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

[will.deacon@arm.com (Will Deacon)]

On Fri, Jan 16, 2015 at 11:54:45PM +0000, Kees Cook wrote:
> On Fri, Jan 16, 2015 at 11:57 AM, Kees Cook <keescook@chromium.org> wrote:
> > On Fri, Jan 16, 2015 at 8:17 AM, Russell King - ARM Linux
> > <linux@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@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@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).

I'm fine either way for the native case, but we should stick with whetever
we end up with. Being compatible with ARM is probably a good idea. Do you
have a preference?

> - compat mode on arm64 _does_ show syscall_restart (via ARM_r7).

That sounds like a bug, then. Any chance you could look into a patch?

Will