Re: [Qemu-devel] [PATCH 2/3] VCPU hotplug support

[Jan Kiszka]

On 2012-01-24 17:24, Igor Mammedov wrote:
> On 01/23/2012 06:55 PM, Jan Kiszka wrote:
>> On 2012-01-23 17:29, Igor Mammedov wrote:
>>> On 01/17/2012 03:17 PM, Jan Kiszka wrote:
>>>
>>> It seams that "env->cpuid_apic_id = cpu" is pointless especcialy
>>> taking in account that in cpu_x86_init cpuid_apic_id is initialized
>>> by cpu_index.
>>> What we gain in having cpuid_apic_id that actually duplicate cpu_index?
>>> May be there is sence to get rid of cpuid_apic_id?
>>
>> cpu_index is for internal counting (I think to remember that,
>> cpuid_apic_id is the ID exposed to the guest. During CPU hotplug, you
>> can control this by virtually plugging the CPU in a specific slot. So we
>> need to pass this ID down the init chain - just not set it in generic code.
> 
> It could be set in target specific new_cpu, it is not necessary to change
> whole cpu_init call chain and affect all other targets that might be not
> interested in this change at all.
> 
>>
>>>
>>> Another question is about how hot-plug/unplug should be designed:
>>> 1st approach:
>>>      With the current code we can't create vcpu with specific index.
>>
>> Forget about index, the apic_id is important to control. And that could
>> become something like -cpu XXX,apid_id=N. Of course, collisions need to
>> be detected and rejected.
>>
>>>      But we can implement xen like approach, where hot-plug command says
>>>      which amount of active vcpus guest should have. This way we can
>>>      leave current cpu_init ->  cpu_x86_init ->  cpu_exec_init call
>>>      chain without change and plug/unplug next/last vcpu.
>>>
>>> 2nd approach:
>>>      Ability to plug/unplug individual vcpus based on their cpu_index.
>>>      to do this we need add cpu_index argument to cpu_init ->
>>>      cpu_x86_init ->  cpu_exec_init call chain. It will look more
>>>      like the real hardware cpu hot-plug, but do virtual guests really
>>>      need it. And what for if this way is more preferrable than the 1st.
>>>
> 
> Jan,
> 
> Am I right in assuming that you are in a favor of 2nd approach with correction
> that some opaque cpu ID (in case of x86 it will be apic_id) will be passed 
> down
> to new_cpu?

I'm in favor of "device_add <cputype>,apic_id=<N>" and that the apic_id
becomes a qdev property. That way you would not need to pass anything
down, the device (cpu) init code could pick it up on its own.

Jan

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