Git SVN

From miki
Revision as of 01:20, 15 December 2011 by Mip (talk | contribs) (→‎Merging)
Jump to navigation Jump to search

This page is dedicated to Subversion integration in Git

References

Configuration

My default .git/config file on Linux: 

[user]
    name = ...
    email = ...
[core]
	repositoryformatversion = 0
	filemode = true
	bare = false
	logallrefupdates = true
	ignorecase = true
	autocrlf = false
	whitespace = -blank-at-eol, -space-before-tab,-blank-at-eof
[apply]
    ignorewhitespace = change
[svn-remote "svn"]
	url = ...
	fetch = ...
	branches = ...
	tags = ...

My default .git/config file on Cygwin: 

[user]
    name = ...
    email = ...
[core]
	repositoryformatversion = 0
	filemode = false
	bare = false
	logallrefupdates = true
	ignorecase = true
	autocrlf = false
	whitespace = -blank-at-eol, -space-before-tab,-blank-at-eof
[apply]
    ignorewhitespace = change
[svn-remote "svn"]
	url = https://www.collabnet.nxp.com/svn/atop
	fetch = trunk:refs/remotes/trunk
	branches = branches/*:refs/remotes/*
	tags = tags/*:refs/remotes/tags/*

Workflow

Init / Clone

Get a clone of subversion repository with: 

# git: git clone => 
# svn: svn checkout =>
git svn clone https://www.mysvn.org/svn/project -T trunk -b branches -t tags --username='USER'
git svn clone https://www.mysvn.org/svn/project --stdlayout --username='USER'                   # Identical
# Clean / compress git DB if necessary...
git gc

Fetch / Pull

Fetch all changes from SVN repo: 

# git: git fetch =>
git svn fetch

Fetch and merge all changes from SVN repo: 

# git: git pull =>
# svn: svn update =>
git svn rebase
git svn rebase -l                   # Only rebase against last fetched commit from upstream SVN
In case of conflicts (see [1]):
  • You are on a (no-branch) branch (with git status showing a .dotest file you can ignore)
  • Use the usual git rebase --abort (to bail out), git rebase --continue (solved conflict), git rebase --skip (no-op patch)

dcommit

Send local changes to SVN repo: 

# git: git push =>
git svn dcommit

Branch

Create a branch using git svn branch. For safety, always use option -n first (dry-run): 

git svn branch -n -m 'Some tricky expirements' mip_tricky                  # DRY-RUN
git svn branch -m 'Some tricky expirements' mip_tricky                     # ... if previous command ok

In case of non-standard project layout with several branch directory specification, we need to use option -d to tell git-svn where to create the branch: 

git config --get-all svn-remote.svn.branches
# branches/*:refs/remotes/*                                              <-- we want to create our branch here
# Release_Branches/Dev/*:refs/remotes/Release_Branches/Dev/*
# Release_Branches/Indus/*:refs/remotes/Release_Branches/Indus/*

git svn branch -n -d branches -m 'Some tricky expirements' mip_tricky      # DRY-RUN
git svn branch -d branches -m 'Some tricky expirements' mip_tricky         # ... if everything's ok

Carefull! — If current commit is not in svn, git svn looks at the nearest svn revision that can be used as parent. This is particularly useful when branching off an existing svn branch or tags. However this fails if the svn tags have been changed into git tags or if they have been deleted: 

git co tags/release_2.3
git svn branch -n -m 'branching off 2.3' mip_tricky                        # DRY-RUN
# Copying https://..../svn/myproj/tags/release_2.3 at r1478 to .... https://..../branches/mip_tricky

git br -r -D tags/release_2.3                                              # Let's see what happen if we delete our remote reference
git svn branch -n -m 'branching off 2.3' mip_tricky                        # DRY-RUN
# Copying https://..../svn/myproj/trunk at r497 to .... https://..../branches/mip_tricky

After deleting our remote reference, git svn does not use revision 'r1478 as starting point on tags release_2.3, but the revision r497 from the trunk.

Miscellaneous

Annotate changes in a file: 

# git: git blame =>
# svn: svn annotate/blame =>
git svn blame fullpath              # Only accept full path! (from project rootdir)
git blame file

Create git .gitignore from svn: 

# Typically in project root:
git-svn show-ignore > .gitignore

More Workflows

Config with non-standard layout

In case the SVN repo does not follow the standard directory layout (i.e. trunk/, branches/ and tags), we can change the default config. Here an example of config: 

[user]
      name = beq06659
      email = michael.peeters@nxp.com
[core]
      repositoryformatversion = 0
      filemode = true
      bare = false
      logallrefupdates = true
      autocrlf = false
      whitespace = -blank-at-eol,-space-before-tab,-blank-at-eof
[apply]
      ignorewhitespace = change
[svn-remote "svn"]
      url = https://www.collabnet.nxp.com/svn/atop
      fetch = trunk:refs/remotes/trunk
      branches = branches/*:refs/remotes/*
#     branches = Release_Branches/Customer/Audi/*:refs/remotes/Release_Branches/Customer/Audi/*
#     branches = Release_Branches/Customer/S1NN/*:refs/remotes/Release_Branches/Customer/S1NN/*
#     branches = Release_Branches/Dev/*:refs/remotes/Release_Branches/Dev/*
#     branches = Release_Branches/Indus/*:refs/remotes/Release_Branches/Indus/*
#     tags = tags/*:refs/remotes/tags/*
      tags = tags/Customer/*:refs/remotes/tags/Customer/*
      tags = tags/Industrial/Telebox/*:refs/remotes/tags/Industrial/Telebox/*
      tags = tags/Industrial/*:refs/remotes/tags/Industrial/*
      tags = tags/Intermediate/*:refs/remotes/tags/Intermediate/*
      tags = tags/Mainline/*:refs/remotes/tags/Mainline/*
Do's and Don't's
  • We see that there are overlapping tag specification (tags/Industrial/Telebox/* and tags/Industrial/*). This is because some tags are created in the parent directory, and some more are to be found in the Telebox sub-directory. This will create a conflict though because git will create a reference named tags/Industrial/Telebox, and later will try to create a reference tags/Industrial/Telebox/... that fails because The first reference prevents creating the sub-directory. To solve, simply remove the reference and restart:
git branch -r -D tags/Industrial/Telebox
git svn fetch
  • Git svn falls back in stupid mode (i.e. fetching the full history of each branch, even if given revision was already fetched in the past) if it is initialized without any option. So always init / clone the repo with option -s (standard layout). If the layout is non-standard, use git svn init and edit the config.
  • Never change the default location of branch. Never use a spec like branches = branches/*:refs/remotes/branches/*. When trying with such a spec, git svn was fetching and fetching again the same revisions (from r1!).

Tips

Remove orphaned svn branch

See [2] (and simple fix at [3])

Convert tag branch to real git tags

See [4]. My script:
#! /bin/bash

# Get .git directory
GITDIR=$(git rev-parse --git-dir)
if [ "$GITDIR" = "" ]; then
    echo 'FAILED! Not in a git repository...'
    exit 1
fi

# Backup current tag references
echo "Saving svn tags reference to $GITDIR/svn_tags_backup..."
SVNTAGBACKUP="$GITDIR/svn_tags_backup"
if [ -a "$SVNTAGBACKUP" ]; then
    read -i N -p "$SVNTAGBACKUP already exists. Overwrite? [y/N] "
    if [ "$REPLY" != "y" -a "$REPLY" != "Y" ]; then
        exit 1
    fi
fi

# Convert the tags
git-for-each-ref refs/remotes/tags > "$SVNTAGBACKUP"
< "$SVNTAGBACKUP" cut -d / -f 4- |
while read REF
do
    # Adding tag comment does not work... go figure why
    # COMMENT=$(git log --pretty=tformat:"%s" -1 "refs/remotes/tags/$REF")
    # git tag -a "svn/$REF" -m "$COMMENT" "refs/remotes/tags/$REF"
    git tag "svn/$REF" "refs/remotes/tags/$REF"
    git branch -r -d tags/$REF
done

Force git svn to re-fetch a branch

See [5]. Basically it amounts to rm .git/svn/git-svn/mynewbranch -Rf && git svn fetch.

A cleaner way might actually be to use git svn reset: 

rm -rf .git/svn/.caches        # Clear the cache
git svn reset -r 300           # set 300 as last fetched revision, use -p to fetch 300 also.
git svn fetch

If given revision must be fetched again, use option -p.

My first reset did fail with some error similar to the one we get when the cache is corrupted. Clearing and resetting/fetching again did the trick though.

Convert git diffs into svn diffs

See [6]

Merging

With the current git svn implementation (1.7.5.1), there is no way to dcommit a merge commit and preserve the merge parents in svn. Although the references to the parents are kept in the git svn repo that pushed the merge, these references will not be seen by other git svn client (the information is simply not there in svn). This means that all revisions created by svn only have a single parent, and merge history is lost.

A way to circumvent this is to immediately create after the merge a revision that will update the svn:mergeinfo properties. For this we will need to have a working copy of the svn repositories.

Let's assume we have - an up-to-date svn repo, working copy of trunk/ - a git svn work tree of the same repo, with master branch tracking trunk/, and a branch topicB, tracking branches/topicB/

(See [7] for original sequence)

  1. First go in the svn repo, and get the latest properties:
    2. svn update
svn propget svn:mergeinfo

    Say you get these info: 

    /branches/topicA:126,183-211
  2. Now go to git svn repo, dcommit everything, don't merge yet, and get the list of revision we'll need to add after the merge
    3. # Assuming we are on topicB already
git stash                         # Save your local modifications if needed

# Update and sync master
git checkout master
git svn rebase                    # We assume we have no changes on master

# Update/sync topicB
git co topicB
git rebase master                 # topicB is now at the tip of master

# Merge
git checkout master
git merge --no-ff work            # Create merge commit

# Check everything
git log --graph --oneline --decorate

git commit --amend                # Add a sensible message
git svn dcommit                   # Push the commit to svn
  3. Now at this stage, git svn as created new revisions in svn. Some in the branch topicB, and one in the trunk. We need to get the list of revisions that were merged into trunk:
    4. git svn log master..topicB | egrep -o "@[0-9]+ "

    Say we get: 

    @350
@349
...
@331
@330
  4. Assume that all revisions are contiguous, in svn notation we have to add the revisions 330-350 to the mergeinfo property. The new property is then:
    5. /branches/topicA:126,183-211
/branches/topicB:330-350
    Attention! It is extremelly important to include ALL revisions in the branch. If one revision is missing, git svn will consider the merge as a cherry-pick and will refuse to mark the revision as a merge.
  5. Update trunk to add the new mergeinfo
    6. cd <svnrepo>
svn update                       # Get the new revision
svn pg svn:mergeinfo .           # Check again the mergeinfo
svn pe svn:mergeinfo .           # Add the new merge info as given above
svn commit                       # Create a new revision with the mergeinfo set
  6. Now check that everything is fine in git svn:
    7. rm -rf .git/svn/.caches
git svn reset -r 330 -p          # Refetch our branch completely
git svn fetch
git svn log --graph --oneline --decorate

Troubleshooting

Byte order not compatible

  • Recently got the following error on cygwin:
Byte order is not compatible at ../../lib/Storable.pm (autosplit into ../../lib/auto/Storable/_retrieve.al) line 380, at /usr/share/perl/5.10.1/Memoize/Storable.pm line 21
[...] 
Could not unmemoize function `lookup_svn_merge', because it was not memoized to begin with at /usr/lib/git-core/git-svn line 3197
Basically git svn cannot read the content of its cache, and crashes. This can happen after an update of Perl, or when using a copy of a repository that has been created on another machine (typ. with different endianness settings) (see [8] and [9]).
The best fix is to simply delete the git svn cache:
rm -r .git/svn/caches

Git svn fetches several times the same revision

If git svn does not detect that a branch is a branch starting from the trunk, it will fetch the complete branch history from the beginning, possibly fetching a revision that was fetched already. This is because svn is very flexible on branch, and git svn has to take the pessimistic choice to make sure no data is lost.

A possible solution is to limit the range of revisions (see [10]): 

git svn log --all -1 | \
  sed -n '2s/r\\([0-9]*\\).*/\\1/p' | \
  xargs --replace=from git svn fetch -r from:HEAD

Another solution would be to fetch by range (0-100, 100-200, etc).

Lots of @rev

Likely to occur when tags have been moved around (see [11]).

A short script to remove them: 

git tag -l svn/* | grep @ | 
while read tag; do 
    DUPL=$(git tag --contains "$tag" | wc -l)
    if [ $DUPL != 0 ]; then 
        git tag -d "$tag"
    fi
done
Retrieved from "https://miki.immie.org/mediawiki/index.php?title=Git_SVN&oldid=3692"