Re: [PATCH v5] scripts/config: honor $KBUILD_OUTPUT by default

[Dmitry Voytik]

Hi Nathan,

Thanks again for looking into this.

On Mon, Oct 5, 2026 at 5:56 PM Nathan Chancellor <nathan@xxxxxxxxxx> wrote:
>
> Hi Dmitry,
>
> > If $KBUILD_OUTPUT is set, use it by default if --file is not passed.
> >
> > Fix the issue when the script silently updates a stale .config in
> > the root directory, while the user expects ${KBUILD_OUTPUT}/.config
> > to be updated.
> > Handle properly if the config file is a relative path.
> >
> > The problem can be worked around with --file, but this is easy to miss,
> > and is inconsistent with how the Makefile and other tools work
> > (e.g., scripts/diffconfig).
> >
> > How to reproduce the issue:
> >       rm .config
> >       export KBUILD_OUTPUT=.out
> >       make defconfig
> >       ./scripts/config \
> >              -e CONFIG_UBSAN
> >       grep: .config: No such file or directory
> >
> > How the patch was tested:
> >       make mrproper
> >       export KBUILD_OUTPUT=.out
> >       make defconfig
> >       grep "CONFIG_UBSAN=y" $KBUILD_OUTPUT/.config || echo "OK"
> >
> > Test the basic usage:
> >       ./scripts/config -e CONFIG_UBSAN
> >       grep "CONFIG_UBSAN=y" $KBUILD_OUTPUT/.config && echo "OK"
> >
> > Test --refresh picks up correct config file by using $KBUILD_OUTPUT:
> >       ./scripts/config -d CONFIG_UBSAN
> >       ./scripts/config --refresh -e CONFIG_UBSAN
> >       grep "CONFIG_UBSAN=y" $KBUILD_OUTPUT/.config && echo "OK"
> >
> > Test --file takes precedence over $KBUILD_OUTPUT and a relative
> > path is handled properly:
> >       ./scripts/config -d CONFIG_UBSAN
> >       mkdir .out2
> >       cp .out/.config .out2/.my_config
> >       grep "CONFIG_UBSAN=y" .out2/.my_config
> >       ./scripts/config --file .out2/.my_config --refresh \
> >              -e CONFIG_UBSAN
> >       grep "CONFIG_UBSAN=y" .out2/.my_config && echo "OK"
> >       ls .out/.out2 || echo "OK"
> >       ls .out/.out || echo "OK"
> >
> > Test --file can accept files with hyphen:
> >       ./scripts/config -d CONFIG_UBSAN
> >       cp .out/.config -- -my_config
> >       grep "CONFIG_UBSAN=y" -- -my_config || echo "OK"
> >       ./scripts/config  --refresh \
> >              -e CONFIG_UBSAN --file -my_config
> >       grep "CONFIG_UBSAN=y" -- -my_config && echo "OK"
> >
> > Signed-off-by: Dmitry Voytik <voytikd@xxxxxxxxx>
> > ---
> > v5:
> >  - Per Nathan Chancellor's suggestion, improved handling $FN path by
> >    realpath
> >  - https://lore.kernel.org/lkml/179097210878.309835.10523307262544578337.b4-review@b4/
> >  - Documented the new test scenario in the git commit message
> > v4:
> >  - Addressed Nicolas Schier's comments - dropped the change how '--refresh'
> >    is handled
> >  - Implemented conversion of $FN to the absolute path which helps with
> >    '--refresh' and relative paths
> >  - Documented testing steps in the commit message
> > v3:
> >  - Added an example in the commit message how to reproduce the issue
> > v2:
> >  - Addressed comments by sashiko:
> >    - https://sashiko.dev/#/patchset/20260920064350.17999-1-voytikd%40gmail.com
> >    - use basename when --refresh is used
> >    - remove echo which can be misleading when --file is passed
> > v1:
> >  - https://lore.kernel.org/all/20260920064350.17999-1-voytikd@xxxxxxxxx/
> > ---
> >  scripts/config | 5 +++++
> >  1 file changed, 5 insertions(+)
> >
> > diff --git a/scripts/config b/scripts/config
> > index 32428ea909c2..7e84c3d4e621 100755
> > --- a/scripts/config
> > +++ b/scripts/config
> > @@ -130,6 +130,10 @@ on_exit() {
> >  trap on_exit EXIT
> >
> >  FN=.config
> > +if [ -n "${KBUILD_OUTPUT}" ]; then
> > +     FN=${KBUILD_OUTPUT}/${FN}
> > +fi
> > +
> >  CMDS=()
> >  while [[ $# -gt 0 ]]; do
> >       if [ "$1" = "--file" ]; then
> > @@ -143,6 +147,7 @@ while [[ $# -gt 0 ]]; do
> >               shift
> >       fi
> >  done
> > +FN=$(realpath -q -- "$FN" || echo "$FN")
>
> Sashiko has some further complaints about this:
>
>   https://sashiko.dev/#/patchset/20261002222119.393508-1-voytikd@xxxxxxxxx

I saw this, but I thought other *NIX environments were not supported
-- only those based on GNU coreutils.
Sashiko's comment on filename that can break echo was a bit contrived IMHO.
Anyhow, lesson learned, next time I'll make sure there is zero
sashiko's complaints :)

> Due to travel and Linux Plumbers,

Have fun!

> I am not sure I will be able to think
> about the best way to resolve this before I finalize the Kbuild tree for
> 7.4, so this may need to be deferred until after 7.4-rc1 is released
> (i.e., it will target 7.5).

I understand. Thank you for your work!

> One option that immediately comes to mind is
> requiring $KBUILD_OUTPUT to exist already and only running realpath on
> $KBUILD_OUTPUT, which would leave the script unchanged for the case of
> $KBUILD_OUTPUT being unset, which seems safer to me:
>
>   if [ -n "$KBUILD_OUTPUT" ]; then
>       if [ ! -d "$KBUILD_OUTPUT" ]; then
>           echo "KBUILD_OUTPUT is set but does not exist or is not a directory, create it before running $myname" >&2
>           exit 1
>       fi
>       FN=$(realpath "$KBUILD_OUTPUT")/$FN
>   fi

I like this approach. Let me try that. If sashiko doesn't complain,
I'll post a v6 patch.

> This would leave a relative '--file' value + KBUILD_OUTPUT + '--refresh'
> in a weird state still but it appears that commit f757f6011c92 ("kbuild:
> fix argument parsing in scripts/config") got that wrong almost two years
> ago so it doesn't feel like it is worth caring about that until someone
> actually complains about it (but I might be missing something).
>
> --
> Cheers,
> Nathan

Thanks and best regards,
Dmitry Voytik.