[Qemu-devel] Re: [PATCH 5/5] scsi-bsg: Add initial support for BSG based SCSIDeviceInfo

[Paolo Bonzini <pbonzini@redhat.com>]

On 11/24/2010 09:40 AM, Nicholas A. Bellinger wrote:
> From: Nicholas Bellinger<nab@linux-iscsi.org>
>
> This patch adds initial support for using the Linux BSG interface with write/read vectored
> AIO as a QEMU backstore (SCSIDeviceInfo) with hw/scsi-bus.c compatible HBA emulation.
>
> So far it has been tested with x86_64 host and guest using hw/megasas.c and TCM_Loop LLD
> Port LUNs.  Because this patch uses struct iovec for struct sg_io_v4->d[out,in]_xferp payloads,
> which currently requires a patch to linux/block/bsg.c:bsg_map_hdr() in order to setup the
> user ->  kernel iovecs.   This patch can be found in lio-core-2.6.git here:
>
> commit fec4e8457c744de50e1ead69a36d5c4ee089d1ac
> Author: Nicholas Bellinger<nab@linux-iscsi.org>
> Date:   Sun Jun 13 23:13:20 2010 -0700
>
>      [BSG]: Add support for struct sg_io_v4->d[out,in]_iovec_count
>
> This also will only currently work with paired user/kernel (eg: 64bit user / 64bit kernel)
> because of different pointer sizes in struct iovec->iov_base.
>
> There are also two FIXMEs in hw/scsi-bsg.c:bsg_generic_initfn() related to extraction of
> SCSI LUN and device type values using BSG and required by QEMU-KVM.

Any reason why this should be a different driver than scsi-generic, i.e. 
why the scsi-generic should not switch automatically between SG and BSG 
depending on the kind of device being opened?

In fact, there is no reason why scsi-generic should use bounce buffers 
rather than sg_iovec_t, so that most of the differences between 
scsi-generic and scsi-bsg would go away.

Paolo