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

[Blue Swirl]

On Thu, Jan 20, 2011 at 7:37 PM, Anthony Liguori
<address@hidden> wrote:
> On 01/20/2011 03:33 AM, Jan Kiszka wrote:
>>
>> On 2011-01-19 20:32, Blue Swirl wrote:
>>
>>>
>>> On Wed, Jan 19, 2011 at 4:57 PM, Anthony Liguori
>>> <address@hidden>  wrote:
>>>
>>>>
>>>> On 01/19/2011 07:15 AM, Markus Armbruster wrote:
>>>>
>>>>>
>>>>> So they interact with KVM (need kvm_state), and they interact with the
>>>>> emulated PCI bus.  Could you elaborate on the fundamental difference
>>>>> between the two interactions that makes you choose the (hypothetical)
>>>>> KVM bus over the PCI bus as device parent?
>>>>>
>>>>>
>>>>
>>>> It's almost arbitrary, but I would say it's the direction that I/Os
>>>> flow.
>>>>
>>>> But if the underlying observation is that the device tree is not really
>>>> a
>>>> tree, you're 100% correct.  This is part of why a factory interface that
>>>> just takes a parent bus is too simplistic.
>>>>
>>>> I think we ought to introduce a -pci-device option that is specifically
>>>> for
>>>> creating PCI devices that doesn't require a parent bus argument but
>>>> provides
>>>> a way to specify stable addressing (for instancing, using a linear
>>>> index).
>>>>
>>>
>>> I think kvm_state should not be a property of any device or bus. It
>>> should be split to more logical pieces.
>>>
>>> Some parts of it could remain in CPUState, because they are associated
>>> with a VCPU.
>>>
>>> Also, for example irqfd could be considered to be similar object to
>>> char or block devices provided by QEMU to devices. Would it make sense
>>> to introduce new host types for passing parts of kvm_state to devices?
>>>
>>> I'd also make coalesced MMIO stuff part of memory object. We are not
>>> passing any state references when using cpu_physical_memory_rw(), but
>>> that could be changed.
>>>
>>
>> There are currently no VCPU-specific bits remaining in kvm_state. It may
>> be a good idea to introduce an arch-specific kvm_state and move related
>> bits over. It may also once be feasible to carve out memory management
>> related fields if we have proper abstractions for that, but I'm not
>> completely sure here.
>>
>> Anyway, all these things are secondary. The primary topic here is how to
>> deal with kvm_state and its fields that have VM-global scope.
>>
>
> The debate is really:
>
> 1) should we remove all passing of kvm_state and just assume it's static
>
> 2) deal with a couple places in the code where we need to figure out how to
> get at kvm_state
>
> I think we've only identified 1 real instance of (2) and it's resulted in
> some good discussions about how to model KVM devices vs. emulated devices.
>  Honestly, (1) just stinks.  I see absolutely no advantage to it at all.

Fully agree.

> In the very worst case scenario, the thing we need to do is just reference
> an extern variable in a few places.  That completely avoids all of the
> modelling discussions for now (while leaving for placeholder FIXMEs so the
> problem can be tackled later).

I think KVMState was designed to match KVM ioctl interface: all stuff
that is needed for talking to KVM or received from KVM are there. But
I think this shouldn't be a design driver.

If the only pieces of kvm_state that are needed by the devices are
irqchip_in_kernel, pit_in_kernel and many_ioeventfds, the problem of
passing kvm_state to devices becomes very different. Each of these are
just single bits, affecting only a few devices. Perhaps they could be
device properties which the board level sets when KVM is used?