Re: Pe: [PATCH v5 1/3] virtio-scsi: first version

["Nicholas A. Bellinger" <nab@linux-iscsi.org>]

Hi Dor, James & Co,

On Mon, 2012-02-13 at 09:57 +0200, Dor Laor wrote:
> On 02/13/2012 09:05 AM, Christian Borntraeger wrote:
> > On 12/02/12 21:16, James Bottomley wrote:
> >> Well, no-one's yet answered the question I had about why.
> >
> > Just to give one example from a different angle:
> > In the big datacenters tape libraries are still very important, and lots
> > of them have a scsi attachement. virtio-blk certainly is not the right
> > way to handle those. Furthermore it seems even pretty hard to craft
> > a virtio-tape since most of those libraries have vendor specific library
> > controls (via sg). We would need to duplicate scsi generic (hint, hint :-)
> >
> >> virtio-scsi seems to be a basic duplication of virtio-blk except that it seems to
> >> fix some problems virtio-blk has.  Namely queue parameter discover,
> >> which virtio-blk doesn't seem to do.  There may also be a reason to cut
> >> the stack lower down.  Error handling is most often cited for this, but
> >> no-one's satisfactorily explaned why it's better to do error handling in
> >> the guest instead of the host.
> >>
> >> Could someone please explain to me why you can't simply fix virtio-blk?
> >
> > I dont think that virtio-scsi will replace virtio-blk everywhere. For non-scsi
> > block devices, image files or logical volumes virtio-blk seems to be the right
> > approach, I think.
> 
> +1
> 
> virtio-scsi is superior w.r.t:
>    - Device support: tapes, cdroms, other

AFAICT any type of non TYPE_DISK struct scsi_device passthrough is going
to currently require virtio-scsi in order to work.

>    - Does guest-host mapped multipath

The logic that comes with target_core_fabric_configfs.c and the native
target control plane gives a host-side (tcm_vhost) fabric driver generic
explict/implict ALUA multipath support by default.

I think there are some interesting possibilities for paravirtualized
ALUA multipath..  8-)

>    - Supports plenty of virtual disks mapped to the guest w/o need for a
>      pci slot per each virtio-blk

Ouch, virtio-blk lacks multi-lun per pci slot support..?

>    - offload fancy/new/sophisticated scsi commands from the guest to the
>      storage array w/o need for qemu implementation. Example XCOPY.
> 

...

> There are some more goodies like ability to support windows guest 
> clustering w/o hacky versions of scsi pass through over virtio-blk.
> virtio-blk is also a candidate to change the request based towards bio 
> based implementation, so sticking to it does not buy us too much.
> 

MSFT cluster guests that require SPC-3 PR support can run today with
tcm_loop LLD SCSI LUNs + SG_IO/BSG + right megasas QEMU HBA emulation,
but I do agree this would be better served by virtio-scsi for guests
that require SPC-3 PR support or passthrough.

--nab