Linux Virtualization - Apr 2020 - Should SEV-ES #VC use IST? (Re: [PATCH] Allow RDTSC and RDTSCP from userspace)

On Sat, Apr 25, 2020 at 1:23 PM Joerg Roedel <joro at 8bytes.org>
wrote:
>
> On Sat, Apr 25, 2020 at 12:47:31PM -0700, Andy Lutomirski wrote:
> > I assume the race you mean is:
> >
> > #VC
> > Immediate NMI before IST gets shifted
> > #VC
> >
> > Kaboom.
> >
> > How are you dealing with this?  Ultimately, I think that NMI will need
> > to turn off IST before engaging in any funny business. Let me ponder
> > this a bit.
>
> Right, I dealt with that by unconditionally shifting/unshifting the #VC IST
entry
> in do_nmi() (thanks to Davin Kaplan for the idea). It might cause
> one of the IST stacks to be unused during nesting, but that is fine. The
> stack memory for #VC is only allocated when SEV-ES is active (in an
> SEV-ES VM).
Blech.  It probably works, but still, yuck.  It's a bit sad that we
seem to be growing more and more poorly designed happens-anywhere
exception types at an alarming rate.  We seem to have #NM, #MC, #VC,
#HV, and #DB.  This doesn't really scale.

--Andy

On Sat, Apr 25, 2020 at 3:10 PM Andy Lutomirski <luto at kernel.org>
wrote:
>
> On Sat, Apr 25, 2020 at 1:23 PM Joerg Roedel <joro at 8bytes.org>
wrote:
> >
> > On Sat, Apr 25, 2020 at 12:47:31PM -0700, Andy Lutomirski wrote:
> > > I assume the race you mean is:
> > >
> > > #VC
> > > Immediate NMI before IST gets shifted
> > > #VC
> > >
> > > Kaboom.
> > >
> > > How are you dealing with this?  Ultimately, I think that NMI will
need
> > > to turn off IST before engaging in any funny business. Let me
ponder
> > > this a bit.
> >
> > Right, I dealt with that by unconditionally shifting/unshifting the
#VC IST entry
> > in do_nmi() (thanks to Davin Kaplan for the idea). It might cause
> > one of the IST stacks to be unused during nesting, but that is fine.
The
> > stack memory for #VC is only allocated when SEV-ES is active (in an
> > SEV-ES VM).
>
> Blech.  It probably works, but still, yuck.  It's a bit sad that we
> seem to be growing more and more poorly designed happens-anywhere
> exception types at an alarming rate.  We seem to have #NM, #MC, #VC,
> #HV, and #DB.  This doesn't really scale.
I have a somewhat serious question: should we use IST for #VC at all?
As I understand it, Rome and Naples make it mandatory for hypervisors
to intercept #DB, which means that, due to the MOV SS mess, it's sort
of mandatory to use IST for #VC.  But Milan fixes the #DB issue, so,
if we're running under a sufficiently sensible hypervisor, we don't
need IST for #VC.

So I think we have two choices:

1. Use IST for #VC and deal with all the mess that entails.

2. Say that we SEV-ES client support on Rome and Naples is for
development only and do a quick boot-time check for whether #DB is
intercepted.  (Just set TF and see what vector we get.)  If #DB is
intercepted, print a very loud warning and refuse to boot unless some
special sev_es.insecure_development_mode or similar option is set.

#2 results in simpler and more robust entry code.  #1 is more secure.

So my question is: will anyone actually use SEV-ES in production on
Rome or Naples?  As I understand it, it's not really ready for prime
time on those chips.  And do we care if the combination of a malicious
hypervisor and malicious guest userspace on Milan can compromise the
guest kernel?  I don't think SEV-ES is really mean to resist a
concerted effort by the hypervisor to compromise the guest.

--Andy

On 27/04/2020 18:37, Andy Lutomirski wrote:
> On Sat, Apr 25, 2020 at 3:10 PM Andy Lutomirski <luto at kernel.org>
wrote:
>> On Sat, Apr 25, 2020 at 1:23 PM Joerg Roedel <joro at 8bytes.org>
wrote:
>>> On Sat, Apr 25, 2020 at 12:47:31PM -0700, Andy Lutomirski wrote:
>>>> I assume the race you mean is:
>>>>
>>>> #VC
>>>> Immediate NMI before IST gets shifted
>>>> #VC
>>>>
>>>> Kaboom.
>>>>
>>>> How are you dealing with this?  Ultimately, I think that NMI
will need
>>>> to turn off IST before engaging in any funny business. Let me
ponder
>>>> this a bit.
>>> Right, I dealt with that by unconditionally shifting/unshifting the
#VC IST entry
>>> in do_nmi() (thanks to Davin Kaplan for the idea). It might cause
>>> one of the IST stacks to be unused during nesting, but that is
fine. The
>>> stack memory for #VC is only allocated when SEV-ES is active (in an
>>> SEV-ES VM).
>> Blech.  It probably works, but still, yuck.  It's a bit sad that we
>> seem to be growing more and more poorly designed happens-anywhere
>> exception types at an alarming rate.  We seem to have #NM, #MC, #VC,
>> #HV, and #DB.  This doesn't really scale.
> I have a somewhat serious question: should we use IST for #VC at all?
> As I understand it, Rome and Naples make it mandatory for hypervisors
> to intercept #DB, which means that, due to the MOV SS mess, it's sort
> of mandatory to use IST for #VC.  But Milan fixes the #DB issue, so,
> if we're running under a sufficiently sensible hypervisor, we don't
> need IST for #VC.
>
> So I think we have two choices:
>
> 1. Use IST for #VC and deal with all the mess that entails.
>
> 2. Say that we SEV-ES client support on Rome and Naples is for
> development only and do a quick boot-time check for whether #DB is
> intercepted.  (Just set TF and see what vector we get.)  If #DB is
> intercepted, print a very loud warning and refuse to boot unless some
> special sev_es.insecure_development_mode or similar option is set.
>
> #2 results in simpler and more robust entry code.  #1 is more secure.
>
> So my question is: will anyone actually use SEV-ES in production on
> Rome or Naples?  As I understand it, it's not really ready for prime
> time on those chips.  And do we care if the combination of a malicious
> hypervisor and malicious guest userspace on Milan can compromise the
> guest kernel?  I don't think SEV-ES is really mean to resist a
> concerted effort by the hypervisor to compromise the guest.
More specifically, it is mandatory for hypervisors to intercept #DB to
defend against CVE-2015-8104, unless they're willing to trust the guest
not to tickle that corner case.

This is believed fixed with SEV-SNP to allow the encrypted guest to use
debugging functionality without posing a DoS risk to the host.? In this
case, the hypervisor is expected not to intercept #DB.

If #DB is intercepted, and #VC doesn't use IST, malicious userspace can
cause problems with a movss-delayed breakpoint over SYSCALL.

The question basically whether it is worth going to the effort of making
#VC IST and all the problems that entails, to cover one corner case in
earlier hardware.

Ultimately, this depends on whether anyone plans to put SEV-ES into
production on pre SEV-SNP hardware, and if developers using pre-SEV-SNP
hardware are happy with "don't run malicious userspace" or
"don't run
malicious kernels and skip the #DB intercept" as a fair tradeoff to
avoid the #VC IST fun.

~Andrew

On 4/27/20 12:37 PM, Andy Lutomirski wrote:
> On Sat, Apr 25, 2020 at 3:10 PM Andy Lutomirski <luto at kernel.org>
wrote:
>>
>> On Sat, Apr 25, 2020 at 1:23 PM Joerg Roedel <joro at 8bytes.org>
wrote:
>>>
>>> On Sat, Apr 25, 2020 at 12:47:31PM -0700, Andy Lutomirski wrote:
>>>> I assume the race you mean is:
>>>>
>>>> #VC
>>>> Immediate NMI before IST gets shifted
>>>> #VC
>>>>
>>>> Kaboom.
>>>>
>>>> How are you dealing with this?  Ultimately, I think that NMI
will need
>>>> to turn off IST before engaging in any funny business. Let me
ponder
>>>> this a bit.
>>>
>>> Right, I dealt with that by unconditionally shifting/unshifting the
#VC IST entry
>>> in do_nmi() (thanks to Davin Kaplan for the idea). It might cause
>>> one of the IST stacks to be unused during nesting, but that is
fine. The
>>> stack memory for #VC is only allocated when SEV-ES is active (in an
>>> SEV-ES VM).
>>
>> Blech.  It probably works, but still, yuck.  It's a bit sad that we
>> seem to be growing more and more poorly designed happens-anywhere
>> exception types at an alarming rate.  We seem to have #NM, #MC, #VC,
>> #HV, and #DB.  This doesn't really scale.
> 
> I have a somewhat serious question: should we use IST for #VC at all?
> As I understand it, Rome and Naples make it mandatory for hypervisors
> to intercept #DB, which means that, due to the MOV SS mess, it's sort
> of mandatory to use IST for #VC.  But Milan fixes the #DB issue, so,
> if we're running under a sufficiently sensible hypervisor, we don't
> need IST for #VC.
> 
> So I think we have two choices:
> 
> 1. Use IST for #VC and deal with all the mess that entails.
> 
> 2. Say that we SEV-ES client support on Rome and Naples is for
> development only and do a quick boot-time check for whether #DB is
> intercepted.  (Just set TF and see what vector we get.)  If #DB is
> intercepted, print a very loud warning and refuse to boot unless some
> special sev_es.insecure_development_mode or similar option is set.
> 
> #2 results in simpler and more robust entry code.  #1 is more secure.
> 
> So my question is: will anyone actually use SEV-ES in production on
> Rome or Naples?  As I understand it, it's not really ready for prime
> time on those chips.  And do we care if the combination of a malicious
Naples was limited in the number of encryption keys available for guests 
(15), but Rome increased that significantly (509). SEV-ES is ready on 
those chips - Rome more so with the increased key count given the 
requirement that SEV and SEV-ES guests have non-overlapping ASID ranges 
(which corresponds to key usage).

Thanks,
Tom
> hypervisor and malicious guest userspace on Milan can compromise the
> guest kernel?  I don't think SEV-ES is really mean to resist a
> concerted effort by the hypervisor to compromise the guest.
> 
> --Andy
>

On Mon, Apr 27, 2020 at 10:37:41AM -0700, Andy Lutomirski
wrote:
> I have a somewhat serious question: should we use IST for #VC at all?
> As I understand it, Rome and Naples make it mandatory for hypervisors
> to intercept #DB, which means that, due to the MOV SS mess, it's sort
> of mandatory to use IST for #VC.  But Milan fixes the #DB issue, so,
> if we're running under a sufficiently sensible hypervisor, we don't
> need IST for #VC.
The reason for #VC being IST is not only #DB, but also SEV-SNP. SNP adds
page ownership tracking between guest and host, so that the hypervisor
can't remap guest pages without the guest noticing.

If there is a violation of ownership, which can happen at any memory
access, there will be a #VC exception to notify the guest. And as this
can happen anywhere, for example on a carefully crafted stack page set
by userspace before doing SYSCALL, the only robust choice for #VC is to
use IST.

Regards,

	Joerg

Hi Andy,

On Mon, Apr 27, 2020 at 10:37:41AM -0700, Andy Lutomirski
wrote:
> 1. Use IST for #VC and deal with all the mess that entails.
With the removal of IST shifting I wonder what you would suggest on how
to best implement an NMI-safe IST handler with nesting support.

My current plan is to implement an IST handler which switches itself off
the IST stack as soon as possible, freeing it for re-use.

The flow would be roughly like this upon entering the handler;

	build_pt_regs();

	RSP = pt_regs->sp;

	if (RSP in VC_IST_stack)
		error("unallowed nesting")

	if (RSP in current_kernel_stack)
		RSP = round_down_to_8(RSP)
	else
		RSP = current_top_of_stack() // non-ist kernel stack

	copy_pt_regs(pt_regs, RSP);
	switch_stack_to(RSP);

To make this NMI safe, the NMI handler needs some logic too. Upon
entering NMI, it needs to check the return RSP, and if it is in the #VC
IST stack, it must do the above flow by itself and update the return RSP
and RIP. It needs to take into account the case when PT_REGS is not
fully populated on the return side.

Alternativly the NMI handler could safe/restore the contents of the #VC
IST stack or just switch to a special #VC-in-NMI IST stack.

All in all it could get complicated, and imho shift_ist would have been
simpler, but who am I anyway...

Or maybe you have a better idea how to implement this, so I'd like to
hear your opinion first before I spend too many days implementing
something.

Regards,

	Joerg