Re: [PATCH v4 2/4] x86/compat: Test both PV and PVH guests for compat mode

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

On 09/02/2015 09:13 AM, Jan Beulich wrote:
>>>> On 02.09.15 at 14:31, <boris.ostrovsky@oracle.com> wrote:
>> On 09/02/2015 04:08 AM, Jan Beulich wrote:
>>>>>> Boris Ostrovsky <boris.ostrovsky@oracle.com> 09/02/15 2:55 AM >>>
>>>> On 08/27/2015 12:01 PM, Jan Beulich wrote:
>>>>>>>> On 13.08.15 at 20:12, <boris.ostrovsky@oracle.com> wrote:
>>>>>> @@ -777,7 +777,7 @@ int arch_set_info_guest(
>>>>>>     
>>>>>>         /* The context is a compat-mode one if the target domain is
>> compat-mode;
>>>>>>          * we expect the tools to DTRT even in compat-mode callers. */
>>>>>> -    compat = is_pv_32bit_domain(d);
>>>>>> +    compat = is_pv_32bit_domain(d) || is_pvh_32bit_domain(d);
>>>>> I continue to think that this should include a v->domain ==
>>>>> current->domain check (to match behavior for HVM guests
>>>>> from the tool stack perspective). Having looked at patch 4, I also
>>>>> can't see how the tool stack is being made expect a non-native
>>>>> guest context record in the 32-bit PVH case (i.e. I'd appreciate
>>>>> if you could point out where that hides).
>>>> For vcpu 0 current->domain is dom0 so I am not sure how this check would
>>>> work.
>>>>
>>>> For a 32-bit PVH guest the toolstack will place data into
>>>> vcpu_guest_context_x86_32_t (in vcpu_x86_32()) and so the hypervisor,
>>>> knowing that the guest is a compat one (based on the test above), will
>>>> access appropriate fields.
>>>>
>>>> This is not how HVM guests are started --- "classic" PVH behaves very
>>>> much like a PV guest, unlike what we are doing with no-dm PVH.
>>> And I believe this to be wrong, and potentially getting in the way of the
>> no-dm
>>> work - Roger?
>>>
>>> As to the reference to vcpu_x86_32() - by its name alone it is already clear
>>> that this is after the determination of what bit width a guest to deal with, and
>>> looking at xc_dom_32_pae I still can't see why PVH guests would (intentionally
>>> and legitimately) be treated like PV rather than HVM ones (not to speak of the
>>> fact that I don't think 32-bit PVH is in any way limited to PAE).
>> The purpose of this series is to get 32-bit guests to parity with
>> classic PVH with minimal changes and then move on to no-dm. Not getting
>> in the way of no-dm is obviously important but making classic behave
>> like no-dm (which is, to certain extent, is what you are suggesting) is
>> out of scope.
> Well, okay. But are you saying then that 64-bit PVH also just
> _happens_ to be treated like PV in the tool stack?

Yes. They both are processed by tollstack's elf parser (and the size is 
determined by xc_dom_guest_type()).

> IOW I'm still missing
> the explicit tool stack adjustment that makes it use a guest-bit-width
> context for PVH guests.
>
>> (I haven't considered non-PAE case, TBH)
> But I'm afraid you need to.
>

Actually, at least for Linux, CONFIG_XEN 'depends on X86_64 || (X86_32 
&& X86_PAE)'. And I don't know whether there are any other OSs (i.e. 
FreeBSD) that are going to ever support 32-bit PVH-classic (Roger might 
correct me on that).

-boris