On Tue, May 24, 2022 at 5:32 AM Borislav Petkov <bp@xxxxxxxxx> wrote:
>
> finally a somewhat final version, lightly tested.
I can't find anything wrong with this, but who knows what
patch-blindness I have from looking at a few different versions of it.
Maybe my eyes just skim over it now.
I do note that the clearing of %rax here:
> +.Lerms_exit:
> + xorl %eax,%eax
> + RET
seems to be unnecessary, since %rax is never modified in the path
leading to this. But maybe just as well just for consistency with the
cases where it *is* used as a temporary.
And I still suspect that "copy_to_user()" is *much* more interesting
than "clear_user()", but I guess we can't inline it anyway due to all
the other overhead (ie access_ok() and stac/clac).
And for a plain "call memcpy/memset", we'd need compiler help to do
this (at a minimum, we'd have to have the compiler use the 'rep
movs/stos' register logic, and then we could patch things in place
afterwards, with objtool creating the alternatives section or
something).
Linus
