Re: [Qemu-devel] [PATCH 1/5] virtio: rename VIRTIO_PCI_QUEUE_MAX to VIRTIO_QUEUE_MAX

[Jason Wang]

On Fri, Feb 6, 2015 at 7:54 PM, Cornelia Huck 
<address@hidden> wrote:
> On Fri,  6 Feb 2015 15:54:09 +0800
> Jason Wang <address@hidden> wrote:
>
> >  VIRTIO_PCI_QUEUE_MAX was not specific to pci, so rename it to 
> > VIRTIO_QUEUE_MAX.
> >  
> >  Cc: Amit Shah <address@hidden>
> >  Cc: Anthony Liguori <address@hidden>
> >  Cc: Michael S. Tsirkin <address@hidden>
> >  Cc: Alexander Graf <address@hidden>
> >  Cc: Richard Henderson <address@hidden>
> >  Cc: Cornelia Huck <address@hidden>
> >  Cc: Christian Borntraeger <address@hidden>
> >  Cc: Paolo Bonzini <address@hidden>
> >  Signed-off-by: Jason Wang <address@hidden>
> >  ---
> >   hw/char/virtio-serial-bus.c  |  2 +-
> >   hw/s390x/s390-virtio-bus.c   |  4 ++--
> >   hw/s390x/s390-virtio-ccw.c   |  2 +-
> >   hw/s390x/virtio-ccw.c        | 10 +++++-----
> >   hw/scsi/virtio-scsi.c        |  4 ++--
> >   hw/virtio/virtio-mmio.c      |  4 ++--
> >   hw/virtio/virtio-pci.c       | 10 +++++-----
> >   hw/virtio/virtio.c           | 26 +++++++++++++-------------
> >   include/hw/s390x/s390_flic.h |  2 +-
> >   include/hw/virtio/virtio.h   |  2 +-
> >   10 files changed, 33 insertions(+), 33 deletions(-)
>
> While I certainly agree with the patch on its own, I read through your
> complete patch series as well and I have some additional comments.
>
> Your patch series bumps up the maximum number of virtqueues for all
> transports, but I think we need to consider carefully what this means
> for each transport.
>
> For virtio-ccw, for example, we'll run into trouble when the guest 
> uses
> classic queue indicators: These are limited to 64 per device (as I
> assumed 64 virtqueues per device initially and they fit quite nicely
> into a 64 bit integer). If the guest uses adapter interrupts with
> two-stage indicators, we should probably be fine. The guest will
> hopefully care about that, but we need to describe failure modes for
> this in the spec.
>
> For s390-virtio, I frankly have no idea what large numbers of queues
> will do :) Maybe we just run out of space for devices much earlier.

I see the issue, thanks. So looks like we need a transport specific 
limitation for this.
> You seem to need some changes for this to work for virtio-pci as well,
> don't you?

PCI allows at most 2048 MSI-X entires, and also allow one entry to be 
shared by several virtqueues which looks fine for even thousands of 
virtqueues. If there's a architecture limitation (e.g x86 only allow 
240 interrupt vectors for MSI-X), the rest work was in the driver. If 
not, probably no much work needs to be done for PCI.

Thanks