On 16 Feb 2023, at 7:09 pm, Marco Elver [off-list ref] wrote:

On Thu, Feb 16, 2023 at 07:12AM +0000, Christophe Leroy wrote:

quoted

Le 16/02/2023 à 06:09, Rohan McLure a écrit :

quoted

KCSAN instruments calls to atomic builtins, and will in turn call these
builtins itself. As such, architectures supporting KCSAN must have
compiler support for these atomic primitives.

Since 32-bit systems are unlikely to have 64-bit compiler builtins,
provide a stub for each missing builtin, and use BUG() to assert
unreachability.

In commit 725aea873261 ("xtensa: enable KCSAN"), xtensa implements these
locally. Move these definitions to be accessible to all 32-bit
architectures that do not provide the necessary builtins, with opt in
for PowerPC and xtensa.

Signed-off-by: Rohan McLure <redacted>
Reviewed-by: Max Filippov <jcmvbkbc@gmail.com>

This series should also be addressed to KCSAN Maintainers, shouldn't it ?

KCSAN
M: Marco Elver [off-list ref]
R: Dmitry Vyukov [off-list ref]
L: kasan-dev@googlegroups.com
S: Maintained
F: Documentation/dev-tools/kcsan.rst
F: include/linux/kcsan*.h
F: kernel/kcsan/
F: lib/Kconfig.kcsan
F: scripts/Makefile.kcsan

quoted

---
Previously issued as a part of a patch series adding KCSAN support to
64-bit.
Link: https://lore.kernel.org/linuxppc-dev/167646486000.1421441.10070059569986228558.b4-ty@ellerman.id.au/T/#t (local)
v1: Remove __has_builtin check, as gcc is not obligated to inline
builtins detected using this check, but instead is permitted to supply
them in libatomic:
Link: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108734
Instead, opt-in PPC32 and xtensa.
---
 arch/xtensa/lib/Makefile                              | 1 -
 kernel/kcsan/Makefile                                 | 2 ++
 arch/xtensa/lib/kcsan-stubs.c => kernel/kcsan/stubs.c | 0
 3 files changed, 2 insertions(+), 1 deletion(-)
 rename arch/xtensa/lib/kcsan-stubs.c => kernel/kcsan/stubs.c (100%)

diff --git a/arch/xtensa/lib/Makefile b/arch/xtensa/lib/Makefile
index 7ecef0519a27..d69356dc97df 100644
--- a/arch/xtensa/lib/Makefile
+++ b/arch/xtensa/lib/Makefile

@@ -8,5 +8,4 @@ lib-y += memcopy.o memset.o checksum.o \
     divsi3.o udivsi3.o modsi3.o umodsi3.o mulsi3.o umulsidi3.o \
     usercopy.o strncpy_user.o strnlen_user.o
 lib-$(CONFIG_PCI) += pci-auto.o
-lib-$(CONFIG_KCSAN) += kcsan-stubs.o
 KCSAN_SANITIZE_kcsan-stubs.o := n

diff --git a/kernel/kcsan/Makefile b/kernel/kcsan/Makefile
index 8cf70f068d92..86dd713d8855 100644
--- a/kernel/kcsan/Makefile
+++ b/kernel/kcsan/Makefile

@@ -12,6 +12,8 @@ CFLAGS_core.o := $(call cc-option,-fno-conserve-stack) \
  -fno-stack-protector -DDISABLE_BRANCH_PROFILING

 obj-y := core.o debugfs.o report.o
+obj-$(CONFIG_PPC32) += stubs.o
+obj-$(CONFIG_XTENSA) += stubs.o

Not sure it is acceptable to do it that way.

There should likely be something like a CONFIG_ARCH_WANTS_KCSAN_STUBS in 
KCSAN's Kconfig then PPC32 and XTENSA should select it.

The longer I think about it, since these stubs all BUG() anyway, perhaps
we ought to just avoid them altogether. If you delete all the stubs from
ppc and xtensa, but do this:

| diff --git a/kernel/kcsan/core.c b/kernel/kcsan/core.c
| index 54d077e1a2dc..8169d6dadd0e 100644
| --- a/kernel/kcsan/core.c
| +++ b/kernel/kcsan/core.c
| @@ -1261,7 +1261,9 @@ static __always_inline void kcsan_atomic_builtin_memorder(int memorder)
|  DEFINE_TSAN_ATOMIC_OPS(8);
|  DEFINE_TSAN_ATOMIC_OPS(16);
|  DEFINE_TSAN_ATOMIC_OPS(32);
| +#ifdef CONFIG_64BIT
|  DEFINE_TSAN_ATOMIC_OPS(64);
| +#endif
|  
|  void __tsan_atomic_thread_fence(int memorder);
|  void __tsan_atomic_thread_fence(int memorder)

Does that work?

This makes much more sense. Rather than assume that kcsan is the only
consumer of __atomic_*_8, and stubbing accordingly, we should just
remove its mention from relevant sub-archs.