Re: [PATCH 3/7] x86/cpu: Disable kernel LASS when patching kernel alternatives

[Peter Zijlstra]

On Mon, Jan 09, 2023 at 09:52:00PM -0800, Yian Chen wrote:
> Most of the kernel is mapped at virtual addresses
> in the upper half of the address range. But kernel
> deliberately initialized a temporary mm area
> within the lower half of the address range
> for text poking, see commit 4fc19708b165
> ("x86/alternatives: Initialize temporary mm
> for patching").
> 
> LASS stops access to a lower half address in kernel,
> and this can be deactivated if AC bit in EFLAGS
> register is set. Hence use stac and clac instructions
> around access to the address to avoid triggering a
> LASS #GP fault.
> 
> Kernel objtool validation warns if the binary calls
> to a non-whitelisted function that exists outside of
> the stac/clac guard, or references any function with a
> dynamic function pointer inside the guard; see section
> 9 in the document tools/objtool/Documentation/objtool.txt.
> 
> For these reasons, also considering text poking size is
> usually small, simple modifications have been done
> in function text_poke_memcpy() and text_poke_memset() to
> avoid non-whitelisted function calls inside the stac/clac
> guard.
> 
> Gcc may detect and replace the target with its built-in
> functions. However, the replacement would break the
> objtool validation criteria. Hence, add compiler option
> -fno-builtin for the file.

Please reflow to 72 characters consistently, this is silly.

> Co-developed-by: Tony Luck <tony.luck@xxxxxxxxx>
> Signed-off-by: Tony Luck <tony.luck@xxxxxxxxx>
> Signed-off-by: Yian Chen <yian.chen@xxxxxxxxx>
> ---
>  arch/x86/include/asm/smap.h      | 13 +++++++++++++
>  arch/x86/kernel/Makefile         |  2 ++
>  arch/x86/kernel/alternative.c    | 21 +++++++++++++++++++--
>  tools/objtool/arch/x86/special.c |  2 ++
>  4 files changed, 36 insertions(+), 2 deletions(-)
> 
> diff --git a/arch/x86/include/asm/smap.h b/arch/x86/include/asm/smap.h
> index bab490379c65..6f7ac0839b10 100644
> --- a/arch/x86/include/asm/smap.h
> +++ b/arch/x86/include/asm/smap.h
> @@ -39,6 +39,19 @@ static __always_inline void stac(void)
>  	alternative("", __ASM_STAC, X86_FEATURE_SMAP);
>  }
>  
> +/* Deactivate/activate LASS via AC bit in EFLAGS register */
> +static __always_inline void low_addr_access_begin(void)
> +{
> +	/* Note: a barrier is implicit in alternative() */
> +	alternative("", __ASM_STAC, X86_FEATURE_LASS);
> +}
> +
> +static __always_inline void low_addr_access_end(void)
> +{
> +	/* Note: a barrier is implicit in alternative() */
> +	alternative("", __ASM_CLAC, X86_FEATURE_LASS);
> +}

Can't say I like the name. Also if you look at bit 63 as a sign bit,
it's actively wrong since -1 is lower than 0.

> +
>  static __always_inline unsigned long smap_save(void)
>  {
>  	unsigned long flags;
> diff --git a/arch/x86/kernel/Makefile b/arch/x86/kernel/Makefile
> index 96d51bbc2bd4..f8a455fc56a2 100644
> --- a/arch/x86/kernel/Makefile
> +++ b/arch/x86/kernel/Makefile
> @@ -7,6 +7,8 @@ extra-y	+= vmlinux.lds
>  
>  CPPFLAGS_vmlinux.lds += -U$(UTS_MACHINE)
>  
> +CFLAGS_alternative.o	+= -fno-builtin
> +
>  ifdef CONFIG_FUNCTION_TRACER
>  # Do not profile debug and lowlevel utilities
>  CFLAGS_REMOVE_tsc.o = -pg
> diff --git a/arch/x86/kernel/alternative.c b/arch/x86/kernel/alternative.c
> index 7d8c3cbde368..4de8b54fb5f2 100644
> --- a/arch/x86/kernel/alternative.c
> +++ b/arch/x86/kernel/alternative.c
> @@ -1530,14 +1530,31 @@ __ro_after_init unsigned long poking_addr;
>  
>  static void text_poke_memcpy(void *dst, const void *src, size_t len)
>  {
> -	memcpy(dst, src, len);
> +	const char *s = src;
> +	char *d = dst;
> +
> +	/* The parameter dst ends up referencing to the global variable
> +	 * poking_addr, which is mapped to the low half address space.
> +	 * In kernel, accessing the low half address range is prevented
> +	 * by LASS. So relax LASS prevention while accessing the memory
> +	 * range.
> +	 */
> +	low_addr_access_begin();
> +	while (len-- > 0)
> +		*d++ = *s++;
> +	low_addr_access_end();
>  }
>  
>  static void text_poke_memset(void *dst, const void *src, size_t len)
>  {
>  	int c = *(const int *)src;
> +	char *d = dst;
>  
> -	memset(dst, c, len);
> +	/* The same comment as it is in function text_poke_memcpy */
> +	low_addr_access_begin();
> +	while (len-- > 0)
> +		*d++ = c;
> +	low_addr_access_end();
>  }

This is horrific tinkering :/

Also, what about the EFI mm? IIRC EFI also lives in the user address
space.