Re: [Qemu-devel] [PATCH 6/8] spapr: move interrupt allocator to xics

[Alexander Graf <agraf@suse.de>]

On 11.04.14 18:01, Alexey Kardashevskiy wrote:
> On 04/12/2014 01:38 AM, Alexander Graf wrote:
>> On 11.04.14 17:27, Alexey Kardashevskiy wrote:
>>> On 04/12/2014 12:58 AM, Alexander Graf wrote:
>>>> On 11.04.14 16:50, Alexey Kardashevskiy wrote:
>>>>> On 04/11/2014 11:58 PM, Alexander Graf wrote:
>>>>>> On 11.04.2014, at 14:38, Alexey Kardashevskiy <aik@ozlabs.ru> wrote:
>>>>>>
>>>>>>> On 04/11/2014 07:24 PM, Alexander Graf wrote:
>>>>>>>> On 10.04.14 16:43, Alexey Kardashevskiy wrote:
>>>>>>>>> On 04/10/2014 11:26 PM, Alexander Graf wrote:
>>>>>>>>>> On 10.04.14 15:24, Alexey Kardashevskiy wrote:
>>>>>>>>>>> On 04/10/2014 10:51 PM, Alexander Graf wrote:
>>>>>>>>>>>> On 14.03.14 05:18, Alexey Kardashevskiy wrote:
>>>>>>>>>>>>> The current allocator returns IRQ numbers from a pool and does not
>>>>>>>>>>>>> support IRQs reuse in any form as it did not keep track of what it
>>>>>>>>>>>>> previously returned, it only had the last returned IRQ.
>>>>>>>>>>>>> However migration may change interrupts for devices depending on
>>>>>>>>>>>>> their order in the command line.
>>>>>>>>>>>> Wtf? Nonono, this sounds very bogus and wrong. Migration shouldn't
>>>>>>>>>>>> change
>>>>>>>>>>>> anything.
>>>>>>>>>>> I put wrong commit message. By change I meant that the default state
>>>>>>>>>>> before
>>>>>>>>>>> the destination guest started accepting migration is different from
>>>>>>>>>>> what
>>>>>>>>>>> the destination guest became after migration finished. And migration
>>>>>>>>>>> cannot
>>>>>>>>>>> avoid changing this default state.
>>>>>>>>>> Ok, why is the IRQ configuration different?
>>>>>>>>> Because QEMU creates devices in the order as in the command line, and
>>>>>>>>> libvirt changes this order - the XML used to create the guest and the
>>>>>>>>> XML
>>>>>>>>> which is sends during migration are different. libvirt thinks it is ok
>>>>>>>>> while it keeps @reg property for (for example) spapr-vscsi devices
>>>>>>>>> but it
>>>>>>>>> is not because since the order is different, devices call IRQ
>>>>>>>>> allocator in
>>>>>>>>> different order and get different IRQs.
>>>>>>>> So your patch migrates the current IRQ configuration, but once you
>>>>>>>> restart
>>>>>>>> the virtual machine on the destination host it will have different IRQ
>>>>>>>> numbering again, right?
>>>>>>> No, why? IRQs are assigned at init time from realize() callbacks (and
>>>>>>> survive reset) or as a part of ibm,change-msi rtas call which happens in
>>>>>>> the same order as it only depends on pci addresses and we do not change
>>>>>>> this either.
>>>>>> Ok, let me rephrase. If I shut the machine down because I'm doing
>>>>>> on-disk hibernate and then boot it back up, will the guest find the same
>>>>>> configuration?
>>>>> I do not understand what you mean by this. Hibernation by the guest OS
>>>>> itself or by QEMU? If this involves QEMU exit and QEMU start - then yes,
>>>> by the guest OS. The host will only see a genuine "shutdown" event. The
>>>> guest OS will expect the machine to look *the exact same* as before the
>>>> shutdown.
>>> Ok. So. I have to implement "irq" property everywhere (PHB is missing
>>> INTA/B/C/D now) and check if they did not change during migration via those
>> Hrm. Not sure. Maybe it'd make sense to join next week's call on platform
>> device creation. The problem seems pretty closely related.
> What are those platform devices and what are you going to discuss exactly?

Devices that don't have a unified interrupt routing scheme like PCI 
where you just link lines A/B/C/D to your controller and you're good to go.

>
>
>>> VMSTATE.*EQUAL. Correct?
>> Why would you need this? I think we already said a couple dozen times that
>> configuration matching is a bigger problem, no?
> For debug! It is not needed in general, yes.
>
>
>>> If so (more or less), I still would like to keep patches 1..7.
>>> In fact, the first one is independent and we need it anyway.
>>> Yes/no?
>> Why?
> IOMMUs do not migrate correctly - they only have a class have and
> instance_id and this instance_it depends on command line arguments order.
> The #1 patch makes it classname + liobn.

Why do we need a bus for that?

>
>
>>>
>>>>> config may be different. If it is "migrate to file" and then "migrate from
>>>>> file" (do not know what you call it when migration goes to a pipe which is
>>>>> "tar") - then config will be the same.
>>>>>
>>>>>
>>>>>>>> I'm not sure that's a good solution to the problem. I guess we should
>>>>>>>> rather aim to make sure that we can make IRQ allocation explicit.
>>>>>>>> Fundamentally the problem sounds very similar to the PCI slot
>>>>>>>> allocation
>>>>>>>> which eventually got solved by libvirt specifying the slots manually.
>>>>>>> We can do that too. Who decides? :)
>>>>>> The better solution wins :)
>>>>> We both know who decides ;) I posted series, I need heads up if it is
>>>>> going
>>>>> the right way or not.
>>>> It's not :). If a guest may not have different IRQ allocation after
>>>> migration, it also must not have different IRQ allocation after shutdown +
>>>> restart.
>>> Ok. That's good answer, thanks. How does x86 work then? IRQs are hardcoded
>>> (some are for sure but I do not know about MSI)? Or in order to support
>> Non-PCI IRQs are hardcoded, yes. PCI IRQs are mapped to one of the 4 PCI
>> interrupts which again are hardcoded to IOAPIC interrupt lines after some
>> PCI line swizzling.
> This is what I meant - I need to have a way to tell PHB IRQ numbers for
> INTA/B/C/D.

Yes, just like platform devices ;).


Alex