Re: Is: PVH dom0 - MWAIT detection logic to get deeper C-states exposed in ACPI AML code. Was:Re: [PATCH v2 10/30] xen/x86: Annotate VM applicability in featureset

[Boris Ostrovsky <boris.ostrovsky@oracle.com>]

On 02/18/2016 10:12 AM, Andrew Cooper wrote:
> On 18/02/16 15:02, Roger Pau Monné wrote:
>> El 17/2/16 a les 20:02, Konrad Rzeszutek Wilk ha escrit:
>>> On Mon, Feb 15, 2016 at 03:41:41PM +0000, Andrew Cooper wrote:
>>>> On 15/02/16 15:02, Jan Beulich wrote:
>>>>>>>> On 15.02.16 at 15:53, <andrew.cooper3@citrix.com> wrote:
>>>>>> On 15/02/16 14:50, Jan Beulich wrote:
>>>>>>>>>> On 15.02.16 at 15:38, <andrew.cooper3@citrix.com> wrote:
>>>>>>>> On 15/02/16 09:20, Jan Beulich wrote:
>>>>>>>>>>>> On 12.02.16 at 18:42, <andrew.cooper3@citrix.com> wrote:
>>>>>>>>>> On 12/02/16 17:05, Jan Beulich wrote:
>>>>>>>>>>>>>> On 05.02.16 at 14:42, <andrew.cooper3@citrix.com> wrote:
>>>>>>>>>>>>   #define X86_FEATURE_MWAITX        ( 3*32+29) /*   MWAIT extension
>>>>>>>>>> (MONITORX/MWAITX) */
>>>>>>>>>>> Why not exposed to HVM (also for _MWAIT as I now notice)?
>>>>>>>>>> Because that is a good chunk of extra work to support.  We would need to
>>>>>>>>>> use 4K monitor widths, and extra p2m handling.
>>>>>>>>> I don't understand: The base (_MWAIT) feature being exposed to
>>>>>>>>> guests today, and kernels making use of the feature when available
>>>>>>>>> suggests to me that things work. Are you saying you know
>>>>>>>>> otherwise? (And if there really is a reason to mask the feature all of
>>>>>>>>> the sudden, this should again be justified in the commit message.)
>>>>>>>> PV guests had it clobbered by Xen in traps.c
>>>>>>>>
>>>>>>>> HVM guests have:
>>>>>>>>
>>>>>>>> vmx.c:
>>>>>>>>      case EXIT_REASON_MWAIT_INSTRUCTION:
>>>>>>>>      case EXIT_REASON_MONITOR_INSTRUCTION:
>>>>>>>> [...]
>>>>>>>>      hvm_inject_hw_exception(TRAP_invalid_op, HVM_DELIVER_NO_ERROR_CODE);
>>>>>>>>          break;
>>>>>>>>
>>>>>>>> and svm.c:
>>>>>>>>      case VMEXIT_MONITOR:
>>>>>>>>      case VMEXIT_MWAIT:
>>>>>>>>          hvm_inject_hw_exception(TRAP_invalid_op, HVM_DELIVER_NO_ERROR_CODE);
>>>>>>>>          break;
>>>>>>>>
>>>>>>>> I don't see how a guest could actually use this feature.
>>>>>>> Do you see the respective intercepts getting enabled anywhere?
>>>>>>> (I don't outside of nested code, which I didn't check in detail.)
>>>>>> Yes - the intercepts are always enabled to prevent the guest actually
>>>>>> putting the processor to sleep.
>>>>> Hmm, you're right, somehow I've managed to ignore the relevant
>>>>> lines grep reported. Yet - how do things work then, without the
>>>>> MWAIT feature flag currently getting cleared?
>>>> I have never observed it being used.  Do you have some local patches in
>>>> the SLES hypervisor?
>>>>
>>>> There is some gross layer violation in xen/enlighten.c to pretend that
>>>> MWAIT is present to trick the ACPI code into evaluating _CST() methods
>>>> to report back to Xen.  (This is yet another PV-ism which will cause a
>>>> headache for a DMLite dom0)
>>> Yes indeed. CC-ing Roger, and Boris.
>> Yes, all this is indeed not very nice, and we would ideally like to get
>> rid of it on PVHv2.

We will have to come up with something else: AIUI the whole point of 
xen_check_mwait() is to come up with proper ECX and EDX values for the 
MWAIT CPUID leaf. Those value are expected to be reported from 
xen_cpuid() pv_op so that acpi_processor_ffh_state_probe_cpu() can set C 
states structures properly.

The problem is that PVH executes CPUID instruction natively. (And so 
this must have been broken on PVHv1 as well).

-boris


>>
>> Could we use the acpica tools (acpidump/acpixtract/acpiexec/...) in
>> order to fetch this information from user-space and send it to Xen using
>> privcmd?
>>
>> AFAIK those tools work on most OSes (or at least the ones we care about
>> as Dom0).
> In general, we can't rely on userspace evaluation of AML.
>
> For trivial AML which evaluates to a constant, it could be interpreted
> by userspace, but anything accessing system resources will need
> evaluating by the kernel.
>
> ~Andrew