Re: [PATCH v13 8/8] x86/vsyscall/64: Fixup Shadow Stack and Indirect Branch Tracking for vsyscall emulation

[Andy Lutomirski <luto@kernel.org>]

On Thu, Oct 1, 2020 at 9:51 AM Yu, Yu-cheng <yu-cheng.yu@intel.com> wrote:
>
> On 9/30/2020 6:10 PM, Andy Lutomirski wrote:
> > On Wed, Sep 30, 2020 at 6:01 PM H.J. Lu <hjl.tools@gmail.com> wrote:
> >>
> >> On Wed, Sep 30, 2020 at 4:44 PM Andy Lutomirski <luto@kernel.org> wrote:
>
> [...]
>
> >>>>>>>    From 09803e66dca38d7784e32687d0693550948199ed Mon Sep 17 00:00:00 2001
> >>>>>>> From: Yu-cheng Yu <yu-cheng.yu@intel.com>
> >>>>>>> Date: Thu, 29 Nov 2018 14:15:38 -0800
> >>>>>>> Subject: [PATCH v13 8/8] x86/vsyscall/64: Fixup Shadow Stack and
> >>>>>>> Indirect Branch
> >>>>>>>     Tracking for vsyscall emulation
> >>>>>>>
> >>>>>>> Vsyscall entry points are effectively branch targets.  Mark them with
> >>>>>>> ENDBR64 opcodes.  When emulating the RET instruction, unwind shadow stack
> >>>>>>> and reset IBT state machine.
> >>>>>>>
> >>>>>>> Signed-off-by: Yu-cheng Yu <yu-cheng.yu@intel.com>
> >>>>>>> ---
> >>>>>>> v13:
> >>>>>>> - Check shadow stack address is canonical.
> >>>>>>> - Change from writing to MSRs to writing to CET xstate.
> >>>>>>>
> >>>>>>>     arch/x86/entry/vsyscall/vsyscall_64.c     | 34 +++++++++++++++++++++++
> >>>>>>>     arch/x86/entry/vsyscall/vsyscall_emu_64.S |  9 ++++++
> >>>>>>>     arch/x86/entry/vsyscall/vsyscall_trace.h  |  1 +
> >>>>>>>     3 files changed, 44 insertions(+)
> >>>>>>>
> >>>>>>> diff --git a/arch/x86/entry/vsyscall/vsyscall_64.c
> >>>>>>> b/arch/x86/entry/vsyscall/vsyscall_64.c
> >>>>>>> index 44c33103a955..30b166091d46 100644
> >>>>>>> --- a/arch/x86/entry/vsyscall/vsyscall_64.c
> >>>>>>> +++ b/arch/x86/entry/vsyscall/vsyscall_64.c
> >>>>>>> @@ -38,6 +38,9 @@
> >>>>>>>     #include <asm/fixmap.h>
> >>>>>>>     #include <asm/traps.h>
> >>>>>>>     #include <asm/paravirt.h>
> >>>>>>> +#include <asm/fpu/xstate.h>
> >>>>>>> +#include <asm/fpu/types.h>
> >>>>>>> +#include <asm/fpu/internal.h>
> >>>>>>>
> >>>>>>>     #define CREATE_TRACE_POINTS
> >>>>>>>     #include "vsyscall_trace.h"
> >>>>>>> @@ -286,6 +289,44 @@ bool emulate_vsyscall(unsigned long error_code,
> >>>>>>>           /* Emulate a ret instruction. */
> >>>>>>>           regs->ip = caller;
> >>>>>>>           regs->sp += 8;
> >>>>>>> +
> >>>>>>> +#ifdef CONFIG_X86_CET
> >>>>>>> +       if (tsk->thread.cet.shstk_size || tsk->thread.cet.ibt_enabled) {
> >>>>>>> +               struct cet_user_state *cet;
> >>>>>>> +               struct fpu *fpu;
> >>>>>>> +
> >>>>>>> +               fpu = &tsk->thread.fpu;
> >>>>>>> +               fpregs_lock();
> >>>>>>> +
> >>>>>>> +               if (!test_thread_flag(TIF_NEED_FPU_LOAD)) {
> >>>>>>> +                       copy_fpregs_to_fpstate(fpu);
> >>>>>>> +                       set_thread_flag(TIF_NEED_FPU_LOAD);
> >>>>>>> +               }
> >>>>>>> +
> >>>>>>> +               cet = get_xsave_addr(&fpu->state.xsave, XFEATURE_CET_USER);
> >>>>>>> +               if (!cet) {
> >>>>>>> +                       /*
> >>>>>>> +                        * This should not happen.  The task is
> >>>>>>> +                        * CET-enabled, but CET xstate is in INIT.
> >>>>>>> +                        */
> >>>>>>
> [...]
> >>>>>>
> >>>>>
> >>>>> For what it's worth, I think there is an alternative.  If you all
> >>>>> (userspace people, etc) can come up with a credible way for a user
> >>>>> program to statically declare that it doesn't need vsyscalls, then we
> >>>>> could make SHSTK depend on *that*, and we could avoid this mess.  This
> >>>>> breaks orthogonality, but it's probably a decent outcome.
> >>>>>
> >>>>
> >>>> Would an arch_prctl(DISABLE_VSYSCALL) work?  The kernel then sets a
> >>>> thread flag, and in emulate_vsyscall(), checks the flag.
> >>>>
> >>>> When CET is enabled, ld-linux will do DISABLE_VSYSCALL.
> >>>>
> >>>> How is that?
> >>>
> >>> Backwards, no?  Presumably vsyscall needs to be disabled before or
> >>> concurrently with CET being enabled, not after.
> >>>
> >>> I think the solution of making vsyscall emulation work correctly with
> >>> CET is going to be better and possibly more straightforward.
> >>>
> >>
> >> We can do
> >>
> >> 1. Add ARCH_X86_DISABLE_VSYSCALL to disable the vsyscall page.
> >> 2. If CPU supports CET and the program is CET enabled:
> >>      a. Disable the vsyscall page.
> >>      b. Pass control to user.
> >>      c. Enable the vsyscall page when ARCH_X86_CET_DISABLE is called.
> >>
> >> So when control is passed from kernel to user, the vsyscall page is
> >> disabled if the program
> >> is CET enabled.
> >
> > Let me say this one more time:
> >
> > If we have a per-process vsyscall disable control and a per-process
> > CET control, we are going to keep those settings orthogonal.  I'm
> > willing to entertain an option in which enabling SHSTK without also
> > disabling vsyscalls is disallowed, We are *not* going to have any CET
> > flags magically disable vsyscalls, though, and we are not going to
> > have a situation where disabling vsyscalls on process startup requires
> > enabling SHSTK.
> >
> > Any possible static vsyscall controls (and CET controls, for that
> > matter) also need to come with some explanation of whether they are
> > properties set on the ELF loader, the ELF program being loaded, or
> > both.  And this explanation needs to cover what happens when old
> > binaries link against new libc versions and vice versa.  A new
> > CET-enabled binary linked against old libc running on a new kernel
> > that is expected to work on a non-CET CPU MUST work on a CET CPU, too.
> >
> > Right now, literally the only thing preventing vsyscall emulation from
> > coexisting with SHSTK is that the implementation eeds work.
> >
> > So your proposal is rejected.  Sorry.
> >
> I think, even with shadow stack/ibt enabled, we can still allow XONLY
> without too much mess.
>
> What about this?
>
> Thanks,
> Yu-cheng
>
> ======
>
> diff --git a/arch/x86/entry/vsyscall/vsyscall_64.c
> b/arch/x86/entry/vsyscall/vsyscall_64.c
> index 8b0b32ac7791..d39da0a15521 100644
> --- a/arch/x86/entry/vsyscall/vsyscall_64.c
> +++ b/arch/x86/entry/vsyscall/vsyscall_64.c
> @@ -48,16 +48,16 @@
>   static enum { EMULATE, XONLY, NONE } vsyscall_mode __ro_after_init =
>   #ifdef CONFIG_LEGACY_VSYSCALL_NONE
>          NONE;
> -#elif defined(CONFIG_LEGACY_VSYSCALL_XONLY)
> +#elif defined(CONFIG_LEGACY_VSYSCALL_XONLY) || defined(CONFIG_X86_CET)
>          XONLY;
> -#else
> +#else
>          EMULATE;
>   #endif

I don't get it.

First, you can't do any of this based on config -- it must be runtime.

Second, and more importantly, I don't see how XONLY helps at all.  The
(non-executable) text that's exposed to user code in EMULATE mode is
trivial to get right with CET -- your code already handles it.  It's
the emulation code (that runs identically in EMULATE and XONLY mode)
that's tricky.