_rtld_global_ro and static dlopen (was: Re: Pending fixes for all-ABIs builds)

[Florian Weimer <fweimer@redhat.com>]

On 11/10/2016 10:27 AM, Andreas Schwab wrote:
> On Nov 10 2016, Florian Weimer <fweimer@redhat.com> wrote:
>
>> Could we put this information into rtld_global_ro, instead of having a
>> separate variable?  (Assuming that rtld_global_ro is indeed in read-only
>> memory after relocation.)
>
> It cannot be read-only, it is set only after ld.so is initialized.

I know that glibc has trouble with multiple definitions, but at least 
one copy of _rtld_global_ro *is* read-only after relocation:

(gdb) print memset (&_rtld_global_ro, 1, 1)
Program received signal SIGSEGV, Segmentation fault.
__memset_sse2 () at ../sysdeps/x86_64/memset.S:110
110		movb	%cl, (%rdi)

It also contains a function pointer to _dl_lookup_symbol_x, so I think 
this data structure should work for locating ld.so functions.

Is the problem that we determine too late that we are running in a 
nested configuration, and the inner ld.so has already been relocated? 
Maybe gdb is lying to me, but it seems that we run the inner libc.so 
with a _rtld_global_ro which has not been initialized beyond what is in 
the default initializer.  It seems that dl_main never runs.

But this means that most uses of GLRO inside libc.so.6 are broken.  And 
indeed, GLRO (dl_auxv) does not work as intended, and results in a NULL 
pointer dereference in getauxval:

   https://sourceware.org/bugzilla/show_bug.cgi?id=20802

How can we fix this?  Could we synthesize a link map which interposes 
the necessary definitions?

Thanks,
Florian