Re: [Qemu-devel] [PATCH] RFC: powerpc: add PVR compatibility check

[Alexander Graf]

On 06.11.2013, at 04:02, Paul Mackerras <address@hidden> wrote:

> On Tue, Nov 05, 2013 at 05:16:33PM +0100, Andreas Färber wrote:
>> Am 05.11.2013 07:05, schrieb Alexander Graf:
>>> 
>>> 
>>> Am 05.11.2013 um 05:00 schrieb Paul Mackerras <address@hidden>:
>>> 
>>>> On Mon, Nov 04, 2013 at 10:05:58AM +0100, Alexander Graf wrote:
>>>>> 
>>>>> Yeah, we really need to check that guest vpcu == host vcpu for HV KVM.
>>>> 
>>>> In general I agree, but the one difficulty I see is that a check for
>>>> exact equality will interact badly with qemu's habit of picking a
>>>> specific processor version when the user specifies something general
>>>> like "POWER7".  So if the user does -cpu POWER7 on a machine with
>>>> (for example) a POWER7 v2.1 processor, but qemu arbitrarily picks the
>>>> PVR for POWER7 v2.3, then it will fail, which will be completely
>>>> puzzling to the user -- "I asked for POWER7, and it is a POWER7,
>>>> what's the problem??".
>>>> 
>>>> Maybe if the user asks for a non-specific processor type, and the
>>>> host's PVR matches the request, then qemu should take the host's PVR
>>>> rather than just picking some arbitrary processor version.
>>> 
>>> Yup.
>> 
>> But then it's no longer generally reproducible: "POWER7" won't be
>> "POWER7" on another machine.
> 
> There aren't any observable differences between POWER7 versions that
> have been sold to customers, as far as I have been able to ascertain
> (other than the PVR value, of course).  So this whole business of
> carefully distinguishing between POWER7 v2.2 and POWER7 v2.3 is
> largely a waste of time as far as I can see.
> 
> I admit that in the past we (IBM) did a silly thing in releasing the
> POWER5+ v3.0 chip with some architecturally new features (64k pages
> and some other MMU changes).  That was a mistake and I don't think
> we'll do it again.
> 
> I think the default assumption should be that versions of a given IBM
> POWER chip (identified by the upper 16 bits of the PVR) are
> architecturally identical, and behaviourally identical at the level
> at which QEMU models the chip.  Differences between chips would
> normally be limited to bug fixes and performance improvements.  Then
> we just need a way to cope with POWER5+ v3.0.
> 
>> One thing I original did iirc was to hide the aliases from QMP. You can
>> always do stupid things on the command line and then we can blame you,
>> but if libvirt and upper layers don't offer "POWER7" to the end user
>> then we don't need to worry about the average user misinterpreting its
>> semantics.
> 
> Given that the only difference between POWER7 v2.2 and POWER7 v2.3
> (say) will be which set of host systems you get an error on, there
> doesn't seem to me to be a lot of point.

Given that for HV KVM the only possible cpu type that is ever guaranteed to 
work is -cpu host, why bother with a database that maintains compatibility 
flags between different versions? Just tell the user to use -cpu host and call 
it a day, no?


Alex