Re: the x86 sysret_rip test fails on the Intel FRED architecture

[Ammar Faizi <ammarfaizi2@gnuweeb.org>]

On 1/23/23 6:45 AM, H. Peter Anvin wrote:
> static enum regs_ok check_regs_syscall(int syscall,
>      unsigned long arg1, unsigned long arg2)
> {
> 
>      register unsigned long r11 asm("%r11");
>      unsigned long rcx, rbx, tmp;

tmp is unused.

>      r11 = r11_sentinel;
>      rcx = rcx_sentinel;
> 
>      asm volatile("push %3; popf; "
>               "lea 1f(%%rip),%2; "
>               "syscall; "
>               "1:"
>               : "+r" (r11), "+c" (rcx), "=b" (rbx)
>               : "g" (rflags_sentinel),
>                 "a" (syscall), "D" (arg1), "S" (arg2));

BTW, I just realized this "push" is unsafe for userspace code if the
compiler decides to inline this inside a leaf function that uses the
redzone.

Reason: Because this "push;" clobbers redzone.

It doesn't always happen, but when that happens it can be confusing to
debug.

A simple workaround is: just compile it with "-mno-red-zone" flag.

Alternative, without using that flag, maybe preserve the value like:

     movq    -8(%rsp), %r12
     pushq   %[rflags_sentinel]
     popf
     movq    %r12, -8(%rsp)
     syscall

with "r12" and "memory" added to the clobber list.

What do you think?

-- 
Ammar Faizi