Git Plan:

This is the procedure used to convert the OpenAFS CVS version control into a git repository. This work was done by Maximilian Cohan and Michael Meffie at the OpenAFS hackathon 2008 at Google, based off the work already done by Derrick Brashear.

Background

The OpenAFS cvs repository poses several unique and interesting challenges in the git conversion. A custom tool called wdelta, combined with a custom set of commit scripts, has been used since the projects inception. The deltas are effectively patch sets made up of rcs revisions. We need to preserve these deltas and all the history associated with the deltas. Unfortuntunately, the not all deltas were created automically. This leads to some issues,

• There are cases where the same file has been changed more than once in a given delta.
• In those cases, the revisions of a given file in a given delta is not always consective.
• In the cases where there are revision gaps, there could be conflicts when trying to collapse the changes into a single patch for the file.
• Also in the cases where there are gaps, it is possible that cvs tags could interleave the revisions of a delta, although it is unknown at this time if such a case exists.

Overview of the process

Where the tools are

/afs/sinenomine.net/public/openafs/projects/git_work/

Set up base Repo

Prep:

Process:

(commit 1) Start from IBM 1.0 + 3.6 docs

• (in a tmp directory) cvs -d:pserver:anonymous@cvs.openafs.org:/cvs -z5 export -ropenafs-stable-1_1_0 openafs/doc
• Make and change to git_base directory
• untar openafs-1.0-src.tar.gz
• move tmp ´doc´ folder into git_base
• untar initial-commit-postapply.tar
• git init
• git add .
• git commit -m "Initial IBM source code import\rDelta initial-20001103" --author="Transarc <afs@transarc.com>"

(commit 2) Apply license changes

• Apply 2.patch
• Cleanup: rm .rej & .orig
• untar 2-postapply.tar
• git commit -m "" --author="Transarc <afs@transarc.com>"

Patch Fixup

• Original codebase had bit errors; tarball used for git base does not have them, remove (most of) these fixes from patchsets

Repo Fixup

• Resolve issues into commit 3
  • Part of the doc tree was not in the initial import
  • README files were missing
  • IBM tree has munged tags; cvsps output does not...
    • Check in IBM tree to CVS, export -kk and use resulting diff -R to make all Revisions in the tree unmunged
• Store commit 3 diff
• Stash commit 2
• Reset to commit 1
• Apply commit 3 diff
• Amend commit 1
• Reapply commit 2

Issues:

Automate patch -> branches

Prep:

Process:

• Clone base repo (git clone git_base git_working)
• Starting with 3.patch
• Loop:
  • Create branch with patch/delta name (git checkout -b)
  • Apply patch (git apply --index)
  • Commit (-m ´Apply patch X´)

Issues:

Handle binaries

Prep:

• Find a list of binaries in cvs, search for '^expand\w+@b|o@'

Process:

• Loop:
  • For patch X
  • Parse file list from patch X
  • If patch has binary files, git check out branch X
  • For each binary file, CVS check out listed revision over existing file
  • Commit (-m ´Adding binaries´)

Merge branches into master branch

Prep:

• Parse patches to make X.msg & X.author

Process:

• Loop:
  • Merge branch X into master (git merge --squash)
  • Commit with message & author for patch X (git commit)
  • ! On failure... stop, and allow for reset
• Push master to origin/master

Validate and release

• Check that checkout of master matches cvs HEAD

Handle tags

• Determine tag->delta relationship

-- MichaelMeffie - 28 Oct 2008