[libvirt] [PATCHv5 00/13] qemu: allow disabling certain virtio revisions
Laine Stump
laine at laine.org
Wed Sep 7 20:34:17 UTC 2016
On 09/07/2016 03:38 PM, Sascha Silbe wrote:
> Dear Laine,
>
> Laine Stump <laine at laine.org> writes:
>
>> On 09/07/2016 02:35 PM, Sascha Silbe wrote:
>>> "Daniel P. Berrange" <berrange at redhat.com> writes:
>>> [...]
>>>> <sound model="virtio"/> == QEMU virtio
>>>> <sound model="virtio1.0"/> == QEMU virtio + disable-legacy
>>> What would this do for devices using the virtio-ccw transport?
>> From libvirt's point of view, the option "disable-legacy=on" would be
>> added to the device's commandline argument.
> Which would break s390x guests. virtio-ccw doesn't have any concept of
> "legacy" or "modern" devices (that's purely a virtio-pci concept), so
> virtio-*-ccw devices don't recognise that switch:
Okay, so you already know what would happen in qemu. Looking at Jan's
code in this patch series, (which I didn't do before, but should have)
when someone tries to set the option for disable-legacy=on when the
device address is anything except PCI , it logs an error and fails.
No code for Dan's suggestion has been written yet, but if there's no
concept of a legacy mode for virtio-*-ccw, then we would do the same
thing. And also I would guess that libosinfo would never suggest that
anyone try to add a "virtio1.0" model device to an s390 virtual machine).
>
> silbe at oc4731375738:~$ ~/build/qemu-devel/x86_64-softmmu/qemu-system-x86_64 -device virtio-blk,help 2>&1 |grep legacy
> virtio-blk-pci.disable-legacy=OnOffAuto (on/off/auto)
> silbe at oc4731375738:~$ ~/build/qemu-devel/s390x-softmmu/qemu-system-s390x -device virtio-blk,help 2>&1 |grep legacy
>
> That nicely illustrates the issue I have with a) mixing virtio-pci
> legacy/modern into the model name and b) conflating it with virtio
> 0.9/1.0 (or transitional/non-transitional for that matter).
>
> FWIW, the thing closest to virtio-pci legacy/modern is virtio-ccw
> max_revision. But I doubt there's any reason to set this beyond
> debugging and testing.
Definitely - once we've added an option to libvirt, we have to keep it
there forever - our backward compatibility guarantee requires it. So we
don't want to add anything unless there's a clear use for it.
More information about the libvir-list
mailing list