Re: [PATCH 4/8] objtool: Add support for intra-function calls

[Peter Zijlstra]

On Thu, Apr 23, 2020 at 04:34:21PM +0200, Miroslav Benes wrote:
> >  /*
> >   * Find the destination instructions for all calls.
> >   */
> > @@ -715,10 +725,7 @@ static int add_call_destinations(struct
> >  				continue;
> >  
> >  			if (!insn->call_dest) {
> > -				WARN_FUNC("unsupported intra-function call",
> > -					  insn->sec, insn->offset);
> > -				if (retpoline)
> > -					WARN("If this is a retpoline, please patch it in with alternatives and annotate it with ANNOTATE_NOSPEC_ALTERNATIVE.");
> > +				WARN_FUNC("intra-function call", insn->sec, insn->offset);
> 
> "unsupported intra-function call"?

Well, I think the thinking was that intra-function calls are actually
supported, 'unannotated' perhaps ?

> >  				return -1;
> >  			}
> >  
> > @@ -741,6 +748,12 @@ static int add_call_destinations(struct
> >  			}
> >  		} else
> >  			insn->call_dest = rela->sym;
> > +
> > +		/*
> > +		 * Whatever stack impact regular CALLs have, should be
> > +		 * undone by the RETURN of the called function.
> 
>  * Annotated intra-function CALLs are treated as JMPs with a stack_op.
>  * See read_intra_function_calls().
> 
> would make it a bit clearer.

That doesn't work for me; we want to explain why it is OK to delete
stack_ops for regular CALLs. The reason this is OK, is because they're
matched by RETURN.

> > +                */
> > +               remove_insn_ops(insn);
> >         }
> >  
> >         return 0;
> > @@ -1416,6 +1429,57 @@ static int read_instr_hints(struct objto
> >         return 0;
> >  }
> > 
> > +static int read_intra_function_calls(struct objtool_file *file)
> > +{
> > +       struct instruction *insn;
> > +       struct section *sec;
> > +       struct rela *rela;
> > +
> > +       sec = find_section_by_name(file->elf, ".rela.discard.intra_function_calls");
> > +       if (!sec)
> > +               return 0;
> > +
> > +       list_for_each_entry(rela, &sec->rela_list, list) {
> > +               unsigned long dest_off;
> > +
> > +               if (rela->sym->type != STT_SECTION) {
> > +                       WARN("unexpected relocation symbol type in %s",
> > +                            sec->name);
> > +                       return -1;
> > +               }
> > +
> > +               insn = find_insn(file, rela->sym->sec, rela->addend);
> > +               if (!insn) {
> > +                       WARN("bad .discard.intra_function_call entry");
> > +                       return -1;
> > +               }
> > +
> > +               if (insn->type != INSN_CALL) {
> > +                       WARN_FUNC("intra_function_call not a direct call",
> > +                                 insn->sec, insn->offset);
> > +                       return -1;
> > +               }
> > +
> > +               /*
> > +                * Treat intra-function CALLs as JMPs, but with a stack_op.
> > +                * Also see how setup_call_dest() strips stack_ops from normal
> > +                * CALLs.
> 
> /*
>  * Treat annotated intra-function CALLs as JMPs, but with a stack_op.
>  * Also see how add_call_destinations() strips stack_ops from normal
>  * CALLs.
>  */
> 
> ? (note added "annotated" and s/setup_call_dest/add_call_destinations/)

Unannotated intra-function calls are not allowed, so I don't see a
reason to make that distinction, but sure.

> > +                */
> > +               insn->type = INSN_JUMP_UNCONDITIONAL;
> 
> [...]
> 
> > @@ -2245,6 +2313,9 @@ static int validate_branch(struct objtoo
> >  				return 0;
> >  		}
> >  
> > +		if (handle_insn_ops(insn, &state))
> > +			return 1;
> > +
> >  		switch (insn->type) {
> >  
> >  		case INSN_RETURN:
> > @@ -2304,9 +2375,6 @@ static int validate_branch(struct objtoo
> >  			break;
> >  
> >  		case INSN_EXCEPTION_RETURN:
> > -			if (handle_insn_ops(insn, &state))
> > -				return 1;
> > -
> >  			/*
> >  			 * This handles x86's sync_core() case, where we use an
> >  			 * IRET to self. All 'normal' IRET instructions are in
> > @@ -2326,8 +2394,6 @@ static int validate_branch(struct objtoo
> >  			return 0;
> >  
> >  		case INSN_STACK:
> > -			if (handle_insn_ops(insn, &state))
> > -				return 1;
> >  			break;
> 
> So we could get rid of INSN_STACK now as Julien proposed, couldn't we? If 
> I am not missing something. handle_insn_ops() is called unconditionally 
> here for all insn types and you remove stack_ops when unneeded.

Yes, INSN_STACK can now go away in favour of NOPs with stack_ops.
Separate patch though.

> We could also go ahead with Julien's proposal to remove 
> INSN_EXCEPTION_RETURN hack and move it to tools/objtool/arch/x86/decode.c. 

I don't immediately see how; we don't have a symbol there.