On Tue, 3 Dec 2019, Nick Desaulniers wrote:

On Fri, Nov 29, 2019 at 1:33 PM Nicolas Pitre nico@fluxnic.net wrote:

...

On Fri, 22 Nov 2019, Nick Desaulniers wrote:

...

From: Kees Cook keescook@chromium.org

There's an implicit dependency on the section ordering of the orphaned section .fixup that can break arm_copy_from_user if the linker places the .fixup section before the .text section. Since .fixup is not explicitly placed in the existing ARM linker scripts, the linker is free to order it anywhere with respect to the rest of the sections.

Multiple users from different distros (Raspbian, CrOS) reported kernel panics executing seccomp() syscall with Linux kernels linked with LLD.

Documentation/x86/exception-tables.rst alludes to the ordering dependency. The relevant quote:

NOTE:
Due to the way that the exception table is built and needs to be ordered,
only use exceptions for code in the .text section.  Any other section
will cause the exception table to not be sorted correctly, and the
exceptions will fail.

Things changed when 64-bit support was added to x86 Linux. Rather than
double the size of the exception table by expanding the two entries
from 32-bits to 64 bits, a clever trick was used to store addresses
as relative offsets from the table itself. The assembly code changed
from::

    .long 1b,3b
  to:
          .long (from) - .
          .long (to) - .

and the C-code that uses these values converts back to absolute addresses
like this::

        ex_insn_addr(const struct exception_table_entry *x)
        {
                return (unsigned long)&x->insn + x->insn;
        }

Since the addresses stored in the __ex_table are RELATIVE offsets and not ABSOLUTE addresses, ordering the fixup anywhere that's not immediately preceding .text causes the relative offset of the faulting instruction to be wrong, causing the wrong (or no) address of the fixup handler to looked up in __ex_table.

This explanation makes no sense.

The above is valid only when ARCH_HAS_RELATIVE_EXTABLE is defined. On ARM32 it is not, nor would it make sense to be.

Hmm...I thought that was the smoking gun. From the description in Documentation, I thought they meant that exception table entry lookup was changed to be homogeneous for 32b AND 64b arch's, but as you point out they're not. Now with the reference to ARCH_HAS_RELATIVE_EXTABLE, I know to look through: include/asm-generic/extable.h include/linux/extable.h lib/extable.c kernel/extable.c arch/arm/mm/extable.c arch/arm/mm/fault.c (__do_kernel_fault() calls fixup_exception(), which is of interest).

Looks like the exception table is sorted by address of faulting instruction, then binary searched when an exception occurs. Seems the exception table is like an array of pairs of addresses (address of faulting instruction from the get_user() call, address of fixup handler) (when !ARCH_HAS_RELATIVE_EXTABLE), IIUC.

Exact. And those are absolute addresses.

To save on memory footprint, this table was made into relative addresses on 64-bit targets with the restriction that the relative offsets have to fit in a 32-bit value and not be negative. On 32-bit targets there is no point using relative addresses as the absolute addresses use only 32 bits already, and then we don't have to compute the absolute address at run time.

Reviewing the logs from the bugreport (https://bugs.chromium.org/p/chromium/issues/detail?id=1020633#c44) indeed the error string looks like the error message from __do_kernel_fault() in arch/arm/mm/fault.c where a call to fixup_exceptions() in arch/arm/mm/extable returned 1 because the call to search_exception_tables() in kernel/extable.c returned NULL.

So I was right that `no address of the fixup handler to (be) looked up in __ex_table`, but not quite right about *why* it fails to be looked up.

search_exception_tables() in kernel/extable.c calls 3 functions:

1. search_kernel_exception_table()
2. search_module_extables()
3. search_bpf_extables()

So the next question is which one of the above should have found the exception table entry, and why did it not when LLD placed the orphaned .fixup section BEFORE .text? All three of the above do some processing on the address but in the end all call search_extable().

I really don't think kernel modules are involved.

Since we're observing the fault via a call to seccomp(), which IIUC takes a BPF program when setting a filter (SECCOMP_SET_MODE_FILTER), I'm curious to look into search_bpf_extables() first. Maybe constructing a BPF program involves insertion of special exception handler? Looks like bpf programs are stored in an rb_tree called bpf_tree. They have an auxiliary field that contains a pointer to an exception table entry. "Auxiliary" makes it sounds optional, and the only assignment I can find to this field is in arch/x86/net/bpf_jit_comp.c, so I doubt it's relevant for arm. It also just looks like it's zero initialized (bpf_prog_alloc_no_stats() in kernel/bpf/core.c) for arch generic code.

That leaves just search_kernel_exception_table(). I wonder if it fails because we put garbage entries in, or sorted it improperly?

Well, if you can reproduce the issue (would be very helpful) then you can add some printk to log search_exception_tables() and figure out for sure actually which category is involved here. Unlike with user space, kernel exceptions are relatively rare events and they shouldn't flood your log.

Oh, and messing around with grep, it looks like entries to the exception table can be sorted at build time rather than boot time via CONFIG_BUILDTIME_EXTABLE_SORT?

I don't see the pr_notice from sort_main_extable() (kernel/extable.c called from start_kernel() in init/main.c) in the bugreports: https://bugs.chromium.org/p/chromium/issues/attachmentText?aid=421842 https://github.com/ClangBuiltLinux/linux/issues/282 (But IIRC, the log level may not be set to display these).

CONFIG_BUILDTIME_EXTABLE_SORT is selected by arch/arm/Kconfig if CONFIG_MMU, which I highly suspect is already selected for the above two reports. (Would an arm32 based chromebook not have an MMU? I doubt it.)

So there may be an ordering dependency in scripts/sortextable.{c|h}? Neither mention `fixup`.

Will continue debugging more tomorrow or later this week, but please stop me if any of the above is also incorrect. Some ideas for further experiments:

• scripts/check_extable.sh sounds like some kind of validator.

Manoj/Nathan, I wonder if you linked the problematic kernel with LLD than ran `./scripts/check_extable.sh vmlinux` if it would warn? Dunno about all those command line flags though.

• if we apply a similar diff to the patch I posted, but force .fixup

to be before .text for BFD (as LLD is placing the orphaned section currently), relinked with BFD and could reproduce the issue, that seems like proof about the implicit section ordering.

diff --git a/arch/arm/kernel/vmlinux.lds.S b/arch/arm/kernel/vmlinux.lds.S
index 319ccb10846a..adfb5297f359 100644
--- a/arch/arm/kernel/vmlinux.lds.S
+++ b/arch/arm/kernel/vmlinux.lds.S
@@ -58,6 +58,7 @@ SECTIONS
 #ifdef CONFIG_ARM_MPU
        . = ALIGN(PMSAv8_MINALIGN);
 #endif
+       .fixup : { *(.fixup) }
        .text : {                       /* Real text segment            */
                _stext = .;             /* Text and read-only data      */
                ARM_TEXT

-- Thanks, ~Nick Desaulniers