Skip to contentSkip to conversation

include/linux: Adjust headers for C23

PATCH In review Jakub Jelinek 2 participants 2 messages Jan 20, 2025

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 include/linux: Adjust headers for C23

GCC 15 changed default from -std=gnu17 to -std=gnu23. In C23 among many other changes bool, false and true are keywords (like in C++), so defining those using typedef or enum is invalid.

The following patch adjusts the include/linux/ headers to be C23 compatible. _Bool and the C23 bool are ABI compatible, false/true have the same values but different types (previously in the kernel case it was an anonymous enum, in C23 it is bool), so if something uses say sizeof(false) or typeof(true), those do change, but I doubt that is used anywhere in the kernel.

The last change is about va_start. In C23 ellipsis can be specified without any arguments before it, like
int foo(...)

{
	va_list ap;
	va_start(ap);
	int ret = va_arg(ap, int);
	va_end(ap);
	return ret;
}

and unlike in C17 and earlier, va_start is macro with variable argument count. Normally one should use it with just one argument or for backwards compatibility with C17 and earlier with two arguments, second being the last named argument. Of course, if there is no last named argument, only the single argument va_start is an option. The stdarg.h change attempts to be compatible with older versions of GCC and with clang as well. Both GCC 13-14 and recent versions of clang define va_start for C23 as
#define va_start(v, ...) __builtin_va_start(v, 0)
The problem with that definition is that it doesn't emit warnings when one uses complete nonsense in there (e.g. va_start(ap, 8) or va_start(ap, +-*, /, 3, 4.0)), so for GCC 15 it uses a different builtin which takes care about warnings for unexpected va_start uses (as suggested by the C23 standard). Hopefully clang will one day implement that too.

Anyway, without these changes, kernel could detect compiler defaulting to C23 and use say -std=gnu17 option instead, but even in that case IMHO this patch doesn't hurt.

---
 include/linux/types.h  |    2 ++
 include/linux/stddef.h |    2 ++
 include/linux/stdarg.h |   10 ++++++++++
 3 files changed, 14 insertions(+)
include/linux/types.h +2 -0
index 2d7b9ae8714c..f62dca96c7f1 100644
@@ -32,7 +32,9 @@ typedef __kernel_timer_t	timer_t;
 typedef __kernel_clockid_t	clockid_t;
 typedef __kernel_mqd_t		mqd_t;
 
+#if !defined(__STDC_VERSION__) || __STDC_VERSION__ < 202311L
 typedef _Bool			bool;
+#endif
 
 typedef __kernel_uid32_t	uid_t;
 typedef __kernel_gid32_t	gid_t;
include/linux/stddef.h +2 -0
index 929d67710cc5..16508c74fca9 100644
@@ -7,10 +7,12 @@
 #undef NULL
 #define NULL ((void *)0)
 
+#if !defined(__STDC_VERSION__) || __STDC_VERSION__ < 202311L
 enum {
 	false	= 0,
 	true	= 1
 };
+#endif
 
 #undef offsetof
 #define offsetof(TYPE, MEMBER)	__builtin_offsetof(TYPE, MEMBER)
include/linux/stdarg.h +10 -0
index c8dc7f4f390c..038214722c6e 100644
@@ -3,7 +3,17 @@
 #define _LINUX_STDARG_H
 
 typedef __builtin_va_list va_list;
+#if defined(__STDC_VERSION__) && __STDC_VERSION__ > 201710L
+#define va_start(v, ...) __builtin_va_start(v, 0)
+#ifdef __has_builtin
+#if __has_builtin(__builtin_c23_va_start)
+#undef va_start
+#define va_start(...)	__builtin_c23_va_start(__VA_ARGS__)
+#endif
+#endif
+#else
 #define va_start(v, l)	__builtin_va_start(v, l)
+#endif
 #define va_end(v)	__builtin_va_end(v)
 #define va_arg(v, T)	__builtin_va_arg(v, T)
 #define va_copy(d, s)	__builtin_va_copy(d, s)
#1 lore raw

Hi Jakub,

Quoted 14 lines · Jakub Jelinek

GCC 15 changed default from -std=gnu17 to -std=gnu23. In C23 among many other changes bool, false and true are keywords (like in C++), so defining those using typedef or enum is invalid.

The following patch adjusts the include/linux/ headers to be C23 compatible. _Bool and the C23 bool are ABI compatible, false/true have the same values but different types (previously in the kernel case it was an anonymous enum, in C23 it is bool), so if something uses say sizeof(false) or typeof(true), those do change, but I doubt that is used anywhere in the kernel.

The last change is about va_start. In C23 ellipsis can be specified without any arguments before it, like
int foo(...)

Quoted 7 lines
{
	va_list ap;
	va_start(ap);
	int ret = va_arg(ap, int);
	va_end(ap);
	return ret;
}

I think this patch should be split into two: one for the _Bool fix and one for the va_start() change. They are related but distinct changes from what I can tell. I would also send this to Andrew Morton <[email protected]>, as it is unlikely somebody will pick this up from LKML directly.

Quoted 14 lines

and unlike in C17 and earlier, va_start is macro with variable argument count. Normally one should use it with just one argument or for backwards compatibility with C17 and earlier with two arguments, second being the last named argument. Of course, if there is no last named argument, only the single argument va_start is an option. The stdarg.h change attempts to be compatible with older versions of GCC and with clang as well. Both GCC 13-14 and recent versions of clang define va_start for C23 as
#define va_start(v, ...) __builtin_va_start(v, 0)
The problem with that definition is that it doesn't emit warnings when one uses complete nonsense in there (e.g. va_start(ap, 8) or va_start(ap, +-*, /, 3, 4.0)), so for GCC 15 it uses a different builtin which takes care about warnings for unexpected va_start uses (as suggested by the C23 standard). Hopefully clang will one day implement that too.

Quoted 3 lines

Anyway, without these changes, kernel could detect compiler defaulting to C23 and use say -std=gnu17 option instead, but even in that case IMHO this patch doesn't hurt.

The kernel already uses '-std=gnu11' in the general case:

  $ rg gnu11 Makefile
  465:                     -O2 -fomit-frame-pointer -std=gnu11
  579:KBUILD_CFLAGS += -std=gnu11

Unfortunately, KBUILD_CFLAGS gets recreated in certain places without it [1], which is how the issue brought up here and in that thread even happens. I had suggested something a little more lofty as a solution there and while Masahiro did not like the variable aspect of it, we could still just add '-std=gnu11' to the places that need it, which might just be this diff (I do not have a copy of GCC 15 readily available to test).

arch/x86/boot/compressed/Makefile +1 -0
index f2051644de94..606c74f27459 100644
@@ -25,6 +25,7 @@ targets := vmlinux vmlinux.bin vmlinux.bin.gz vmlinux.bin.bz2 vmlinux.bin.lzma \
 # avoid errors with '-march=i386', and future flags may depend on the target to
 # be valid.
 KBUILD_CFLAGS := -m$(BITS) -O2 $(CLANG_FLAGS)
+KBUILD_CFLAGS += -std=gnu11
 KBUILD_CFLAGS += -fno-strict-aliasing -fPIE
 KBUILD_CFLAGS += -Wundef
 KBUILD_CFLAGS += -DDISABLE_BRANCH_PROFILING
drivers/firmware/efi/libstub/Makefile +1 -1
index ed4e8ddbe76a..1141cd06011f 100644
@@ -11,7 +11,7 @@ cflags-y			:= $(KBUILD_CFLAGS)
 
 cflags-$(CONFIG_X86_32)		:= -march=i386
 cflags-$(CONFIG_X86_64)		:= -mcmodel=small
-cflags-$(CONFIG_X86)		+= -m$(BITS) -D__KERNEL__ \
+cflags-$(CONFIG_X86)		+= -m$(BITS) -D__KERNEL__ -std=gnu11 \
 				   -fPIC -fno-strict-aliasing -mno-red-zone \
 				   -mno-mmx -mno-sse -fshort-wchar \
 				   -Wno-pointer-sign \

Your solution does not seem too bad either though, especially since we will need it if/when the kernel moves to gnu23.

[1]: https://lore.kernel.org/20241119041550.GA573925@thelio-3990X/

Quoted 43 lines
---
 include/linux/types.h  |    2 ++
 include/linux/stddef.h |    2 ++
 include/linux/stdarg.h |   10 ++++++++++
 3 files changed, 14 insertions(+)
include/linux/types.h +2 -0
index 2d7b9ae8714c..f62dca96c7f1 100644
@@ -32,7 +32,9 @@ typedef __kernel_timer_t	timer_t;
 typedef __kernel_clockid_t	clockid_t;
 typedef __kernel_mqd_t		mqd_t;
 
+#if !defined(__STDC_VERSION__) || __STDC_VERSION__ < 202311L
 typedef _Bool			bool;
+#endif
 
 typedef __kernel_uid32_t	uid_t;
 typedef __kernel_gid32_t	gid_t;
include/linux/stddef.h +2 -0
index 929d67710cc5..16508c74fca9 100644
@@ -7,10 +7,12 @@
 #undef NULL
 #define NULL ((void *)0)
 
+#if !defined(__STDC_VERSION__) || __STDC_VERSION__ < 202311L
 enum {
 	false	= 0,
 	true	= 1
 };
+#endif
 
 #undef offsetof
 #define offsetof(TYPE, MEMBER)	__builtin_offsetof(TYPE, MEMBER)
include/linux/stdarg.h +0 -0
index c8dc7f4f390c..038214722c6e 100644

include/linux/stdarg.h line 8

Quoted 6 lines
include/linux/stdarg.h +3 -0
@@ -3,7 +3,17 @@
 #define _LINUX_STDARG_H
 
 typedef __builtin_va_list va_list;
+#if defined(__STDC_VERSION__) && __STDC_VERSION__ > 201710L
+#define va_start(v, ...) __builtin_va_start(v, 0)
+#ifdef __has_builtin

This #ifdef should not be necessary because __has_builtin is defined to 0 in compiler_types.h if not defined by the compiler, which should be included via the command line for every C file:

  $ rg compiler_types.h scripts/Makefile.lib
  244:             -include $(srctree)/include/linux/compiler_types.h
Quoted 13 lines
include/linux/stdarg.h +7 -0
+#if __has_builtin(__builtin_c23_va_start)
+#undef va_start
+#define va_start(...)	__builtin_c23_va_start(__VA_ARGS__)
+#endif
+#endif
+#else
 #define va_start(v, l)	__builtin_va_start(v, l)
+#endif
 #define va_end(v)	__builtin_va_end(v)
 #define va_arg(v, T)	__builtin_va_arg(v, T)
 #define va_copy(d, s)	__builtin_va_copy(d, s)