Re: [Qemu-devel] Do you have a use for a tester of virtio-scsi with CD drives ?

[Paolo Bonzini <pbonzini@redhat.com>]

On 11/02/2011 12:25 PM, Stefan Hajnoczi wrote:
>> Now i have two drives by one option.
>> xorriso -devices reports
>>   0  -dev '/dev/sr0' rwrw-- :  'TSSTcorp' 'CDDVDW SH-S223B'
>>   1  -dev '/dev/sr1' rwrw-- :  'QEMU    ' 'QEMU DVD-ROM'
>> with /dev/sr1 being an empty drive. (Is this a known bug ?)
>
> Off the top of my head I'd say this is an empty IDE CD-ROM drive which
> is added by default.  Would need to check the code to be sure though.

Yes, that's correct.

That empty IDE CD-ROM was the same that you were getting with 
virtio-blk.  With virtio-blk you get a /dev/vda drive that will respond 
to SG_IO, so it doesn't look like a pass-through CD-ROM but it is one.

>> The window where i ran qemu says:
>>
>>   qemu: ./qemu_dir/qemu-0.15.1/hw/lsi53c895a.c:540:
>>         lsi_do_dma: Assertion `s->current' failed.
>
> That's a bug in the LSI SCSI HBA emulation or a generic SCSI emulation
> problem in QEMU.  I have CCed Paolo and Kevin who have recently been
> working on SCSI and IDE much more than me and know more.

I suggest trying with 1.0-rc (origin/master).  No guarantee that it 
works, but definitely a lot of bugs have been fixed.  It is also 
possible that a few new ones were introduced.  I can also provide an 
updated version of my vmw_pvscsi device model, which is simpler and more 
stable than the lsi device model.

Also, with 1.0-rc you can do passthrough in two ways: with scsi-generic 
as you did, which always uses SG_IO, and with scsi-block.  scsi-block 
uses read+write for READ/WRITE CDBs and SG_IO for everything else.  You 
can define a scsi-block disk like this:

   -drive file=/dev/sr0,if=none,id=hostcd -device scsi-block,drive=hostcd

The advantage of scsi-block is that it doesn't need a bounce buffer as 
big as the whole request; in virtio-scsi it will not require a bounce 
buffer at all.  From the spec I don't understand how this will fail to 
burn CDs.  It's possible that some modes work and others don't.  I 
planned to test this on test day (Nov 7th).

> Thanks for the extensive testing.  I actually suggest focussing just
> on qemu.git/master because that is where developers mostly invest
> their time.  The data points on older versions can be interesting but
> my guess is folks won't have time to debug them - instead let's make
> qemu.git/master work.

Agreed.

Paolo