add_new_section_symbols() sizes its symndx_map VLA as sh_size /
sh_entsize - 1 elements.  libelf bounds symbol accesses by sh_size
divided by the size of one symbol, and gelf_getsymshndx() rejects only
indices past that bound.  When the stripped file's SHT_SYMTAB has an
sh_entsize larger than the symbol size, the VLA is smaller than the
number of symbols libelf accepts.  A leading run of STT_SECTION symbols
then makes the loop write symndx_map[i - 1] past the end of the array,
on the stack.

With AddressSanitizer this shows as a dynamic-stack-buffer-overflow,
WRITE of size 8, at unstrip.c in add_new_section_symbols.  It is
reached by an ET_REL stripped file with sh_entsize 48 and 501 or more
sections, and by the four-section file in the new test.

Add valid_symtab_entsize () and use it where unstrip derives a symbol
count from sh_entsize: the stripped SYMTAB and DYNSYM copy loop, and
the merge path for the stripped and unstripped SYMTAB.  A symbol table
whose sh_entsize is not the symbol size is malformed, and is now
rejected.  This extends the sh_size < sh_entsize check of commit
364419bc.  Symbol tables with the usual entsize are not affected.

add_new_section_symbols() also refuses an old_shnum larger than nsym
before it writes symndx_map, as defense in depth.  The entsize check
makes that case unreachable from input, but the bound stays next to
the array it protects.

Related: PR tools/34705.

* src/unstrip.c (valid_symtab_entsize): New function.
(add_new_section_symbols): Reject old_shnum larger than nsym.
(copy_elided_sections): Use valid_symtab_entsize for the stripped
SYMTAB and DYNSYM copy loop, and for the stripped and unstripped
SYMTAB merge.
* tests/run-unstrip-symtab-entsize.sh: New test script.
* tests/testfile-unstrip-symtab2.stripped.bz2: New test file.
* tests/testfile-unstrip-symtab2.debug.bz2: Likewise.
* tests/Makefile.am (TESTS): Add run-unstrip-symtab-entsize.sh.
(EXTRA_DIST): Add new test script and bz2 files.

Signed-off-by: vivekjha <[email protected]>

Attachment: 0001-unstrip-reject-SYMTAB-sh_entsize-that-is-not-the-sym.patch
Description: Binary data

Reply via email to