Dealing with broken ELF libraries

[Jeffrey A Law <law@cygnus.com>]

Ho hum.  In typical fashion HP made some goof's in their ELF implementation
which apparently nobody's noticed because nobody else in their right mind is
writing ELF tools to work on hpux11 :-)

Can someone spot the bug in these two header dumps?

readelf -S libc.so 
There are 42 section headers, starting at offset 0x1227b8:

Section Headers:
  [Nr] Name              Type            Addr     Off    Size   ES Flg Lk Inf 
Al
  [ 0]                   NULL            00000000 ffffffff 000000 00      0   
0 0
  [ 1] .dynamic          DYNAMIC         000011c8 0001c8 000150 10  A   0   0 8
  [ 2] .dynsym           DYNSYM          00001318 000318 00a050 18  A   3   0 8
  [ 3] .dynstr           STRTAB          0000b368 00a368 00482f 00  A   0   0 1

[ ... ]

[law@portal /scr/hp-elf-native/ld/test] readelf -S libdl.so
There are 35 section headers, starting at offset 0x51c8:

Section Headers:
  [Nr] Name              Type            Addr     Off    Size   ES Flg Lk Inf 
Al
  [ 0]                   NULL            00000000 ffffffff 000000 00      0   
0 0
  [ 1] .dynamic          DYNAMIC         000011c8 0001c8 000120 10  A   0   0 8
  [ 2] .dynsym           DYNSYM          000012e8 0002e8 000360 18  A   3   0 8
  [ 3] .dynstr           STRTAB          00001648 000648 000203 00  A   0   0 1


Stumped?  The link for the .dynamic field points back to section #0 instead
of pointing to section #3.  Opps.

This causes gld/bfd some problems:

elflink.c::elf_link_add_object_symbols

      s = bfd_get_section_by_name (abfd, ".dynamic");
[ ... blah blah blah ... ]

          elfsec = _bfd_elf_section_from_bfd_section (abfd, s);
          if (elfsec == -1)
            goto error_return;
          link = elf_elfsections (abfd)[elfsec]->sh_link;

So we get a link of 0.  Of course section #0 is not a string table.  This
causes grief later when we do something like call
bfd_elf_string_from_elf_section.


So what do y'all think the best workaround might be?  Yes, I know HP should fix
their tools & libraries, but the libraries with the bogus link fields are
fairly widespread already.  And while HP may fix their tools, all the existing
sites lose if we can't find a good workaround.

We might be able to do something like recognize that LINK refers to a section
that is not of type STRTAB (and then fall back to looking for .dynsym's
string table).

Other options?

jeff