[PATCH 0/2] kbuild: Fallout from moving '-mfentry' check into Kconfig stage

[PATCH 0/2] kbuild: Fallout from moving '-mfentry' check into Kconfig stage

[Nathan Chancellor]

Hi all,

Ricardo reported an issue from moving the compiler test for '-mfentry'
into Kconfig. In testing that patch, I noticed that the kernel.org GCC
builds for s390 were not probing correctly for '-mfentry' support.

The first patch moves the check for '-mfentry' back into Kbuild where we
have the proper information for whether this flag is actually supported.

The second patch documents the reasons for these issues for future
travelers to hopefully avoid these problems.

I plan to take this via kbuild-fixes after sufficient review/comments.

---
Nathan Chancellor (2):
      kbuild: Move testing of '-mfentry' back to Makefile from Kconfig
      scripts/Kconfig.toolchain: Highlight cc-option differences between Kconfig and Kbuild

 Makefile                  |  3 ++-
 scripts/Kconfig.toolchain | 23 +++++++++++++++++++----
 2 files changed, 21 insertions(+), 5 deletions(-)
---
base-commit: 997208ce9f43f01febac956cf9aeffe89945e3c1
change-id: 20261020-kconfig-fix-mfentry-check-50525bc5a6ca

Best regards,
--  
Cheers,
Nathan

[PATCH 1/2] kbuild: Move testing of '-mfentry' back to Makefile from Kconfig

[Nathan Chancellor]

Unlike cc-option in Kbuild, cc-option in Kconfig does not test options
with either bit ('-m32' / '-m64') or endianness ('-mbig-endian' /
'-mlittle-endian') flags because which flag to use may depend on the
result of the configuration stage. This difference changes the result of
testing for '-mfentry' support when using the kernel.org s390 cross
compiler, which is configured for 32-bit by default.

  $ s390-linux-gcc -mfentry -o /dev/null -x c -c /dev/null
  cc1: error: '-mfentry' is supported only for 64-bit CPUs

arch/s390/Makefile sets '-m64' in KBUILD_CPPFLAGS early, so '-mfentry'
support was properly tested before the move to Kconfig.

  $ s390-linux-gcc -m64 -mfentry -o /dev/null -x c -c /dev/null

Ricardo reported another issue with moving this check from Kbuild to
Kconfig, as '-mfentry' is not supported for x86_32 without '-fno-pic',
which affects GCC builds that are configured with PIE enabled by
default. arch/x86/Makefile adds '-fno-pic' to KBUILD_CFLAGS, so this
works properly in Kbuild.

Move the '-mfentry' check back into Kbuild to avoid adding additional
complexity to the Kconfig entry.

Fixes: 454e00cf8ee5 ("kbuild: avoid re-running compiler and linker probes")
Reported-by: Ricardo Ribalda <ribalda@chromium.org>
Closes: https://lore.kernel.org/20261020-pie-v1-1-9be09f2324b1@chromium.org/
Signed-off-by: Nathan Chancellor <nathan@kernel.org>
---
Cc: linux-s390@vger.kernel.org
---
 Makefile                  | 3 ++-
 scripts/Kconfig.toolchain | 4 ----
 2 files changed, 2 insertions(+), 5 deletions(-)

diff --git a/Makefile b/Makefile
index 8d23af898d44..f123ac983e62 100644
--- a/Makefile
+++ b/Makefile
@@ -1057,7 +1057,8 @@ ifdef CONFIG_FTRACE_MCOUNT_USE_RECORDMCOUNT
   endif
 endif
 ifdef CONFIG_HAVE_FENTRY
-  ifdef CONFIG_CC_HAS_MFENTRY
+  # s390-linux-gnu-gcc did not support -mfentry until gcc-9.
+  ifeq ($(call cc-option-yn, -mfentry),y)
     CC_FLAGS_FTRACE	+= -mfentry
     CC_FLAGS_USING	+= -DCC_USING_FENTRY
   endif
diff --git a/scripts/Kconfig.toolchain b/scripts/Kconfig.toolchain
index 254393d4f98a..01b49fa5d6be 100644
--- a/scripts/Kconfig.toolchain
+++ b/scripts/Kconfig.toolchain
@@ -169,10 +169,6 @@ config CC_OPT_NO_ALLOW_STORE_DATA_RACES
 config CC_HAS_MNOP_MCOUNT
 	def_bool $(cc-option,-mnop-mcount)
 
-# s390-linux-gnu-gcc did not support -mfentry until gcc-9.
-config CC_HAS_MFENTRY
-	def_bool $(cc-option,-mfentry)
-
 config LD_OPT_FORCE_GROUP_ALLOCATION
 	def_string "$(ld-option-str,--force-group-allocation)"
 

-- 
2.56.0

[PATCH 2/2] scripts/Kconfig.toolchain: Highlight cc-option differences between Kconfig and Kbuild

[Nathan Chancellor]

Performing cc-option checks in Kconfig instead of Kbuild can help speed
up incremental builds where the configuration does not change, as the
results of the probes can be easily cached. However, there are two
potential issues that folks need to consider when using cc-option in
Kconfig compared to Kbuild.

The first is that in Kbuild, cc-option additions are cumulative, so

  KBUILD_CFLAGS += $(call cc-option,-mfoo)
  KBUILD_CFLAGS += $(call cc-option,-mbar)

results in '-mbar' being tested with '-mfoo' if it was supported. In
Kconfig, '-mbar' would be tested without '-mfoo', which could change the
result if '-mbar' depends on '-mfoo' being present, so it would need to
test support with

  $(cc-option,-mbar -mfoo)

The second is that the Kconfig stage is where questions like "is the
target architecture 32-bit or 64-bit" or "is the target architecture big
or little endian" are answered, which very much matters for testing
compiler options. Toolchains may be configured for either option, so the
kernel needs to pass compiler flags like '-m32', '-m64',
'-mlittle-endian', and '-mbig-endian' to properly match a user's
configuration choices. Without these flags, '-mfoo' is tested against a
toolchain's configured default, resulting in an inconsistent answer to
whether the flag is actually supported for the requested configuration.
This could manifest as a build breakage from an unexpectedly included
but unsupported flag or missing functionality from an unexpected omitted
flag.

Add a big comment summarizing these two issues to help future travelers.

Signed-off-by: Nathan Chancellor <nathan@kernel.org>
---
 scripts/Kconfig.toolchain | 19 +++++++++++++++++++
 1 file changed, 19 insertions(+)

diff --git a/scripts/Kconfig.toolchain b/scripts/Kconfig.toolchain
index 01b49fa5d6be..a99d875489ee 100644
--- a/scripts/Kconfig.toolchain
+++ b/scripts/Kconfig.toolchain
@@ -144,6 +144,25 @@ config CC_HAS_ASSUME
 config CC_HAS_NO_PROFILE_FN_ATTR
 	def_bool $(success,echo '__attribute__((no_profile_instrument_function)) int x();' | $(CC) -x c - -c -o /dev/null -Werror)
 
+# Beware of two problems when using cc-option in Kconfig compared to Kbuild
+#
+# 1. cc-option in Kbuild tests the provided flags with all previously added
+#    CFLAGS, whereas cc-option in Kconfig only tests the provided flags with a
+#    small, constant set of additional flags.  If a flag (e.g., '-mfoo')
+#    depends on another (e.g, '-fbar') to function properly, the dependency
+#    must be included in the cc-option call (e.g, '-fbar -mfoo').
+#
+# 2. Bit and endianness flags ('-m32', '-m64', '-mbig-endian', and
+#    '-mlittle-endian') are not present in Kconfig cc-option checks because
+#    they may be included in KBUILD_CFLAGS / KBUILD_CPPFLAGS based on the
+#    finalized configuration result. This can result in inconsistent results
+#    when probing toolchains that support either option, as the toolchain's
+#    configured default tested in Kconfig may not support the tested flag but
+#    the post-configuration result in Kbuild would.
+#
+#    In general, prefer adding truly generic compiler flags to avoid this
+#    pitfall.
+
 config CC_OPT_ZERO_INIT_PADDING_BITS
 	def_string "$(cc-option-str,-fzero-init-padding-bits=all)"
 

-- 
2.56.0