Re: [Qemu-devel] [PATCH/RFC 2/3] s390x/ais: enable ais when migration is available

[Pierre Morel <pmorel@linux.vnet.ibm.com>]

On 22/09/2017 16:07, Christian Borntraeger wrote:
> 
> 
> On 09/22/2017 04:02 PM, Pierre Morel wrote:
>> On 22/09/2017 14:40, Christian Borntraeger wrote:
>>>
>>>
>>> On 09/22/2017 02:13 PM, Pierre Morel wrote:
>>>> On 22/09/2017 10:38, Christian Borntraeger wrote:
>>>>> Instead of unconditionally enabling the KVM AIS capability
>>>>> in the kvm arch init function, do this in the flic realize function
>>>>> when we know if migration is available. This requires to initialize
>>>>> flic before the CPUs.
>>>>
>>>> I am not sure to agree.
>>>>
>>>> AIS facility is used for PCI (currently only PCI)
>>>> We want to support PCI emulation and PCI VFIO
>>>>
>>>>
>>>> Not having AIS support in the host kernel or not supporting AISM in the host kernel does not affect the emulation.
>>>> Neither virtio-pci nor TCG.
>>>> The only devices, (currently), which can not work without AIS and is not migratable without AISM are PCI VFIO devices.
>>>
>>> This patch enable the conditional enablement facility for the KVM host mask. The cpu model enablement is done
>>> differently for KVM and TCG anyway.
>>> Right now AIS is only enabled for KVM. For TCG AIS is not implemented at all and disabled. So for whenever this
>>> is fixed in TCG it can be handled then.
>>>
>>> And for emulated devices under KVM you still need the kernel support - otherwise migration is really broken for
>>> the nimm/dimm values.
>>>
>>
>> Yes, that is right.
>> I do not pretend that it is working directly.
>>
>> To support AIS emulation we have some work to do there (the migration), in kvm_s390_inject_airq() and other functions using ais_supported, to use emulated values.
>>
>> Don't you think that we could use emulation instead of faulting when the kernel can not handle values needed for pass-through?
>>
>> We can fault at the moment a VFIO device is realized.
>> Migration will be stopped on realizing VFIO PCI devices instead of flic.
>>
>> I can miss something, is there a reason not to do so?
> 
> The problem is not vfio vs emulation. The problem is that for KVM the kernel is handling the interrupts.
> and it holds the state of the interrupt controller (nimm/dimm) and without an interface to read/write these,
> the PCI core will be in a wrong state after migration regarding AIS.

I am aware of this.

> 
> Really, this patch just moves the supported host kernel from >=4.12 to >=4.13. It makes absolutely no
> sense to start having a mixed interrupt stack (qemu + kernel) just to make this work on 4.12.
> 
> 

In fact it would have been for much more than just the 4.12
But OK, I can understand that it makes no sense to discuss this now.
The urgency is to fence 4.12.
Eventually, we can discuss this later.

-- 
Pierre Morel
Linux/KVM/QEMU in Böblingen - Germany