Re: [Qemu-devel] [PATCH 28/35] kvm: x86: Introduce kvmclock device to save/restore its state

[Jan Kiszka]

On 2011-01-18 17:37, Anthony Liguori wrote:
> On 01/18/2011 10:17 AM, Jan Kiszka wrote:
>> On 2011-01-18 17:04, Anthony Liguori wrote:
>>    
>>>>>>> A KVM device should sit on a KVM specific bus that hangs off of
>>>>>>> sysbus.
>>>>>>> It can get to kvm_state through that bus.
>>>>>>>
>>>>>>> That bus doesn't get instantiated through qdev so requiring a pointer
>>>>>>> argument should not be an issue.
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>              
>>>>>> This design is in conflict with the requirement to attach KVM-assisted
>>>>>> devices also to their home bus, e.g. an assigned PCI device to the PCI
>>>>>> bus. We don't support multi-homed qdev devices.
>>>>>>
>>>>>>
>>>>>>            
>>>>> The bus topology reflects how I/O flows in and out of a device.  We do
>>>>> not model a perfect PC bus architecture and I don't think we ever intend
>>>>> to.  Instead, we model a functional architecture.
>>>>>
>>>>> I/O from an assigned device does not flow through the emulated PCI bus.
>>>>> Therefore, it does not belong on the emulated PCI bus.
>>>>>
>>>>> Assigned devices need to interact with the emulated PCI bus, but they
>>>>> shouldn't be children of it.
>>>>>
>>>>>          
>>>> You should be able to find assigned devices on some PCI bus, so you
>>>> either have to hack up the existing bus to host devices that are, on the
>>>> other side, not part of it or branch off a pci-kvm sub-bus, just like
>>>> you would have to create a sysbus-kvm.
>>>>        
>>> Management tools should never transverse the device tree to find
>>> devices.  This is a recipe for disaster in the long term because the
>>> device tree will not remain stable.
>>>
>>> So yes, a management tool should be able to enumerate assigned devices
>>> as they would enumerate any other PCI device but that has almost nothing
>>> to do with what the tree layout is.
>>>      
>> I'm probably misunderstanding you, but if the bus topology as the guest
>> sees it is not properly reflected in an object tree on the qemu side, we
>> are creating hacks again.
>>    
> 
> There is no such thing as the "bus topology as the guest sees it".
> 
> The guest just sees a bunch of devices.  The guest can only infer things 
> like ISA busses.  The guest sees a bunch of devices: an i8254, i8259, 
> RTC, etc.  Whether those devices are on an ISA bus, and LPC bus, or all 
> in a SuperI/O chip that's part of the southbridge is all invisible to 
> the guest.
> 
> The device model topology is 100% a hidden architectural detail.

This is true for the sysbus, it is obviously not the case for PCI and
similarly discoverable buses. There we have a guest-explorable topology
that is currently equivalent to the the qdev layout.

> 
>> Management and analysis tools must be able to traverse the system buses
>> and find guest devices this way.
> 
> We need to provide a compatible interface to the guest.  If you agree 
> with my above statements, then you'll also agree that we can do this 
> without keeping the device model topology stable.
> 
> But we also need to provide a compatible interface to management tools.  
> Exposing the device model topology as a compatible interface 
> artificially limits us.  It's far better to provide higher level 
> supported interfaces to give us the flexibility to change the device 
> model as we need to.

How do you want to change qdev to keep the guest and management tool
view stable while branching off kvm sub-buses? Please propose such
extensions so that they can be discussed. IIUC, that would be second
relation between qdev and qbus instances besides the physical topology.
What further use cases (besides passing kvm_state around) do you have in
mind?

> 
>>   If they create a device on bus X, it
>> must never end up on bus Y just because it happens to be KVM-assisted or
>> has some other property.
> 
> Nope.  This is exactly what should happen.
> 
> 90% of the devices in the device model are not created by management 
> tools.  They're part of a chipset.  The chipset has well defined 
> extension points and we provide management interfaces to create devices 
> on those extension points.  That is, interfaces to create PCI devices.
> 

Creating kvm irqchips via the management tool would be one simple way
(not the only one, though) to enable/disable kvm assistance for those
devices.

Jan

-- 
Siemens AG, Corporate Technology, CT T DE IT 1
Corporate Competence Center Embedded Linux