Re: [PATCH v11 6/9] xen: Add ring 3 vmware_port support

[George Dunlap <george.dunlap@eu.citrix.com>]

On 06/03/2015 04:58 PM, Andrew Cooper wrote:
> On 03/06/15 16:26, George Dunlap wrote:
>> On 05/22/2015 04:50 PM, Don Slutz wrote:
>>> Summary is that VMware treats "in (%dx),%eax" (or "out %eax,(%dx)")
>>> to port 0x5658 specially.  Note: since many operations return data
>>> in EAX, "in (%dx),%eax" is the one to use.  The other lengths like
>>> "in (%dx),%al" will still do things, only AL part of EAX will be
>>> changed.  For "out %eax,(%dx)" of all lengths, EAX will remain
>>> unchanged.
>>>
>>> This instruction is allowed to be used from ring 3.  To
>>> support this the vmexit for GP needs to be enabled.  I have not
>>> fully tested that nested HVM is doing the right thing for this.
>>>
>>> Enable no-fault of pio in x86_emulate for VMware port
>>>
>>> Also adjust the emulation registers after doing a VMware
>>> backdoor operation.
>>>
>>> Add new routine hvm_emulate_one_gp() to be used by the #GP fault
>>> handler.
>>>
>>> Some of the best info is at:
>>>
>>> https://sites.google.com/site/chitchatvmback/backdoor
>>>
>>> Signed-off-by: Don Slutz <dslutz@verizon.com>
>> So let me get this straight.
>>
>> VMWare allows ring3 to access the magic port regardless of whether the
>> guest OS has enabled access to that IO port or not.
>>
>> In order to emulate this, we need to:
>> * Trap to Xen on #GPs rather than just letting the hardware handle it
>> * Emulate all instructions which cause a #GP, just to see if they might
>> be an IO instruction accessing the magic port.
>> * If it is an IO instruction, and it's accessing the magic port, then we
>> skip the ioport access checks (which will cause the instruction to
>> execute as though it had been given access).
>> * Under all other circumstances (we hope) the emulator in Xen will do
>> exactly what the hardware just did, and deliver a #GP to the guest.
>>
>> In an attempt to make this more safe, emulation ops that write (such as
>> write and cmpxchg) are replaced with stubs which always return an error.
>>
>> Is that about right?
>>
>> That sounds completely insane.  It opens up an almost infinite surface
>> of attack onto the Xen emulator.
>>
>> I understand that having the "VMWare compatible" is a nice tick-box to
>> have, but seriously, I cannot imagine that having unprivileged
>> user-space tools know the real clock frequency without having to involve
>> the OS is anywhere close to worth the risk involved.
> 
> The attack surface sadly is not enlarged in the slightest by this change.
> 
> We already trap and emulate all #UD exceptions in an attempt to support
> migration of VMs between Intel and AMD hardware.  See XSA-105.  (There
> is a good argument to be made for not trapping #UD, but that doesn't
> completely close the hole)

So at the moment, an attacker on Intel can force the emulation of any
AMD-only instruction (and vice versa), is that right?

This would allow an attacker to force the emulation of every #GP
condition of every instruction we emulate.

Those two sets may be within an order of magnitude of each other, but
they will only overlap a little bit.  So my guess is that enabling this
would double the surface of attack (give or take).

I'd be a lot happier with this patch if we could make it so that on a
#GP the only instruction that could get emulated would be an IO instruction.

 -George