(In reply to comment #0)  
> Actual results:
> after step 1, it will fail to boot up, and qemu will prompt like:
> qemu-system-x86_64: -drive
> file=/dev/sg1,if=none,id=drive-data-disk,format=raw,werror=stop,rerror=stop,
> cache=none,aio=native: could not open disk image /dev/sg1: Invalid argument
> 
Ignore above prompt, as this prompt is about the cache=none prompt. The right one as following:
e.g:...-drive file=/dev/sg2,if=none,id=drive-data-disk,format=raw,werror=stop,rerror=stop,aio=native -device virtio-scsi-pci,id=scsi1,addr=0x5 -device scsi-generic,bus=scsi1.0,drive=drive-data-disk,id=data-disk

qemu-kvm: -device scsi-generic,bus=scsi1.0,drive=drive-data-disk,id=data-disk: Device doesn't support drive option werror
qemu-kvm: -device scsi-generic,bus=scsi1.0,drive=drive-data-disk,id=data-disk: Device 'scsi-generic' could not be initialized

--------------------------------------------------------------------------------

Paolo, 

    If i boot guest with the cache=none for scsi-generic drive, it will fail to boot up, does it the same issue, does the scsi-generic support the cache=none for drive ? should we need to open a new bug for it ? btw, the "writeback", "unsafe", "directsync" and "writethrough" is ok for cache value of scsi-generic drive. 
e.g:...-drive file=/dev/sg2,if=none,id=drive-data-disk,format=raw,cache=none,aio=native -device virtio-scsi-pci,id=scsi1,addr=0x5 -device scsi-generic,bus=scsi1.0,drive=drive-data-disk,id=data-disk 

qemu-kvm: -drive file=/dev/sg2,if=none,id=drive-data-disk,format=raw,cache=none,aio=native: could not open disk image /dev/sg2: Invalid argument

Best Regards.
sluo

Thanks for the report.  Yes, you can open a separate bug.  QEMU should ignore cache=none for scsi-generic because it always _behaves_ like cache=none but _using_ cache=none fails (O_DIRECT is not supported by /dev/sgN devices).

(In reply to comment #2)
> Thanks for the report.  Yes, you can open a separate bug.  QEMU should
> ignore cache=none for scsi-generic because it always _behaves_ like
> cache=none but _using_ cache=none fails (O_DIRECT is not supported by
> /dev/sgN devices).
see bug 920096, thx for your kindly checking.

scsi-generic.c needs to call scsi_handle_rw_error() (or equivalent function) like
scsi-disk.c does (see commit ea8a5d7 and commit 5dba48a).  Upstream first.

We refined rerror/werror upstream to not apply when the LBA range is
invalid, and are backporting that to RHEL-7 (bug 1085232).

Stupid question: how could we similarly filter out invalid LBA ranges
for scsi-generic?

rerror/werror require support in the underlying device model.
Currently, only ide-hd, ide-cd, ide-drive (legacy), scsi-hd, scsi-cd,
scsi-disk (legacy), scsi-block (partially[*]), virtio-blk-device do.

Device models that do not support it: isa-fdc, cfi.pflash01,
scsi-generic, nvme, ...  This bug is only about scsi-generic.

See also bug 1123355 - Document -drive rerror/werror restrictions.


[*] Should work as long as you stick to plain read/write.  Fancy stuff
like write same gets tricky.

The obvious way to fully implement rerror/werror for scsi-generic
involves understanding all SCSI commands.  Even vendor-specific
commands.  Defeats the purpose device pass-through.

Perhaps something covering just werror=enospc (and the default
=report, of course) could be hacked up with just understanding sense
codes.

Unless we can come up with workable ideas *and* a business case
justifying the work, this feels like a WONTFIX to me.

However, Paolo would like to address this upstream.  His rough ETA is
in one year.  He's prefer to keep this bug open for now.