Re: [PATCH 0/4] x86: enable User-Mode Instruction Prevention

From: Ricardo Neri
Date: Thu Nov 10 2016 - 01:46:09 EST


On Wed, 2016-11-09 at 03:05 -0800, Andy Lutomirski wrote:
> On Tue, Nov 8, 2016 at 8:31 PM, Ricardo Neri
> <ricardo.neri-calderon@xxxxxxxxxxxxxxx> wrote:
> > On Tue, 2016-11-08 at 07:34 -0800, Andy Lutomirski wrote:
> >> > Would it not be better to emulate these instructions for them? What
> >> way
> >> > we can verify they're not malicious.
> >>
> >> Forget malice -- if they are really needed for some silly vm86-using
> >> program, let's trap them and emulate them so they return dummy values.
> >>
> >> Also, keep in mind that vm86 is already effectively gated behind a
> >> sysctl for non-root. I think the default should be that, if root has
> >> enabled vm86, it should work.
> >
> > Then should I keep UMIP enabled by default and still provide an option
> > to disable it via a kernel parameter?
>
> Probably, but clearcpuid might be good enough. There might be some
> unexpected breakage.
>
> >
> > Also, a third option, umip=novm86 would "disable" UMIP in vm86 tasks.
> > Under the new approach (of emulating the impacted instructions), this
> > option, a #GP fault would still be generated but the actual values of
> > GDT/LDT/IDT/MSW would be passed to user space. Does this make sense?
>
> I don't think so. As far as I know, there is no legitimate reason for
> a vm86-using program to care about what these instructions spit out.
> Heck, in real mode and vm86 mode, there aren't segment descriptors at
> all, so the GDT is really quite useless even if it were readable.

I took a closer look at the dosemu code. It appears that it does not
purposely utilize SGDT to obtain the descriptor table while in vm86. It
does use SGDT (in protected mode) to emulate certain functionality such
as the Virtual xxx Driver. In such a case, UMIP needs to be disabled.
However, this code seems to be disabled [1]. dosemu includes an i386
emulator that in some cases uses the actual instructions of the host
system. In such cases, UMIP might be needed to be disabled.

So, yes, I agree now that UMIP does not need to be disabled specifically
for vm86 tasks but via clearcpuid.

Thanks and BR,
Ricardo
[1].
https://sourceforge.net/p/dosemu/code/ci/master/tree/src/dosext/dpmi/vxd.c#l731
>
> I would suggest having all of these instructions return compile-time
> constants in vm86 mode.