Re: [PATCH 4/9] x86/alternative: Implement .retpoline_sites support

[Sami Tolvanen]

On Tue, Oct 19, 2021 at 3:17 AM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
>
> + Sami
>
> (Sami, for context:
>
>   https://git.kernel.org/pub/scm/linux/kernel/git/peterz/queue.git/log/?h=objtool/core
>
> which contains the following code:
>
> +       void (*target)(void);
> +       int reg, i = 0;
> +
> +       target = addr + insn->length + insn->immediate.value;
> +       reg = (target - &__x86_indirect_thunk_rax) /
> +             (&__x86_indirect_thunk_rcx - &__x86_indirect_thunk_rax);
> +
> +       if (WARN_ON_ONCE(reg & ~0xf))
> +               return -1;
>
> which blows up something fierce on clang-cfi)
>
> On Tue, Oct 19, 2021 at 09:47:26AM +0000, Alexander Lobakin wrote:
>
> > Oh okay, it's because of ClangCFI:
> >
> > SMP alternatives: You were looking for __typeid__ZTSFvvE_global_addr+0x370/0x1410 at 0xffffffffa523cd60,>
> > SMP alternatives: rax is __typeid__ZTSFvvE_global_addr+0x360/0x1410 at 0xffffffffa523cd50
> >
> > Sorry for confusing, seems like it's a side effect of using it on
> > Clang 12 while the original series supports only 13+. I'll double
> > check and let know if find something.
>
> I'm thinking CFI will totally screw this up regardless, seeing how a
> function pointer is taken, and the CFI magicks will turn that into one
> of those weird trampolines instead of the actual symbol.
>
> The compiler could of course deduce that these addresses are never
> called and don't escape the function, and therefore doesn't need to do
> the CFI transformation on then, but I'm guessing it isn't quite that
> clever.

Yes, it's unfortunately not that clever.

> Also doing CFI on retpoline thunks seems 'weird', they have a very
> particular calling convention, excplicitly very much not the standard C
> one. Can't we mark them using asmlinkage or something to tell the
> compiler to politely 'bugger off' or somesuch ;-)

I confirmed that using an opaque type for the thunk declaration fixes
this issue with CFI. It also makes it obvious that these are not
callable from C code.

Sami