Re: [PATCH v1 3/8] x86/entry/clearregs: Clear registers for 64bit SYSCALL

[Josh Poimboeuf <jpoimboe@redhat.com>]

On Tue, Jan 09, 2018 at 05:03:23PM -0800, Andi Kleen wrote:
> From: Andi Kleen <ak@linux.intel.com>
> 
> We clear all the non argument registers for 64bit SYSCALLs
> to minimize any risk of bad speculation using user values.
> 
> So far unused argument registers still leak. To be addressed
> in future patches.
> 
> Signed-off-by: Andi Kleen <ak@linux.intel.com>
> ---
>  arch/x86/entry/entry_64.S | 9 +++++++++
>  1 file changed, 9 insertions(+)
> 
> diff --git a/arch/x86/entry/entry_64.S b/arch/x86/entry/entry_64.S
> index bbdfbdd817d6..632081fd7086 100644
> --- a/arch/x86/entry/entry_64.S
> +++ b/arch/x86/entry/entry_64.S
> @@ -236,6 +236,14 @@ GLOBAL(entry_SYSCALL_64_after_hwframe)
>  	pushq	%r11				/* pt_regs->r11 */
>  	sub	$(6*8), %rsp
>  	SAVE_EXTRA_REGS
> +	/* Sanitize registers against speculation attacks */

This comment isn't necessary, though it would be good to add comments
above the CLEAR macros themselves explaining why they're needed.

> +	/* r10 is cleared later, arguments are handled in san_args* */

What is san_args?

> +	CLEAR_R11_TO_R15
> +#ifndef CONFIG_FRAME_POINTER
> +	xor	%ebp, %ebp
> +#endif

Why is %rbp not cleared with CONFIG_FRAME_POINTER?  Is it because it
will get clobbered by the first called function?

> +	xor	%ebx, %ebx
> +	xor	%ecx, %ecx

I think clearing %ecx isn't needed, it gets clobbered below for the fast
path, and gets clobbered by do_syscall_64() for the slow path.

>  
>  	UNWIND_HINT_REGS extra=0
>  
> @@ -263,6 +271,7 @@ entry_SYSCALL_64_fastpath:
>  #endif
>  	ja	1f				/* return -ENOSYS (already in pt_regs->ax) */
>  	movq	%r10, %rcx
> +	xor	%r10, %r10
>  
>  #ifdef CONFIG_RETPOLINE
>  	movq	sys_call_table(, %rax, 8), %rax

Now that the fast path is getting slower, I wonder if it still makes
sense to have a "fast path"?  It would be good to see measurements
comparing the fast and slow paths.

-- 
Josh