Re: [Qemu-ppc] [PATCH] spapr: Reduce advertised max LUNs for spapr_vscsi

[David Gibson]

On Thu, Sep 10, 2015 at 08:12:47AM +0200, Thomas Huth wrote:
> On 10/09/15 03:24, David Gibson wrote:
> > On Wed, Sep 09, 2015 at 09:29:18AM +0200, Thomas Huth wrote:
> >> On 09/09/15 09:19, David Gibson wrote:
> >>> On Wed, Sep 09, 2015 at 08:25:34AM +0200, Thomas Huth wrote:
> >>>> On 09/09/15 03:22, David Gibson wrote:
> >>>>> The implementation of the PAPR paravirtual SCSI adapter currently
> >>>>> allows up to 32 LUNs (max_lun == 31).  However the adapter isn't really
> >>>>> designed to support lots of devices - the PowerVM implementation only
> >>>>> ever puts one disk per vSCSI controller.
> >>>>
> >>>> Do you know how many LUNs are advertised by PowerVM?
> >>>
> >>> Well, what do you mean by "advertised".  AFAIK from the point of view
> >>> of the guest, the number of LUNs is advertised per-target, not per
> >>> controller.
> >>
> >> I mean, what's the highest LUN number that can be seen by a guest under
> >> PowerVM? Is it always using only one LUN per controller, or is there a
> >> way to change the amount of LUNs? (Sorry if I ask dumb questions ... I
> >> do not have much experience with PowerVM yet)
> > 
> > Um.. I'm not sure, I have very little experience with PowerVM too.  I
> > think with PowerVM it's usually real SCSI devices being passed
> > through, rather than disk images, so presumably the SCSI target itself
> > reports however many LUNs it has.  There may be a limitation in
> > PowerVM, or in the AIX VIO server I think it typically backends onto,
> > but I don't know what it is.
> > 
> > Since that limit has been in the guest side driver forever, presumbly
> > no-one has hit LUNs > 8 in practice.
> > 
> >>>>> More specifically, the Linux guest side vscsi driver (the only one we
> >>>>> really care about) is hardcoded to allow a maximum of 8 LUNs.
> >>>>
> >>>> So what about changing the vscsi driver in Linux instead to support more
> >>>> LUNs?
> >>>
> >>> Doesn't help for existing guests.  Basically what I'm trying to
> >>> achieve is for qemu to reject up-front configurations that are
> >>> unlikely to actually work in the guest.
> >>
> >> I just wonder whether it makes sense to change the guest instead. In the
> >> future, if we ever have guests that support more LUNs than 8 (maybe some
> >> non-Linux guests like FreeBSD?), we've got to change QEMU back again...
> >> OTOH, since this is just a one-line fix, it's likely ok to limit this to
> >> 8 now - it's easy to revert if we ever need to, so I'm fine with that
> >> change, I just wanted to discuss the other possibilites.
> > 
> > Remember that the spapr-vscsi device exists pretty much entirely to
> > make transition simpler for existing PowerVM guests.  New guests
> > (Linux or otherwise) intended to run under KVM should be using
> > virtio-blk or virtio-scsi.
> 
> FWIW, I had a quick look at FreeBSD sources here:
> 
> https://svnweb.freebsd.org/base/stable/10/sys/powerpc/pseries/phyp_vscsi.c?revision=259204&view=markup
> 
> ... and as far as I can see, they do not limit the LUNs to 8.
> (I only spotted a "cpi->max_lun = ~(lun_id_t)(0);" in there).
> So there indeed might also be older guests that support more than 8 LUNs.

Fair enough, you've convinced me.

I still think it makes sense downstream only, though.

-- 
David Gibson                    | I'll have my music baroque, and my code
david AT gibson.dropbear.id.au  | minimalist, thank you.  NOT _the_ _other_
                                | _way_ _around_!
http://www.ozlabs.org/~dgibson

pgpOSY60Kjn9C.pgp
Description: PGP signature