Gentoo Archives: gentoo-portage-dev

From: Sebastian Pipping <sping@g.o>
To: gentoo-portage-dev@l.g.o
Subject: [gentoo-portage-dev] Handling merge issues (on the case of prefix)
Date: Fri, 26 Mar 2010 05:24:35

As the prefix branch is not the last case where people may run into
problems with merging due to the conversion from SVN to Git I feel like
writing about it here.

In this mail
- The problem
  - Merges in SVN
  - Merges in Git
  - Consequences

- Dealing with it
  - The concept
  - Realization

The problem

Merges in SVN
In SVN people select what commits are merged from where to where
manually: SVN, merge commits 3 to 10 to branch "prefix" please.  What
has been merge is tracked in the mind of the person merging and in log

Merges in Git
In Git there's another place where merges are saved: in the DAG of commits:

  - A commit with several children/successors is a branch point
  - A commit with several parents is a merge point

When merging one branch into another Git runs the DAG up (into the past)
from both commits until it finds a shared merge or branch point: that's
the point up to where both branches were synced last time.  History
after that point is then merged over, nothing before.

So Git relies on the existence of merge commits to detect what has been
merged already.  Creating all these merge commits is a tough job for a
conversion tool (like svn2git) as it would have to distinguish between
cases where just a few commist have been cherry-picked over and cases
where all previous commits have made it over.

So the portage Git history does not have merge commits at these points
but plain single-parent commits.

Dealing with it

The concept
So to not get diffs way bigger than needed when merging what we can do
is we can manually (and permanently) teach Git what we know more about
portage's history.
Let's look the case of the prefix branch.

The current head on prefix has this log message:

  [head on prefix]
  "Merged from trunk -r15842:15843"

Looking a few commits back (using gitg or git log) on branch master I
find this commit:

  [commit f52e83b0982c9c18d96757ab55109d43a9873b3f on master]
  install_qa_check: make sure init.d and conf.d files do not have
    syntax errors in them #310805
  svn path=/main/trunk/; revision=15843

So that's the commit where grobian merged trunk into prefix last time:

How do we teach Git that all that stuff has been merged already?
By creating a merge commit with two parents:

  1) f52e83b0982c9c18d96757ab55109d43a9873b3f
  2) head on prefix

This is how to do it on the shell

  # Checkout prefix locally
  git checkout -b prefix overlays-gentoo-org/prefix

  # Create merge commit manually (_NOT_ something to do usually)
  MERGE_COMMIT=$(echo \
    'Teach Git that trunk@15843 has been merged into prefix' \
    | git commit-tree 'prefix^{tree}' -p prefix -p

  # Inspect result
  echo ${MERGE_COMMIT}

  # Move prefix head forward to include that commit
  git merge ${MERGE_COMMIT}

  # Inspect what we have done visually

That's where we leave the land of dirty hacks and Git's so-called
plumbings. We can can continue merging the few remaining commits from
here as usual.

  # Get a feeling for what would be merged
  # Should list ~10 commits (instead of ~8000 without the hack before)
  git cherry -v prefix overlays-gentoo-org/master

  # Merge master into prefix
  git merge overlays-gentoo-org/master  # Fails with conflicts, fine
  git status
  git mergetool
  git commit
  git push overlays-gentoo-org prefix

I hope this mail was helpful to you.