On Tue, Oct 13, 2026 at 10:38:37PM +0200, Thomas Weißschuh wrote:
On 2026-10-11 19:36:41-0400, Sasha Levin wrote:
(...)
Assisted-by: LLM
Signed-off-by: Sasha Levin <[email protected]>
---
.gitignore | 5 +++
Makefile | 8 ++++-
lib/Kconfig.debug | 2 ++
tools/Kconfig | 72 +++++++++++++++++++++++++++++++++++++++
tools/Makefile.kbuild | 79 +++++++++++++++++++++++++++++++++++++++++++
5 files changed, 165 insertions(+), 1 deletion(-)
create mode 100644 tools/Kconfig
create mode 100644 tools/Makefile.kbuild
(...)
diff --git a/tools/Kconfig b/tools/Kconfig
new file mode 100644
index 0000000000000..bf0643b39b86e
--- /dev/null
+++ b/tools/Kconfig
@@ -0,0 +1,72 @@
+# SPDX-License-Identifier: GPL-2.0-only
+
+config TOOLS_BUILD_ABI
+ def_bool X86_64 || ARM64
+
+config TOOLS_BUILD_GLIBC
+ def_bool $(success,echo '#include <features.h>' | $(CC) $(CLANG_FLAGS)
$(USERCFLAGS) -E -dM -x c - | grep -q __GLIBC__)
+
+menuconfig TOOLS_BUILD
+ bool "Build userspace tools from tools/"
+ depends on CC_CAN_LINK && !UML
Why !UML?
Building with ARCH=um was reaching the tools, which then set "-I
tools/arch/um/include" but that dir doesn't exist.
I worked around this for the RFC, but I think I can address it when I resent
the patches.
+ depends on TOOLS_BUILD_ABI && TOOLS_BUILD_GLIBC
These dependencies could be on separate lines, it will reduce churn when
they get modified.
ack
+ help
+ Build the selected programs under tools/ for the target together
+ with the kernel, as part of "make all", or alone with
+ "make tools_build". Each program is built by its own build system
+ into tools-build/ in the object tree. The tools then always match
+ the kernel they were built with, and a change that breaks one of
+ them shows up in the same build that compiles the kernel.
+
+ This needs a compiler that can link glibc programs for the target,
+ and is currently limited to x86_64 and arm64.
+
+ If unsure, say N.
(...)
diff --git a/tools/Makefile.kbuild b/tools/Makefile.kbuild
new file mode 100644
index 0000000000000..820a2d9a3893b
--- /dev/null
+++ b/tools/Makefile.kbuild
@@ -0,0 +1,79 @@
+# SPDX-License-Identifier: GPL-2.0-only
+#
+# Builds the userspace programs selected in tools/Kconfig for the target.
+# Invoked by "make tools_build", which "make all" runs when CONFIG_TOOLS_BUILD
+# is set. Each program is built by its own build system, the way it is built
+# standalone, into $(objtree)/tools-build/.
+
+PHONY := __tools_build
+__tools_build:
+
+include include/config/auto.conf
+
+abs_srctree := $(realpath $(srctree))
+toolsdir := $(abs_srctree)/tools
+outdir := $(CURDIR)/tools-build
+
+# kbuild's state and the user's compiler flags, from the environment or the
+# command line, would change how these build
+unexport srctree objtree srcroot VPATH building_out_of_srctree sub_make_done
+unexport CLANG_FLAGS INSTALL_HDR_PATH O
+unexport CFLAGS CPPFLAGS LDFLAGS EXTRA_CFLAGS EXTRA_LDFLAGS USERCFLAGS
USERLDFLAGS
+space := $(subst ,, )
+overrides := $(subst \$(space),\@space@,$(MAKEOVERRIDES))
+MAKEOVERRIDES := $(subst \@space@,\$(space),$(filter-out O=% CFLAGS=%
CPPFLAGS=% \
+ LDFLAGS=% EXTRA_CFLAGS=% EXTRA_LDFLAGS=% USERCFLAGS=%
USERLDFLAGS=%,$(overrides)))
+
+# Keep Python's bytecode caches out of the source tree
+export PYTHONDONTWRITEBYTECODE := 1
This should just be done in the toplevel Makefile IMO.
I originally added this because perf was silly, but apparently that was a bug
instead of intended behavior, which was fixed in 5479e1a13b759 ("perf tools:
Put Python bytecode in output directory"), so I think I can just drop this
altogether.
--
Thanks,
Sasha
