[PATCH v5 3/5] x86: Split syscall_trace_enter into two phases

[Kees Cook]

On Thu, Feb 5, 2015 at 6:38 PM, Andy Lutomirski <luto at amacapital.net> wrote:
> On Thu, Feb 5, 2015 at 6:32 PM, Dmitry V. Levin <ldv at altlinux.org> wrote:
>> On Thu, Feb 05, 2015 at 04:09:06PM -0800, Andy Lutomirski wrote:
>>> On Thu, Feb 5, 2015 at 3:49 PM, Kees Cook <keescook at chromium.org> wrote:
>>> > On Thu, Feb 5, 2015 at 3:39 PM, Dmitry V. Levin <ldv at altlinux.org> wrote:
>> [...]
>>> >> There is a clear difference: before these changes, SECCOMP_RET_ERRNO used
>>> >> to keep the syscall number unchanged and suppress syscall-exit-stop event,
>>> >> which was awful because userspace cannot distinguish syscall-enter-stop
>>> >> from syscall-exit-stop and therefore relies on the kernel that
>>> >> syscall-enter-stop is followed by syscall-exit-stop (or tracee's death, etc.).
>>> >>
>>> >> After these changes, SECCOMP_RET_ERRNO no longer causes syscall-exit-stop
>>> >> events to be suppressed, but now the syscall number is lost.
>>> >
>>> > Ah-ha! Okay, thanks, I understand now. I think this means seccomp
>>> > phase1 should not treat RET_ERRNO as a "skip" event. Andy, what do you
>>> > think here?
>>>
>>> I still don't quite see how this change caused this.
>>
>> I have a test for this at
>> http://sourceforge.net/p/strace/code/ci/HEAD/~/tree/test/seccomp.c
>>
>>> I can play with
>>> it a bit more.  But RET_ERRNO *has* to be some kind of skip event,
>>> because it needs to skip the syscall.
>>>
>>> We could change this by treating RET_ERRNO as an instruction to enter
>>> phase 2 and then asking for a skip in phase 2 without changing
>>> orig_ax, but IMO this is pretty ugly.
>>>
>>> I think this all kind of sucks.  We're trying to run ptrace after
>>> seccomp, so ptrace is seeing the syscalls as transformed by seccomp.
>>> That means that if we use RET_TRAP, then ptrace will see the
>>> possibly-modified syscall, if we use RET_ERRNO, then ptrace is (IMO
>>> correctly given the current design) showing syscall -1, and if we use
>>> RET_KILL, then ptrace just sees the process mysteriously die.
>>
>> Userspace is usually not prepared to see syscall -1.
>> For example, strace had to be patched, otherwise it just skipped such
>> syscalls as "not a syscall" events or did other improper things:
>> http://sourceforge.net/p/strace/code/ci/c3948327717c29b10b5e00a436dc138b4ab1a486
>> http://sourceforge.net/p/strace/code/ci/8e398b6c4020fb2d33a5b3e40271ebf63199b891
>>
>
> The x32 thing is a legit ABI bug, I'd argue.  I'd be happy to submit a
> patch to fix that (clear the x32 bit if we're not x32).
>
>> A slightly different but related story: userspace is also not prepared
>> to handle large errno values produced by seccomp filters like this:
>> BPF_STMT(BPF_RET, SECCOMP_RET_ERRNO | SECCOMP_RET_DATA)
>>
>> For example, glibc assumes that syscalls do not return errno values greater than 0xfff:
>> https://sourceware.org/git/?p=glibc.git;a=blob;f=sysdeps/unix/sysv/linux/x86_64/sysdep.h#l55
>> https://sourceware.org/git/?p=glibc.git;a=blob;f=sysdeps/unix/sysv/linux/x86_64/syscall.S#l20

To save others the link reading: "Linus said he will make sure the no
syscall returns a value in -1 .. -4095 as a valid result so we can
savely test with -4095."

Strictly speaking (ISO C, "man 3 errno"), errno is supposed to be a
full int, though digging around I find this in include/linux/err.h:

/*
 * Kernel pointers have redundant information, so we can use a
 * scheme where we can return either an error code or a normal
 * pointer with the same return value.
 *
 * This should be a per-architecture thing, to allow different
 * error and pointer decisions.
 */
#define MAX_ERRNO       4095

#ifndef __ASSEMBLY__

#define IS_ERR_VALUE(x) unlikely((x) >= (unsigned long)-MAX_ERRNO)

But no architecture overrides this.

>> If it isn't too late, I'd recommend changing SECCOMP_RET_DATA mask
>> applied in SECCOMP_RET_ERRNO case from current 0xffff to 0xfff.

I'm not opposed to this. I would want to more explicitly document the
4095 max value in man pages, though.

> I think this is solidly in the "don't do that" category.  Seccomp lets
> you tamper with syscalls.  If you tamper wrong, then you lose.
>
> Kees, what do you think about reversing the whole thing so that
> ptracers always see the original syscall?

What do you mean by "reversing"? The interactions I see here are:

PTRACE_SYSCALL
SECCOMP_RET_ERRNO
SECCOMP_RET_TRACE
SECCOMP_RET_TRAP

Both ptrace and seccomp will trigger via _TIF_WORK_SYSCALL_ENTRY. Only
ptrace will trigger via _TIF_WORK_SYSCALL_EXIT.

For SECCOMP_RET_ERRNO to work, we must skip the syscall, as mentioned earlier:

arch/x86/kernel/entry_32.S ...
syscall_trace_entry:
        movl $-ENOSYS,PT_EAX(%esp)
        movl %esp, %eax
        call syscall_trace_enter
        /* What it returned is what we'll actually use.  */
        cmpl $(NR_syscalls), %eax
        jnae syscall_call
        jmp syscall_exit
END(syscall_trace_entry)

Both before and after the 2-phase change, syscall_trace_enter would
return -1 if it hit SECCOMP_RET_ERRNO, before calling
tracehook_report_syscall_entry. On exit, if PTRACE_SYSCALL, we'd hit
tracehook_report_syscall_exit during syscall_trace_leave, which means
a ptracer will see a syscall-exit-stop without a matching
syscall-enter-stop.

Using SECCOMP_RET_TRACE with PTRACE_SYSCALL in place seems totally
crazy, as the ptracer would need to be the same program, and if it
chose to skip a syscall, it would be in the same place: it would see
PTRACE_EVENT_SECCOMP, then no syscall-enter-stop, then a
syscall-exit-stop. I think we can ignore this pathological case.

Using SECCOMP_RET_TRAP with PTRACE_SYSCALL also results in a skip,
which produces the same "only syscall-exit-stop seen" problem.

In the SECCOMP_RET_ERRNO case, the syscall nr doesn't change (and
isn't executed). In the SECCOMP_RET_TRAP case, the syscall nr doesn't
change (and isn't executed). In the SECCOMP_RET_TRACE, the syscall nr
_could_ change, but the ptracer would be doing it, so the crazy
situation around PTRACE_SYSCALL is probably safe to ignore (as long as
we document what is expected to happen).

So, the question is: should PTRACE_SYSCALL see a syscall that is _not_
being executed (due to seccomp)? Audit doesn't see it currently, and
neither does ptrace. I would argue that it should continue to not see
the syscall. That said, if it shouldn't be shown, we also shouldn't
trigger syscall-exit-stop. If you can convince me it should see
syscall-enter-stop, then I have two questions:

1) Do we accept that a ptracer can interfere with SECCOMP_RET_ERRNO? I
think we probably must, since it can already interfere via
syscall-exit-stop and change the errno. And especially since a ptracer
can change syscalls during syscall-enter-stop to any syscall it wants,
bypassing seccomp. This condition is already documented.

2) What do we do with audit? Suddenly we have ptrace seeing a syscall
that audit doesn't?

And an unrelated thought:

3) Can't we find some way to fix the inability of a ptracer to
distinguish between syscall-enter-stop and syscall-exit-stop?

-Kees

-- 
Kees Cook
Chrome OS Security