Re: [PATCH] selftests: livepatch: unset sub_make_done in case top level Makefile be overwritten

[Petr Mladek <pmladek@suse.com>]

On Tue 2026-06-30 18:12:46, Thomas Weißschuh wrote:
> Hi Zelin,
> 
> On Tue, Jun 30, 2026 at 04:32:16PM +0200, Petr Mladek wrote:
> > Added some people involved in "filechk_makefile" feature into Cc.
> > 
> > For the new people, see the whole thread at
> > https://lore.kernel.org/all/20260525083721.27857-1-zelin.deng@linux.alibaba.com/
> 
> Thanks for the report and forwarding it.
> 
> > On Sun 2026-06-21 22:36:24, Zelin Deng wrote:
> > > 
> > > 在 2026/6/19 22:42, Miroslav Benes 写道:
> > > > On Mon, 25 May 2026, Zelin Deng wrote:
> > > Here're how I reproduce the issue:
> > > 
> > > 1. pull linux-next, reset to HEAD, for example
> > > 3ce97bd3c4f18608335e709c24d6a40e7036cab8 (tag next-20260619)
> > > 
> > > 2. at linux-next tree: make all -j$(nproc) && make modules_install
> > > headers_install -j$(nproc) && make install && reboot
> > > 
> > > 3. at linux-next tree: make kselftest-all
> > > 
> > > 4. top level Makefile in linux-next has been overwritten by
> > > 
> > >     export KBUILD_OUTPUT = .
> > >     export KBUILD_EXTMOD = /home/shiyu.dzl/linux-next
> > >     export KBUILD_EXTMOD_OUTPUT = /home/shiyu.dzl/linux-next
> > >     include /home/shiyu.dzl/linux-next/Makefile
> > > 
> > > it is a stub generated by filechk_makefile.
> > 
> > It seems that "filechk_makefile" has reached mainline in v7.1-rc1,
> > see the commit c9bb03ac2c66bc5aa81b ("kbuild: reduce output spam when
> > building out of tree").
> > 
> > The commit message says:
> > 
> > <paste>
> >     The Makefile is now created even if the build is aborted due to an
> >     unclean working tree. That should not make a difference in practice.
> >  </paste>
> > 
> > It seems that is actually makes a difference in the livepatch selftest
> > case.
> > 
> > Sigh, the Makefile rules are a maze to me. It is possible that
> > we do something wrong in
> > tools/testing/selftests/livepatch/test_modules/Makefile. Namely
> > with:
> > 
> > ifneq ("$(wildcard $(KDIR))", "")
> > 	$(Q)$(MAKE) -C $(KDIR) modules KBUILD_EXTMOD=$(TESTMODS_DIR)
> > endif
> > 
> > But I see "$(MAKE) -C ..." called in many other makefiles.
> > 
> > Best Regards,
> > Petr
> > 
> > > I'm not quite sure that it could be related to my toolchain (like make
> > > version ?), I briefed analysis the root cause on my environment (KDIR
> > > '/lib/modules/7.1.0-next-20260619/build' which actually a symbol link to my
> > > kernel source -> '/home/shiyu.dzl/linux-next'):
> > > 
> > >   1. sub_make_done leaks via environment. The top-level Makefile sets export
> > > sub_make_done := 1 after its first-pass initialization.
> > >   Because it is exported, every child make process inherits it.
> > >   2. livepatch test_modules re-invokes the top-level Makefile. The call
> > > chain is: top-level Makefile → kselftest-% pattern rule →
> > >   tools/testing/selftests/ → livepatch test_modules/Makefile → $(MAKE) -C
> > > $(KDIR) modules KBUILD_EXTMOD=..., which re-enters the top-level Makefile to
> > >   build an external module.
> > >   3. The inherited sub_make_done=1 skips critical initialization. The
> > > top-level Makefile's first-pass block (ifneq ($(sub_make_done),1)) is
> > >   skipped entirely. This block is responsible for correctly parsing
> > > command-line variables and setting up KBUILD_EXTMOD-related paths.
> > >   4. Path mismatch triggers false out-of-tree detection. With the
> > > initialization skipped, srcroot and CURDIR end up with mismatched values
> > > (e.g.,
> > >   absolute vs. relative). The comparison (ifeq ($(srcroot),$(CURDIR)))
> > > fails, so building_out_of_srctree is incorrectly set to 1.
> > >   5. outputmakefile overwrites the source tree's Makefile. Because
> > > building_out_of_srctree is set, the filechk_makefile rule fires and replaces
> > > the
> > >   real top-level Makefile with a generated stub containing KBUILD_OUTPUT,
> > > KBUILD_EXTMOD, and an include directive.
> > > 
> > >   Why make -C tools/testing/selftests/ all is unaffected: it enters the
> > > selftests directory directly without ever executing the top-level Makefile
> > >   first, so sub_make_done is never exported into the environment. When
> > > livepatch test_modules later invokes $(MAKE) -C $(KDIR), the top-level
> > > Makefile
> > >   runs its full initialization normally.
> 
> Can you try this:
> 
> diff --git a/Makefile b/Makefile
> index b9c5792c79e0..f717a4dc96d6 100644
> --- a/Makefile
> +++ b/Makefile
> @@ -1591,10 +1591,10 @@ tools/%: FORCE
>  
>  PHONY += kselftest
>  kselftest: headers
> -	$(Q)$(MAKE) -C $(srctree)/tools/testing/selftests run_tests
> +	$(Q)unset sub_make_done; $(MAKE) -C $(srctree)/tools/testing/selftests run_tests
>  
>  kselftest-%: headers FORCE
> -	$(Q)$(MAKE) -C $(srctree)/tools/testing/selftests $*
> +	$(Q)unset sub_make_done; $(MAKE) -C $(srctree)/tools/testing/selftests $*
>  
>  PHONY += kselftest-merge
>  kselftest-merge:

Heh, this is exactly the same change which Zelin Deng proposed as
well, see
https://lore.kernel.org/all/20260525083721.27857-1-zelin.deng@linux.alibaba.com/

> The same is done for cmd_install.

I guess that this is related to the commit 14ccc638b02f9ec ("kbuild:
cancel sub_make_done for the install target to fix DKMS"). Which
is related to the commit bcf637f54f6d2515d ("kbuild: parse C= and M=
before changing the working directory").

Honestly, it all looks like a wild magic to me. I would need much
better understanding of the Makefile maze to get on top of it.

But it seems that it is a reasonable fix after all. It solves the
original problem. It does not break my workflow. So, I tend to add:

Acked-by: Petr Mladek <pmladek@suse.com>

Best Regards,
Petr