Am 04.06.2014 17:46, schrieb Johannes Schindelin:
> Hi kusma,
> On Wed, 4 Jun 2014, Johannes Schindelin wrote:
>> The problem arises whenever git.exe calls subprocesses. You can pollute
>> the environment by setting HOME, I do not recall the details, but I
>> remember that we had to be very careful *not* to do that, hence the patch.
>> Sorry, has been a long time.
> Actually, a quick search in my Applegate vaults^W^Wmail archives suggests
> that we had tons of troubles with non-ASCII characters in the path.
After a bit of digging in the history and the old googlegroups issue tracker, I
think this patch is completely unrelated to the non-ASCII problems.
In summary, this patch fixes 'git config' for the portable version only, and it
only does so partially. Thus I don't think its ready for upstream, at least not
in its current form. See below for the nasty details.
> I would be strongly
> in favor of fixing the problem by the root: avoiding to have Git rely on
> the HOME environment variable to be set, but instead add a clean API call
> that even says what it is supposed to do: gimme the user's home
> directory's path. And that is exactly what the patch does.
By that argument we'd have to introduce API abstractions for every environment
variable that could possibly resemble a path (PATH, TMPDIR, GIT_DIR,
GIT_WORK_DIR, GIT_TRACE* etc.).
We already have similar fallback logic for TMPDIR that is completely
non-intrusive to core git code (fully encapsulated in mingw.c, see mingw_getenv
(upstream) or mingw_startup (msysgit)). IMO such a solution would be hugely
preferable over adding an additional get_home_directory() API (and continuously
checking that no new upstream code accidentally introduces another
Analysis of $HOME-realted issues:
1. mangled non-ASCII characters in environment variables
E.g. issue 491 , reportedly fixed in v1.7.10 ( comment #10).
This is actually a bug in msys.dll, and there's nothing that can be done about
it from within git.exe. It is also not a problem if git is launched from
The root cause is that the msys environment is initialized using
GetEnvironmentStringsA(), which returns GetOEMCP()-encoded strings (e.g.
cp850), rather than GetACP() (e.g. cp1252) as all other *A API functions do
. This adds one level of mangling whenever a native Windows program starts
an msys program (so e.g. the call chain bash->git->bash->wish would mangle
twice, see  comment #3).
For the fixed GetEnvironmentStringsA(), see  lines 459ff.
(As a side note, $HOMEDRIVE and $HOMEPATH originally did not have this problem,
as they were separately initialized from NetUserGetInfoA(). This was changed in
v1.6.3, however, at that time etc/profile was still using the broken
$USERPROFILE. See , .)
2. 'git config' doesn't work with disconnected network drives
Issues 259 , 497  and 512 , fixed in v126.96.36.199 for bash and v188.8.131.52
Apparently, $HOMEDRIVE$HOMEPATH is the home directory on the network, and
$USERPROFILE is local. To be able to work offline, we need to check if
$HOMEDRIVE$HOMEPATH exists and fall back to $USERPROFILE otherwise.
Note that git-wrapper does _not_ check if $HOMEDRIVE$HOMEPATH actually exists,
as the original git.cmd did. This is probably a regression wrt issue 259.
3. HOME is not set when using the portable version
Issue 482 , partially fixed in v184.108.40.206 by this patch.
- there's no fallback to $USERPROFILE, so it doesn't work with disconnected
network drives (see problem 2.)
- it doesn't setenv(HOME) for child processes (at least git-gui accesses
$env(HOME) directly, but I haven't checked what happens if HOME is not set)
Incidentally, this patch was first released with v220.127.116.11, which also sets
$HOME correctly in both etc/profile and git.cmd. So I suspect that this patch
has always been essentially dead code (except perhaps for the portable version,
I've never used that).
To unsubscribe from this list: send the line "unsubscribe git" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html