Re: [Qemu-devel] [PATCH 1/7] Introduce a new bus "ICC" to connect APIC

[Igor Mammedov <imammedo@redhat.com>]

On 02/17/2012 06:02 PM, Igor Mammedov wrote:
> On 02/16/2012 01:42 PM, Anthony Liguori wrote:
>> On 02/16/2012 05:25 AM, Jan Kiszka wrote:
>>> On 2012-02-16 00:16, Igor Mammedov wrote:
>>>> Introduce a new structure CPUS as the controller of ICC (INTERRUPT
>>>> CONTROLLER COMMUNICATIONS), and new bus "ICC" to hold APIC,instead
>>>> of sysbus. So we can support APIC hot-plug feature.
>>>>
>>>> This is repost of original patch for qemu-kvm rebased on current qemu:
>>>> http://lists.nongnu.org/archive/html/qemu-devel/2011-11/msg01478.html
>>>> All credits to Liu Ping Fan for writing it.
>>>>
>>>> V2 changes:
>>>> - cpusockets_init: cpu_sockets is not yet initialized, use cpus that
>>>> we got as input param instead for qbus_create, this makes cpus
>>>> apics visible in "info qtree" monitor command
>>>> - fix format error spotted by Jan and missed by checkpatch
>>>> - cpu_has_apic_feature: return bool instead of int
>>>>
>>>
>>> This patch surely no longer applies. And the ICC requires QOM conversion.
>>
>> Also, post-QOM, I don't think having an ICC bus makes a whole lot of sense.
>>
>> The LAPIC can be made a child of the CPU device with a bidirectional link.
>>
>> I would simply create a fixed set of CPU links<> hung off of /devices somewhere and use that as the hotplug mechanism. This matches well the way we
>> model this to the guest (we expose a fixed number of pluggable sockets).
>
> I've just QOM-ified it, but in light of what you just said it may be ignored.
> ICC bus was used on pre Pentium 4 smp systems. And whole thing with introducing
> it was to provide hot-plugable bus for cpus, since hot-plug on sysbus is disabled
> and people argued that sysbus shouldn't be hot-plugable. However it depends on
> what we choose to model, we can use pre P4 ICC bus for inter-apic/ioapic communications
> or use P4 model allowing hot-plug on sysbus and use it for inter-apic/ioapic
> communications if needed.
>
> So I'd rather drop ICC patch and try your approach with CPU links<>, I see no
> point in introducing new bus providing we have an alternative model and existing
> bus for the task.

I've looked at device_add command and qdev_device_add it uses for doing actual work
and in current state it requires (based on Andreas' qom_cpu branch):

For approach where apic and cpu hot-plugged to sysbus.
   1. created object must be descendant of TYPE_DEVICE. So QOM TYPE_CPU should be inherited
      from TYPE_DEVICE at least or TYPE_SYS_BUS_DEVICE.
   2. hot-plug on the bus should be allowed. if we ditch icc bus then we should allow
      hot-plug on sysbus. Can we do this? (i.e. it seems that for P4 and later cpus
      sysbus should be hot-plugable).
   3. should DeviceClass.init be used for cpu initialization or should .instance_init
      do all the job and make DeviceClass.init nop?

Another approach that tries to re-use device_add interface:
   1. allow run-time type detection in qdev_device_add and execute separate branch for
      TYPE_CPU. This way we could easily use links<> on sysbus
   2. device_del will require the same hacking as device_add
   3. apic now is sysbus device, question is what will be lost if it is attached to link and
      won't be sysbus_device_type anymore?
   4. will reset called on sysbus reach apic/cpu if it is on the link?

Any opinions on direction I should look more closely?