Re: [Qemu-devel] virtio scsi host draft specification, v2

[Michael S. Tsirkin]

On Thu, Jun 02, 2011 at 02:02:52PM +0200, Paolo Bonzini wrote:
> On 06/02/2011 01:41 PM, Michael S. Tsirkin wrote:
> >Now to our problem:
> >As far as I can tell there are two input buffers in each request: sense
> >and data. Right?
> >
> >If sense is fixed length, we can simply put it first, have device write
> >sense then data.  This does not seem too limiting, if you want a lot of
> >flexibility sense length can be in device config.  If we don't want to
> >limit ourselves to fixed length sense, we would have driver use two
> >heads for a request.  This is possible but one needs to be careful in
> >the driver to make sure there's enough space for both requests. Maybe
> >add_bufs API to add multiple bufs might be a good idea here.
> 
> I should be on holiday today so I'll answer this quickly.  Sounds
> like we can converge, I'll put data at the end and define the length
> of sense in the config: the device writes a default (defined by the
> spec to be always 96) and the driver can modify it.  The _single_
> head would contain:
> 
> - read-only: command etc.
> 
> followed by:
> 
> - write-only: sense, status etc.
> 
> followed by:
> 
> - read-only: data to device
> - write-only: data from device

Yes, this works.

> IIUC, qemu only sees a bunch of read-only and write-only buffers.
> It doesn't see the relative ordering of read-only vs. write-only.

In virtio write is always before read.

> But it knows the sizes of read-only and write-only data, so it can
> figure out datain_size and dataout_size.  sense_size is in the
> config, so neither of the three needs to be in the request.
> 
> sense_len needs to stay, since any number of bytes can be written in
> the sense buffer.

I think this means sense_len needs to go into the in buffer
as well. As head is first it's an out buffer.

>  The used-length field should be usable for
> uni-directional commands, but I'm not sure about commands that have
> both datain and dataout.  I'll read the SCSI spec about it tomorrow.

used-length is the part of in buffer actually written.
actual data len is thus used-length - 96
(sense is assumed to be fixed length by virtio ring,
 sense_len tells driver how many bytes are
 actually valid).


> 
> Making qemu support arbitrarily partitioned buffers may require some
> extra utility functions to work on iovecs, but nothing too complex.
> If your patches already contain something like that, please dig them
> up so I can avoid duplicate work!
> 
> Paolo

Yes. Will do.

-- 
MST