[PATCH 1/2] ARM: entry-common: fix forgotten set of thread_info->syscall

[Kees Cook]

On Sun, Jan 18, 2015 at 9:58 PM, Roman Peniaev <r.peniaev at gmail.com> wrote:
> On Sat, Jan 17, 2015 at 8:54 AM, Kees Cook <keescook at chromium.org> wrote:
>> On Fri, Jan 16, 2015 at 11:57 AM, Kees Cook <keescook at chromium.org> wrote:
>>> On Fri, Jan 16, 2015 at 8:17 AM, Russell King - ARM Linux
>>> <linux at arm.linux.org.uk> wrote:
>>>> On Sat, Jan 17, 2015 at 01:08:11AM +0900, Roman Peniaev wrote:
>>>>> On Sat, Jan 17, 2015 at 12:59 AM, Russell King - ARM Linux
>>>>> <linux at arm.linux.org.uk> wrote:
>>>>> > On Sat, Jan 17, 2015 at 12:57:02AM +0900, Roman Peniaev wrote:
>>>>> >> On Fri, Jan 16, 2015 at 7:54 AM, Kees Cook <keescook at chromium.org> wrote:
>>>>> >> > One interesting thing I noticed (which is unchanged by this series),
>>>>> >> > but pulling ARM_r7 during the seccomp ptrace event shows __NR_poll,
>>>>> >> > not __NR_restart_syscall, even though it was a __NR_restart_syscall
>>>>> >> > trap from seccomp. Is there a better place to see the actual syscall?
>>>>> >>
>>>>> >> As I understand we do not push new r7 to the stack, and ptrace uses the
>>>>> >> old value.
>>>>> >
>>>>> > And why should we push r7 to the stack?  ptrace should be using the
>>>>> > recorded system call number, rather than poking about on the stack
>>>>> > itself.
>>>>>
>>>>> Probably we should not, but the behaviour comparing arm to x86 is different.
>>>>
>>>> We definitely should not, because changing the stacked value changes the
>>>> value in r7 after the syscall has returned.  We have guaranteed that the
>>>> value will be preserved across syscalls for years, so we really should
>>>> not be changing that.
>>>
>>> Yeah, we can't mess with the registers. I was just asking for
>>> clarification on how this is visible to userspace.
>>>
>>>>
>>>>> Also there is no any way from userspace to figure out what syscall was
>>>>> restarted, if you do not trace each syscall enter and exit from the
>>>>> very beginning.
>>>>
>>>> Thinking about ptrace, that's been true for years.
>>>>
>>>> It really depends whether you consider the restart syscall a userspace
>>>> thing or a kernelspace thing.  When you consider that the vast majority
>>>> of syscall restarts are done internally in the kernel, and we just
>>>> re-issue the syscall, it immediately brings up the question "why is
>>>> the restart block method different?" and "should the restart block
>>>> method be visible to userspace?"
>>>>
>>>> IMHO, it is prudent not to expose kernel internals to userspace unless
>>>> there is a real reason to, otherwise they become part of the userspace
>>>> API.
>>>
>>> I couldn't agree more, but restart_syscall is already visible to
>>> userspace: it can be called directly, for example. And it's visible to
>>> tracers.
>>>
>>> Unfortunately, the difference here is the visibility during trace
>>> trap. On x86, it's exposed but on ARM, there's no way (that I can
>>> find) to query the "true" syscall, even though the true syscall is
>>> what triggers the tracer. The syscall number isn't provided by any
>>> element of the ptrace event system, nor through siginfo, and must be
>>> examined on a per-arch basis from registers.
>>>
>>> Seccomp does, however, provide a mechanism to pass arbitrary event
>>> data on a TRACE event, so poll vs restart_syscall can be distinguished
>>> that way.
>>>
>>> It seems even strace doesn't know how to find this information. For example:
>>>
>>> x86:
>>> poll([{fd=3, events=POLLIN}], 1, 4294967295
>>> ) = ? ERESTART_RESTARTBLOCK (Interrupted by signal)
>>> --- SIGSTOP {si_signo=SIGSTOP, si_code=SI_USER, si_pid=994, si_uid=1000} ---
>>> --- stopped by SIGSTOP ---
>>> --- SIGCONT {si_signo=SIGCONT, si_code=SI_USER, si_pid=994, si_uid=1000} ---
>>> restart_syscall(<... resuming interrupted call ...>
>>>
>>> ARM:
>>> poll([{fd=3, events=POLLIN}], 1, -1
>>> )    = ? ERESTART_RESTARTBLOCK (Interrupted by signal)
>>> --- SIGSTOP {si_signo=SIGSTOP, si_code=SI_USER, si_pid=20563, si_uid=0} ---
>>> --- stopped by SIGSTOP ---
>>> --- SIGCONT {si_signo=SIGCONT, si_code=SI_USER, si_pid=20563, si_uid=0} ---
>>> poll([{fd=3, events=POLLIN}], 1, -1
>>>
>>> Would it make sense to add REGSET_SYSTEM_CALL to ARM? (Though this
>>> begs the question, "Is restart_syscall visible during a trace on
>>> arm64?", which I'll have to go check...)
>>
>> So, some further testing:
>> - native arm64 presents "poll" again even to seccomp when
>> restart_syscall is triggered (both via regs[8] and
>> NT_ARM_SYSTEM_CALL).
>> - compat mode on arm64 _does_ show syscall_restart (via ARM_r7).
>>
>> Which of these behaviors is intentional? :)
>>
>
>
> Just want to summarize the difference.
> (please, correct me if i am mistaken)
>
> Userspace has two ways to see actual syscall number:
>    1. /proc/pid/syscall file
>    2. ptrace
>
> So the following is the table showing what syscall number
> userspace sees using proc file or doing ptrace in case of restarted poll:
>
>                          x86           ARM              ARM64     ARM64 compat
> cat /proc/pid/syscall:   NR_restart    Not supported    ?????     ?????
>                ptrace:   NR_restart    NR_poll          NR_poll   NR_restart
>
>
> Not supported - should be fixed by these two patches, the behaviour should
>                 be similar to x86, i.e. userspace will see NR_restart
>
>          ???? - I do not have ARM64 for testing.
>                 Kees, could you please cat /proc/pid/syscall for those two?
>                 I took a quick look into arm64 syscall.h/entry.S and seems it
>                 is supported fine and the result should be equal to ptrace.

Yup, checking this directly agrees, /proc/$pid/syscall for me:

native arm64 shows NR_poll
arm64 compat shows NR_restart

> So, yes, compatibility is important, but /proc/pid/syscall never works on ARM
> and ptrace output is different even among ARM architectures.

-Kees

-- 
Kees Cook
Chrome OS Security