On Sat, 2019-05-04 at 08:49 +0200, Christoph Hellwig wrote: > On Fri, May 03, 2019 at 10:00:54PM +0300, Max Gurtovoy wrote: > > Don't see a big difference of taking NVMe queue and namespace/partition to > > guest OS or to P2P since IO is issued by external entity and pooled outside > > the pci driver. > > We are not going to the queue aside either way.. That is where the > last patch in this series is already working to, and which would be > the sensible vhost model to start with. Why are you saying that? I actualy prefer to use a sepearate queue per software nvme controller, tat because of lower overhead (about half than going through the block layer) and it better at QoS as the separate queue (or even few queues if needed) will give the guest a mostly guaranteed slice of the bandwidth of the device. The only drawback of this is some code duplication but that can be worked on with some changes in the block layer. The last patch in my series was done with 2 purposes in mind which are to measure the overhead, and to maybe utilize that as a failback to non nvme devices. Best regards, Maxim Levitsky
