[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20250527-fearless-impala-of-poetry-8acc3a@l-nschier-aarch64>
Date: Tue, 27 May 2025 16:01:56 +0200
From: Nicolas Schier <nicolas.schier@...ux.dev>
To: Thomas Weißschuh <thomas.weissschuh@...utronix.de>
Cc: Masahiro Yamada <masahiroy@...nel.org>,
Nathan Chancellor <nathan@...nel.org>,
Andrew Morton <akpm@...ux-foundation.org>, Willy Tarreau <w@....eu>,
Thomas Weißschuh <linux@...ssschuh.net>,
Brendan Higgins <brendan.higgins@...ux.dev>,
David Gow <davidgow@...gle.com>, Rae Moar <rmoar@...gle.com>,
Shuah Khan <shuah@...nel.org>, Jonathan Corbet <corbet@....net>,
Christophe Leroy <christophe.leroy@...roup.eu>,
linux-kbuild@...r.kernel.org, linux-kernel@...r.kernel.org,
linux-kselftest@...r.kernel.org, kunit-dev@...glegroups.com,
linux-doc@...r.kernel.org
Subject: Re: [PATCH v2 01/11] kbuild: userprogs: add nolibc support
On Mon, May 26, 2025 at 04:40:17PM +0200, Thomas Weißschuh wrote:
> On Mon, May 26, 2025 at 04:19:53PM +0200, Nicolas Schier wrote:
> > On Mon, Apr 07, 2025 at 09:42:38AM +0200, Thomas Weißschuh wrote:
> > > Userprogs are built with the regular kernel compiler $CC.
> > > A kernel compiler does not necessarily contain a libc which is required
> > > for a normal userspace application.
> > > However the kernel tree does contain a minimal libc implementation
> > > "nolibc" which can be used to build userspace applications.
> > >
> > > Introduce support to build userprogs against nolibc instead of the
> > > default libc of the compiler, which may not exist.
> > >
> > > Signed-off-by: Thomas Weißschuh <thomas.weissschuh@...utronix.de>
> > >
> > > ---
> > > This could probably be moved out of the generic kbuild makefiles.
> > > I think the ergonimics would suffer and this functionality could be
> > > used by other users of userprogs.
> > >
> > > Also this does currently not support out-of-tree builds.
> >
> > (out-of-tree == external kmods; out-of-source == build-dir != source-dir)
> >
> > you probably meant out-of-source.
>
> I *did* mean out-of-tree.
>
> Out-of-source already works with the current patchset. It is the default setup of kunit.py.
>
> > > For that tools/include/nolibc/*.h and usr/include/*.h would need to be
> > > installed into the build directory.
> >
> > Out-of-source builds could be achieved by adding 'headers' as
> > dependency, see below.
> >
> > > ---
> > > Documentation/kbuild/makefiles.rst | 12 ++++++++++++
> > > scripts/Makefile.userprogs | 16 +++++++++++++---
> > > 2 files changed, 25 insertions(+), 3 deletions(-)
> > >
> > > diff --git a/Documentation/kbuild/makefiles.rst b/Documentation/kbuild/makefiles.rst
> > > index 3b9a8bc671e2e92126857059e985d6e5b2c43fd4..f905a6f77c965311c491cfd7ab3103185af7e82e 100644
> > > --- a/Documentation/kbuild/makefiles.rst
> > > +++ b/Documentation/kbuild/makefiles.rst
> > > @@ -970,6 +970,18 @@ When linking bpfilter_umh, it will be passed the extra option -static.
> > >
> > > From command line, :ref:`USERCFLAGS and USERLDFLAGS <userkbuildflags>` will also be used.
> > >
> > > +Bulding userprogs against nolibc
> >
> > Bulding -> Building
>
> Ack.
>
> > > +--------------------------------
> > > +
> > > +Not all kernel toolchains provide a libc.
> > > +Simple userprogs can be built against a very simple libc call "nolibc" provided
> > > +by the kernel source tree.
> > > +
> > > +Example::
> > > +
> > > + # lib/kunit/Makefile
> > > + uapi-preinit-nolibc := $(CONFIG_ARCH_HAS_NOLIBC)
> > > +
> > > When userspace programs are actually built
> > > ------------------------------------------
> > >
> > > diff --git a/scripts/Makefile.userprogs b/scripts/Makefile.userprogs
> > > index f3a7e1ef3753b54303718fae97f4b3c9d4eac07c..a1447c02b948901631098b585f5cf4d3ea383a57 100644
> > > --- a/scripts/Makefile.userprogs
> > > +++ b/scripts/Makefile.userprogs
> > > @@ -16,10 +16,20 @@ user-csingle := $(addprefix $(obj)/, $(user-csingle))
> > > user-cmulti := $(addprefix $(obj)/, $(user-cmulti))
> > > user-cobjs := $(addprefix $(obj)/, $(user-cobjs))
> > >
> > > +user-libgcc := $(call try-run,$(CC) -Werror $(KBUILD_USERCFLAGS) -lgcc -x c -shared /dev/null -o "$$TMP",-lgcc)
> > > +
> > > +user_nolibc_ccflags := -nostdlib -nostdinc -static -fno-ident -fno-asynchronous-unwind-tables \
> > > + -ffreestanding -fno-stack-protector \
> > > + -isystem $(objtree)/usr/include -include $(srctree)/tools/include/nolibc/nolibc.h -isystem $(srctree)/tools/include/nolibc/
> > > +user_nolibc_ldflags := -nostdlib -nostdinc -static
> > > +user_nolibc_ldlibs := $(user-libgcc)
> > > +
> > > user_ccflags = -Wp,-MMD,$(depfile) $(KBUILD_USERCFLAGS) $(userccflags) \
> > > - $($(target-stem)-userccflags)
> > > -user_ldflags = $(KBUILD_USERLDFLAGS) $(userldflags) $($(target-stem)-userldflags)
> > > -user_ldlibs = $(userldlibs) $($(target-stem)-userldlibs)
> > > + $($(target-stem)-userccflags) $(if $($(target-stem)-nolibc),$(user_nolibc_ccflags))
> > > +user_ldflags = $(KBUILD_USERLDFLAGS) $(userldflags) $($(target-stem)-userldflags) \
> > > + $(if $($(target-stem)-nolibc),$(user_nolibc_ldflags))
> > > +user_ldlibs = $(userldlibs) $($(target-stem)-userldlibs) \
> > > + $(if $($(target-stem)-nolibc),$(user_nolibc_ldlibs))
> > >
> > > # Create an executable from a single .c file
> > > quiet_cmd_user_cc_c = CC [U] $@
> >
> > Adding another hunk for scripts/Makefile.userprogs would allow to build
> > out-of-source:
> >
> > @@ -39,5 +49,5 @@ $(call multi_depend, $(user-cmulti), , -objs)
> > # Create .o file from a .c file
> > quiet_cmd_user_cc_o_c = CC [U] $@
> > cmd_user_cc_o_c = $(CC) $(user_ccflags) -c -o $@ $<
> > -$(user-cobjs): $(obj)/%.o: $(src)/%.c FORCE
> > +$(user-cobjs): $(obj)/%.o: $(src)/%.c headers FORCE
> > $(call if_changed_dep,user_cc_o_c)
> >
> > But I am unsure if it is ok to add 'headers' as a build dependency for
> > userprogs. For me, it feels a bit odd, but I think it really makes
> > sense here.
>
> Currently this dependency is encoded in Kconfig.
> If CONFIG_HEADERS_INSTALL=y then the headers are installed in the 'prepare'
> phase and already available when building any userprog.
> To me this seems like the easier and nicer implementation.
I am sure, I had an out-of-source test build that failed due to missing
header files -- but I can't reproduce it any more and yes,
CONFIG_HEADERS_INSTALL should really be enough and better in several
ways.
Sorry for the noice.
Kind regards,
Nicolas
Powered by blists - more mailing lists