Re: [Qemu-devel] QEMU -netdev vhost=on + -device virtio-net-pci bug

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

On Fri, Mar 08, 2013 at 03:48:04PM +1100, Alexey Kardashevskiy wrote:
> Michael,
> 
> Thanks for the fix.
> 
> There was another question which was lost in the thread.
> 
> I am testing virtio-net in two ways:
> 
> Old -net interface:
> -net tap,ifname=tap0,script=qemu-ifup.sh \
> -net nic,model=virtio,addr=0:0:0
> 
> (qemu) info network
> hub 0
>  \ virtio-net-pci.0:
> index=0,type=nic,model=virtio-net-pci,macaddr=52:54:00:12:34:56
>  \ tap.0: index=0,type=tap,ifname=tap0,script=qemu-ifup.sh,downscript=/etc/qemu-ifdown
> 
> New -netdev interface:
> -netdev tap,id=tapnet,ifname=tap0,script=qemu-ifup.sh \
> -device virtio-net-pci,netdev=tapnet,addr=0.0
> 
> (qemu) info network
> virtio-net-pci.0:
> index=0,type=nic,model=virtio-net-pci,macaddr=52:54:00:12:34:56
>  \ tapnet: index=0,type=tap,ifname=tap0,script=qemu-ifup.sh,downscript=/etc/qemu-ifdown
> 
> 
> I get very different virtio0 device features and speed (70MB/s vs.
> 700MB/s). I guess somehow the "hub 0" is responsible but there is no
> way to avoid it.
> 
> Is there any way to speed up the virtio-net using the old -net interface?

Not at the moment. Why do you want to use it?

> 
> On 06/03/13 21:31, Michael S. Tsirkin wrote:
> >On Wed, Mar 06, 2013 at 09:57:40AM +1100, Alexey Kardashevskiy wrote:
> >>On 06/03/13 01:23, Michael S. Tsirkin wrote:
> >>>On Wed, Mar 06, 2013 at 12:21:47AM +1100, Alexey Kardashevskiy wrote:
> >>>>On 05/03/13 23:56, Michael S. Tsirkin wrote:
> >>>>>>The patch f56a12475ff1b8aa61210d08522c3c8aaf0e2648 "vhost: backend
> >>>>>>masking support" breaks virtio-net + vhost=on on PPC64 platform.
> >>>>>>
> >>>>>>The problem command line is:
> >>>>>>1) -netdev tap,id=tapnet,ifname=tap0,script=qemu-ifup.sh,vhost=on \
> >>>>>>-device virtio-net-pci,netdev=tapnet,addr=0.0 \
> >>>>>
> >>>>>I think the issue is irqfd in not supported on kvm ppc.
> >>>>
> >>>>How can I make sure this is the case? Some work has been done there
> >>>>recently but midnight is quite late to figure this out :)
> >>>
> >>>Look in virtio_pci_set_guest_notifiers, what is the
> >>>value of with_irqfd?
> >>>   bool with_irqfd = msix_enabled(&proxy->pci_dev) &&
> >>>         kvm_msi_via_irqfd_enabled();
> >>>
> >>>Also check what each of the values in the expression above is.
> >>
> >>Yes, ppc does not have irqfd as kvm_msi_via_irqfd_enabled() returned "false".
> >>
> >>>>>Could you please check this:
> >>>>>
> >>>>>+        /* If guest supports masking, set up irqfd now.
> >>>>>+         * Otherwise, delay until unmasked in the frontend.
> >>>>>+         */
> >>>>>+        if (proxy->vdev->guest_notifier_mask) {
> >>>>>+            ret = kvm_virtio_pci_irqfd_use(proxy, queue_no, vector);
> >>>>>+            if (ret < 0) {
> >>>>>+                kvm_virtio_pci_vq_vector_release(proxy, vector);
> >>>>>+                goto undo;
> >>>>>+            }
> >>>>>+        }
> >>>>>
> >>>>>
> >>>>>Could you please add a printf before "undo" and check whether the
> >>>>>error path above is triggered?
> >>>>
> >>>>
> >>>>Checked, it is not triggered.
> >>>>
> >>>>
> >>>>--
> >>>>Alexey
> >>>
> >>>I think I get it.
> >>>Does the following help (probably not the right thing to do, but just
> >>>for testing):
> >>
> >>
> >>It did not compile (no "queue_no") :) I changed it a bit and now
> >>vhost=on works fine:
> >>
> >>diff --git a/hw/virtio-pci.c b/hw/virtio-pci.c
> >>index a869f53..df1e443 100644
> >>--- a/hw/virtio-pci.c
> >>+++ b/hw/virtio-pci.c
> >>@@ -798,6 +798,10 @@ static int
> >>virtio_pci_set_guest_notifiers(DeviceState *d, int nvqs, bool
> >>assign)
> >>          if (r < 0) {
> >>              goto assign_error;
> >>          }
> >>+
> >>+        if (!with_irqfd && proxy->vdev->guest_notifier_mask) {
> >>+            proxy->vdev->guest_notifier_mask(proxy->vdev, n, !assign);
> >>+        }
> >>      }
> >>
> >>      /* Must set vector notifier after guest notifier has been assigned */
> >>
> >>
> >
> >I see, OK, the issue is that vhost now starts in a masked state
> >and no one unmasks it. While the patch will work I think,
> >it does not benefit from backend masking, the right thing
> >to do is to add mask notifiers, like what the irqfd path does.
> >
> >Will look into this, thanks.
> >
> 
> 
> -- 
> Alexey