On 09/16/2011 05:06 AM, David Gibson wrote:
>
> 1) do you mean, in Linux terms, one target per SCSI _host_ or one
> target per SCSI _channel_? i.e. if you looks at
> /sys/bus/scsi/devices, right now it looks like
>
> 0:0:0:0 0:0:1:0 (two targets on the same host and channel)
>
> Should it be?
>
> 0:0:0:0 0:1:0:0 (one target per channel)
>
> or
>
> 0:0:0:0 1:0:0:0 (one target per host)
>
> If it is the former, then you are simply hitting a limitation of the
> SCSI layer in QEMU and I do have patches to make assignment more
> flexible. Based on the Linux VSCSI driver, and based on what SLOF
> does, I'd guess that's what you mean.
Well, now I'm confused. I had a look at a pHyp machine, and Linux
seemed to see it as multiple targets on a single channel, but I'm sure
the PAPR spec says you shouldn't have that. So I'm going to have to
look closer now.
If this is the case, there might be a bug in SLOF's probing of SCSI devices.
SLOF probes target 0/LUN 0 on eight channels, i.e. from 0:0:0 to 7:0:0.
Linux however shows them the same as pHyp, i.e. from 0:0:0 to 0:7:0.
The reason this works is because LUN parsing in QEMU is completely
broken (by Ben's admission in spapr_vscsi.c :)) and so SLOF's x:0:0 and
Linux's 0:x:0 end up referring to the same device.
Now, when implementing SCSI addressing I had two choices:
(1) leave them where Linux sees them. This seems correct according to
your experiments with pHyp, but then SLOF could only see 0:0:0;
(2) move the devices so that both SLOF and Linux see them as x:0:0 (one
target per channel). This would be inconsistent with pHyp, but it
doesn't break either SLOF or Linux.
So, I would like to agree on a plan for merging the SCSI addressing
series. Right now I am doing (2), because it lets me use the current
version of SLOF. Is it okay for you to merge the feature with these
semantics?
If you want to change to (1), that can be done easily. However, it
requires fixing SLOF, so it would have to go preferably through Alex's
PPC tree.
(Again, that would be just the defaults---the addressing can always be
overridden by using -device explicitly).
Paolo
