hitting home directory's parent

Hi:

I just installed git in my Cygwin installation for the first time.  The 
git version is 1.6.1.2, which is the latest version they have.  When I 
enter one of the following commands:
    git --help
    git config --global user.name "Daniel Convissor"

I get this error:
    fatal: Cannot change to /home/danielc/..: Permission denied

Though "git --version" executes fine.

I've been using Cygwin for many years and no other program has ever produced this issue for me. For example, svn --help, ls --help, etc work just fine.

Why is git venturing into to the home directory's parent directory? Is this a bug? If not, is this really necessary?

Thanks,

--Dan

PS:  I'm not subscribed to the list.

On Sat, Aug 22, 2009 at 3:05 AM, Daniel Convissor<danielc@analysisandsolutions.com> wrote:

I guess it tried to find .git directory upward. I think you can set GIT_CEILING_DIRECTORIES to make it stop at /home/danielc. Excerpt from git.txt

-'GIT_CEILING_DIRECTORIES'::
        This should be a colon-separated list of absolute paths.
        If set, it is a list of directories that git should not chdir
        up into while looking for a repository directory.
        It will not exclude the current working directory or
        a GIT_DIR set on the command line or in the environment.
        (Useful for excluding slow-loading network directories.)
-
Duy

Hi Duy:

On Sat, Aug 22, 2009 at 11:10:18AM +0700, Nguyen Thai Ngoc Duy wrote:

That didn't change the situation.

I did read a bit farther in the manual and initialized a new repository. Issuing "git --help" once inside that new repository works. Requiring the --help command to be called from inside a repository, or even that it's looking for a repository at all, seems unwise. All it should do is display the usage information and exit.

Thanks,

--Dan

On Sat, Aug 22, 2009 at 10:05 PM, Daniel Convissor<danielc@analysisandsolutions.com> wrote:

I'd suggest you to degrade Git to a version earlier than 1.6.1, when the error was not added. :-D

But did it at least change the error message? Here is what I got when "chmod 000 /tmp/a":

/tmp/a/b/c $ cd ../.. bash: cd: ../..: Permission denied /tmp/a/b/c $ git --help fatal: Cannot change to '/tmp/a/b/c/..': Permission denied /tmp/a/b/c $ GIT_CEILING_DIRECTORIES=/tmp/a/b git --help fatal: Cannot come back to cwd: Permission denied

Looks like absolute path will cause troubles anyway this case.

For simple things like --help, I agree Git should not do extra work such as searching for Git repository, which is the cause. That was on Jeff's plan IIRC.

On Sat, Aug 22, 2009 at 11:20 PM, Nguyen Thai Ngoc Duy<pclouds@gmail.com> wrote:

Note: I tested it on Linux, not Cygwin so the behavior may differ.

On Sat, Aug 22, 2009 at 11:20:39PM +0700, Nguyen Thai Ngoc Duy wrote:

Without looking, I would not be surprised if it is a side effect of git trying to look up help-related config. So I don't think it is a problem exactly that it checks to see if it is in a repo (which would impact config lookup), but that inability to find a repo is a hard error.

Probably it is calling "setup_git_directory_gently" and ignoring an error return, but there is a die() inside that function. The bug then is that the _gently form is calling die().

As for fixing the setup code, I am sadly way behind on my git queue and not likely to catch up to this anytime soon. So if anybody feels like taking a stab at it, that would be very welcome.

-Peff

On Sun, Aug 23, 2009 at 1:16 AM, Jeff King<peff@peff.net> wrote:

It is (and should be worked around with GIT_CEILING_DIRECTORIES). Unfortunately in my test, it could not chdir() back when it failed to find gitdir. chdir() was called with an absolute directory, and one directory in that path was inaccesible, leading another die("Cannot come back to cwd"). This one is fatal and should not be ignored. I don't know whether having an inaccesible parent directory is a real scenario, as lots of tools would break.

On Sun, Aug 23, 2009 at 11:07:39AM +0700, Nguyen Thai Ngoc Duy wrote:

Hmm, good point. Some of those die()s really can't be removed, because we have munged state in a way that can't be recovered (or is hard to recover).

Also, thinking about it a bit more, I'm not sure it is correct to turn such a die() into a "return -1" without passing more information along to the caller. Because there are really three possible outcomes:

  1. we are in a git repo
  2. we are not in a git repo
  3. some error occurred and we could not determine which
     of (1) and (2) is the case

And the "gently" bit really just refers to (1) versus (2). It is tempting to fold (3) into (2), but I'm not sure that is the right thing to do; some callers (including script callers not under our control) may care about the distinction (though I think we already blur it; for example, a non-existent '.git' regular file is treated the same as one which we couldn't stat for permission problems).

So perhaps the right thing is to either pass back a tri-state status, or to have an extra_gently mode, either of which could be used by "git help". The problem with either is that the current setup process is not amenable to exiting in the middle -- it really may leave us in a weird and buggy state due to the chdir.

So short of major surgery on the setup procedure, I think we may have to leave this bug be for now (especially because I'm not convinced the setup of a non-executable parent directory isn't a little bit crazy).

-Peff

Hi Jeff:

On Sat, Aug 22, 2009 at 02:16:07PM -0400, Jeff King wrote:

That explains what's happening. cmd_help() calls git_config(). Though it seems it would be more, uh, helpful, if help produced basic usage output no matter what, and then produce the more detailed config based output if possible.

Yep.  setup.c, lines 391 and 392 (master branch):
		if (chdir(".."))
			die_errno("Cannot change to '%s/..'", cwd);

Is there a bug tracker y'all use where I can enter a ticket?

Thanks,

--Dan

On Sun, Aug 23, 2009 at 01:12:44AM -0400, Daniel Convissor wrote:

I agree, but the way the code is structured right now means doing this right is probably a pretty big change (see my other message to Duy). So I think it will probably go unfixed for the time being.

We don't have a formal bug tracker, but this is really just a sub-part of another issue, so it will get addressed when the parent issue is fixed.

-Peff

Hi Duy:

On Sat, Aug 22, 2009 at 11:20:39PM +0700, Nguyen Thai Ngoc Duy wrote:

:)  Might be easier to use "svn ci" instead. :)

No.  It produced the same exact error message:
    fatal: Cannot change to /home/danielc/..: Permission denied

Thanks,

--Dan