[PATCH] arm: ptrace: fix syscall modification under PTRACE_O_TRACESECCOMP

[Andy Lutomirski]

On Thu, Jul 3, 2014 at 9:32 AM, Will Deacon <will.deacon at arm.com> wrote:
> On Thu, Jul 03, 2014 at 05:13:50PM +0100, Andy Lutomirski wrote:
>> On Thu, Jul 3, 2014 at 9:11 AM, Will Deacon <will.deacon at arm.com> wrote:
>> > On Thu, Jul 03, 2014 at 04:39:21PM +0100, Andy Lutomirski wrote:
>> >> On Thu, Jul 3, 2014 at 3:24 AM, Will Deacon <will.deacon at arm.com> wrote:
>> >> > On Thu, Jul 03, 2014 at 08:43:07AM +0100, AKASHI Takahiro wrote:
>> >> >> So it seems to me that we can't change a system call by ptrace().
>> >> >> Do I misunderstand anything?
>> >> >
>> >> > No, it looks like you have a point here. I don't think userspace has any
>> >> > business with orig_x0, but changing syscallno is certainly useful. I can
>> >> > think of two ways to fix this:
>> >> >
>> >> >   (1) Updating syscallno based on w8, but this ties us to the current ABI
>> >> >       and could get messy if this register changes in the future.
>> >> >
>> >> >   (2) Adding a PTRACE_SET_SYSCALL request, like we have for arch/arm/,
>> >> >       but that means adding arch-specific stuff to arch_ptrace (which
>> >> >       currently goes straight to ptrace_request on arm64).
>> >> >
>> >> > It looks like x86 uses orig_ax, which I *think* means we would go with
>> >> > (1) above if we followed their lead.
>> >>
>> >> w8 is a real register, right?  On x86, at least orig_ax isn't a real
>> >> register, so it's quite unlikely to conflict with hardware stuff.
>> >
>> > Yeah, w8 is the hardware register which the Linux ABI uses for the system
>> > call number. I was thinking We could allow the debugger/tracer to update
>> > the syscall number by updating that register, or do you see an issue with
>> > that? (other than tying us to the current ABI).
>>
>> Not immediately, but I'm not super-familiar with ptrace.
>>
>> Is w8 clobbered or otherwise changed by syscalls?  Using w8 for this
>> has the odd effect that tracers can't force a return with a specific
>> value of w8 without executing the corresponding syscall.  If that's a
>> meaningful limitation, then presumably some other channel should be
>> used.
>
> Hmm, that's true. Currently w8 is preserved across a syscall by the kernel,
> so it would be pretty bizarre for somebody to try and modify it but I guess
> they could do it if they wanted to. However, they could just as easily
> modify it on the syscall return path and have the same effect...
>
> Furthermore, glibc unconditionally emits a mov into w8 prior to the svc
> instruction, so from a user's perspective that register always contains
> the system call number.

That means that, if someone uses a seccomp trace action to skip or
emulate a syscall by writing -1 to w8, then user code will see an
unexpected -1 in w8.  I don't know how much this matters.

>
> FWIW: the role of w8 in the PCS is `Indirect result location register', so
> I'd expect it to be saved across the syscall anyway.
>
> Will



-- 
Andy Lutomirski
AMA Capital Management, LLC