On Mon, Jan 28, 2013 at 02:33:44PM +0100, Paolo Bonzini wrote: > Il 28/01/2013 14:36, Michael S. Tsirkin ha scritto: > > On Mon, Jan 28, 2013 at 02:29:23PM +0100, Paolo Bonzini wrote: > >> Il 28/01/2013 14:11, Michael S. Tsirkin ha scritto: > >>>> I asked for a standalone device because the configuration mechanism > >>>> (configfs vs. command-line) and the feature set are completely > >>>> different. Unlike virtio-net, it's not possible to switch one to the > >>>> other at run time. > >>> > >>> Exactly the same applies to any other frontend option. > >>> For example if you have two qemu instances with > >>> different num_queues values you can not migrate one > >>> to the other. > >>> So in this sense it is not different from any other > >>> frontend option, right? > >> > >> Indeed, in this sense it is not. > >> > >> Actually in this case migrating one to the other could succeed, and make > >> all disks disappear on the destination (because of the different > >> configuration mechanism). That however could be overcome with vhost=on > >> registering a migration blocker. > > > > Or better add a subsection if vhost is set: vhost=on to vhost=on > > can migrate, right? > > I think it's not yet supported by the kernel. You have no guarantee > that I/O is quiescent at the time the VM starts on the destination. > You'd need a ioctl to do the equivalent of bdrv_drain_all(). > > Once you have that, a subsection would do the job, yes. > > Paolo
OK once that's in it would be easy to probe for. > >> I won't really block the patch with the vhost=on/off frontend option if > >> it is properly done (e.g. the QEMU SCSI bus should not be created for > >> vhost=on) and minimally invasive to the non-vhost code. > >> > >> Paolo
