[PATCH] A little do-rel.h fix (take 2)

[Jakub Jelinek]

On Fri, Aug 24, 2001 at 09:17:39AM -0700, Ulrich Drepper wrote:
> Jakub Jelinek <jakub@redhat.com> writes:
> 
> > The first one is only if you think it is a good idea to put it in now
> > (otherwise I'll keep it in my prelink patchset where it is obvisously very
> > useful e.g. for dlopening prelinked libs).
> 
> Is this always correct for rela platforms?  Are the relocation
> addresses required to contain the same value as the r_addend?  I'd
> expect not and then it patch is wrong.
> 
> What does m68k do for instance?

You're right.
For REL this should work always though, for RELA the check will have to be
something like
    && (l_addr || map->l_info[VALIDX(DT_GNU_PRELINKED)])
(if prelinking works, the RELATIVE memory really needs to be set up
properly).
But I was especially interested in the r = endrel; line since that matters
even for non-prelinked binaries (there are couple dozens of RELATIVE relocs
in ld.so).

2001-08-24  Jakub Jelinek  <jakub@redhat.com>

	* elf/do-rel.h (elf_dynamic_do_rel): If l_addr is 0, skip relative
	records.  Set r unconditionally to first non-relative record.

--- libc/elf/do-rel.h.jj	Fri Aug 24 18:15:24 2001
+++ libc/elf/do-rel.h	Fri Aug 24 18:18:23 2001
@@ -70,11 +70,17 @@ elf_dynamic_do_rel (struct link_map *map
 	 _dl_rtld_map, which is incompatible with a weak decl in the same
 	 file.  */
       weak_extern (_dl_rtld_map);
-      if (map != &_dl_rtld_map) /* Already done in rtld itself.  */
+      if (map != &_dl_rtld_map /* Already done in rtld itself.  */
+#ifndef DO_RELA
+	  && l_addr
+#endif
+	 )
 #endif
 	for (; r < endrel; ++r)
 	  elf_machine_rel_relative (l_addr, r,
 				    (void *) (l_addr + r->r_offset));
+      r = endrel;
 
       if (map->l_info[VERSYMIDX (DT_VERSYM)])
 	{


	Jakub