reimplement package patching and update-patches logic

This works by using git as backend for all the dirty work. This means
that patches are not just applied, but committed separately on top of the
base sources (which are put into an initial commit). A final empty commit
marks the end of the applied patch series, which allows to have multiple
sets of patches to apply on top of each other. So a git history might
look like this:

- OpenADK patch marker: 0000
  (this is the initial commit, containing the unpatched sources)
- patch 1 of series 1
- patch 2 of series 1
- patch 3 of series 1
- OpenADK patch marker: 0001
- patch 1 of series 2
- patch 2 of series 2
- OpenADK patch marker: 0002

In addition to the separating empty commits, for every patch series
metadata files are added (which are used for update-patches):

__patchfiles__: A list of the patches' file names
__patchdir__: The directory containing the applied patches

Since patches might have to be unzipped first and in order to allow
calling git-am just once for each patch series, the patches (along with
above metadata files) are cached in dedicated directories:
.git/patch_tmp/NNNN (where NNNN is the series number with leading zeroes
[so shell globbing returns them in the right order]).

In case update-patches is called later, update_patches.sh works it's way
reverse through the git history, searching for commits named 'OpenADK
patch marker: NNNN'. For each one it finds, it uses the metadata info to
first remove all source patch files, then export the history in between
using git-format-patch.

To change patches or add new ones, the user has to use git-rebase in
order to get things where they need to be for update_patches.sh to put
stuff at the right place. For an example, here is how to change patch 3
of series 1 in the sample history above:

- make desired code changes
- commit them, ideally using --fixup option
- call 'git rebase -i --autosquash <hash of OpenADK patch marker: 0000>'

Using --fixup and --autosquash is convenient, since it automatically
edits the rebase todo as intended. It's optional though, editing the todo
manually will do just fine as well.

Signed-off-by: Phil Sutter <phil@nwl.cc>

1 files changed, 31 insertions, 0 deletions

diff --git a/scripts/update-patches-git b/scripts/update-patches-git
new file mode 100644
index 000000000..8337fa847
--- /dev/null
+++ b/scripts/update-patches-git
@@ -0,0 +1,31 @@
+#!/usr/bin/env bash
+#
+# Update patches using git-format-patch from a source tree prepared by
+# patch_git.sh.
+#
+# (c) 2016 Phil Sutter <phil@nwl.cc>
+
+wrkdist=$1
+wd=$(pwd)
+
+cd "$wrkdist"
+top=""
+top_series=""
+git log --grep="^OpenADK patch marker:" --oneline | while read hash subject; do
+	[ -n "$top" ] || {
+		top=$hash
+		top_series="${subject#OpenADK patch marker: }"
+		continue
+	}
+	bottom=$hash
+	bottom_series="${subject#OpenADK patch marker: }"
+
+	patchdir=$(<.git/patch_tmp/${top_series}/__patchdir__)
+	while read patchfile; do
+		rm ${patchdir}/$patchfile
+	done < .git/patch_tmp/${top_series}/__patchfiles__
+	git format-patch -N -o "$patchdir" ${bottom}..${top}
+
+	top=$bottom
+	top_series=$bottom_series
+done