Re: [PATCH] bind_textdomain_codeset: don't return failure unless encoding isn't UTF-8

[Rich Felker <dalias@libc.org>]

On Sat, Dec 03, 2016 at 09:04:42PM -0600, Laine Gholson wrote:
> returning null broke a vlc media player built with gettext support

> >From 2f79aa294db5d9230ad71298e3de4b5561b441be Mon Sep 17 00:00:00 2001
> From: Laine Gholson <laine.gholson@gmail.com>
> Date: Wed, 9 Nov 2016 20:19:00 -0600
> Subject: [PATCH] bind_textdomain_codeset: don't return failure unless encoding isn't UTF-8
> 
> VLC isn't happy when bind_textdomain_codeset returns NULL
> ---
>  src/locale/bind_textdomain_codeset.c | 4 +++-
>  1 file changed, 3 insertions(+), 1 deletion(-)
> 
> diff --git a/src/locale/bind_textdomain_codeset.c b/src/locale/bind_textdomain_codeset.c
> index 5ebfd5e..e5f3f52 100644
> --- a/src/locale/bind_textdomain_codeset.c
> +++ b/src/locale/bind_textdomain_codeset.c
> @@ -5,7 +5,9 @@
>  
>  char *bind_textdomain_codeset(const char *domainname, const char *codeset)
>  {
> -	if (codeset && strcasecmp(codeset, "UTF-8"))
> +	if (codeset && ((strcasecmp(codeset, "UTF-8") == 0) || (strcasecmp(codeset, "UTF8") == 0))) {
> +		return "UTF-8";
> +	} else if (codeset)
>  		errno = EINVAL;
>  	return NULL;
>  }
> -- 
> 2.10.2

I think this needs some more thought. The documentation of the API is
that a null pointer argument/result means "the locale's character
encoding", and that the default is null; presumably even when the
locale's codeset is "foo", null (default) and "foo" are still
different states.

I don't actually like that, and don't think we should copy it --
especially since, now that we also have a C locale with "ASCII" as the
codeset, we _can't_ provide a codeset matching the locale in all cases
-- but I also don't think it's right for the return value (null or
"UTF-8") to depend on the argument rather than on the "previous state"
like it's documented to.

There seem to be two possible reasonable behaviors:

1. Diverge from the GNU behavior and treat textdomains as always-bound
   to "UTF-8", regardless of whether bind_textdomain_codeset has been
   called. The function would then return a null pointer with EINVAL
   set for strings other than "UTF-8"/"UTF8", and would return "UTF-8"
   for a valid or null-pointer argument.

2. Keep a 1-bit state for each textdomain reflecting whether its
   nominally in "default" mode or "UTF-8" mode. Either way the
   original UTF-8 string would be returned; the only point of the
   state would be providing a return value for bind_textdomain_codeset
   that reflects how it was previously called.

Being that 2 is gratuitous complexity to do something stupid and
meaningless, I'd lean towards 1, but I don't want to break anything
that works. Does this seem safe to do?

Rich