Re: [Qemu-discuss] qemu and lio virtio?

[Mike Lovell]

On 05/20/2014 10:05 AM, Phillip Susi wrote:
> I've been reading up on lio / linux scsi target, and they claim to
> have a virtio fabric module that implements a virtio disk for the
> purpose of kvm-qemu.  This makes no sense to me.  My understanding of
> virtio is that qemu implements it in user space to present an
> interface to the guest os, where it shows up as a pci device.  This
> seems to have nothing whatsoever to do with lio, which seems to be all
> about emulating a scsi disk on a non virtual system.

this is actually the combination of a couple pieces. first off, virtio 
is just a way to send data to and from the guest. there are several 
things which build components using it such as virtio-net which provides 
a virtual ethernet interface that passes ethernet frames between the 
guest and the host. i think you are using the term virtio to refer to 
virtio-blk (or block) which is a virtual block device. when the guest 
wants to read a block from the device, it sends some data across virtio 
to qemu which interprets the command, performs the actual data 
retrieval, and returns the data to the guest through virtio. that is a 
very high level simplification of it.

virtio-blk has a few limitations and some of the developers wrote a new 
virtio based mechanism called virtio-scsi. this uses virtio to present a 
virtual scsi controller to the guest and the driver for the guest has 
been in the upstream kernel source for about 2 years. this driver is 
just the guest side of it.

since the data being passed over virtio is now scsi commands, there 
needs to be something on the host side to interpret and act on those 
commands. there are a couple methods for this. the first one that qemu 
supported was that qemu would do all the work itself and is what would 
be used for things like qcow2 files. another method is to use the lio 
scsi target and this is possible because the guest and host are using 
the scsi command set to communicate. vhost-scsi provides a way for qemu 
to connect the virtual scsi controller (or initiator in scsi parlance) 
to the scsi target in the host using the vhost infrastructure. there is 
actually a third method that i know of which is to the use libiscsi 
support in qemu to connect the guest to an iscsi target entirely in 
user-space with no iscsiadm commands required on the host.

unfortunately, it seems that there isn't much documentation around how 
to use virtio-scsi with vhost-scsi and tcm_vhost. using this requires 
configuring the lio target and the qemu. the documentation on the lio 
wiki describes how to configure the target [1]. a while ago, i found a 
post to lkml that shows some benchmarks but also has a 
qemu-system-x86_64 command example [2]. it is a couple years old so 
there might be some details that have changed but it might still work.

hopefully that gives you a better idea of how all of this fits together.

mike

[1] http://linux-iscsi.org/wiki/VHost
[2] https://lkml.org/lkml/2012/8/10/347