From: Ard Biesheuvel ardb@kernel.org
GCC and Clang both implement stack protector support based on Thread Local Storage (TLS) variables, and this is used in the kernel to implement per-task stack cookies, by copying a task's stack cookie into a per-CPU variable every time it is scheduled in.
Both now also implement -mstack-protector-guard-symbol=, which permits the TLS variable to be specified directly. This is useful because it will allow to move away from using a fixed offset of 40 bytes into the per-CPU area on x86_64, which requires a lot of special handling in the per-CPU code and the runtime relocation code.
However, while GCC is rather lax in its implementation of this command line option, Clang actually requires that the provided symbol name refers to a TLS variable (i.e., one declared with __thread), although it also permits the variable to be undeclared entirely, in which case it will use an implicit declaration of the right type.
The upshot of this is that Clang will emit the correct references to the stack cookie variable in most cases, e.g.,
10d: 64 a1 00 00 00 00 mov %fs:0x0,%eax 10f: R_386_32 __stack_chk_guard
However, if a non-TLS definition of the symbol in question is visible in the same compilation unit (which amounts to the whole of vmlinux if LTO is enabled), it will drop the per-CPU prefix and emit a load from a bogus address.
Work around this by using a symbol name that never occurs in C code, and emit it as an alias in the linker script.
Fixes: 3fb0fdb3bbe7 ("x86/stackprotector/32: Make the canary into a regular percpu variable") Signed-off-by: Ard Biesheuvel ardb@kernel.org Signed-off-by: Brian Gerst brgerst@gmail.com Signed-off-by: Borislav Petkov (AMD) bp@alien8.de Reviewed-by: Nathan Chancellor nathan@kernel.org Tested-by: Nathan Chancellor nathan@kernel.org Cc: stable@vger.kernel.org Link: https://github.com/ClangBuiltLinux/linux/issues/1854 Link: https://lore.kernel.org/r/20241105155801.1779119-2-brgerst@gmail.com (cherry picked from commit 577c134d311b9b94598d7a0c86be1f431f823003) --- arch/x86/Makefile | 3 ++- arch/x86/entry/entry.S | 15 +++++++++++++++ arch/x86/include/asm/asm-prototypes.h | 3 +++ arch/x86/kernel/cpu/common.c | 2 ++ arch/x86/kernel/vmlinux.lds.S | 3 +++ 5 files changed, 25 insertions(+), 1 deletion(-)
diff --git a/arch/x86/Makefile b/arch/x86/Makefile index 9c09bbd390ce..f8a7d2a65434 100644 --- a/arch/x86/Makefile +++ b/arch/x86/Makefile @@ -81,7 +81,8 @@ ifeq ($(CONFIG_X86_32),y)
ifeq ($(CONFIG_STACKPROTECTOR),y) ifeq ($(CONFIG_SMP),y) - KBUILD_CFLAGS += -mstack-protector-guard-reg=fs -mstack-protector-guard-symbol=__stack_chk_guard + KBUILD_CFLAGS += -mstack-protector-guard-reg=fs \ + -mstack-protector-guard-symbol=__ref_stack_chk_guard else KBUILD_CFLAGS += -mstack-protector-guard=global endif diff --git a/arch/x86/entry/entry.S b/arch/x86/entry/entry.S index f4419afc7147..23f9efbe9d70 100644 --- a/arch/x86/entry/entry.S +++ b/arch/x86/entry/entry.S @@ -48,3 +48,18 @@ EXPORT_SYMBOL_GPL(mds_verw_sel);
.popsection
+#ifndef CONFIG_X86_64 +/* + * Clang's implementation of TLS stack cookies requires the variable in + * question to be a TLS variable. If the variable happens to be defined as an + * ordinary variable with external linkage in the same compilation unit (which + * amounts to the whole of vmlinux with LTO enabled), Clang will drop the + * segment register prefix from the references, resulting in broken code. Work + * around this by avoiding the symbol used in -mstack-protector-guard-symbol= + * entirely in the C code, and use an alias emitted by the linker script + * instead. + */ +#ifdef CONFIG_STACKPROTECTOR +EXPORT_SYMBOL(__ref_stack_chk_guard); +#endif +#endif diff --git a/arch/x86/include/asm/asm-prototypes.h b/arch/x86/include/asm/asm-prototypes.h index 5cdccea45554..390b13db24b8 100644 --- a/arch/x86/include/asm/asm-prototypes.h +++ b/arch/x86/include/asm/asm-prototypes.h @@ -18,3 +18,6 @@ extern void cmpxchg8b_emu(void); #endif
+#if defined(__GENKSYMS__) && defined(CONFIG_STACKPROTECTOR) +extern unsigned long __ref_stack_chk_guard; +#endif diff --git a/arch/x86/kernel/cpu/common.c b/arch/x86/kernel/cpu/common.c index f0cc4c616ceb..5db433cfaaa7 100644 --- a/arch/x86/kernel/cpu/common.c +++ b/arch/x86/kernel/cpu/common.c @@ -2000,8 +2000,10 @@ EXPORT_PER_CPU_SYMBOL(cpu_current_top_of_stack);
#ifdef CONFIG_STACKPROTECTOR DEFINE_PER_CPU(unsigned long, __stack_chk_guard); +#ifndef CONFIG_SMP EXPORT_PER_CPU_SYMBOL(__stack_chk_guard); #endif +#endif
#endif /* CONFIG_X86_64 */
diff --git a/arch/x86/kernel/vmlinux.lds.S b/arch/x86/kernel/vmlinux.lds.S index 351c604de263..ab36dacb4cc5 100644 --- a/arch/x86/kernel/vmlinux.lds.S +++ b/arch/x86/kernel/vmlinux.lds.S @@ -490,6 +490,9 @@ SECTIONS ASSERT(SIZEOF(.rela.dyn) == 0, "Unexpected run-time relocations (.rela) detected!") }
+/* needed for Clang - see arch/x86/entry/entry.S */ +PROVIDE(__ref_stack_chk_guard = __stack_chk_guard); + /* * The ASSERT() sink to . is intentional, for binutils 2.14 compatibility: */
base-commit: 0a51d2d4527b43c5e467ffa6897deefeaf499358