Re: git rebase chokes on directory -> symlink -> directory

[Junio C Hamano <junkio@cox.net>]

Junio C Hamano <junkio@cox.net> writes:

> However.
>
> I usually have "[apply] whitespace = strip" in my ~/.gitconfig,
> but during this verification run, I disabled it to keep rebase
> from falling back to 3-way merge using merge-recursive.  If I
> turn it on, rebase still fails and I strongly suspect "rebase
> -m" would fail the same way, although I haven't tried it (it
> takes too much time).
>
> I'll be somewhat busy this weekend, so I would welcome anybody
> else beating me to fixing the problem in merge-recursive.

Yuck.  merge-recursive does seem to have problem with D/F
conflict in general; I suspect this is not limited to cases that
involve symbolic links.

First, the setup.

$ git clone git://git.kernel.org/pub/scm/linux/kernel/git/hpa/linux-2.6-newsetup.git/
$ cd linux-2.6-newsetup.git
$ git fetch git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux-2.6.git/
$ git branch other b2ad90f4969226fe8cf3edc5330711ed5fc20105
$ git branch ancestor other^1
$ git reset --hard ancestor
$ echo >>Makefile ; git add Makefile ; git commit -m 'change one'
$ git branch try

To see that the underlying read-tree is working as expected (and
do not have to be debugged), we can first try the usual three-way
read tree:

$ git read-tree -m -u ancestor HEAD other
$ git ls-files -u --abbrev arch/x86-64

This would show...

    120000 ad3f146 3       arch/x86_64/boot
    100644 495f20c 1       arch/x86_64/boot/.gitignore
    100644 495f20c 2       arch/x86_64/boot/.gitignore
    100644 ee6f650 1       arch/x86_64/boot/Makefile
    100644 ee6f650 2       arch/x86_64/boot/Makefile
    ...

The current HEAD (the one that we tagged as 'try') has the
directory arch/x86_64/boot/, but all of its files are unchanged
from the "common ancestor", so they all have identical stages 1
and 2, with stage 3 missing.  The other tree has boot/ as a
symlink.  These "One-side removes other side does not touch" and
"one-side adds" cases are left unmerged by "read-tree -m -u", as
that is how merge-recursive can find renames to begin with.

Now we've seen what the read-tree (which is the same machinery
used as git_merge_trees() in mege-recursive) does with these
three trees, let's see how merge-recursive finishes this off:

    $ git reset --hard try
    $ git merge-recursive ancestor -- HEAD other
    CONFLICT (directory/file): There is a directory with name arch/x86_64/boot in HEAD. Added arch/x86_64/boot as arch/x86_64/boot~other
    Removed arch/x86_64/boot/.gitignore
    Removed arch/x86_64/boot/Makefile
    Removed arch/x86_64/boot/bootsect.S
    Removed arch/x86_64/boot/compressed/Makefile
    Removed arch/x86_64/boot/compressed/head.S
    Removed arch/x86_64/boot/compressed/misc.c
    Removed arch/x86_64/boot/compressed/vmlinux.lds
    Removed arch/x86_64/boot/compressed/vmlinux.scr
    Removed arch/x86_64/boot/install.sh
    Removed arch/x86_64/boot/mtools.conf.in
    Removed arch/x86_64/boot/setup.S
    Removed arch/x86_64/boot/tools/.gitignore
    Removed arch/x86_64/boot/tools/build.c

There is no rename, so rename detection does not interfere, but
the D/F conflict detection code in merge-recursive thinks that
arch/x86_64/boot is a directory in one tree (yes, it is true in
the original tree, but it goes away as all files under it), and
a non-directory in another.  And instead of replacing the boot/
directory that becomes empty with a new symlink boot/, it
creates the boot~funny-name symlink and fails the operation.

When it actually is checking out arch/x86_64/boot out of the
resolved index, it should notice that (1) arch/x86_64/boot
directory is unnecessary to house anything in the resulting
index anymore, and that (2) there is no locally created
untracked file that is 'precious' [*1*].  Then it can rmdir()
and create the boot/ symlink in its place.  To fix this, it may
be necessary to update its checkout code to perform the "remove
first then create" two-pass process git-apply does.

It is unfortunate that for merge-recursive it is probably too
cumbersome to always do the right thing, but I think it should
at least notice that arch/x86_64/boot whose files all disappear
does not conflict with creation of the new symlink.