Re: [BUG/PATCH] setup: Copy an environment variable to avoid overwrites

From: Junio C Hamano <gitster@pobox.com>
Date: Jan 5, 2013, 01:17 UTC
Message-ID: <7vsj6gqvhc.fsf@alter.siamese.dyndns.org>
In-Reply-To: <CAEvUa7niTJVfp8_kuWs50kvhfZ59F-yAuAmeOXEduHXOq-tRFA@mail.gmail.com>

David Michael <fedora.dm0@gmail.com> writes:

Show 16 quoted lines

> I have encountered an issue with consecutive calls to getenv
> overwriting earlier values.  Most notably, it prevents a plain "git
> clone" from working.
>
> Long story short: This value of GIT_DIR gets passed around setup.c
> until it reaches check_repository_format_gently.  This function calls
> git_config_early, which eventually runs getenv("HOME").  When it
> returns back to check_repository_format_gently, the gitdir variable
> contains my home directory path.  The end result is that I wind up
> with ~/objects/ etc. and a failed repository clone.  (Simply adding a
> bare getenv("GIT_DIR") afterwards to reset the pointer also corrects
> the problem.)
>
> Since other platforms are apparently working, yet this getenv behavior
> is supported by the standards, I am left wondering if this could be a
> symptom of something else being broken on my platform (z/OS).

The execve(2) function

       int execve(const char *filename, char *const argv[],
                  char *const envp[]);

takes a NULL terminated array of NUL terminated strings of form "VAR=VAL" in envp[], and this is kept in:

	extern char **environ;

of the new image that runs.

The most naive and straight-forward way to implement getenv(3) is to iterate over this environ[] array to look for an element that begins with "GIT_DIR=", and return the pointer pointing at the location one byte past that equal sign. So even if the standard allowed the returned value to be volatile across calls to getenv(3), it will take *more* work for implementations if they want to break our use pattern. They have to deliberately return a string that they will overwrite in subsequent calls to getenv(3).

Also the natural way to implement putenv(3) and setenv(3) is to replace the pointer in the environ[] array (not overwrite the existing string in environ[] that holds the "VAR=VAL" string that represents the current value, which might be shorter than the new value of the enviornment variable), hence even calling these functions is unlikely to invalidate the result you previously received from getenv(3).

I am not at all surprised that nobody from other platforms has seen this breakage.

In fact,

    http://pubs.opengroup.org/onlinepubs/9699919799/functions/getenv.html

says that only setenv(), unsetenv() and putenv() may invalidate previous return values. Note that getenv() is not listed as a function that is allowed to break return values from a previous call to getenv().

So in short, your platform's getenv(3) emulation may be broken. We have other calls to getenv() where we rely on the result of it being stable, and you might discover these places also break.

Having said that, we do have codepaths to update a handful of environment variables ourselves (GIT_DIR is among them), so I think your patch is a good safety measure in general.

Show 25 quoted lines

>  setup.c | 8 ++++++--
>  1 file changed, 6 insertions(+), 2 deletions(-)
>
> diff --git a/setup.c b/setup.c
> index f108c4b..64fb160 100644
> --- a/setup.c
> +++ b/setup.c
> @@ -675,8 +675,12 @@ static const char
> *setup_git_directory_gently_1(int *nongit_ok)
>       * validation.
>       */
>      gitdirenv = getenv(GIT_DIR_ENVIRONMENT);
> -    if (gitdirenv)
> -        return setup_explicit_git_dir(gitdirenv, cwd, len, nongit_ok);
> +    if (gitdirenv) {
> +        gitdirenv = xstrdup(gitdirenv);
> +        ret = setup_explicit_git_dir(gitdirenv, cwd, len, nongit_ok);
> +        free(gitdirenv);
> +        return ret;
> +    }
>
>      if (env_ceiling_dirs) {
>          string_list_split(&ceiling_dirs, env_ceiling_dirs, PATH_SEP, -1);
> --
> 1.7.11.7