Re: [PATCH 03/11] config: prefer git_config_string_dup() for temp variables

[Jeff King <peff@peff.net>]

On Sat, Apr 06, 2024 at 09:50:23PM -0400, Eric Sunshine wrote:

> On Sat, Apr 6, 2024 at 9:00 PM Jeff King <peff@peff.net> wrote:
> > In some cases we may use git_config_string() or git_config_pathname() to
> > read a value into a temporary variable, and then either hand off memory
> > ownership of the new variable or free it. These are not potential leaks,
> > since we know that there is no previous value we are overwriting.
> >
> > However, it's worth converting these to use git_config_string_dup() and
> > git_config_pathname_dup(). It makes it easier to audit for leaky cases,
> > and possibly we can get rid of the leak-prone functions in the future.
> > Plus it lets the const-ness of our variables match their expected memory
> > ownership, which avoids some casts when calling free().
> >
> > Signed-off-by: Jeff King <peff@peff.net>
> > ---
> > diff --git a/builtin/config.c b/builtin/config.c
> > @@ -282,11 +282,11 @@ static int format_config(struct strbuf *buf, const char *key_,
> > -                       const char *v;
> > -                       if (git_config_pathname(&v, key_, value_) < 0)
> > +                       char *v = NULL;
> > +                       if (git_config_pathname_dup(&v, key_, value_) < 0)
> >                                 return -1;
> >                         strbuf_addstr(buf, v);
> > -                       free((char *)v);
> > +                       free(v);
> 
> This revised code implies that git_config_pathname_dup() doesn't
> assign allocated memory to `v` in the "failure" case, thus it is safe
> to `return` immediately without calling free(v). However, the
> documentation for git_config_pathname_dup() and cousins doesn't state
> this explicitly, which means the caller needs to peek into the
> implementation of git_config_pathname_dup() to verify that it is safe
> to write code such as the above. Hence, should the documentation be
> updated to explain that `v` won't be modified in the "failure" case?

That's nothing new with my patch, though, is it? The same would apply
for git_config_pathname(). And git_config_string() for that matter. I'm
also struggling to imagine what it _would_ copy into "v" on failure
anyway.

-Peff