Re: virtio-vsock requires 'disable-legacy=on' in QEMU 5.1

[Cornelia Huck]

On Thu, 13 Aug 2020 12:24:30 +0200
Stefano Garzarella <sgarzare@redhat.com> wrote:

> On Thu, Aug 13, 2020 at 11:28:20AM +0200, Cornelia Huck wrote:
> > On Thu, 13 Aug 2020 11:16:56 +0200
> > Stefano Garzarella <sgarzare@redhat.com> wrote:
> >   
> > > Hi,
> > > 
> > > Qinghua discovered that virtio-vsock-pci requires 'disable-legacy=on' in
> > > QEMU 5.1:
> > >     $ ./qemu-system-x86_64 ... -device vhost-vsock-pci,guest-cid=5
> > >     qemu-system-x86_64: -device vhost-vsock-pci,guest-cid=5:
> > >     device is modern-only, use disable-legacy=on
> > > 
> > > Bisecting I found that this behaviour starts from this commit:
> > > 9b3a35ec82 ("virtio: verify that legacy support is not accidentally on")  
> > 
> > Oh, I had heard that from others already, was still trying to figure
> > out what to do.
> >   
> > > 
> > > IIUC virtio-vsock is modern-only, so I tried this patch and it works:
> > > 
> > > diff --git a/hw/virtio/vhost-user-vsock-pci.c 
> > > b/hw/virtio/vhost-user-vsock-pci.c
> > > index f4cf95873d..6e4cc874cd 100644
> > > --- a/hw/virtio/vhost-user-vsock-pci.c
> > > +++ b/hw/virtio/vhost-user-vsock-pci.c
> > > @@ -40,6 +40,7 @@ static void vhost_user_vsock_pci_realize(VirtIOPCIProxy 
> > > *vpci_dev, Error **errp)
> > >      VHostUserVSockPCI *dev = VHOST_USER_VSOCK_PCI(vpci_dev);
> > >      DeviceState *vdev = DEVICE(&dev->vdev);
> > > 
> > > +    virtio_pci_force_virtio_1(vpci_dev);
> > >      qdev_realize(vdev, BUS(&vpci_dev->bus), errp);
> > >  }
> > > 
> > > diff --git a/hw/virtio/vhost-vsock-pci.c b/hw/virtio/vhost-vsock-pci.c
> > > index a815278e69..f641b974e9 100644
> > > --- a/hw/virtio/vhost-vsock-pci.c
> > > +++ b/hw/virtio/vhost-vsock-pci.c
> > > @@ -44,6 +44,7 @@ static void vhost_vsock_pci_realize(VirtIOPCIProxy 
> > > *vpci_dev, Error **errp)
> > >      VHostVSockPCI *dev = VHOST_VSOCK_PCI(vpci_dev);
> > >      DeviceState *vdev = DEVICE(&dev->vdev);
> > > 
> > > +    virtio_pci_force_virtio_1(vpci_dev);
> > >      qdev_realize(vdev, BUS(&vpci_dev->bus), errp);
> > >  }
> > > 
> > > 
> > > Do you think this is the right approach or is there a better way to
> > > solve this issue?  
> > 
> > We basically have three possible ways to deal with this:
> > 
> > - Force it to modern (i.e., what you have been doing; would need the
> >   equivalent changes in ccw as well.)  
> 
> Oo, thanks for pointing out ccw!
> I don't know ccw well, in this case should we set dev->max_rev to 1 or 2
> to force to modern?

No, ->max_rev is the wrong side of the limit :) You want

    ccw_dev->force_revision_1 = true;

in _instance_init() (see e.g. virtio-ccw-gpu.c).

> 
> >   Pro: looks like the cleanest approach.
> >   Con: not sure if we would need backwards compatibility support,
> >   which looks hairy.  
> 
> Not sure too.

Yes, I'm not sure at all how to handle user-specified values for
legacy/modern.

> 
> > - Add vsock to the list of devices with legacy support.
> >   Pro: Existing setups continue to work.
> >   Con: If vsock is really virtio-1-only, we still carry around
> >   possibly broken legacy support.  
> 
> I'm not sure it is virtio-1-only, but virtio-vsock was introduced in
> 2016, so I supposed it is modern-only.

Yes, I would guess so as well.

> 
> How can I verify that? Maybe forcing legacy mode and run some tests.

Probably yes. The likeliest area with issues is probably endianness, so
maybe with something big endian in the mix?

> 
> > - Do nothing, have users force legacy off. Bad idea, as ccw has no way
> >   to do that on the command line.
> > 
> > The first option is probably best.
> >  
> 
> Yeah, I agree with you!

Yes, it's really a pity we only noticed this after the release; this
was supposed to stop new devices with legacy support creeping in, not
to break existing command lines :(