Re: [Qemu-devel] If Qemu support NVMe over Fabrics ?y

[Keith Busch <keith.busch@intel.com>]

On Fri, Jan 11, 2019 at 05:07:26PM +0100, Paolo Bonzini wrote:
> On 11/01/19 16:58, Stefan Hajnoczi wrote:
> > Before investing time in doing that, what is the goal?
> > 
> > Is this for test and bring-up of NVMe-oF?  Or why does the guest need to
> > know that the storage is NVMe-oF?
> > 
> > As I mentioned before, if your host supports NVMe-oF you can simply give
> > the block device to QEMU and let the guest access it via virtio-blk,
> > virtio-scsi, NVMe, etc.
> 
> NVMe-OF is an RDMA protocol, basically a different transport for the
> NVMe command set and queue abstraction.  It would allow QEMU to access
> the device directly (similar to what block/nvme.c does for PCI using
> VFIO) without going through the host kernel.  The guest would see the
> device as virtio-blk/scsi, NVMe or anything else.
> 
> Paolo

I think there's multiple ways to interpret the original request, and I'm
not sure which is the right one.

If he just wants a guest's backing storage to be an nvmeof target,
then Stefan's response sounds right with qemu abstracted from the
underlying protocol as long as the host supports it.

The other way I interpreted this is he wants a guest to connect to an
nvmeof target. If so, then the guest just needs to be provisioned with an
RDMA NIC or fibre channel, or even a virtio-nic may work if the fabric
target is NVMe-over-TCP. You'd just need a capable nvmeof driver stack
in the guest.

If the request is really to implement a vfio based solution to bypass
the host kernel ... that sounds like an interesting project. :)