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

[Kees Cook]

On Fri, Jun 20, 2014 at 10:36 AM, Kees Cook <keescook at chromium.org> wrote:
> On Fri, Jun 20, 2014 at 10:23 AM, Will Deacon <will.deacon at arm.com> wrote:
>> On Fri, Jun 20, 2014 at 05:44:52PM +0100, Kees Cook wrote:
>>> On Fri, Jun 20, 2014 at 3:22 AM, Will Deacon <will.deacon at arm.com> wrote:
>>> > I'm struggling to see the bug in the current code, so apologies if my
>>> > questions aren't helpful.
>>> >
>>> > On Wed, Jun 18, 2014 at 09:27:48PM +0100, Kees Cook wrote:
>>> >> An x86 tracer wanting to change the syscall uses PTRACE_SETREGS
>>> >> (stored to regs->orig_ax), and an ARM tracer uses PTRACE_SET_SYSCALL
>>> >> (stored to current_thread_info()->syscall). When this happens, the
>>> >> syscall can change across the call to secure_computing(), since it may
>>> >> block on tracer notification, and the tracer can then make changes
>>> >> to the process, before we return from secure_computing(). This
>>> >> means the code must respect the changed syscall after the
>>> >> secure_computing() call in syscall_trace_enter(). The same is true
>>> >> for tracehook_report_syscall_entry() which may also block and change
>>> >> the syscall.
>>> >
>>> > I don't think I understand what you mean by `the code must respect the
>>> > changed syscall'. The current code does indeed issue the new syscall, so are
>>> > you more concerned with secure_computing changing ->syscall, then the
>>> > debugger can't see the new syscall when it sees the trap from tracehook?
>>> > Are these even supposed to inter-operate?
>>>
>>> The problem is the use of "scno" in the call -- it results in ignoring
>>> the value that may be set up in ->syscall by a tracer:
>>>
>>> syscall_trace_enter(regs, scno):
>>>     current_thread_info()->syscall = scno;
>>>     secure_computing(scno):
>>>         [block on ptrace]
>>>         [ptracer changes current_thread_info()->syscall vis PTRACE_SET_SYSCALL]
>>>     ...
>>>     return scno;
>>>
>>> This means the tracer's changed syscall value will be ignored
>>> (syscall_trace_enter returns original "scno" instead of actual
>>> current_thread_info()->syscall). In the original code, failure cases
>>> were propagated correctly, but not tracer-induced changes.
>>>
>>> Is that more clear? It's not an obvious state (due to the external
>>> modification of process state during the ptrace blocking). I've also
>>> got tests for this, if that's useful to further illustrate:
>>>
>>> https://github.com/kees/seccomp/commit/bd24e174593f79784b97178b583f17e0ea9d2aa7
>>
>> Right, gotcha. Thanks for the explanation. I was confused, because
>> tracehook_report_syscall does the right thing (returns
>> current_thread_info()->syscall), but if we don't have TIF_SYSCALL_TRACE set,
>> then updates during the secure_computing callback will be ignored.
>>
>> However, my fix to this is significantly smaller than your patch, so I fear
>> I'm still missing something.
>
> Oh, yes, that's much smaller. Nice! I will test this and report back.

Yup, I can confirm this works. Thanks!

Tested-by: Kees Cook <keescook at chromium.org>

-Kees

-- 
Kees Cook
Chrome OS Security