Re: [PATCH] x86/alternatives: Fix optimize_nops() checking

[Josh Poimboeuf]

On Wed, Jan 10, 2018 at 12:26:25PM -0800, Linus Torvalds wrote:
> On Wed, Jan 10, 2018 at 12:15 PM, Josh Poimboeuf <jpoimboe@xxxxxxxxxx> wrote:
> >
> > I think .altinstruction relocations *do* work if they're for the first
> > instruction, and it's a jump or a call.
> 
> Yes - for the alternative that is in-line - not in the "altinstruction" section.
> 
> Because then the alternative is in the right spot at link-time already.
> 
> But the "altinstruction" section definitely should not have
> relocations.

I misspoke, it's really .altinstr_replacement which has the replacement
instructions.  And it has a bunch of relocations:

  Relocation section [ 8] '.rela.altinstr_replacement' for section [ 7] '.altinstr_replacement' at offset 0x14439710 contains 355 entries:

> I guess you could hack them up by hand by explicitly
> trying to take the difference between the non-altinstruction and the
> altinstruction into account, but it would be error-prone and fragile
> as hell.

apply_alternatives() already does that today.  It actually seems pretty
solid, except for the whole "only works on the first instruction" thing.

> > I think Boris had a patch floating around to add an instruction decoder
> > to alternatives, so you can do a call/jmp anywhere.
> 
> .. and no, we're not doing that. Christ.
> 
> People, we need to try to be *robust* here. That's doubly (triply!)
> true of things like altinstructions where people - very much by design
> - won't even *test* the alternatives very much, because very much by
> design the altinstructions are only used on certain architectures or
> in certain situations.
> 
> And we almost certainly don't actuially _need_ relocations. But we
> need to protect against the "oops, I didn't realize" issue, exactly
> because testing won't actually catch the odd cases.

If we need objtool to detect them, it's certainly possible.  But maybe I
missed the previous discussion -- what's the, um, alternative to
relocations, when we have calls and jumps being patched in?

> Because we don't want to be in the situation where some random poor
> user hits it because they have an old CPU that no developer has, and
> then the relocation will basically do completely random things.
> 
> Imagine just how crazy that would be to debug. You'd be basically
> executing insane code, and looking at the sources - or even the
> binaries - it would _look_ completely sane.

Well, I think we already made that deal with the devil when we added
alternatives/paravirt/smp_locks/jump_labels/kprobes/ftrace/bpf, etc.

-- 
Josh