]> git.ipfire.org Git - thirdparty/kernel/linux.git/commitdiff
arm64: Fix non-atomic __READ_ONCE() with CONFIG_LTO=y
authorMarco Elver <elver@google.com>
Fri, 30 Jan 2026 13:28:24 +0000 (14:28 +0100)
committerWill Deacon <will@kernel.org>
Mon, 2 Feb 2026 15:40:54 +0000 (15:40 +0000)
The implementation of __READ_ONCE() under CONFIG_LTO=y incorrectly
qualified the fallback "once" access for types larger than 8 bytes,
which are not atomic but should still happen "once" and suppress common
compiler optimizations.

The cast `volatile typeof(__x)` applied the volatile qualifier to the
pointer type itself rather than the pointee. This created a volatile
pointer to a non-volatile type, which violated __READ_ONCE() semantics.

Fix this by casting to `volatile typeof(*__x) *`.

With a defconfig + LTO + debug options build, we see the following
functions to be affected:

xen_manage_runstate_time (884 -> 944 bytes)
xen_steal_clock (248 -> 340 bytes)
  ^-- use __READ_ONCE() to load vcpu_runstate_info structs

Fixes: e35123d83ee3 ("arm64: lto: Strengthen READ_ONCE() to acquire when CONFIG_LTO=y")
Cc: stable@vger.kernel.org
Reviewed-by: Boqun Feng <boqun@kernel.org>
Signed-off-by: Marco Elver <elver@google.com>
Tested-by: David Laight <david.laight.linux@gmail.com>
Signed-off-by: Will Deacon <will@kernel.org>
arch/arm64/include/asm/rwonce.h

index 78beceec10cda47b319db29d9f79d2a5df35e92d..fc0fb42b0b6411680d821036ce38ade52088d5b0 100644 (file)
@@ -58,7 +58,7 @@
        default:                                                        \
                atomic = 0;                                             \
        }                                                               \
-       atomic ? (typeof(*__x))__u.__val : (*(volatile typeof(__x))__x);\
+       atomic ? (typeof(*__x))__u.__val : (*(volatile typeof(*__x) *)__x);\
 })
 
 #endif /* !BUILD_VDSO */