Re: [RFC][PATCH v4 18/32] objtool: mcount: Move nop_mcount()

[Matt Helsley <mhelsley@vmware.com>]

On Fri, Jun 12, 2020 at 03:26:56PM +0200, Peter Zijlstra wrote:
> On Tue, Jun 02, 2020 at 12:50:11PM -0700, Matt Helsley wrote:
> > +static int nop_mcount(struct section * const rels,
> > +		      const char *const txtname)
> > +{
> > +	struct reloc *reloc;
> > +	struct section *txts = find_section_by_index(lf, rels->sh.sh_info);
> > +	unsigned mcountsym = 0;
> > +	int once = 0;
> > +
> > +	list_for_each_entry(reloc, &rels->reloc_list, list) {
> > +		int ret = -1;
> > +
> > +		if (!mcountsym)
> > +			mcountsym = get_mcountsym(reloc);
> > +
> > +		if (mcountsym == GELF_R_INFO(reloc->sym->idx, reloc->type) && !is_fake_mcount(reloc)) {
> 
> This makes no sense to me; why not have mcountsym be a 'struct symbol
> *' and have get_mcountsym() return one of those.
> 
> 	if (reloc->sym == mcountsym && ... )
> 
> is much nicer, no?

Indeed! I'll change it from returning an unsigned long to struct symbol * before I
move it out of the wrapper code.

> 
> > +			if (make_nop) {
> > +				ret = make_nop(txts, reloc->offset);
> > +				if (ret < 0)
> > +					return -1;
> > +			}
> > +			if (warn_on_notrace_sect && !once) {
> > +				printf("Section %s has mcount callers being ignored\n",
> > +				       txtname);
> > +				once = 1;
> > +				/* just warn? */
> > +				if (!make_nop)
> > +					return 0;
> > +			}
> > +		}
> > +
> > +		/*
> > +		 * If we successfully removed the mcount, mark the relocation
> > +		 * as a nop (don't do anything with it).
> > +		 */
> > +		if (!ret) {
> > +			reloc->type = rel_type_nop;
> > +			rels->changed = true;
> 
> I have an elf_write_rela(), I'll make sure to Cc you.

Thanks! I might also make use of your patch to rewrite instructions. We
need a way to turn certain prologue instructions into nops. Would it be
more widely useful to move that functionality out of mcount and into
the objtool ELF/per-arch code or do you think it's better inside the
mcount subcommand code?

Cheers,
     -Matt