MSysGit:MSysGitHerald5

OBSOLETE CONTENT

This wiki has been archived and the content is no longer updated. Please visit git-scm.com/doc for up-to-date documentation.

Good morning git land!

This lovely dark 4am (see http://youtube.com/watch?v=yXi6hg90LUU) is as good an occasion as any to offer to you the 5th issue of the msysGit Herald, the not-quite-biweekly news letter to keep you informed about msysGit, the effort to bring one of the most powerful Source Code Management systems to the poor souls stuck with Windows.

These are the covered topics:

• git-gui and git-fetch

• git-svn

• cvsps

• git.git, mingw.git and 4msysgit.git closing ranks

Seems like this edition is mostly about transport... which is a good sign. It means that the installer has no more issues and we can concentrate on getting git on Windows to work.

git-gui and git-fetch

Some time ago, git-gui could call git-fetch to update the remotes, but it popped up a new cmd window. This was ugly, and at some stage Shawn Pearce, the maintainer of git-gui, changed some parts to avoid that. Unfortunately, this broke fetching from within git-gui.

Warning: technical explanation ahead!

The problem is this: apparently there are two different kinds of processes on Windows (actually three, but let's not make things complicated, and for the same reason, let's not play the obvious pun on Microsoft): graphical and console processes.

Whenever you start a console process from a graphical process, you have to provide it with a console instance, which would be inherited from the calling process if that were a console process, too. But in our case, a graphical process calls a console process (this works), which in turn calls another console process (this doesn't).

So the symptom is that git-fetch, called from git-gui, does not pop up a cmd window, but ssh, which is called from git-fetch for ssh:// urls, does.

I suspected a bug in Tcl/Tk, which might have been fixed in version 8.5b2, so I compiled it and tested it, but the behaviour is still the same. At least we now have a script in our repository to conveniently compile and install Tcl/Tk, should the need ever arise again!

Steffen Prohaska (see interview in the fourth Herald) and Johannes Sixt (see interview in the second Herald) are currently hard at work to flesh out this issue, and maybe the solution lies in using Putty's plink instead of MSys' ssh.

Or maybe we will find some clever way to replace the spawnvpe() call with a CreateProcess() call which forces the console instance to be inherited by the called process. Hannes suggested that a possible reason is our heavy usage of dup2(), which might be helped by using CreateProcess(), but only time, and a fix, will tell what happens here.

git-svn

It seems that quite a few people want to have a working git-svn in msysGit. In our bug tracker, this problem comes in disguise such as "Digest.pm is missing".

Alas, it is not as easy as that... Perl is an MSys program (remember, MSys gives you Unix-like paths relative to the install location), and compiling that came with its own set of surprises: you need to use a different compiler than for MinGW programs!

As I wrote in the third Herald, Simon Sasburg was hard at work to get it to compile, and finally succeeded. But dynamically linked modules (otherwise known as DLLs) would not work, instead resulting in Windows' version of a segmentation fault.

A few more days of trial and error led to the discovery of a source package of Perl for MSys, which was successfully munged into a quilt-like setup: we now have a script to download Perl's sources, patch them with our set of patches, build, install and prepare everything for commit.

The next steps are easy: git-svn's dependencies MIME-Base64, Digest and Digest-MD5 are nice to build and install.

The bigger chunk will be to get subversion and its Perl module to compile, especially given the dependency of the Apache Runtime (APR): subversion's configure complains that it was not found, and the documentation suggest to -- you guessed it -- checking it out with subversion.

Stay tuned.

cvsps

While working at git-svn, I also though about git-cvsimport. This tool heavily depends on cvsps, a program which was inspired by the PatchSets of the source code management system git set out to replace. It is not all that trivial to compile cvsps on MSys, though, since it really loves our good old friend, fork().

Alas, Peer Sommerlund already worked on compiling this on Windows, and was friendly enough to export his Mercurial commits as plain patches, which soon found their way into a git repository. Now cvsps compiles fine, but cvsimport still fails to work, probably because Peer's patches make a few assumptions about the cvs program he used: cvsnt.

It's likely not hard to work out the kinks until it works, but the time to do so successfully avoided finding me.

git.git, mingw.git and 4msysgit.git closing ranks

The official repository for git is git.git, maintained by Junio C Hamano. It does not contain any support for MinGW, because the port is not polished enough yet. Johannes "Hannes" Sixt maintains the official port to MinGW in the mingw.git fork of Junio's repository, and since the msysGit project started during Hannes' vacation, mingw.git was forked yet again, as 4msysgit.git.

It is not totally superfluous to have separate repositories for these projects, because they differ in workflow and focus. For example, one of the big goals of the msysGit project was to provide an integrated installer which comes with everything you will need to run git on Windows (including MSys with its bash and perl), while the MinGW port tries to avoid the dependency on MSys as much as possible.

But recently Hannes has succeeded in polishing a few topics from mingw.git to the point where Junio included them in git.git. There is no MinGW specific code yet, but the foundations are laid.

In related news, Steffen Prohaska has undertaken some tremendous effort to merge mingw.git's 'master' branch into 4msysgit.git's 'devel' branch. Maybe we succeed at some stage to close in totally on mingw.git, and then we'll probably move our history to some "historical" branch, and just keep with mingw.git's master.