On 08/09/2016 08:13 AM, Michal Privoznik wrote:
On 25.07.2016 22:48, Eric Farman wrote:
In the case of other SCSI host devices, the device that is created within the guest will be within the usual SCSI namespace (e.g., host:bus:target:lun). But for vhost-scsi, the actual naming structure will come from the host configuration, and will be invisible to both QEMU and Libvirt. So specifying one with an <address type='drive' ...> tag (as is often done for virtio-scsi) doesn't make sense in this case, as it will be ignored.
Nevertheless, we need something to identify our vhost-scsi device, so for that we'll drop back to the devno mapping that we use for other disks. (This option exists only for vhost-scsi-ccw, but not vhost-scsi-pci. Hrm...) This becomes <address type='ccw' ...> in the case of s390 systems, and puts the device number on the resulting QEMU command line. Thus, existing device number conflict detection can be used across other devices that may or may not be specified in the guest XML.
If one is not specified, we need to be sure to allocate one so that we can avoid polluting the device numbers with silently-created entries.
Signed-off-by: Eric Farman <farman@linux.vnet.ibm.com> Reviewed-by: Bjoern Walk <bwalk@linux.vnet.ibm.com> Reviewed-by: Boris Fiuczynski <fiuczy@linux.vnet.ibm.com> --- src/conf/domain_conf.c | 19 +++++++++++++++++-- src/qemu/qemu_command.c | 7 +++++++ src/qemu/qemu_domain_address.c | 10 ++++++++++ 3 files changed, 34 insertions(+), 2 deletions(-) I don't feel confident enough to review this one. The code looks okay, but from higher perspective I have no idea what it does. Sorry.
In testing my rework to Daniel Berrange's comment to patch 10 (https://www.redhat.com/archives/libvir-list/2016-August/msg00464.html) this patch ends up going away. So, no worries on the uncertainty here! - Eric