On 9/10/26 12:59 PM, Helge Deller wrote:
> On 9/9/26 01:08, Karl Mehltretter wrote:
>> apply_relocations() skips relocation sections whose sh_info target index
>> is outside the section table. ARM, ARM64, LoongArch, PA-RISC and RISC-V
>> use sh_info earlier in module_frob_arch_sections(), before this check.
>>
>> ARM, ARM64, LoongArch and RISC-V use the unchecked index to read
>> sh_flags outside the section header table. PA-RISC uses it to index an
>> e_shnum-sized heap array for a read and an update. QEMU reproduced
>> page-fault Oopses on ARM, ARM64, LoongArch and RISC-V, and a Data TLB
>> miss on the PA-RISC array read.
>>
>> Validate sh_info for SHT_REL and SHT_RELA sections in
>> elf_validity_cache_sechdrs(). Reject the module with ENOEXEC before
>> architecture code can use the index.
>>
>> Fixes: c298be74492b ("parisc: fix module loading failure of large kernel
>> modules")
>> Fixes: 7d485f647c1f ("ARM: 8220/1: allow modules outside of bl range")
>> Fixes: fd045f6cd98e ("arm64: add support for module PLTs")
>> Fixes: ab1ef68e5401 ("RISC-V: Add sections of PLT and GOT for kernel module")
>> Fixes: fcdfe9d22bed ("LoongArch: Add ELF and module support")
>> Cc: [email protected]
>> Assisted-by: LLM
>> Signed-off-by: Karl Mehltretter <[email protected]>
>> ---
>>
>> A custom harness for upstream Frama-C 33.0 (Arsenic) Eva found the
>> ARM32 instance in a source-identical ARM module_frob_arch_sections()
>> slice. Eva reported the out-of-range section-table pointer and sh_flags
>> access.
>>
>> The analysis and ARM32 A/B test ran at Linux b9b3e33b70b7 ("Merge tag
>> 'trace-v7.2-rc6' of
>> git://git.kernel.org/pub/scm/linux/kernel/git/trace/linux-trace"). The
>> PA-RISC, ARM64, RISC-V, LoongArch and x86_64 A/B tests ran at the
>> declared base commit, 28924df2a08f. arch/arm/kernel/module-plts.c and
>> the touched loop in kernel/module/main.c are identical between the two
>> commits.
>>
>> Each A/B test changed only a relocation section's sh_info to
>> 0x10000000. The same configurations and modules were used before and
>> after the change. All controls loaded before and after the change. The
>> fixed kernels rejected the malformed modules with ENOEXEC.
>>
>> Original-kernel results with QEMU 10.2.1 TCG:
>>
>> - ARM32, virt/Cortex-A15, GCC 15.2.0, multi_v7_defconfig plus
>> VMSPLIT_2G: page fault at module_frob_arch_sections()+0x160.
>> - ARM64, virt/Cortex-A57, GCC 15.2.0, defconfig: page fault at
>> module_frob_arch_sections()+0x110.
>> - PA-RISC, B160L, hppa-linux-gcc 8.1.0, binutils 2.30,
>> generic-32bit_defconfig: Data TLB miss at
>> module_frob_arch_sections()+0x11c on the stub_entries read for a
>> counted R_PARISC_PCREL17F relocation.
>> - RISC-V, virt, GCC 15.2.0, defconfig plus RELOCATABLE with
>> MODULE_SECTIONS enabled: page fault at
>> module_frob_arch_sections()+0xe4.
>> - LoongArch, virt/LA464, LLVM 21.1.8, loongson64_defconfig: page fault
>> at module_frob_arch_sections()+0x1b8.
>>
>> On x86_64, which has no vulnerable early sh_info access, the original
>> kernel loaded both modules. The fixed kernel loaded the control and
>> rejected the malformed module with ENOEXEC. The test used pc/qemu64,
>> x86_64_defconfig and GCC 15.2.0.
>
> Interesting.
> So, this patch helps to prevent loading modules with buggy entries,
> even if the module was e.g. loaded on x86-64 before.
>
> For me the patch is OK, but it only adds one random sanitizing check,
> and where would we stop to check?
Loading a module should normally at least get through the signature and
blacklist checks without crashing due to a corrupted module ELF file.
After that point, I believe the ELF data should be trusted, similar to
how the actual module code is expected to be correct.
module_frob_arch_sections() is called later in the module-loading
process, after the signature and blacklist checks, so I think this patch
is not strictly necessary.
--
Thanks,
Petr
