On Wed, Sep 21, 2016 at 04:01:23PM +0100, Daniel P. Berrange wrote:
I'm not a fan of the idea of silently picking a different device for the guest behind the applications back. By not exposing the different device types with a "model" attribute, we miss a way to report to the application which models are supported by the QEMU they're using - eg via domain capabilities.This in turn means the application doesn't know whether they're getting the new or old version, and so don't know if they're going to have working migration or not. If we expanded the XML to include model, then application can explicitly request the new model (which supports migration) and know that they'll get a startup failure if not supported, as opposed to silently falling back to the non-migratable version. Also, it makes life hard for us if the ivshmem-plain device wants to support use of the 'server' attribute in the future, as we will then not know which to create. We've often been burnt in the past by trying todo magic like this, instead of explicitly representing stuff in the XML, so I think we should be being explicit about ivshmem models here. Of course, if we do add <model> support, we have to allow for it to be missing for sake of upgrades. So there's a question of which model we should select as the default, if not seen in the XML.
If selecting the newest one whenever the element is missing is fine, then I'm OK with that. But that would change the device when upgrading libvirt (without user intervention), which you didn't like IIUC.
Regards, Daniel -- |: http://berrange.com -o- http://www.flickr.com/photos/dberrange/ :| |: http://libvirt.org -o- http://virt-manager.org :| |: http://autobuild.org -o- http://search.cpan.org/~danberr/ :| |: http://entangle-photo.org -o- http://live.gnome.org/gtk-vnc :|
Description: Digital signature
-- libvir-list mailing list email@example.com https://www.redhat.com/mailman/listinfo/libvir-list