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

[Ammar Faizi]

On Mon, Jan 23, 2023 at 11:43:35AM -0800, H. Peter Anvin wrote:
> Good spotting. %rax needs to be marked clobbered, too.

Yeah, that 'syscall' variable should be using a "+a" constraint.

But anyway, I found something wrong with this. I was playing with your
code and I found it failed to assert that %r11 == rflags_sentinel on a
non-FRED system. It happens because "popf" doesn't set the %rflags
to the expected value.

I tried to simplify it like this:

    pushq  $0x200893
    popf              # This popf sets %rflags to 0x200a93, not to 0x200893.
    pushf
    popq   %r11

    # Now %r11 == 0x200a93,
    # but the expected value is %r11 == 0x200893.

Looking into their bits:

    (gdb) p/t 0x200893
    $1 = 1000000000100010010011
    (gdb) p/t 0x200a93
    $2 = 1000000000101010010011

Align them to spot differences:

    0x200893 = 0b1000000000100010010011
    0x200a93 = 0b1000000000101010010011
                             ^

Or just xor them to find the differences:

    (gdb) p/x 0x200893 ^ 0x200a93
    $3 = 0x200

** Checks my Intel SDM cheat sheets. **

Then, I was like "Oh, that's (1 << 9) a.k.a. IF. Of course we can't
change rflags[IF] from userspace!!!".

In short, we can't use 0x200893 as the rflags_sentinel value because it
clears the interrupt flag.

-- 
Ammar Faizi