Re: [PATCH RFC] kvm: enable irq injection from interrupt context

["Michael S. Tsirkin" <mst@redhat.com>]

On Thu, Sep 16, 2010 at 05:43:26PM +0200, Gleb Natapov wrote:
> On Thu, Sep 16, 2010 at 05:24:11PM +0200, Michael S. Tsirkin wrote:
> > On Thu, Sep 16, 2010 at 04:51:17PM +0200, Gleb Natapov wrote:
> > > > > What for? Device emulation should do de-assert.
> > > > 
> > > > Sorry, but at this point I have no idea what you call device emulation.
> > > The same thing everyone calls device emulation. In case of virtio-net it
> > > is in hw/virtio-net.c. If vhost-net is in use device emulation is split
> > > between userspace and kernel, but it is still just device emulation.
> > 
> > 
> > case in point, virtio net does not know about pci at all,
> > so it can not deassert.
> > 
> I don't see what PCI has to do with it. virtio-net device implementation is
> split between several files. virtio-net/virtio-ring/virtio-pci. All of
> them still implement one emulated device. Whoever implemented virtio-net
> initially decided to put irq ack register into PCI config space, so code
> that does de-assert is in virtio-pci,
> but it is still part of emulated
> device.
> 
> > > > qemu has code to de-assert. vhost has code to assert.
> > > Good. So qemu will de-assert. So what do you mean by
> > > "KVM would need to find all irqfd objects mapped to gsi and notify
> > > them on deassert"
> > > 
> > > > I would like to optimize level interrupts and stop driving
> > > > scheduler insane if at all possible.
> > > > 
> > > Worthy goal. Do it in irqfd. Irqfd shouldn't call kvm_set_irq() if irq
> > > level hasn't changed.
> > 
> > Right. Then it needs to know about deasserts. It does not get this
> Dessert should be done by qemu writing 0 into irqfd.

writing 0 to eventfd does nothing. The way to deassert irq
is currently through ioctl and it seems entirely sensible
to me to keep it that way.

Also - which irqfd :)
We need multiple irqfds to map to a single gsi, with kvm
doing OR on them.

> Assertion and
> de-assertion should be done through irqfd.
> > information, so when kvm gets deassert ioctl it should
> > locate irqfd object and set level to 0.
> > See?
> No.
> 
> --
> 			Gleb.