On Tue, Nov 26, 2013 at 06:24:02PM +0800, Osier Yang wrote:
[[ TO libvir-list ]]
Hi, Daniel,
I'm going to share the thread to public list for further discussion. Hope you don't mind.
On 26/11/13 02:37, Daniel Erez wrote:
Hi Osier,
It seems there's a limitation in libvirt that allows up to six disks in a virtio-scsi controller. I.e. when sending more than six disks, libvirt automatically creates a new controller but of type virtual LSI Logic SCSI. Is this behavior a known issue?
For narrow SCSI bus, we allow 6 disks indeed.
For wide SCSI bus, we allow 15 disks (not including the controller itself on unit 7).
I'm doubting if we have problem on detecting if it supports wide SCSI bus though, since as far as I see from the user cases, it's always narrow SCSI bus.
Shouldn't libvirt allow up to 256 disks per controller or at least create a new controller of type virtio-scsi when needed?
The controller model for virtio-scsi controller is lsilogic, which we can't change simply, since it might affect the existing guests.
There was the similar discussion in libvir-list before [1].
But auto generation for controller is quite old, which I'm also not quite clear about. I'd like see another discussion to make it more clear whether we should do some work for upper layer app (e.g. oVirt).
Basicly two points:
* Should we do some changes on the maximum units for a SCSI controller, I.e. Should 7 (narrow bus) 16 (wide bus) be changed to other numbers? I'm afraid the changes could affect existing guests though.
* Do we really want to put the burden on users, I.E, let them create the controller explicitly. For use cases like one wants to add many disks for a guest, they need to know whether it's narrow SCSI bus or wide SCSI bus first (which we don't expose outside), and then do the calculation to know when to create a new SCSI controller.
@Daniel, am I correct on your problems? Please comments if it doesn't cover all your thoughts.
The logic for auto-assignment of disks to controllers is never going to make everyone happy. If people really care about the mapping they should define it explicitly by providing a controller address with their disk XML. 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 :|