % entire delta queue applied
/ / / 3.0 (quilt) has debian/patches
/ / 3* "master" on Debian git servers
/ / /
2* 2* 2
/ / /
1 1 1 breakwater branch, merging baseline
/ / / unmodified upstream code
---@-----@--A----@--B--C plus debian/ (but no debian/patches)
/ / / no ref refers to this: we
--#-----#-------#-----> upstream reconstruct its identity by
inspecting interchange branch
Key:
1,2,3 commits touching upstream files only
A,B,C commits touching debian/ only
B3 mixed commit (eg made by an NMUer)
# upstream releases
-@- anchor merge, takes contents of debian/ from the
/ previous `breakwater' commit and rest from upstream
-/- pseudomerge; contents are identical to
/ parent lower on diagram.
% dgit- or git-debrebase- generated commit of debian/patches.
`3.0 (quilt)' only; generally dropped by git-debrebase.
* Maintainer's HEAD was here while they were editing,
before they said they were done, at which point their
tools made -/- (and maybe %) to convert to
the fast-forwarding interchange branch.
! NMUer's HEAD was here when they said `dgit push'.
Rebase branch launderer turns each ! into an
equivalent *.
BRANCHES AND BRANCH STATES - OVERVIEW
git-debrebase has one primary branch, the interchange branch. This branch is found on Debian contributors’ workstations (typically, a maintainer would call it master), in the Debian dgit git server as the suite branch (dgit/dgit/sid) and on other git servers which support Debian work (eg master on salsa).
The interchange branch is fast-forwarding (by virtue of pseudomerges, where necessary).
It is possible to have multiple different interchange branches for the same package, stored as different local and remote git branches. However, divergence should be avoided where possible - see " OTHER MERGES" .
A suitable interchange branch can be used directly with dgit. In this case each dgit archive suite branch is a separate interchange branch.
Within the ancestry of the interchange branch, there is another important, implicit branch, the breakwater. The breakwater contains unmodified upstream source, but with Debian’s packaging superimposed (replacing any "debian/" directory that may be in the upstream commits). The breakwater does not contain any representation of the delta queue (not even debian/patches). The part of the breakwater processed by git-debrebase is the part since the most recent anchor, which is usually a special merge generated by git-debrebase.
When working, locally, the user’s branch can be in a rebasing state, known as unstitched. While a branch is unstitched, it is not in interchange format. The previous interchange branch tip is recorded, so that the previous history and the user’s work can later be stitched into the fast-forwarding interchange form.
An unstitched branch may be in laundered state, which means it has a more particular special form convenient for manipulating the delta queue.
BRANCH CONTENTS - DETAILED SPECIFICATION
It is most
convenient to describe the breakwater branch first. A
breakwater is fast-forwarding, but is not usually
named by a ref. It contains in this order (ancestors
first):
Anchor
An anchor commit, which is usually a special two-parent merge:
The first parent contains the most recent version, at that point, of the Debian packaging (in debian/); it also often contains upstream files, but they are to be ignored. Often the first parent is a previous breakwater tip.
The second parent is an upstream source commit. It may sometimes contain a debian/ subdirectory, but if so that is to be ignored. The second parent’s upstream files are identical to the anchor’s. Anchor merges always contain "[git-debrebase anchor: ...]" as a line in the commit message.
Alternatively, an anchor may be a single-parent commit which introduces the "debian/" directory and makes no other changes: ie, the start of Debian packaging.
Packaging
Zero or more single-parent commits containing only packaging changes. (And no quilt patch changes.)
The
laundered branch state is rebasing. A
laundered branch is based on a breakwater but also contains,
additionally, after the breakwater, a representation
of the delta queue:
Delta queue commits
Zero or more single-parent commits containing only changes to upstream files.
The merely
unstitched (ie, unstitched but unlaundered) branch
state is also rebasing. It has the same contents as
the laundered state, except that it may contain,
additionally, in any order but after the breakwater:
Linear commits to the source
Further commit(s) containing changes to to upstream files and/or to packaging, possibly mixed within a single commit. (But not quilt patch changes.)
Quilt patch addition for ’3.0 (quilt)’
Commit(s) which add patches to debian/patches/, and add those patches to the end of series.
These are only necessary when working with packages in ".dsc 3.0 (quilt)" format. For git-debrebase they are purely an output; they are deleted when branches are laundered. git-debrebase takes care to make a proper patch series out of the delta queue, so that any resulting source packages are nice.
Finally, an
interchange branch is fast forwarding. It has
the same contents as an unlaundered branch state, but may
(and usually will) additionally contain (in some order,
possibly intermixed with the extra commits which may be
found on an unstitched unlaundered branch):
Pseudomerge to make fast forward
A pseudomerge making the branch fast forward from previous history. The contributing parent is itself in interchange format. Normally the overwritten parent is a previous tip of an interchange branch, but this is not necessary as the overwritten parent is not examined.
If the two parents have identical trees, the one with the later commit date (or, if the commit dates are the same, the first parent) is treated as the contributing parent.
dgit dsc import pseudomerge
Debian .dsc source package import(s) made by dgit (during dgit fetch of a package most recently uploaded to Debian without dgit, or during dgit import-dsc).
git-debrebase requires that each such import is in the fast-forwarding format produced by dgit: a two-parent pseudomerge, whose contributing parent is in the non-fast-forwarding dgit dsc import format (not described further here), and whose overwritten parent is the previous interchange tip (eg, the previous tip of the dgit suite branch).
STITCHING, PSEUDO-MERGES, FFQ RECORD
Whenever the branch "refs/B" is unstitched, the previous head is recorded in the git ref "refs/ffq-prev/B".
Unstiched branches are not fast forward from the published interchange branches [1]. So before a branch can be pushed, the right pseudomerge must be reestablished. This is the stitch operation, which consumes the ffq-prev ref.
When the user has an unstitched branch, they may rewrite it freely, from the breakwater tip onwards. Such a git rebase is the default operation for git-debrebase. Rebases should not go back before the breakwater tip, and certainly not before the most recent anchor.
Unstitched branches must not be pushed to interchange branch refs (by the use of "git push -f" or equivalent). It is OK to share an unstitched branch in similar circumstances and with similar warnings to sharing any other rebasing git branch.
[1] Strictly, for a package which has never had a Debian delta queue, the interchange and breakwater branches may be identical, in which case the unstitched branch is fast forward from the interchange branch and no pseudomerge is needed.
When ffq-prev is not present, "refs/debrebase-last/B" records some ancestor of refs/B, (usually, the result of last stitch). This is used for status printing and some error error checks - especially for printing guesses about what a problem is. To determine whether a branch is being maintained in git-debrebase form it is necessary to walk its history.
OTHER MERGES
Note that the representation described here does not permit general merges on any of the relevant branches. For this reason the tools will try to help the user avoid divergence of the interchange branch.
See dgit-maint-debrebase(7) for a discussion of what kinds of behaviours should be be avoided because they might generate such merges.
Automatic resolution of divergent interchange branches (or laundering of merges on the interchange branch) is thought to be possible, but there is no tooling for this yet:
Nonlinear (merging) history in the interchange branch is awkward because it (obviously) does not preserve the linearity of the delta queue. Easy merging of divergent delta queues is a research problem.
Nonlinear (merging) history in the breakwater branch is in principle tolerable, but each of the parents would have to be, in turn, a breakwater, and difficult questions arise if they don’t have the same anchor.
We use the commit message annotation to distinguish the special anchor merges from other general merges, so we can at least detect unsupported merges.
LEGAL OPERATIONS
The following
basic operations follow from this model (refer to the
diagram above):
Append linear commits
No matter the branch state, it is always fine to simply git commit (or cherry-pick etc.) commits containing upstream file changes, packaging changes, or both.
(This may make the branch unlaundered.)
Launder branch
Record the previous head in ffq-prev, if we were stitched before (and delete debrebase-last).
Reorganise the current branch so that the packaging changes come first, followed by the delta queue, turning "-@-A-1-2-B3" into "...@-A-B-1-2-3".
Drop pseudomerges and any quilt patch additions.
Interactive rebase
With a laundered branch, one can do an interactive git rebase of the delta queue.
New upstream rebase
Start rebasing onto a new upstream version, turning "...#..@-A-B-1-2-3" into "(...#..@-A-B-, ...#'-)@'-1-2".
This has to be a wrapper around git-rebase, which prepares @’ and then tries to rebase 1 2 onto @’. If the user asks for an interactive rebase, @’ doesn’t appear in the commit list, since @’ is the newbase of the rebase (see git-rebase(1)).
Note that the construction of @’ cannot fail because @’ simply copies debian/ from B and and everything else from #’. (Rebasing A and B is undesirable. We want the debian/ files to be non-rebasing so that git log shows the packaging history.)
Stitch
Make a pseudomerge, whose contributing parent is the unstitched branch and whose overwritten parent is ffq-prev, consuming ffq-prev in the process (and writing debrebase-last instead). Ideally the contributing parent would be a laundered branch, or perhaps a laundered branch with a quilt patch addition commit.
Commit quilt patches
To generate a tree which can be represented as a 3.0 (quilt) .dsc source package, the delta queue must be reified inside the git tree in debian/patches/. These patch files can be stripped out and/or regenerated as needed.
ILLEGAL OPERATIONS
Some git
operations are not permitted in this data model. Performing
them will break git-debrebase.
General merges
See " OTHER MERGES" , above.
git-rebase starting too soon, or without base argument
git-rebase must not be invoked in such a way that the chosen base is before the anchor, or before the last pseudomerge. This is because git-rebase mangles merges. git rebase --preserve-merges is also dangerous.
git-rebase without a base argument will often start too early.
For these reasons, it is better to use git-debrebase and let it choose the base for your rebase. If you do realise you have made this mistake, it is best to use the reflog to recover to a suitable good previous state.
Editing debian/patches
debian/patches is an output from git-debrebase, not an input. If you edit patches git-debrebase will complain and refuse to work. If you add patches your work is likely to be discarded.
Instead of editing patches, use git-debrebase to edit the corresponding commits.
Renaming (etc.) branch while unstitched
The previous HEAD, which will be pseudomerged over by operations like git-debrebase stitch, is recorded in a ref name dervied from your branch name.
If you rename unstitched branches, this information can get out of step.
Conversely, creating a new branch from an unstitched branch is good for making a branch to play about in, but the result cannot be stitched.
COMMIT MESSAGE ANNOTATIONS
git-debrebase makes annotations in the messages of commits it generates.
The general form is
[git-debrebase COMMIT-TYPE [ ARGS...]: PROSE, MORE PROSE]
git-debrebase treats anything after the colon as a comment, paying no attention to PROSE.
The full set of
annotations is:
[git-debrebase split: mixed commit, debian part]
[git-debrebase split: mixed commit, upstream-part]
[git-debrebase convert dgit import: debian changes]
[git-debrebase anchor: convert dgit import, upstream
changes]
[git-debrebase upstream-combine . PIECE[ PIECE...]: new upstream] [git-debrebase anchor: new upstream NEW-UPSTREAM-VERSION, merge] [git-debrebase changelog: new upstream NEW-UPSTREAM-VERSION] [git-debrebase make-patches: export and commit patches] [git-debrebase convert-from-gbp: drop patches] [git-debrebase anchor: declare upstream] [git-debrebase pseudomerge: stitch] [git-debrebase merged-breakwater: constructed from vanilla merge] [git-debrebase convert-to-gbp: commit patches] [git-debrebase convert-from-dgit-view upstream-import-convert: VERSION] [git-debrebase convert-from-dgit-view drop-patches]
Only anchor merges have the "[git-debrebase anchor: ...]" tag. Single-parent anchors are not generated by git-debrebase, and when made manually should not contain any "[git-debrebase ...]" annotation.
The "split mixed commit" and "convert dgit import" tags are added to the pre-existing commit message, when git-debrebase rewrites the commit.
APPENDIX - DGIT IMPORT HANDLING
The dgit .dsc import format is not documented or specified (so some of the following terms are not defined anywhere). The dgit import format it is defined by the implementation in dgit, of which git-debrebase has special knowledge.
Consider a non-dgit NMU followed by a dgit NMU:
interchange --/--B3!--%--//----D*-->
/ /
% 4
/ 3
/ 2
/ 1
2 &_
/ /| \
1 0 00 =XBC%
/
/
--@--A breakwater
/
--#--------> upstream
Supplementary key:
=XBC% dgit tarball import of .debian.tar.gz containing
Debian packaging including changes B C and quilt patches
0 dgit tarball import of upstream tarball
00 dgit tarball import of supplementary upstream piece
&_ dgit import nearly-breakwater-anchor
// dgit fetch / import-dsc pseudomerge to make fast forward
&' git-debrebase converted import (upstream files only)
C' git-debrebase converted packaging change import
* ** before and after HEAD
We want to
transform this into:
I. No new upstream version
(0 + 00 eq #)
--/--B3!--%--//-----D*-------------/-->
/ / /
% 4 4**
/ 3 3
/ 2 2
/ 1 1
2 &_ /
/ /| \ /
1 0 00 =XBC% /
/ /
/ /
--@--A-----B---------------------C'---D
/
--#----------------------------------------->
II. New upstream
(0 + 00 neq #)
--/--B3!--%--//-----D*-------------/-->
/ / /
% 4 4**
/ 3 3
/ 2 2
/ 1 1
2 &_ /
/ /| \ /
1 0 00 =XBC% /
/ /
/ /
--@--A-----B-----------------@---C'---D
/ /
--#--------------------- - - / - - --------->
/
&'
/|
0 00