> > This could be documented, > > It is documented AFAIK. Pls take a look at the spec documentation.
Found it now. It's not under GET_VRING_BASE, it's under "starting and stopping rings"---fair enough. In the case of vhost-user-scsi, however, QEMU also must not proceed until vhost-user-scsi has drained the pending I/O---and this pending I/O would be completed _after_ QEMU has sent GET_VRING_BASE. Is this handled by VHOST_USER_PROTOCOL_F_REPLY_ACK already? If so, migration would be denied if the server lacks that protocol feature. Paolo > > but perhaps it's best to add a START_STOP > > feature and message to the vhost-user protocol? > > We just never need to GET_VRING_BASE if ring keeps going - > makes no sense since base gets invalidated immediately. > > > > > The feature then can be optional for vhost-user-net and mandatory for > > vhost-user-scsi. When this is done we can remove .unmigratable. > > > > Thanks, > > > > Paolo > > If vhost-user-scsi does not stop the ring after responding to > GET_VRING_BASE, it's just a bug that needs to be fixed. > > > > > > > > > Can you please send a version of your patch that uses .unmigratable? > > > > > > Sure I can do that. We can work on the migration later on. > > > > > > > > > > > I'll send a v6 that momentarily drops vhost-scsi, but I intend to > > > > include it again in the next pull request. > > > > > > Sounds good to me. > > > > > > Felipe > > > > > > > > > > > Thanks, > > > > > > > > Paolo > > > > > > >