Re: [musl] RELRO vs deferred binding

[Rich Felker <dalias@libc.org>]

On Wed, Mar 31, 2021 at 05:51:26PM +0300, Alexander Monakov wrote:
> On Wed, 31 Mar 2021, Rich Felker wrote:
> 
> > Thanks for raising this. I think deferred binding needs to be updated
> > either to ignore RELRO if there are outstanding relocations (possibly
> > deferring it until they are all resolved)
> 
> This seems undesirable as it leaves GOT unprotected for the rest of
> run time if unresolved relocations remain.

Yes, but in practice this is only for broken xorg modules and the
unresolved relocations are resolved by the time any attack-surface
code runs, no? Still I agree it's better to avoid this.

> > or to unprotect and
> > reprotect on every incremental link. (This could be optimized out and
> > preserve some further safety by scanning the outstanding relocation
> > table and skipping the unprotect/reprotect if none of them lie in the
> > RELRO range.)
> 
> Even better might be to do relocation normally and lazily unprotect RELRO
> on first relocation that needs that, then reprotect once done with that DSO.
> (i.e. without doing an additional scan, like your parenthesized statement
> seems to suggest)

That puts the additional branch/logic inside the hot path used by all
relocation processing rather than a path that's relegated to just
outstanding relocations on libraries that didn't declare their
dependencies properly.

My version looks something like, inside the for loop in
redo_lazy_relocs:

	need_unprotect = 0;
	for (i=0; i<size; i+=3);
		if ((uintptr_t)laddr(p, p->lazy[i])-relro_start < relro_end)
			need_unprotect = 1;
	if (need_unprotect) mprotect(...);
	do_relocs(...);
	if (need_unprotect) mprotect(...);

Does that look reasonable?

Rich