Migration from CVS to Git looses merges

I want to convert several old CVS repositories to Git. Some of these CVS repositories contain branches, which have later been merged to the main trunk. When I try to convert using cvs2git or git cvsimport the branches appear in the new git repository but they are not merged to the master branch.

Here is an example of how the branches in the CVS repository were created and merged:

    cd /tmp
    export CVSROOT=$PWD/CVS

    cvs init
    mkdir CVS/foo
    cvs co foo

    cd foo
    (date; seq 10; date) > bar
    cvs add bar
    cvs ci -m msg1                              # rev 1.1

    sleep 1
    printf "1c\n%s\n.\nwq\n" "`date`" | ed bar
    cvs ci -m msg2                              # rev 1.2

    sleep 1
    cvs tag -b a-branch

    sleep 1
    printf "1c\n%s\n.\nwq\n" "`date`" | ed bar
    cvs ci -m msg3                              # rev 1.3

    sleep 1
    cvs up -r a-branch
    printf "12c\n%s\n.\nwq\n" "`date`" | ed bar
    cvs ci -m msg-b1                            # rev 1.2.2.1

    sleep 1
    printf "12c\n%s\n.\nwq\n" "`date`" | ed bar
    cvs ci -m msg-b2                            # rev 1.2.2.2

    sleep 1
    cvs up -A
    cvs up -j a-branch
    cvs ci -m "Merge branch a-branch"           # rev 1.4

Now I have tried 2 ways to convert this to git:

1. mkdir g; cd g; git cvsimport -A <file> -m foo
2. mkdir g; cd g; git init;
   cvs2git --blobfile=foo.blob --dumpfile=foo.dump --username=urs ../CVS/foo
   cat foo.blob foo.dump | git fast-import

In both cases, the branch "a-branch" is in the git repository but is not merged with the master branch, i.e. rev 1.4 has only parent 1.3 but not 1.2.2.2. I also tried cvsimport with several regexes passed using -M to match "Merge branch a-branch", but still the same result.

How should the CVS repository be converted to git, so that the commit corresponding to rev 1.4 has two parents, 1.3 and 1.2.2.2?

urs

On Jun 25 2017, Urs Thuermann <urs@isnogud.escape.de> wrote:

Merges are recognized purely by matching the commit message, and the regexp must capture the branch name in the first subexpr. The -m option enables some default regexps but won't match your example. You can use -M 'Merge branch ([-\w]+)' to match it.

Andreas.

Andreas Schwab <schwab@linux-m68k.org> writes:

Thanks, I tried again and now it works. I misread the man page and also tried enclosing the regexp in //, also tried to escape the parentheses (as in BRE), and I thought I also tried 'Merge branch (.*)'. Obviously, I missed that one since it now works.

BTW, I now looked up the regexes for -m in git-cvsimport.perl. Should probably in the man page.

I also suggest the following patch, so that -m would work with my not so uncommon merge commit message: