[libvirt] qemu: hotplug virtio_scsi over lsilogic when sicsi controller is not present?

[Peter Krempa]

On Fri, Jul 14, 2017 at 08:52:35 -0400, Liang Yan wrote:
> On 7/14/17 3:56 AM, Martin Kletzander wrote:
> > On Tue, Jul 11, 2017 at 03:47:55PM -0400, Liang Yan wrote:
> >> Hi,
> >>
> >> We hit some problems when we attached some lun devices in our vm, turns
> >> out that libvirt created lsilogic scsi controller automatically for
> >> these scsi devices, however these device works well under virtio_scsi
> >> controller.
> >>
> >> the current code logic is check lsilogic first, if qemu could not
> >> support it, then check virtio_scsi, however is it better to check
> >> virtio_scsi earlier? since it is supported better in qemu level?
> >>
> >
> > It is not in older QEMUs which we need to stay compatible with.
> >
> Yes, Peter said similar things that kernel may not include newer driver.
> >> I am wondering which solution is better?
> >> 1. simple switch sequence and check virtio_scsi first
> >>
> >
> > We can't do that.  We have to keep parsing older XMLs (that did not have
> > the model in them) the same way as we were before so that we keep stable
> > guest ABI.
> >
> >> 2. add extra option for "virsh attach-disk" to choose specific
> >> controller type
> >>
> >
> > Or you can first attach the controller and then attach-device with the
> > xml that specifies that particular controller.  attach-disk is a
> > syntactic sugar for attach-device IIRC.
> >
> Yes, this is our current workaround.  Some users are just wondering if
> could finish this process by command line(attach-disk) in one step. I
> checked the options of attach-disk that "-targetbus" is a close one but
> focus on high level bus controller. 

I don't think that will be easy to implement. Adding a new controller is
a rather complex step which can fail and having a two step operation
crammed into a single API is always painful in error cases, since you
can end up in a partially modified state which can't be rolled back.

The current state was to make it just work, but as you can see, this
can't be guessed correctly.

The problem also is that one parameter may be enough to get the
controller type for now, but in the future it may not be enough as it is
not enough just to pick the default one now.

I think it would be much more usefull to add a virsh convenience command
to attach controllers, so that users can do this in a simple way rather
than trying to do it in a single API call.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 833 bytes
Desc: not available
URL: <http://listman.redhat.com/archives/libvir-list/attachments/20170714/ff6ddbc0/attachment-0001.sig>