Git-SVN on Cygwin: svn+ssh good, https awkward

Hi Eric,

=== Short story ===

git-svn on Cygwin works well with svn+ssh://, but needs some fixes around tempfile and/or subprocess handling for https:// - please help.

=== Long story ===

First the good news: git-svn on Cygwin seems to work just fine with svn+ssh://... style Subversion URLs. I've been doing this for quite a while now with various repos, never had issues: I can init, clone, fetch rebase, dcommit just fine.

I'm looking for some help https:// access on Cygwin (Windows XP SP3 underneath; more version info at the end of the message).

I've debugged it a little, but am not sufficiently acquainted with the git-svn Perl source to do changes myself; I hope I can describe the issue well enough that you (or somebody else) can suggest code changes to improve the experience.

0. PROBLEM/SYMPTOM:

login@mycomp ~/compnet2.gitsvn
$ git svn fetch
         M       file1
         A       file2
         A       file3
         A       file4
r15 = 70a56f818a057442e29161170e28c5df4b38a811 (git-svn)
Permission denied: Can't open  
'/cygdrive/c/DOKUME~1/login/LOKALE~1/Temp/tempfile.tmp': Permission denied  
at /usr/local/libexec/git-core/git-svn line 2540

which is, in my version:

   2537                  }
   2538                  $ed = SVN::Git::Fetcher->new($self);
   2539          }
+ 2540          unless ($self->ra->gs_do_update($last_rev, $rev, $self,  
$ed)) {
   2541                  die "SVN connection failed somewhere...\n";
   2542          }
   2543          $self->make_log_entry($rev, \@parents, $ed);

Apparently, we die when fetching a new revision for the 2nd time. I can restart git svn fetch and it will pick up where it left off. It will make progress, but is painful.

git svn rebase has the same issues. I haven't tested other operations such as dcommit yet.

1. WORKAROUND for fetch/rebase:

while ! git svn fetch ; do : ; done git svn rebase

2. A SHORT LOG OF FINDING THE CAUSE:

$ perl -d /usr/local/libexec/git-core/git-svn fetch
...
main::(/usr/local/libexec/git-core/git-svn:10):
10:     $AUTHOR = 'Eric Wong <normalperson@yhbt.net>';
   DB<1> c 2540
Git::SVN::do_fetch(/usr/local/libexec/git-core/git-svn:2540):
2540:           unless ($self->ra->gs_do_update($last_rev, $rev, $self,  
$ed)) {
   DB<2> c 2540
...
r16 = d7829595222a013e703fe4377996984b52921789 (git-svn)
Git::SVN::do_fetch(/usr/local/libexec/git-core/git-svn:2540):
2540:           unless ($self->ra->gs_do_update($last_rev, $rev, $self,  
$ed)) {

Now watch this (we're immediately before the problem):

$ ls -l /cygdrive/c/DOKUME~1/login/LOKALE~1/Temp/ ??????????? ? ? ? ? ? tempfile.tmp $ ls -l /cygdrive/c/DOKUME~1/login/LOKALE~1/Temp/tempfile.tmp ls: cannot access /cygdrive/c/DOKUME~1/login/LOKALE~1/Temp/tempfile.tmp: No such file or directory

This zombie behaviour of trashing the file flags happens if Cygwin is asked to unlink() a file to which there are still open handles. It can't delete the file right away (due to the open handles), so it marks it delete-on-close instead, which then happens when the process(es) terminate.

$ ~/Desktop/handle.exe tempfile.tmp ... git.exe pid: 2236 62C: C:\DOKUME~1\mandree\LOKALE~1\Temp\tempfile.tmp git.exe pid: 5552 62C: C:\DOKUME~1\mandree\LOKALE~1\Temp\tempfile.tmp git-hash-object.exe pid: 5320 62C: C:\DOKUME~1\mandree\LOKALE~1\Temp\tempfile.tmp

So we have three processes that have the file open. I cannot say if they mean the same file, or a different one, but this is the problem.

Here's the pstree dump to help identify the users of this file, with cmdline arguments:

perl,5660 -d -I... /usr/local/libexec/git-core/git-svn fetch
   |-git,2236  --batch
   |-git,5552 git-hash-object -w --stdin-paths
   |   `-git-hash-object,5320
   `-git,1128 update-index -z --index-info

Let's continue in the Perl debugger and crash:

   DB<3> c
Permission denied: Can't open  
'/cygdrive/c/DOKUME~1/login/LOKALE~1/Temp/tempfile.tmp': Permission denied  
at /usr/local/libexec/git-core/git-svn line 2540

Bottom line: the second time we hit this "unless ($self->ra->gs_do_update($last_rev, $rev, $self, $ed))" line, we die because there are still processes that have the tempfile open and thus the former unlink() on the tempfile.tmp couldn't become effective.

3. UNDERLYING CYGWIN LIMITATIONS

<http://wiki.osdev.org/Cygwin_Issues#unlink.28.29> documents now  
(2009-04-23 16:15 UTC):
* unlink isn't atomic (as on other platforms also)
...
On cygwin unlink() is deferred to the end of the process, when unlink  
fails because of locks or delete on close (ERROR_SHARING_VIOLATION).

* unlink a file and subsequent creation of a new file in the same location  
is not supported.

* "delete on close" (unlinking an open file) is supported.

I can provide a short C program to demonstrate this behaviour.

4. FIX SUGGESTIONS AND QUESTIONS

Can we defer unlink()ing the tempfile until all handles to it are closed?

Can we let complete and wait for all processes that hold handles to the tempfile.tmp file before attempting to create a new copy?

Can we use unique tempfile names (timestamps, counter, random number generator) instead of the hardcoded "tempfile.tmp"? This is probably a good idea anyways to evade symlink attacks.

Either should work to fix the issue.

I'm willing to test patches.

5. VERSION INFORMATION:

Cygwin: CYGWIN_NT-5.1 balu 1.5.25(0.156/4/2) 2008-06-12 19:34 i686 Cygwin

Git: git version 1.6.3.rc1.51.gea0b7

Subversion:
svn, version 1.6.1 (r37116)
    compiled Apr 11 2009, 12:07:06

Copyright (C) 2000-2009 CollabNet. Subversion is open source software, see http://subversion.tigris.org/ This product includes software developed by CollabNet (http://www.Collab.Net/).

The following repository access (RA) modules are available:

* ra_neon : Module for accessing a repository via WebDAV protocol using  
Neon.
   - handles 'http' scheme
   - handles 'https' scheme
* ra_svn : Module for accessing a repository using the svn network  
protocol.
   - with Cyrus SASL authentication
   - handles 'svn' scheme
* ra_local : Module for accessing a repository on local disk.
   - handles 'file' scheme
* ra_serf : Module for accessing a repository via WebDAV protocol using  
serf.
   - handles 'http' scheme
   - handles 'https' scheme

Thanks for any efforts to help!

As a followup to my own message,

here's one more link to explain Cygwin/Windows limitations that seem to break git-svn here: http://markmail.org/message/4o4yk6b4u6e44wyh

Thanks in advance for any help offered with git-svn.

Matthias Andree <matthias.andree <at> gmx.de> writes:

These last two points are old information. I would first recommend that you try running the new cygwin 1.7 (still in beta), where unlink() semantics have been greatly improved (you can actually unlink() an in-use file and recreate a new file by the same name while the old handle is still open).

http://cygwin.com/#beta-test

Although newer cygwin may fix your particular problem, this is not a bad suggestion for msysgit, so I suspect someone may be bothered enough by this to write a patch. I, however, will not be the one doing it.

Am 23.04.2009, 19:58 Uhr, schrieb Eric Blake <ebb9@byu.net>:

Hi Eric,

thanks for the hints and pointers to Cygwin 1.7 beta, however I'd suggest to make such changes upstream (i. e. in git.git) for the benefit of everybody else.

It also looks to me as though minor changes in the actual baseline Git code may be sufficient to get git-svn working on Cygwin 1.5 (stable) already. I'd prefer such a minor intrusion to hacking away on forks. :-)

Otherwise I'd start looking at how git-svn handles https:// and svn+ssh:// methods differently from each other next week or so.

Looking forward to further insights.

Am 23.04.2009, 19:58 Uhr, schrieb Eric Blake <ebb9@byu.net>:

Indeed, with a fresh Cygwin 1.7 install, all this seems to work smoothly. I'd rather not bet on Cygwin release dates though, and can't judge on short notice what else will break with Cygwin 1.7... so I'd still appreciate help with fixing git-svn on Cygwin 1.5.

Matthias Andree <matthias.andree@gmx.de> wrote:

Hi Matthias,

Marcus Griep did a lot of work for more efficiently handling of tempfiles in Git.pm a few months ago, so maybe he has more insight into how things work...

git-svn used to use IO::File->new_tmpfile() which was much simpler and probably less prone to portability problems, but cycled through inodes too quickly for Marcus (and probably some other people).

Speaking only for myself and not other contributors:

  I have no interest in dealing with crazy systems like Windows.  Ever.

  However, I'll gladly accept patches to make things work if they don't
  break existing use cases.

  Keep in mind that I'm an extremist.  I dislike Windows (and some other
  crazy software systems) to the point that I refuse to deal with them
  even in paid contracts and employment.

Am 23.04.2009, 21:03 Uhr, schrieb Eric Wong <normalperson@yhbt.net>:

I debugged this a bit further, and the damage (i. e. removal of the tempfile) apparently happens in $pool->clear; in line 4355. I single-stepped it, and the apr_pool_clear(...) is the culprit, it unlinks() the tempfile, making this location unusable.

The temp file is generated when the Reporter object is created through $self->do_update in line 4336.

It remains unclear to me who generates the non-unique filename (it's ...\Temp\tempfile.tmp for me), I've not found the code that generates the file names.

Questions:

- How can I either make sure that the temporary file name for the reporter  
gets either a unique name (near line 4336, through SVN::Ra...)

- or is that the temp file truncated, rather than deleted, near line 4355  
(through SVN::Pool::clear)?

- Is there any way to influence how the SVN::Ra::Reporter obtains  
temporary files?
I seem to be unable to trace this down to the actual functions, but then  
again, my perlboot is rather holey...

Any help?

   4132  package Git::SVN::Ra;
   4133  use vars qw/@ISA $config_dir $_log_window_size/;
   4134  use strict;
   4135  use warnings;
   4136  my ($ra_invalid, $can_do_switch, %ignored_err, $RA);
   4137
   ....
   4324  sub gs_do_update {
   4325          my ($self, $rev_a, $rev_b, $gs, $editor) = @_;
   4326          my $new = ($rev_a == $rev_b);
   4327          my $path = $gs->{path};
   4328
   4329          if ($new && -e $gs->{index}) {
   4330                  unlink $gs->{index} or die
   4331                    "Couldn't unlink index: $gs->{index}: $!\n";
   4332          }
   4333          my $pool = SVN::Pool->new;
   4334          $editor->set_path_strip($path);
   4335          my (@pc) = split m#/#, $path;
: 4336          my $reporter = $self->do_update($rev_b, (@pc ? shift @pc :  
''),
   4337                                          1, $editor, $pool);
   4338          my @lock = $SVN::Core::VERSION ge '1.2.0' ? (undef) : ();
   4339
   4340          # Since we can't rely on svn_ra_reparent being available,  
we'll
   4341          # just have to do some magic with set_path to make it so
   4342          # we only want a partial path.
   4343          my $sp = '';
   4344          my $final = join('/', @pc);
   4345          while (@pc) {
   4346                  $reporter->set_path($sp, $rev_b, 0, @lock, $pool);
   4347                  $sp .= '/' if length $sp;
   4348                  $sp .= shift @pc;
   4349          }
   4350          die "BUG: '$sp' != '$final'\n" if ($sp ne $final);
   4351
   4352          $reporter->set_path($sp, $rev_a, $new, @lock, $pool);
   4353
   4354          $reporter->finish_report($pool);
: 4355          $pool->clear;
   4356          $editor->{git_commit_ok};
   4357  }