On Thu, 2008-11-06 at 14:02 -0600, Anthony Liguori wrote:
> Hollis Blanchard wrote:
> > I wanted to make sure people on non-x86 architectures couldn't run into
> > vring-size related problems that didn't also appear on x86.
> >
>
> Having a VRING_SHIFT and a VRING_PAGE_SIZE where VRING_PAGE_SIZE != (1
> << VRING_SHIFT) is almost certainly going to break things in unexpected
> ways because it's going against common understanding of the relationship
> between shift and page_size.
In the patch I sent, they are named VIRTIO_PFN_SHIFT and
VRING_PAGE_SIZE. In fact, the first could really be named
VIRTIO_PCI_PFN_SHIFT or even more specific, which hopefully would
alleviate the confusion.
Anyways, I've been trying to implement your other idea, which was to
advertise a "virtio page size" from the host to the guest. Since the
virtio IO space is only 20 bytes long, and it's full, we need to offer a
feature bit to notify the guest that the IO space has been expanded.
virtio's PCI IO resource
------------------------------------
| 20 bytes | |
------------------------------------
virtio-pci virtio-<device>
specific specific
data data
If guests don't acknowledge the feature, the host pretends the
VIRTIO_PAGE_SIZE register (which would be the new IO offset 20) doesn't
exist, and reads to offset 20 instead return the device-specific config
data.
Problem: the guest has not yet acknowledged the advertised features
before accessing device-specific data. So in the virtio-blk/virtio-pci
case, virtio_dev_probe() calls virtblk_probe() *before*
vp_finalize_features(). virtblk_probe() uses virtio_config_val(), which
returns the wrong data because the new yet unknown feature flag is still
set. (Of course, we can't just fix the ordering guest-side, because that
breaks backwards compatibility.)
It seems to me that virtio-pci configuration is not very flexible. I
think I'm going to continue revising the hardcoded page size patches I
sent earlier...
--
Hollis Blanchard
IBM Linux Technology Center
--
To unsubscribe from this list: send the line "unsubscribe kvm-ia64" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
