Re: [PATCH] fix consistency issues in set_program_name()

[Pádraig Brady]

On 12/01/2025 18:40, Bruno Haible wrote:
> Hi Pádraig,
>
> > I noticed a discrepancy in the program name output by error(),
> > which looks to be due to whether the gnulib or glibc version is used.
> > If we use gnulib's error() it uses getprogname()
> > which defaults to program_invocation_short_name.
> > The glibc error() instead uses program_invocation_name.
> >
> > I don't know why gnulib differs in that regard
>
> Here's the explanation:
>
>    - We want error() to be usable in library code or, at least, in code
>      that does not impose requirements on the main() function.
>      This was done on 2016-08-08 by Pino Toscano.
>
>    - The function that this uses, getprogname(), returns the base name,
>      because we like to have the same behaviour across platforms, and on
>      most platforms the full program name is not available. (Look at the
>      amount of hacks needed to get only the base name!)
>
> The comments in progname.c indicate the intention to keep glibc's
> variables in sync with what the caller is passing to set_program_name().
>
> > I've not looked into why coreutils is using the gnulib error()
> > given that REPLACE_ERROR=0 and HAVE_ERROR=1
>
> It's because of
>
>    m4_ifdef([gl_HAVE_MODULE_VERROR],
>      [COMPILE_ERROR_C=1],
>      ...
>
> in m4/error_h.m4. coreutils apparently uses the module 'verror'.

Ah right thanks.

coreutils has used verror for a long time, but this new dependency
on error is recent and not yet released. It was introduced with:
https://git.sv.gnu.org/cgit/gnulib.git/commit/?id=959152ba37

It will mean a change in behavior for coreutils
with the base name being used for error(),
but the full name being used elsewhere.  For example:

$ $PWD/src/mktemp 1 2
mktemp: too many templates
Try '/home/padraig/git/coreutils/src/mktemp --help' for more information.

$ ~/coreutils-9.5/src/mktemp 1 2
/home/padraig/coreutils-9.5/src/mktemp: too many templates
Try '/home/padraig/coreutils-9.5/src/mktemp --help' for more information.

That's not a problem I think, since any invocation examples
continue to use the full program_name, it's just something to be aware of.

> > but in any case program_invocation_name and program_invocation_short_name
> > should be consistent with each other.
>
> I agree.
>
> > This is not the case because set_program_name()
> > only sets program_invocation_name normally.
>
> Can you describe the circumstances where you observed the discrepancy?

It impacts the tests when we `./configure --enable-single-binary=shebangs`.
In that case "coreutils" will be initialised by glibc to
program_invocation_short_name, and then when we override with
set_program_name(actual_utility), it wasn't updated.
Then various error messages checked in the tests failed to match.

> > The attached patch also sets program_invocation_short_name
> > and results in coreutils' tests passing in all cases.
>
> The patch would look cleaner, I think, if the fix would be placed
> around line 92, not around line 71. Also there's an indentation problem.

Yes that's cleaner.
Updated version attached.

cheers,
Pádraig

progname.patch
Description: Text Data