[PATCH v2 0/2] ld: fix retargeted-output stub-file and hash-table accesses

Harshit Kumar harshit7kr@gmail.com
Fri Nov 6 14:28:26 GMT 2026


A linker script can retarget the output BFD with OUTPUT_FORMAT.  The
AVR, m68hc1x and VxWorks emulation hooks run after that and assume the
output backend created a stub file or an ELF link hash table.  The
generic elf32-gen/elf64-gen backends do neither, so the hooks either
dereference a NULL stub_file or read past a generic hash table.

Base commit: a729341631f5ddea915cebe722e4ec198ff96c38

The two patches are independent.  They are sent together because they
share the same retargeting trigger and the same review area.  I checked
the other emulation hooks in ld/emultempl/ and ld/*.c for the same
patterns (stub_file, elf_hash_table, bfd_target_elf_flavour) and found
no other unguarded sites.

1. ld: refuse the link when no stub file was created

   AVR and m68hc1x dereference stub_file unconditionally.  AVR tested
   the output flavour, which is ELF in the failing case; m68hc1x warned
   and then continued into the dereference.  Both now refuse the link.
   ldelf_after_open_output is called before the check, which safely
   returns early for non-ELF outputs.  Refusing here also ensures the
   two later unguarded stub_file uses in avrelf.em are not reached.

   `ld-x86-64/stub-file-retarget.exp` covers five emulations in four
   generic ELF formats: 20/20 fail on the unpatched tree (linker crash)
   and 20/20 pass on the patched tree (clean refusal message).  The
   test is placed in ld-x86-64/ because it uses an x86-64 relocatable
   object as the foreign input to test retargeting across generic ELF
   formats.

2. ld: guard the vxworks ELF link hash table read

   `vxworks_after_open` reads `dynamic_sections_created` after checking
   only the output flavour.  A generic ELF output allocates a generic
   bfd_link_hash_table, which does not contain the elf_link_hash_table
   fields.  The patch tests `is_elf_hash_table` first, matching the
   equivalent guards in ld/ldelf.c and bfd/elflink.c.

   This patch has no testsuite test: the only plain-build symptom is a
   false warning whose appearance depends on an uninitialised heap byte
   and the input object.  With AddressSanitizer, all seven VxWorks
   emulations report a heap-buffer-overflow in `vxworks_after_open`
   before the fix and no sanitizer report after it.

Changes in v2:
- Patch 1: Unchanged from v1; resent for completeness as the series
  is not yet applied.
- Patch 2: Dropped the explanatory comment in vxworks.em (Jan).
  The commit message covers the reasoning.

Harshit Kumar (2):
  ld: refuse the link when no stub file was created
  ld: guard the vxworks ELF link hash table read

 ld/emultempl/avrelf.em                        |  13 +-
 ld/emultempl/m68hc1xelf.em                    |  11 +-
 ld/emultempl/vxworks.em                       |   1 +
 ld/testsuite/ld-x86-64/stub-file-retarget.exp | 124 ++++++++++++++++++
 ld/testsuite/ld-x86-64/stub-file-retarget.s   |  14 ++
 5 files changed, 158 insertions(+), 5 deletions(-)
 create mode 100644 ld/testsuite/ld-x86-64/stub-file-retarget.exp
 create mode 100644 ld/testsuite/ld-x86-64/stub-file-retarget.s

-- 
2.55.0




More information about the Binutils mailing list