Re: [PATCH V2 9/9] x86/speculation: Remove all ANNOTATE_NOSPEC_ALTERNATIVE directives

[Alexandre Chartre]

On 4/7/20 6:28 PM, Josh Poimboeuf wrote:
> On Tue, Apr 07, 2020 at 06:18:51PM +0200, Alexandre Chartre wrote:
> > On 4/7/20 4:32 PM, Alexandre Chartre wrote:
> > > On 4/7/20 3:34 PM, Josh Poimboeuf wrote:
> > > > On Tue, Apr 07, 2020 at 03:28:37PM +0200, Peter Zijlstra wrote:
> > > > > Josh, we should probably have objtool verify it doesn't emit ORC entries
> > > > > in alternative ranges.
> > > >
> > > > Agreed, it might be as simple as checking for insn->alt_group in the
> > > > INSN_STACK check or in update_insn_state().
> > >
> > > We could do that only for the "objtool orc generate" command. That way
> > > "objtool check" would still check the alternative, but "objtool orc generate"
> > > will just use the first half of the alternative (like it does today with
> > > ANNOTATE_NOSPEC_ALTERNATIVE). We can even keep all ANNOTATE_NOSPEC_ALTERNATIVE
> > > but only use them for "objtool orc generate".
> >
> > I have checked and objtool doesn't emit ORC entries for alternative:
> > decode_instructions() doesn't mark such section with sec->text = true
> > so create_orc_sections() doesn't emit corresponding ORC entries.
> >
> > So I think we can remove the ANNOTATE_NOSPEC_ALTERNATIVE directives,
> > this will allow objtool to check the instructions but it still won't
> > emit ORC entries (same behavior as today). In the future, if ORC
> > eventually supports alternative we will be ready to have objtool emit
> > ORC entries.
>
> What's the benefit of removing ANNOTATE_NOSPEC_ALTERNATIVE if there's no
> ORC support to go along with it?

To have the code validated by objtool like any other alternative code
(which is not tagged with ANNOTATE_NOSPEC_ALTERNATIVE).

> Also I want to avoid adding "ORC alternatives".  ORC is nice and simple
> and we should keep it that way as much as possible.
>
> Again, we should warn on stack changes inside alternatives, and then
> look at converting RSB and retpolines to use static branches so they
> have deterministic stacks.
objtool doesn't currently warn on stack changes inside alternatives.
The RSB/retpoline alternatives have warning because objtool doesn't
support retpoline ret and intra-function calls. If you have an alternative
doing stack changes that objtool understand (like push/pop, add/remove
to sp) then you won't have a warning.

I think that's the case with smap_save:

static __always_inline unsigned long smap_save(void)
{
        unsigned long flags;

        asm volatile (ALTERNATIVE("", "pushf; pop %0; " __ASM_CLAC,
                                  X86_FEATURE_SMAP)
                      : "=rm" (flags) : : "memory", "cc");

        return flags;
}

The alternative does change the stack but objtool won't complain
because it handles the pushf and pop instruction.

alex.