Migrating Bricolage CVS and SVN to Git

Now that I've successfully migrated the old Bricolage SourceForge CVS repository to Git, and also migrated Subversion to Git, it's time to stitch the two repositories together into one with all history intact. I'm glad to say that figuring out how to do so took substantially less time than the first two steps, thanks in large part to the great help from “doener,” “Ilari,” and “Fissure” on the Freenode #git channel.

Actually, they helped me with a bit more tweaking of my CVS and Subversion conversions. One thing I realized after writing yesterday's post was that, after running git filter-branch, I had twice as many commits as I should have had. It turns out that git filter-branch rewrites all commits, but keeps the old ones around in case you mess something up. doener also pointed out that I wasn't having all grafts properly applied, because git filter-branch only applies to the currently checked-out branch. To get all of the branches, he suggested that I read the git-filter-branch documentation, where I'll find that git filter-branch --tag-name-filter cat -- --all would hit all branches. Actually, such was not clear to me from the documentation, but I took his word for it. Once I did that, to get rid of the dupes, all I had to do was git clone the repository to a new repository. And that was that.

This worked great for my CVS migration, but I realized that I also wanted to clean out metadata from the Subversion migration. Of course, git clone throws out most of the metadata, but git svn also stores some metadata at the end of every commit log message, like this:

git-svn-id: file:///Users/david/svn/bricolage/trunk@8581 e630fb3e-2914-11de-beb6-f300316f8eb1

This had been very handy as I looked through commits in GitX to find parents to set up for grafts, but with that done and everything grafted, I no longer needed it. Ilari helped me to figure out how to properly use git filter-branch to get rid of those. To do it, all I had to do was add a filter for commit messages, like so:

git filter-branch --msg-filter \
'perl -0777 -pe "s/\r?\ngit-svn-id:.+\n//ms"' \
--tag-name-filter cat -- --all

This properly strips out that ugly bit of metadata and finalizes the grafts all at the same time. Very nice.

Now it was time to combine these two repositories for a single unified history. I wasn't able to find a good tutorial for this on the web, other than one that used a third-party Debian utility and only hooked up the master branch, using a bogus intermediary commit to do it. On the other hand, simply copying the pack files, as mentioned in the Git Wiki--and demonstrated by the scripts linked from there--also appeared to be suboptimal: The new commits were not showing up in GitX! And besides, Ilari said, “just copying packs might not suffice. There can also be loose objects.” Well, we can't have that, can we?

Ilari suggested git-fetch, the documentation for which says that it will “download objects and refs from another repository.” Perfect! I wanted to copy the objects from my CVS migration to the Subversion migration.

My first attempt failed: some commits showed up, but not others. Ilari pointed out that it wouldn't copy remote branches unless you asked it to do so, via “refspecs.” Since I'd cloned the repositories to get rid of the duplicate commits created by git filter-branch, all of my lovingly recreated local branches were now remote branches. Actually, this is what I want for the final repository, so I just had to figure out how to copy them. What I came up with was this:

chdir $cvs;
my @branches = map { s/^\s+//; s/\s+$//; $_ } `git branch -r`;

chdir $svn;
system qw(git fetch --tags), $cvs;

for my $branch (@branches) {
    next if $branch eq 'origin/HEAD';
    my $target = $branch =~ m{/master|rev_1_[68]$} ? "$branch-cvs" : $branch;
    system qw(git fetch --tags), $cvs,

It took me a while to figure out the proper incantation for referencing and creating remote branches. Once I got the refs/remotes part figured out, I found that the master, rev_1_6, and rev_1_8 branches from CVS were overwriting the Subversion branches with the same names. What I really needed was to have the CVS branches grafted as parents to the Subversion branches. The #git channel again came to my rescue, where Fissure suggested that I rename those branches when importing them, do the grafts, and then drop the renamed branches. Hence the line above that adds “-cvs” to the names of those branches.

Once the branches were imported, I simply looked for the earliest commits to those branches in Subversion and mapped it to the latest commits to the same branches in CVS, then wrote their SHA1 IDs to .git/info/grafts, like so:

open my $fh, '>', ".git/info/grafts" or die "Cannot open grafts: $!\n";
print $fh '77a35487f18d68b96d294facc1f1a41745ad914c '
       => "835ff47ee1e3d1bf228b8d0976fbebe3c7f02ae6\n", # rev_1_6
          '97ef646f5c2a7c6f47c2046c8d289c1dfc30a73d '
       => "2b9f3c5979d062614ef54afd0a01631f746fa3cb\n", # rev_1_8
          'b3b2e7f53d789bea962fe8047e119148e28865c0 '
       => "8414b64a6a434b2117294c0568c1012a17bc863b\n", # master
close $fh;

With the branches all imported and the grafts created, I simply had to run git filter-branch to make them permanent and drop the temporary CVS branches:

system qw(git filter-branch --tag-name-filter cat -- --all);
unlink '.git/info/grafts';
system qw(git branch -r -D), "origin/$_-cvs" for qw(rev_1_6 rev_1_8 master);

Now I had a complete repository, but with duplicate commits left over by git-filter-branch. To get rid of those, I need to clone the repository. But before I clone, I need the remote branches to be local branches, so that the clone will see them as remotes. For this, I wrote the following function:

sub fix_branches {
    for my $remote (map { s/^\s+//; s/\s+$//; $_ } `git branch -r`) {
        (my $local = $remote) =~ s{origin/}{};
        next if $local eq 'master';
        next if $local eq 'HEAD';
        system qw(git checkout), $remote;
        system qw(git checkout -b), $local;
    system qw(git checkout master);

It's important to skip the master and HEAD branches, as they'll automatically be created by git clone. So then I call the function and and run git gc to take out trash, and then clone:


run qw(git gc);
chdir '..';
run qw(git clone), "file://$svn", 'git_final';

It's important to use the file:/// URL to clone so as to get a real clone; just pointing to the directory instead makes hard links.

Now I that I had the final repository with all history intact, I was ready to push it to GitHub! Well, almost ready. First I needed to make the branches local again, and then see if I could get the repository size down a bit:

chdir 'git_final';
system qw(git remote rm origin);
system qw(git remote add origin git@github.com:bricoleurs/bricolage.git);
system qw(git gc);
system qw(git repack -a -d -f --depth 50 --window 50);

And that's it! My new Bricolage Git repository is complete, and I've now pushed it up to its new home on GitHub. I pushed it like this:

git push origin --all
git push origin --tags

Damn I'm glad that's done! I'll be getting the Subversion repository set to read-only next, and then writing some documentation for my fellow Bricoleurs on how to work with Git. For those of you who already know, fork and enjoy!


Aristotle Pagaltzis wrote:

One thing I noticed is your commit messages. You would get more from git in general if you adopt the style it encourages, which is to treat the first line of the commit message like a mail subject, summarising the commit, and leave more long-winded explanations to the “body” below that. Of course, this is most easily done when your commits are small and focussed.

That has many other rewards too, f.ex. it makes bisecting more helpful: if you bisect a codebase looking for a bug, and you pin it down to a commit that changed 17 different things at once, which of them is the culprit? If you had split it up into 17 commits, git bisect would likely have found the exact source of the problem and saved you the effort.

And the index and stash make it easy to pull out distinct changes from the working copy and make them into separate focussed commits, even across several branches. You can concentrate on the task at hand while coding and need only decide what it was that you did afterwards. In many ways I find that git is humane.

Theory wrote:

Thanks for your note, Aristotle. I've been (usually) following the convention I created for SVN::Notify for a long time now, where the first sentence of a message is a sort of subject. It'll be easy for me t switch to making it the first line.

git-bisect, eh? I hadn't even noticed that one program. There are so many git programs to learn!

And the stash and focused commits…so much to learn!