[RFC] syscalls: Restore address limit after a syscall

Russell King - ARM Linux linux at armlinux.org.uk
Fri Feb 10 13:49:54 PST 2017


On Fri, Feb 10, 2017 at 12:49:34PM -0800, Kees Cook wrote:
> On Fri, Feb 10, 2017 at 11:22 AM, Russell King - ARM Linux
> <linux at armlinux.org.uk> wrote:
> > On Thu, Feb 09, 2017 at 06:42:34PM -0800, Andy Lutomirski wrote:
> >> On Thu, Feb 9, 2017 at 3:41 PM, Thomas Garnier <thgarnie at google.com> wrote:
> >> > So by default it is in the wrapper. If selected, an architecture can
> >> > disable the wrapper put it in the best places. Understood correctly?
> >>
> >> Sounds good to me.
> >>
> >> Presumably the result should go through -mm.  Want to cc: akpm and
> >> linux-arch@ on the next version?
> >>
> >> I've also cc'd arm and s390 folks -- those are the other arches that
> >> try to be on top of hardening.
> >
> > The best place for this on ARM is in the assembly code, rather than in
> > the hundreds of system calls - having it in one place is surely better
> > for reducing the cache impact.
> >
> > This (untested) patch should be sufficient for ARM - there's two choices
> > which I think make sense to do this:
> > 1. Immediately after returning the syscall
> > 2. Immediately before any returning to userspace
> >
> > (1) has the advantage that the address limit will be forced for the
> > exit-path works that we do, preventing those making accesses to kernel
> > space.
> >
> > (2) has the advantage that we'd guarantee that the address limit will
> > be forced while userspace is running for the next entry into kernel
> > space.
> >
> > There's actually a third option as well:
> >
> > (3) forcing the address limit on entry to the kernel from userspace.
> >
> > This patch implements option 1.
> >
> >  arch/arm/kernel/entry-common.S | 6 ++++++
> >  1 files changed, 6 insertions(+)
> >
> > diff --git a/arch/arm/kernel/entry-common.S b/arch/arm/kernel/entry-common.S
> > index eb5cd77bf1d8..6a717a2ccb88 100644
> > --- a/arch/arm/kernel/entry-common.S
> > +++ b/arch/arm/kernel/entry-common.S
> > @@ -39,6 +39,8 @@
> >  ret_fast_syscall:
> >   UNWIND(.fnstart       )
> >   UNWIND(.cantunwind    )
> > +       mov     r1, #TASK_SIZE
> > +       str     r1, [tsk, #TI_ADDR_LIMIT]
> >         disable_irq_notrace                     @ disable interrupts
> >         ldr     r1, [tsk, #TI_FLAGS]            @ re-check for syscall tracing
> >         tst     r1, #_TIF_SYSCALL_WORK | _TIF_WORK_MASK
> > @@ -64,6 +66,8 @@ ENDPROC(ret_fast_syscall)
> >  ret_fast_syscall:
> >   UNWIND(.fnstart       )
> >   UNWIND(.cantunwind    )
> > +       mov     r1, #TASK_SIZE
> > +       str     r1, [tsk, #TI_ADDR_LIMIT]
> >         str     r0, [sp, #S_R0 + S_OFF]!        @ save returned r0
> >         disable_irq_notrace                     @ disable interrupts
> >         ldr     r1, [tsk, #TI_FLAGS]            @ re-check for syscall tracing
> > @@ -262,6 +266,8 @@ ENDPROC(vector_swi)
> >         b       ret_slow_syscall
> >
> >  __sys_trace_return:
> > +       mov     r1, #TASK_SIZE
> > +       str     r1, [tsk, #TI_ADDR_LIMIT]
> >         str     r0, [sp, #S_R0 + S_OFF]!        @ save returned r0
> >         mov     r0, sp
> >         bl      syscall_trace_exit
> >
> 
> That looks pretty great! If I'm reading the macros correctly, this'll
> only happen on _actual_ syscall exit, right? So all the crazy OABI
> stuff won't suddenly break? e.g.:

Correct.

> Is there a similarly good place to do this for arm64?

I'd imagine similar places exist in arm64:

ret_fast_syscall
ret_to_user

maybe?

-- 
RMK's Patch system: http://www.armlinux.org.uk/developer/patches/
FTTC broadband for 0.8mile line: currently at 9.6Mbps down 400kbps up
according to speedtest.net.



More information about the linux-arm-kernel mailing list