On 9/5/26 1:08 PM, liuc63 wrote: > From: Liu Chao <[email protected]> > > elf_validity_cache_secstrings() checks that every section name offset is > inside the section name table, but skips headers of type SHT_NULL, since > ELF leaves their field values undefined. > > However, the section lookup helpers do not check sh_type. This is > find_any_unique_sec(), which load_module() uses to locate ".modinfo": > > for (i = 1; i < info->hdr->e_shnum; i++) { > if (strcmp(info->secstrings + info->sechdrs[i].sh_name, > name) == 0) { > > find_any_sec() and module_enforce_rwx_sections() read sh_name the same > way, the latter from index 0. find_sec() is only shielded by its > SHF_ALLOC test, which a crafted module can satisfy. > > ELF also permits SHT_NULL entries above index 0. A module that carries > one with a large sh_name passes validation and then reads out of bounds: > > BUG: unable to handle page fault for address: ffffc9004052f988 > Oops: 0000 [#1] SMP KASAN PTI > CPU: 0 UID: 0 PID: 261 Comm: insmod Not tainted 6.18.0-rc7 #2 > RIP: 0010:strcmp+0xc/0x40 > Call Trace: > ? find_any_unique_sec+0x89/0xf0 > load_module+0x516/0x3d30 > init_module_from_file+0xf5/0x180 > __x64_sys_finit_module+0x91/0x100 > > The read happens in elf_validity_cache_copy(), before the blacklist check > in early_mod_check(). > > Bounds check sh_name for every section header. The name table is already > known to be non-empty, so the sh_name of 0 that ELF requires at index 0 > still passes, and a name offset that points outside the name table is > malformed regardless of section type. Neither an Ubuntu 20.04 x86_64 > build (GCC 9.4, 6025 modules) nor an aarch64 vendor build (GCC 13.2, 127 > modules) has a header with an out-of-bounds sh_name, or any SHT_NULL > header above index 0. > > Link: > https://lore.kernel.org/linux-modules/[email protected]/ > Cc: [email protected] > Signed-off-by: Liu Chao <[email protected]> > --- > The same report lists two more SHT_NULL gaps, both from > elf_validity_cache_sechdrs skipping offset and size validation: > move_module will memcpy such a section, and elf_validity_cache_index_mod > rejects SHT_NOBITS but not SHT_NULL. Sending this one on its own first, > happy to respin all three as a series if you would rather have them > together. > > The Oops line says KASAN, but the access faults rather than producing a > KASAN report, because lib/string.o is built with KASAN_SANITIZE disabled. > > No Fixes tag. 3c5700aeabd8 ("module: Factor out > elf_validity_cache_secstrings") is where the exemption was added, but the > problem predates it: before that commit sh_name was only checked for > SHF_ALLOC sections and the ".modinfo" strcmp() ran ahead of the check, so > find_any_sec() could already walk off the end. Same conclusion as > 9a5ff4568932 ("module: validate string table section types"), which went > in with Cc: stable and no Fixes tag. checkpatch complains about that > combination. > > The scan behind that last paragraph applies exactly the predicate this > patch introduces, sh_name >= strhdr->sh_size, to every section header > from index 0. On the x86_64 host it covered 17546 modules across three > kernel trees (5.15.0-181, 5.15.0-190, 5.4.0-182), all built by GCC 9.4.0; > those trees are the same toolchain recompiling much the same driver set, > so the changelog quotes one tree rather than the total. The aarch64 side > is a 5.10 vendor kernel built with Arm GNU Toolchain 13.2. Both are GNU > toolchains - no clang-built module set was available to me. > > Reproducer, needs root and an unsigned module on a kernel without > CONFIG_MODULE_SIG_FORCE. Build any minimal module, then append a > SHT_NULL section header with an out-of-bounds sh_name: > > python3 - <<'EOF' > import struct > SHDR = 0x40 > data = bytearray(open('dummy.ko', 'rb').read()) > shoff = struct.unpack_from('<Q', data, 0x28)[0] > shnum = struct.unpack_from('<H', data, 0x3c)[0] > old = bytes(data[shoff:shoff + shnum * SHDR]) > evil = bytearray(SHDR) > struct.pack_into('<I', evil, 0, 0x40000000) > data += b'\0' * ((8 - len(data) % 8) % 8) > new_shoff = len(data) > data += old + bytes(evil) > struct.pack_into('<Q', data, 0x28, new_shoff) > struct.pack_into('<H', data, 0x3c, shnum + 1) > open('evil.ko', 'wb').write(data) > EOF > > insmod evil.ko > > kernel/module/main.c | 3 --- > 1 file changed, 3 deletions(-) > > diff --git a/kernel/module/main.c b/kernel/module/main.c > index d0e1e0bd2ad0..7d5943cd6543 100644 > --- a/kernel/module/main.c > +++ b/kernel/module/main.c > @@ -2060,9 +2060,6 @@ static int elf_validity_cache_secstrings(struct > load_info *info) > > for (i = 0; i < info->hdr->e_shnum; i++) { > shdr = &info->sechdrs[i]; > - /* SHT_NULL means sh_name has an undefined value */ > - if (shdr->sh_type == SHT_NULL) > - continue; > if (shdr->sh_name >= strhdr->sh_size) { > pr_err("Invalid ELF section name in module (section %u > type %u)\n", > i, shdr->sh_type);
The ELF specification [1] says: | SHT_NULL | | This value marks the section header as inactive; it does not have an | associated section. Other members of the section header have | undefined values. The patch may cause the module loader to read undefined values if the section has type SHT_NULL. I think the module loader should handle SHT_NULL properly by skipping such sections during all processing. Alternatively, it could reject any module with an SHT_NULL section other than the one at index 0. [1] https://gabi.xinuos.com/v42/elf/03-sheader.html#section-type -- Thanks, Petr
