Re: [Qemu-devel] [PATCH 05/13] pci: Add pci_device_route_intx_to_irq

[Jan Kiszka]

On 2012-06-10 11:55, Michael S. Tsirkin wrote:
> On Thu, Jun 07, 2012 at 06:46:38PM +0200, Jan Kiszka wrote:
>> On 2012-06-07 18:28, Michael S. Tsirkin wrote:
>>> On Thu, Jun 07, 2012 at 05:10:17PM +0200, Jan Kiszka wrote:
>>>> On 2012-06-07 16:32, Michael S. Tsirkin wrote:
>>>>> On Mon, Jun 04, 2012 at 10:52:13AM +0200, Jan Kiszka wrote:
>>>>>> @@ -1089,6 +1093,14 @@ static void pci_set_irq(void *opaque, int 
>>>>>> irq_num, int level)
>>>>>>      pci_change_irq_level(pci_dev, irq_num, change);
>>>>>>  }
>>>>>>  
>>>>>> +PCIINTxRoute pci_device_route_intx_to_irq(PCIDevice *dev, int pin)
>>>>>> +{
>>>>>> +    PCIBus *bus = dev->host_bus;
>>>>>> +
>>>>>> +    assert(bus->route_intx_to_irq);
>>>>>> +    return bus->route_intx_to_irq(bus->irq_opaque, 
>>>>>> dev->host_intx_pin[pin]);
>>>>>> +}
>>>>>> +
>>>>>>  /***********************************************************/
>>>>>>  /* monitor info on PCI */
>>>>>>  
>>>>>
>>>>> Just an idea: can devices cache this result, bypassing the
>>>>> intx to irq lookup on data path?
>>>>
>>>> That lookup is part of set_irq which we don't bypass so far and where
>>>> this is generally trivial. If we want to cache the effects of set_irq as
>>>> well, I guess things would become pretty complex (e.g. due to vmstate
>>>> compatibility), and I'm unsure if it would buy us much.
>>>
>>> This is less for performance but more for making
>>> everyone use the same infrastructure rather than
>>> assigned devices being the weird case.
>>
>> Device assignment is weird. It bypasses all state updates as it does not
>> have to bother about migratability.
>>
>> Well, of course we could cache the host bridge routing result as well,
>> for every device. It would have to be in addition to host_intx_pin. But
>> the result would look pretty strange to me.
>>
>> In any case, I would prefer to do this, if at all, on top of this
>> series, specifically as it will require to touch all host bridges.
> 
> I'd like to ponder this a bit more then.
> 
> If the claim is that device assignment is only needed for
> piix anyway, then why not make it depend on piix *explicitly*?
> Yes ugly but this will make it very easy to find and
> address any missing pieces.

Because it is conceptually independent of the PIIX, we will need it for
successors of that x86 chipset as well, and I won't add the ugly hack of
qemu-kvm upstream

> 
> As it is you are adding APIs that in theory address
> non PIIX issues but in practice don't implement for non
> PIIX. So we never really know.

I once hacked q35 to make it work with device assignment. This really
requires something like this. It actually requires something generic,
independent of PCI, but that's too much for this round.

Jan

signature.asc
Description: OpenPGP digital signature