Re: [virtio] Groups - Action Item "Create text version of virtio 0.9.5 document" added

[Anthony Liguori <aliguori@us.ibm.com>]

Pawel Moll <pawel.moll@arm.com> writes:

> On Wed, 2013-07-31 at 14:25 +0100, Anthony Liguori wrote:
>> I'm not sure where this discussion took place, but there is nothing in
>> QEMU that requires a dummy device to be used.
>
> http://thread.gmane.org/gmane.linux.kernel.virtualization/20144

N.B.  If you aren't discussing something on qemu-devel, then you aren't
getting the full input of the community.

>> The analogy would be a PCI slot that doesn't have a device plugged into
>> it.  Per the PCI spec, if a device is absent, any reads to the slot
>> will return a well defined value (all ones).  Those values are invalid
>> for things like a device id.
>
> It is very unsafe to "probe" memory mapped peripherals.

It simply depends on the devices.  Prior to PnP, probing was common
practice on the PC.  Yes, sometimes badness ensued.

>> I really dislike the design of virtio-mmio.  I don't understand why it
>> limits itself to a single device.  
>
> It is modelled after simple memory mapped peripherals, so it is simple
> itself. And if I was doing it today from scratch again, I would still
> focus on doing one thing only. After all there are discoverable buses
> out there, I don't see any reason to try to invent my own discovery
> mechanism.

Right, I've already argued with Peter that ya'll should simply use PCI
for mach-virt...  We don't need to repeat that here though.

>> virtio-blk is a separate device that can be connected to the transport.
>
> As in: block device of zero size? This has also been suggested in the
> thread linked above.

That's irrelevant here.

>> And yes, there should be a way to specify "this transport is not
>> connected to anything".
>
> In a "virtio generic" way or mmio-specific one?

If we can do it in a backwards compatible fashion, I think it would be a
very good idea to reserve a device ID to indicate that there is not a
valid device connected to the transport.

> Maybe instead of
> defining zero as a do-nothing-device we should simply make this value
> reserved or illegal?

Yes, I think that's a good idea.

> This would make PCI situation clear (no device will
> be ever allowed to have it) and mmio driver could "overload" its meaning
> as "ignore".

Agreed.

Regards,

Anthony Liguori

>
> Paweł