[PATCH] qemu: do not add model when actual iface type is hostdev

[Laine Stump]

On 6/17/20 4:54 AM, Paulo de Rezende Pinatti wrote:
>
> On 17/06/20 06:01, Laine Stump wrote:
>> On 6/16/20 10:32 AM, Paulo de Rezende Pinatti wrote:
>>> No default model should be added to the interface
>>> entry at post parse when its actual network type is hostdev
>>> as doing so might cause a mismatch between the interface
>>> definition and its actual device type.
>>
>>
>> Have you encountered a real problem from this? (I have, and have been 
>> thinking about the issue for awhile, but only late at night when I'm 
>> not near my keyboard to do something about it. I'm just wondering 
>> what problem you've had :-))
>
> Yes I have. I first defined a passthrough network out of a PF and then 
> created the domain with an interface definition '<source 
> network=passthrough/>' and no model (basically the process outlined in 
> https://wiki.libvirt.org/page/Networking#Assignment_from_a_pool_of_SRIOV_VFs_in_a_libvirt_%3Cnetwork%3E_definition) 
> , but after saving it libvirt automatically added <model 
> type='virtio'> (that's the default model on S390) which led to a qemu 
> error when trying to start the vm (see below).
>
>>
>>
>> The reason it's been on my mind is this: The default model is 
>> rtl8139. When libvirt is auto-assigning PCI addresses to devices at 
>> parse time, it decides whether to assign a network device to a 
>> conventional PCI or PCI Express slot based on the model, and rtl8139 
>> is conventional PCI. So if you have <interface type='network'> where 
>> the network is a pool of hostdevs, and if you don't assign a "fake" 
>> model like "virtio" or "e1000e", then the hostdev device (which is 
>> 100% certainly a PCIe device) will be assigned to a conventional PCI 
>> slot. That works, but is.... "sub-optimal" :-)
>>
>
> For me it didn't work at all because model virtio on S390 leads to a 
> 'devno' property being specified which of course is wrong for this 
> type of device (i.e. '-device 
> vfio-pci,host=0100:00:08.2,id=hostdev0,devno=fe.0.0006').


Okay, so that's a similar problem to mine - in my case the 
address-assignment code makes the choice of "conventional PCI" instead 
of "PCI Express", which isn't great, but still works. In your case it 
mistakenly thinks "CCW" instead of "PCI". Since that's not been reported 
before (that I know of) I guess that means you're the first to use a 
network pool of hostdevs on s390!


I just sent a patch to auto-assign these devices to PCIe on platforms 
that have PCIe. I Cc'ed you so that you might apply the patch locally 
and verify that I didn't break anything for you.


>
>
>>
>> I think we will still need to add a bit to 
>> qemuDomainDeviceCalculatePCIConnectFlags() in order to get the right 
>> type of slot set, but this is a good start.


I just sent a patch to auto-assign these devices to PCIe on platforms 
that have PCIe. I Cc'ed you so that you might apply the patch locally 
and verify that I didn't break anything for you.