Re: [PATCH] Allow direct access relocations referencing a protected function symbol

["H.J. Lu" <hjl.tools@gmail.com>]

On Monday, June 14, 2021, Michael Matz <matz@suse.de> wrote:

> Hello,
>
> On Mon, 14 Jun 2021, Alan Modra via Binutils wrote:
>
> > On Sun, Jun 13, 2021 at 02:54:00PM -0700, Fangrui Song via Binutils
> wrote:
> > > This fixes the bogus "relocation R_* against protected symbol `foo'
> > > can not be used when making a shared object" for function symbols for
> > > at least aarch64/i386/x86-64.
> > >
> > > The controversial "copy relocations on protected data symbols" (which
> has some
> > > fragile glibc support) is irrelevant to function symbols.
> >
> > No, this patch doesn't do that.  What you are doing here will disable
> > dynamic relocations on protected function symbols in shared libraries.
> > That will break function pointer comparison for architectures that
> > implement non-pic executables, where a function that is undefined in
> > the executable is given a fixed address in the executable, that of its
> > plt call code.
>
> Correct.  But I'm oscillating between thinking that this would be a
> problem and thinking the opposite :-/
>
> One could always say that function addresses of protected functions aren't
> comparable.  Taking an address and using it as indirect call target will
> always work, you just wouldn't be able to compare them usefully.  Or one
> could require address references from outside components to a protected
> (function) symbol to always be via a GOT(-like structure).
>
> Or (the other extremum of my oscillations) one says that function address
> comparisons absolutely need to work even in absence of
> indirection-via-GOT, at which point we basically talk about the same
> problem like protected data symbols and copy relocations.  If you then
> don't have relocations differing between calls and address taking, you
> effectively throw out the usefullness of protected visibility.
>
> The problem with the latter stance is that it's not really justifiable
> from the ELF gABI: nothing says that getting at "the" function address
> must be possible without a GOT indirection.
>
> (Adding to that is that what actually works in practice changed over time
> and depends on the architecture (and well, yeah, compiler and linker) so
> that now it's impossible to say "look there, that's how it should work
> and how everyone expects it to work" :-/ )
>
>
> Ciao,
> Michael.
>

I'd to use GOT for function address and remove copy
relocation in both PIE and non-PIE.   The question is
that if we should detect binary incompatibility at link-time
and run-time.



-- 
H.J.