Re: [Qemu-devel] [RFC 0/4] virtio-mmio transport

[Pawel Moll]

On Mon, 2011-12-12 at 13:12 +0000, Michael S. Tsirkin wrote:
> On Mon, Dec 12, 2011 at 12:28:27PM +0000, Pawel Moll wrote:
> > On Mon, 2011-12-12 at 12:14 +0000, Stefan Hajnoczi wrote:
> > > I noticed the virtio-mmio spec has an interrupt status register.  On
> > > x86 and virtio-pci things are moving towards Message Signalled
> > > Interrupts and virtqueues having their own interrupts for better
> > > performance and flexibility.  Any thoughts on how 1 interrupt per
> > > virtqueue works for virtio-mmio?
> > 
> > This could be done by either creating devices with more then one
> > interrupt (platform device can take any number of resources) and
> > declaring that first queue uses the first one etc.
> 
> We currently support mapping from virtqueues to interrupt
> vectors in virtio core. Only virtio pci uses that
> but mmio can too. It's better than fixed mapping
> IMO as driver can control resources per queue.

I'll keep that in mind.

> > > My feeling is that the interrupt details are board-specific and can't
> > > be described in virtio-mmio, 
> > 
> > It's just the the "design pattern" in the "embedded world" that devices
> > usually have one interrupt output, shared between its internal
> > functions. And - of course - there is no in-band signalling (like MSI)
> > possible - interrupt lines are just "wires" :-) In a boundary case
> > scenario we may face a situation when total amount of interrupts for all
> > queues may actually exceed amount of interrupt inputs available in the
> > interrupt controller...
> > 
> > There may be a half-way solution - one interrupt per device but the
> > "active" queue number notified via the interrupt status register (as a
> > FIFO) so the driver wouldn't have to enumerate all the queues.
> 
> We could use a queue for this certainly.

Hm, yes, I suppose so :-) This would be a "system-level" queue rather
than the normal one, but I guess we could do that.

> Why do you have so many queues?

I assume this a question about the example I gave above? The answer is:
I obviously don't :-) This was just to point out that there _may_ be a
problem if we wanted to allocate an interrupt per queue.

Cheers!

Paweł