Re: [PATCH 1/3] Revert "x86/retpoline: Remove .text..__x86.return_thunk section"

[Peter Zijlstra]

On Wed, Oct 11, 2023 at 09:28:43AM -0700, Josh Poimboeuf wrote:
> On Wed, Oct 11, 2023 at 09:41:42AM +0200, Peter Zijlstra wrote:
> > > +++ b/tools/objtool/check.c
> > > @@ -1610,6 +1610,15 @@ static int add_jump_destinations(struct objtool_file *file)
> > >  			return -1;
> > >  		}
> > >  
> > > +		/*
> > > +		 * Since retpolines are in the same section as the return
> > > +		 * thunk, they might not use a relocation when branching to it.
> > > +		 */
> > > +		if (jump_dest->sym && jump_dest->sym->return_thunk) {
> > > +			add_return_call(file, insn, true);
> > > +			continue;
> > > +		}
> > 
> > *urgh*... I mean, yes, that obviously works, but should we not also have
> > the retpoline thingy for consistency? That case makes less sense though
> > :/
> 
> Consistency with what? 

the reloc case; specifically, I was thinking something along these
lines:

		if (jump-dest->sym && jump_dest->sym->retpoline_thunk) {
			add_retpoline_call(file, insn);
			continue;
		}

Then both reloc and immediate versions are more or less the same.

> The extra section seems pointless but maybe I'm missing something.

By having the section things are better delineated I suppose, be it
retpolines or rethunks, all references should be to inside the section
(and thus have a reloc) while within the section there should never be a
reference to itself.

I'm not sure it's worth much, but then we can have the above two cases
issue a WARN instead of fixing it up.

I don't care too deeply, I can't make up my mind either way. But perhaps
keeping the section is easier on all the backports, it's easy to forget
a tiny objtool patch like this.