Clang and GCC behave a little differently when it comes to the
__no_sanitize_thread attribute, which has valid reasons, and depending
on context either one could be right.
Traditionally, user space ThreadSanitizer [1] still expects instrumented
builtin atomics (to avoid false positives) and __tsan_func_{entry,exit}
(to generate meaningful stack traces), even if the function has the
attribute no_sanitize("thread").
[1] https://clang.llvm.org/docs/ThreadSanitizer.html#attribute-no-sanitize-thread
GCC doesn't follow the same policy (for better or worse), and removes
all kinds of instrumentation if no_sanitize is added. Arguably, since
this may be a problem for user space ThreadSanitizer, we expect this may
change in future.
Since KCSAN != ThreadSanitizer, the likelihood of false positives even
without barrier instrumentation everywhere, is much lower by design.
At least for Clang, however, to fully remove all sanitizer
instrumentation, we must add the disable_sanitizer_instrumentation
attribute, which is available since Clang 14.0.
Signed-off-by: Marco Elver <elver@xxxxxxxxxx>
---
v3:
* New patch.
---
include/linux/compiler_types.h | 13 ++++++++++++-
lib/Kconfig.kcsan | 2 +-
2 files changed, 13 insertions(+), 2 deletions(-)
diff --git a/include/linux/compiler_types.h b/include/linux/compiler_types.h
index 1d32f4c03c9e..3c1795fdb568 100644
--- a/include/linux/compiler_types.h
+++ b/include/linux/compiler_types.h
@@ -198,9 +198,20 @@ struct ftrace_likely_data {
# define __no_kasan_or_inline __always_inline
#endif
-#define __no_kcsan __no_sanitize_thread
#ifdef __SANITIZE_THREAD__
+/*
+ * Clang still emits instrumentation for __tsan_func_{entry,exit}() and builtin
+ * atomics even with __no_sanitize_thread (to avoid false positives in userspace
+ * ThreadSanitizer). The kernel's requirements are stricter and we really do not
+ * want any instrumentation with __no_kcsan.
+ *
+ * Therefore we add __disable_sanitizer_instrumentation where available to
+ * disable all instrumentation. See Kconfig.kcsan where this is mandatory.
+ */
+# define __no_kcsan __no_sanitize_thread __disable_sanitizer_instrumentation
# define __no_sanitize_or_inline __no_kcsan notrace __maybe_unused
+#else
+# define __no_kcsan
#endif
#ifndef __no_sanitize_or_inline
diff --git a/lib/Kconfig.kcsan b/lib/Kconfig.kcsan
index e4394ea8068b..63b70b8c5551 100644
--- a/lib/Kconfig.kcsan
+++ b/lib/Kconfig.kcsan
@@ -198,7 +198,7 @@ config KCSAN_WEAK_MEMORY
# We can either let objtool nop __tsan_func_{entry,exit}() and builtin
# atomics instrumentation in .noinstr.text, or use a compiler that can
# implement __no_kcsan to really remove all instrumentation.
- depends on STACK_VALIDATION || CC_IS_GCC
+ depends on STACK_VALIDATION || CC_IS_GCC || CLANG_VERSION >= 140000
help
Enable support for modeling a subset of weak memory, which allows
detecting a subset of data races due to missing memory barriers.
--
2.34.0.rc2.393.gf8c9666880-goog
