Re: [Xen-devel] [PATCH] x86/hvm: Intercept RDPMC when vPMU is disabled

[Andrew Cooper <andrew.cooper3@xxxxxxxxxx>]

On 25/02/2019 13:11, Jan Beulich wrote:
>>>> On 23.02.19 at 00:48, <boris.ostrovsky@xxxxxxxxxx> wrote:
>> On 2/22/19 5:44 PM, Andrew Cooper wrote:
>>> On 22/02/2019 21:58, Boris Ostrovsky wrote:
>>>> On 2/22/19 4:13 PM, Andrew Cooper wrote:
>>>>> vPMU isn't security supported, and in general guests can't access any of 
>>>>> the
>>>>> performance counter MSRs.  However, the RDPMC instruction isn't 
>>>>> intercepted,
>>>>> meaning that guest software can read the instantaneous counter values.
>>>>>
>>>>> When vPMU isn't configured, intercept RDPMC and unconditionally fail it 
>>>>> as 
>> if
>>>>> software has requested a bad counter index (#GP fault).  It is model 
>> specific
>>>>> as to which counters are available to begin with, and in levelled 
>>>>> scenarios,
>>>>> this information may not be accurate in the first place.
>>>>>
>>>>> This change isn't expected to have any impact on VMs.  Userspace is not
>>>>> usually given access to RDPMC (Windows appear to completely prohibit it; 
>> Linux
>>>>> is restricted to root), and kernels won't be executing RDPMC instructions 
>>>>> if
>>>>> their PMU drivers have failed to start.
>>>>>
>>>>> Signed-off-by: Andrew Cooper <andrew.cooper3@xxxxxxxxxx>
>>>>> ---
>>>>> CC: Jan Beulich <JBeulich@xxxxxxxx>
>>>>> CC: Wei Liu <wei.liu2@xxxxxxxxxx>
>>>>> CC: Roger Pau Monné <roger.pau@xxxxxxxxxx>
>>>>> CC: Jun Nakajima <jun.nakajima@xxxxxxxxx>
>>>>> CC: Kevin Tian <kevin.tian@xxxxxxxxx>
>>>>> CC: Boris Ostrovsky <boris.ostrovsky@xxxxxxxxxx>
>>>>> CC: Suravee Suthikulpanit <suravee.suthikulpanit@xxxxxxx>
>>>>> CC: Brian Woods <brian.woods@xxxxxxx>
>>>>> CC: Juergen Gross <jgross@xxxxxxxx>
>>>>>
>>>>> This should be taken into Xen 4.12 and backported to the stable releases.
>>>>> While it isn't an XSA itself, it is an information leak (Xen's NMI 
>>>>> watchdog in
>>>>> particular) which could be advantagous to an attacker trying to exploit a 
>>>>> race
>>>>> condition.
>>>>>
>>>>> The only other option is to emulate the reported family and offer back 
>>>>> all 0's
>>>>> for the accessable counters.  Obviously this is a non-starter.
> I don't really understand why you say this - Boris certainly has a point ...
>
>>>> When VPMU is off MSR reads return zero.
>>> That behaviour isn't long for this world.
>>>
>>>> While it is debatable whether this the right action, shouldn't rdpmc 
>>>> behave in the same fashion?
>>> I specifically don't want to propagate the "lets complete with zero"
>>> behaviour further, because it takes away #GP faults which the guest
>>> would otherwise get.
>> The guest should get a #GP on Intel if CPUID is not reporting any
>> counters but not on AMD where the first 4 counters are architectural.
> ... here.

No - just because something is architectural doesn't mean the guest gets
to play with it.

Especially not for vPMU where the code is of such bad quality we had to
disable in a security fix, and re-disable it again later in another
security fix.

> For Intel, afaics, we indeed produce a blank CPUID leaf in
> all cases, so the behavior looks reasonably consistent. I would
> question though whether a blank CPUID leaf / the absence of any
> counters wouldn't call for #UD instead of #GP(0).

RDPMC hasn't #UD'd in a quarter of a century, but does #GP in userspace
outside of developer profiling scenarios.

> Otherwise,
> along the lines of AMD, aren't the first two indexes uniformly valid
> for Intel?

No - its model specific behaviour.  The only difference for more modern
systems is that they have agreed on a common behaviour.

And that is specifically why implementing 0's is a non-starter - it is
not a remotely sensible use of time to build enough infrastructure to
provide correct model-specific behaviour just for a corner case which
operating systems don't encounter in practice.

> Additionally aren't you invoking vpmu_available() before the data it
> examines actually got set? Afaics vpmu_initialise() gets called after
> hvm_vcpu_initialise(), yet the latter is where you add the intercept
> enabling.

Hmm yes - that does appear to break the vpmu case.  The order of
initialisation will need tweaking as well.

~Andrew

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxxx
https://lists.xenproject.org/mailman/listinfo/xen-devel