git-svn: broken symlink check fetches wrong revision from svn (possible regression)

git-svn: broken symlink check fetches wrong revision from svn (possible regression)

[Anton Gyllenberg]

The check for broken symlinks introduced in dbc6c74d0 "git-svn: handle
empty files marked as symlinks in SVN" changed git-svn fetch behavior
for a real-life repository I need to access. It doesn't have broken
symlinks. The repository had a branch deleted and recreated off of a
different branch point. The new branch point has empty files not in
the original branch point. It seems the symlink check
_mark_empty_symlinks fetches the wrong SVN revision of the empty files
in this special case. git-svn fetch stops with an error like

Filesystem has no item: File not found: revision 4, path
'/branches/branch-c/b1' at git-svn line 3288

Which is true, there is no such file, but in my case the empty file b1
is in revision 6 (new branch-c copied from branch-b@r5 where the
original branch-c was deleted), not revision 4 (original branch-c
copied from trunk in r4).


I used the following shell script to produce a test repository that
displays the symptom:
#!/bin/bash

SVNREPO=file:///$PWD/test.db
svnadmin create test.db
svn co $SVNREPO testreposvn
cd testreposvn
mkdir trunk branches
svn add trunk branches
svn ci -m "init standard layout"
svn cp trunk branches/branch-b
svn ci -m 'branch-b off trunk'
touch branches/branch-b/b1
svn add branches/branch-b/b1
svn ci -m 'add empty file b1'
svn cp -m 'branch-c' $SVNREPO/{trunk,branches/branch-c}
svn del -m 'oops, wrong branchpoint' $SVNREPO/branches/branch-c
svn cp -m 'branch-c off of branch-b' $SVNREPO/branches/{branch-b,branch-c}

----
This leaves the svn repository in ./test.db, svn working copy in
./testreposvn. And the problem:

$ git init -s file:///$PWD/test.db testrepogit
Initialized empty Git repository in /tmp/test/testrepogit/.git/
$ cd testrepogit
$ git --version
git version 1.6.2.rc0
$ git svn fetch
r1 = ff1aea02906c970caa49055db564158d6911a512 (trunk)
Found possible branch point: file:////tmp/test/test.db/trunk =>
file:////tmp/test/test.db/branches/branch-b, 1
Found branch parent: (branch-b) ff1aea02906c970caa49055db564158d6911a512
Following parent with do_switch
Successfully followed parent
r2 = d0f256cbf64dc1e49597e54a55538758ca70d65f (branch-b)
	A	b1
r3 = 48d511b172a95a609d5e10f1275664e6b4ca4e62 (branch-b)
Found possible branch point: file:////tmp/test/test.db/trunk =>
file:////tmp/test/test.db/branches/branch-c, 3
Found branch parent: (branch-c) ff1aea02906c970caa49055db564158d6911a512
Following parent with do_switch
Successfully followed parent
r4 = 57923c03b201b2dbbb9d52b81e198a1a78b9de44 (branch-c)
Found possible branch point:
file:////tmp/test/test.db/branches/branch-b =>
file:////tmp/test/test.db/branches/branch-c, 5
Found branch parent: (branch-c) 48d511b172a95a609d5e10f1275664e6b4ca4e62
Index mismatch: 934e48f4addc4882a9b7b99615348569d531c873 !=
4b825dc642cb6eb9a060e54bf8d69288fbee4904
rereading 48d511b172a95a609d5e10f1275664e6b4ca4e62
Following parent with do_switch
Filesystem has no item: File not found: revision 4, path
'/branches/branch-c/b1' at git-svn line 3288

----

Breaking out of _mark_empty_symlinks() before SVN access makes the
git-svn fetch complete without errors but does record the original
branch-c as parent which I did not expect but it kind of makes sense.
I tried to figure out where the revision to fetch in
_mark_empty_symlinks() comes from but eventually gave up as I am not
that fluent in perl.

Anton