patch[RFC] git on Mac OS and precomposed unicode

Purpose: This patch is a suggestion to work around the unpleasenties when Mac OS is decomposing unicode filenames.

The suggested change:
a) is only used under Mac OS
b) can be switched off by a configuration variable
c) is optimized to handle ASCII only filename
d) will improve the interwork between Mac OS, Linux and Windows*
   via git push/pull, using USB sticks (technically speaking VFAT)
   or mounted network shares using samba.

* (Not all Windows versions support UTF-8 yet:
   Msysgit needs the unicode branch, cygwin supports UTF-8 since 1.7)

Runtime configuration: A new confguration variable is added: "core.precomposedunicode" This variable is only used on Mac OS. If set to false, git behaves exactly as older versions of git. When a new git version is installed and there is a repository where the configuration "core.precomposedunicode" is not present, the new git is backward compatible.

When core.precomposedunicode=true, all filenames are stored in precomposed unicode in the index (technically speaking precomposed UTF-8). Even when readdir() under Mac OS returns filenames as decomposed.

Implementation: Two files are added to the "compat" directory, darwin.h and darwin.c. They implement basically 3 new functions: darwin_opendir(), darwin_readdir() and darwin_closedir().

Compile time configuration: A new compiler option PRECOMPOSED_UNICODE is introduced in the Makefile, so that the patch can be switched off completely at compile time.

No decomposed file names in a git repository: In order to prevent that ever a file name in decomposed unicode is entering the index, a "brute force" attempt is taken: all arguments into git (technically argv[1]..argv[n]) are converted into precomposed unicode. This is done in git.c by calling argv_precompose() for all commands except "git commit". This function is actually a #define, and it is only defined under Mac OS. Nothing is converted on any other OS.

Implementation details: The main work is done in darwin_readdir() and argv_precompose(). The conversion into precomposed unicode is done by using iconv, where decomposed is denoted by "UTF-8-MAC" and precomposed is "UTF-8". When already precomposed unicode is precomposed, the string is returned unchanged.

Thread save: Since there is no need for argv_precompose()to be thread-save, one iconv instance is created at the beginning and kept for all conversions. Even readdir() is not thread-save, so that darwin_opendir() will call iconv_open() once and keep the instance for all calls of darwin_readdir() until darwin_close() is called.

Auto sensing: When creating a new git repository with "git init" or "git clone", the "core.precomposedunicode" will be set automatically to "true" or "false".

Typically core.precomposedunicode is "true" on HFS and VFAT. It is even true for file systems mounted via SAMBA onto a Linux box, and "false" for drives mounted via NFS onto a Linux box.

New test case: The new t3910-mac-os-precompose.sh is added to check if a filename can be reached either in precomposed or decomposed unicode (NFC or NFD).

 Documentation/config.txt     |    9 ++
 Makefile                     |    3 +
 builtin/init-db.c            |   22 +++++
 compat/darwin.c              |  200 ++++++++++++++++++++++++++++++++++++++++++
 compat/darwin.h              |   31 +++++++
 git-compat-util.h            |    8 ++
 git.c                        |    1 +
 t/t0050-filesystem.sh        |    1 +
 t/t3910-mac-os-precompose.sh |  104 ++++++++++++++++++++++
 9 files changed, 379 insertions(+), 0 deletions(-)
 create mode 100644 compat/darwin.c
 create mode 100644 compat/darwin.h
 create mode 100755 t/t3910-mac-os-precompose.sh

Torsten Bögershausen <tboegi@web.de> writes:

I haven't looked at the patch yet but that sounds exactly the right way to go about this. Nice.

That also sounds sensible, but...

... I think it generally is a bad idea to say "all except foo". There may be a reason why "foo" happens to be special in today's code, but who says there won't be another command "bar" that shares the same reason with "foo" to be treated specially? Or depending on the options, perhaps some codepath of "foo" may not want the special casing and want to go through the argv_precompose(), no?

After all, "git commit -- pathspec" will have to get the pathspec from the command line, and match them against the paths in the index, the latter of which you are keeping in the canonical form, so you would want the argv[] also be in the same form, and applying your argv_precompose() would be a sensible way to do so, no?

I would also suspect that the cleanest way to implement it is to replace the main() entry point (see how compat/mingw.h does this).

On 08.01.12 03:46, Junio C Hamano wrote:

Thanks Junio for catching this. I added a new test case as well as fixed the code.

We only need to that argv conversion in git.c, (and not in daemon.c), so I sticked to the old model for V1. I send a new patch soon /Torsten

Torsten Bögershausen <tboegi@web.de> writes:

I think you are sidestepping the real issue I raised, which is:

    What is the reason why you do not want to feed the precompose helper
    with some arguments to 'git commit', while it is OK to pass all
    arguments to other commands through precomposition?

I admit it was my fault that I did not spell it out clearly in my response.

I understand that arguments other than pathspec and revs could be left in decomposed form, but is there any harm in canonicalizing any and all command line parameters given in decomposed form consistently into precomposed form? What problem are you trying to solve by special casing "git commit"? That is the real question to be answered, as there may be other commands some of whose arguments may not want to be canonicalized due to the same reason, but you simply overlooked them. When other people need to fix that oversight, they need a clearly written criterion what kind of arguments should not be fixed and why.

And the reason cannot be a desire to pass the value to "--message" argument intact [*1*]; it is not like osx cannot handle text in precomposed form, right?

In general, I do not want to see ugly code that says "this one potentially names a path so we add a call to fix it from decomposed form, but that other one is not a path and we take it intact" sprinkled all over in the codebase, without a good reason.

It may seem that one alternative to munging argv[] is to have the precomposition [*2*] applied inside get_pathspec() and have it take effect only on the pathspecs, which after all ought to be the only place where this matters, but I doubt it would result in maintainable code. The names of branches and tags taken from the command line that are used as revision names will also be compared with results from readdir in $GIT_DIR/refs/ and need to be canonicalized, for example. So I tend to agree with your "brute force" approach to canonicalize argv[] before any real part of git sees them; that is where my suggestion to wrap main() to do so came from.

Also some commands (e.g. "rev-list --stdin") take pathspecs and revs from their standard input stream, so you would need to be careful about them.

[Footnotes]

*1* Also as other commands like "git merge" also take textual message, and you do pass the helper to canonicalize it. No, I am not suggesting you to special case "git merge".

*2* By the way, this may need a better name if the patch touches anywhere outside compat/osx --- it is about "canonicalize pathname and pathspec given from the command line into the form used internally by git", and from an osx person's point of view, the only difference might be decomposed vs precomposed, but on other odd systems it might be that pathnames on the filesystem may be using a different encoding from what is used for pathnames in the index).

On 01/09/2012 08:29 PM, Junio C Hamano wrote:

The short answer for treating "git commit" special:
   The test suite didn't pass any more. (t4201-shortlog.sh)
   This seems more and more to be a bad excuse...
The long answer:
   I have to look into that more deeply.

Thanks for your replies. /Torsten

     (And yes, Mac OS can handle precomposed unicode (at least the
      western european code points))

[snip]