Re: [PATCH] elf: Refuse to dlopen PIE objects [BZ #24323]

[Carlos O'Donell <codonell at redhat dot com>]

On 4/29/19 5:35 PM, Florian Weimer wrote:
> * Carlos O'Donell:
>
> > >      ifeq (yes,$(build-shared))
> > >    # NB: Please keep cet-built-dso in sysdeps/x86/Makefile in sync with
> > > diff --git a/elf/dl-load.c b/elf/dl-load.c
> > > index 2bbef81b6e..9b0c1406c9 100644
> > > --- a/elf/dl-load.c
> > > +++ b/elf/dl-load.c
> > > @@ -1158,6 +1158,10 @@ _dl_map_object_from_fd (const char *name, const char *origname, int fd,
> > >    	goto call_lose;
> > >          }
> > >    +    /* dlopen of an executable is not valid because it is not
> > > possible
> > > +       to perform proper relocations and run the ELF constructors.
> > > +       For PIE, the check needs the dynamic section, so there is
> > > +       another check below.  */
> >
> > Suggest you mention exactly which relocations are the problem.
> > I assume it's COPY relocs into the main executable which can't be
> > done twice, because both execs will want to reference their own
> > copy. Likewise the problem with ELF constructors should be expanded
> > with more details.
>
> I assume that static TLS is also negatively impacted.  TLS variables
> from the dlopen'ed PIE binary and the real main program will probably
> overlap.

Absolutely. They will both use constant offsets from TP and that will
be directly encoded in the program code with no way to change it.

--
Cheers,
Carlos.