Git rebase using "wrong" commit

Hello all,
	We have a specific branch (integration/bug-fixes) which is under
continual rebase, shared amongst several developers.  The tracked
instances have "branch.integration/bug-fixes.rebase true" and
"branch.integration/bug-fixes.mergeoptions --no-ff" - this is working
well.  Individual developers merge their topic branches (in this case
bug fixes) into integration/bug-fixes and push it out.
	Occasionally, we want to release some bug fixes, but not others.  IOW,
when integration/bug-fixes comes to be rebased onto master, we wish to
preserve some of the topic branch merges, but not others.  This usually
works fine, but an issue has cropped up where there is a conflict
between two fixes, and only one of the two bug fixes has been released;
now, when we come to perform the rebase, it is not working as desired.

Output of "git log --first-parent master..integration/bug-fixes":

-----
commit 9f79ca3928d645435d1608e22297dacd2be94e3b
Merge: 2bc19f9 cd8273f
Author: Philip Allison <philip.allison@smoothwall.net>
Date:   Tue Oct 20 10:11:06 2009 +0100

    Merge branch 'bug/2' into integration/bug-fixes
    
    * bug/2:
      Unify format of files 1, 2 and 3
    
    Conflicts:
        file1

commit 2bc19f9af94e03e06fbe5a1034820c606d13c3d3
Merge: c785da2 b17a93c
Author: Philip Allison <philip.allison@smoothwall.net>
Date:   Tue Oct 20 10:07:37 2009 +0100

    Merge branch 'bug/1' into integration/bug-fixes
    
    * bug/1:
      Fix file 1
-----

bug/2 has, effectively, been merged into master; bug/1 has not; hence, there is not (on master) any equivalent change to that which was made to resolve the conflict.

"git rebase -i master" output (whilst on HEAD of integration/bug-fixes):

----- pick b17a93c Fix file 1

# Rebase cd8273f..9f79ca3 onto cd8273f -----

The trouble here is that b17a93c is the head of bug/1 - the rebase has
picked up the changes from that branch and is trying to apply them as a
fast-forward commit in the newly rebased integration/bug-fixes, instead
of trying to rebase the *merge* of bug/1 (i.e. 2bc19f9).
	The "-m" and "-p" options to rebase don't seem to be helping.  I have
tried getting rebase to "pick 2bc19f9" via -i, but that isn't working
either:

----- fatal: Commit 2bc19f9af94e03e06fbe5a1034820c606d13c3d3 is a merge but no -m option was given. Could not apply 2bc19f9... -----

Interestingly, rebase does *not* seem to be trying to recreate 9f79ca3 (or its second parent), which is the problem I was originally expecting to run into.

At the moment I don't know where to go from here, short of manually recreating the branch as I want it, which I am loathe to do. Ideas and suggestions much appreciated.

Please find attached a working copy of the repository, just before any attempted rebase.

Regards

Philip Allison wrote:

[...]

Ok, so I gather the (simplified) history looks like

  *------A  (master)
   \\__   \
    \  \   \
     B--M1--M2  (topic)

where A and B conflict, so M2 is nontrivial. B is the *second* parent to M1, which I originally thought would affect the 'rebase -p', but it doesn't; see below. (You're talking about bug/2, but the repository shows bug/2 == master, so there are only two branches involved.)

The inherent difficulty with doing a rebase here is, what about the merges? By definition, git-rebase needs to somehow "rebuild" the commits, as defined by the changes they do, on the new base.

There are several ways how merges can be rebased, and you tried some:

* Not at all; discard the merges.  The result would be

    *---A  (master)
         \
          \
           B'  (topic)

  B' will differ substantially from B because there will be a (rebase)
  conflict.  This is the normal mode of operation, as you noticed:

* Attempt to rebuild merges; this is the -p flag.  Assume for a second
  that you have a different history

    *---A  (master)
     \   
      \   
       C---M2  (topic2)
          /
         *

  where M2 is a merge with some other branch.  'git rebase -i -p
  master topic2' will attempt to build

    *---A  (master)
         \   
          \   
           C---M2  (topic2)
              /
             *

  Going back to your scenario, this means building

  *------A--.  (master)
   \      \  \
    \      \  \
     B-----M1--M2  (topic)

  Now there appears to be some bug with rebase -p, because it insists
  that there is no work at all to do (the buffer is just 'noop').
  However, rebase -p is known to be somewhat ill-defined and broken.
  I think that in this case, what actually happens is that the merge
  from master to topic confuses the initialisation of $REWRITTEN in

	# line 668
	for c in $(git merge-base --all $HEAD $UPSTREAM)
	do
		echo $ONTO > "$REWRITTEN"/$c || die "blah"
        done

  because suddenly the merge-base is A, meaning that the later test

	# line 712
	for p in $(git rev-list --parents -1 $sha1 | cut -d' ' -s -f2-)
	do
		if test -f "$REWRITTEN"/$p -a \( $p != $ONTO -o $sha1 = $first_after_upstream \)

  fails for all commits except M2.  So what's *actually* rewritten is
  M2; but doing so is pointless in rebase's eyes, because it assumes
  the topic will be based on master eventually, so merging master into
  it is a no-op.  Hence there is nothing to do.

-m means something entirely different, namely that git-rebase should use an (internal) merge instead of format-patch|am, which has some advantages but a big speed disadvantage. This doesn't matter for rebase -i as it never uses format-patch|am.

As for manually picking merge commits, that fails because you're just trying to fool git-cherry-pick into doing something it can't.

[...]

Note that there is no real reason to rebase continually unless you want to kick out flawed topics or old versions of branches. For example, git.git only rebases 'next' immediately after a release. Furthermore, even if you go for the 'pu' model which is rebuilt all the time, you can automate this with a script. There is one in the 'todo' branch of git.git, called PU.

(I myself just use a list of topics I want merged, and a simple 'git merge $topic' loop.)

Thomas Rast <trast@student.ethz.ch> writes:

I do not use PU nor RB from todo these days, by the way. I instead rebuild my branches with Reintegrate (also from todo).