We have a need to override the definition of barrier_before_unreachable() for MIPS, which means we either need to add architecture-specific code into linux/compiler-gcc.h or we need to allow the architecture to provide a header that can define the macro before the generic definition. The latter seems like the better approach.
A straightforward approach to the per-arch header is to make use of asm-generic to provide a default empty header & adjust architectures which don't need anything specific to make use of that by adding the header to generic-y. Unfortunately this doesn't work so well due to commit a95b37e20db9 ("kbuild: get <linux/compiler_types.h> out of <linux/kconfig.h>") which moved the inclusion of linux/compiler.h to cflags using the -include compiler flag. Because the -include flag is present for all C files we compile, we need the architecture-provided header to be present before any C files are compiled. If any C files can be compiled prior to the asm-generic header wrappers being generated then we hit a build failure due to missing header. Such cases do exist - one pointed out by the kbuild test robot is the compilation of arch/ia64/kernel/nr-irqs.c, which occurs as part of the archprepare target . This leaves us with a few options: 1) Use generic-y & fix any build failures we find by enforcing ordering such that the asm-generic target occurs before any C compilation, such that linux/compiler_types.h can always include the generated asm-generic wrapper which in turn includes the empty asm-generic header. This would rely on us finding all the problematic cases - I don't know for sure that the ia64 issue is the only one. 2) Add an actual empty header to each architecture, so that we don't need the generated asm-generic wrapper. This seems messy. 3) Give up & add #ifdef CONFIG_MIPS or similar to linux/compiler_types.h. This seems messy too. 4) Include the arch header only when it's actually needed, removing the need for the asm-generic wrapper for all other architectures. This patch allows us to use approach 4, by including an asm/compiler_types.h header using the -include flag in the same way we do for linux/compiler_types.h, but only if the header actually exists.  https://lists.01.org/pipermail/kbuild-all/2018-August/051175.html Signed-off-by: Paul Burton <paul.bur...@mips.com> Cc: Arnd Bergmann <a...@arndb.de> Cc: James Hogan <jho...@kernel.org> Cc: Masahiro Yamada <yamada.masah...@socionext.com> Cc: Ralf Baechle <r...@linux-mips.org> Cc: linux-a...@vger.kernel.org Cc: linux-kbu...@vger.kernel.org Cc: linux-m...@linux-mips.org --- Any thoughts anyone? This isn't the prettiest it could possibly be but it's a small change & clearly shouldn't break anything, which are good qualities for a patch fixing build failures that we'd ideally backport as far as 4.16. Changes in v8: - New patch. scripts/Makefile.lib | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/scripts/Makefile.lib b/scripts/Makefile.lib index 1bb594fcfe12..4e7b41ef029b 100644 --- a/scripts/Makefile.lib +++ b/scripts/Makefile.lib @@ -151,8 +151,11 @@ __a_flags = $(call flags,_a_flags) __cpp_flags = $(call flags,_cpp_flags) endif +c_includes = $(wildcard $(srctree)/arch/$(SRCARCH)/include/asm/compiler_types.h) +c_includes += $(srctree)/include/linux/compiler_types.h + c_flags = -Wp,-MD,$(depfile) $(NOSTDINC_FLAGS) $(LINUXINCLUDE) \ - -include $(srctree)/include/linux/compiler_types.h \ + $(addprefix -include ,$(c_includes)) \ $(__c_flags) $(modkern_cflags) \ $(basename_flags) $(modname_flags) -- 2.18.0