Re: [PATCH v5 01/16] x86/stackprotector: Work around strict Clang TLS symbol requirements
From: Nathan Chancellor
Date: Tue Nov 05 2024 - 14:31:06 EST
On Tue, Nov 05, 2024 at 10:57:46AM -0500, Brian Gerst wrote:
> From: Ard Biesheuvel <ardb@xxxxxxxxxx>
>
> 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 us 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")
> Cc: <stable@xxxxxxxxxxxxxxx>
> Cc: Fangrui Song <i@xxxxxxxxxx>
> Cc: Uros Bizjak <ubizjak@xxxxxxxxx>
> Cc: Nathan Chancellor <nathan@xxxxxxxxxx>
> Cc: Andy Lutomirski <luto@xxxxxxxxxx>
> Link: https://github.com/ClangBuiltLinux/linux/issues/1854
> Signed-off-by: Ard Biesheuvel <ardb@xxxxxxxxxx>
> Signed-off-by: Brian Gerst <brgerst@xxxxxxxxx>
>From https://lore.kernel.org/20241016021045.GA1000009@thelio-3990X/:
Reviewed-by: Nathan Chancellor <nathan@xxxxxxxxxx>
Tested-by: Nathan Chancellor <nathan@xxxxxxxxxx>
> ---
> arch/x86/Makefile | 5 +++--
> arch/x86/entry/entry.S | 16 ++++++++++++++++
> 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, 27 insertions(+), 2 deletions(-)
>
> diff --git a/arch/x86/Makefile b/arch/x86/Makefile
> index cd75e78a06c1..5b773b34768d 100644
> --- a/arch/x86/Makefile
> +++ b/arch/x86/Makefile
> @@ -142,9 +142,10 @@ 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
> + KBUILD_CFLAGS += -mstack-protector-guard=global
> endif
> endif
> else
> diff --git a/arch/x86/entry/entry.S b/arch/x86/entry/entry.S
> index 324686bca368..b7ea3e8e9ecc 100644
> --- a/arch/x86/entry/entry.S
> +++ b/arch/x86/entry/entry.S
> @@ -51,3 +51,19 @@ EXPORT_SYMBOL_GPL(mds_verw_sel);
> .popsection
>
> THUNK warn_thunk_thunk, __warn_thunk
> +
> +#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 25466c4d2134..3674006e3974 100644
> --- a/arch/x86/include/asm/asm-prototypes.h
> +++ b/arch/x86/include/asm/asm-prototypes.h
> @@ -20,3 +20,6 @@
> 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 8f41ab219cf1..9d42bd15e06c 100644
> --- a/arch/x86/kernel/cpu/common.c
> +++ b/arch/x86/kernel/cpu/common.c
> @@ -2091,8 +2091,10 @@ void syscall_init(void)
>
> #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 410546bacc0f..d61c3584f3e6 100644
> --- a/arch/x86/kernel/vmlinux.lds.S
> +++ b/arch/x86/kernel/vmlinux.lds.S
> @@ -468,6 +468,9 @@ SECTIONS
> . = ASSERT((_end - LOAD_OFFSET <= KERNEL_IMAGE_SIZE),
> "kernel image bigger than KERNEL_IMAGE_SIZE");
>
> +/* needed for Clang - see arch/x86/entry/entry.S */
> +PROVIDE(__ref_stack_chk_guard = __stack_chk_guard);
> +
> #ifdef CONFIG_X86_64
> /*
> * Per-cpu symbols which need to be offset from __per_cpu_load
> --
> 2.47.0
>