Re: [PATCH 1/2] x86/entry/64: Do not clear %rbx under Xen

[Andy Lutomirski <luto@kernel.org>]

On Sat, Jul 21, 2018 at 12:49 PM, M. Vefa Bicakci <m.v.b@runbox.com> wrote:
> Commit 3ac6d8c787b8 ("x86/entry/64: Clear registers for
> exceptions/interrupts, to reduce speculation attack surface") unintendedly
> broke Xen PV virtual machines by clearing the %rbx register at the end of
> xen_failsafe_callback before the latter jumps to error_exit.
> error_exit expects the %rbx register to be a flag indicating whether
> there should be a return to kernel mode.
>
> This commit makes sure that the %rbx register is not cleared by
> the PUSH_AND_CLEAR_REGS macro, when the macro in question is instantiated
> by xen_failsafe_callback, to avoid the issue.

Seems like a genuine problem, but:

> diff --git a/arch/x86/entry/entry_64.S b/arch/x86/entry/entry_64.S
> index c7449f377a77..96e8ff34129e 100644
> --- a/arch/x86/entry/entry_64.S
> +++ b/arch/x86/entry/entry_64.S
> @@ -1129,7 +1129,7 @@ ENTRY(xen_failsafe_callback)
>         addq    $0x30, %rsp
>         UNWIND_HINT_IRET_REGS
>         pushq   $-1 /* orig_ax = -1 => not a system call */
> -       PUSH_AND_CLEAR_REGS
> +       PUSH_AND_CLEAR_REGS clear_rbx=0
>         ENCODE_FRAME_POINTER
>         jmp     error_exit

The old code first set RBX to zero then, if frame pointers are on,
sets it to some special non-zero value, then crosses its fingers and
hopes for the best.  Your patched code just skips the zeroing part, so
RBX either contains the ENCODE_FRAME_POINTER result or is
uninitialized.

How about actually initializing rbx to something sensible like, say, 1?