Il 12/12/2012 18:53, Andreas Färber ha scritto: > Am 12.12.2012 15:25, schrieb Peter Maydell: >> How will the PCI transport's PCI vendor/device/class IDs be >> set (a) when a virtio-blk backend is created and separately >> plugged into a virtio-pci transport (b) for the legacy >> virtio-pci-blk? [ideally the answer to (b) should be "in the >> same way as for (a)"] > > The obvious answer would be that PCI properties need to be set on the > PCI device, not an a VirtioDevice sitting on a virtio-bus.
Yes, but the question is *how*... if there will be no usable "-device virtio-pci", the value of this refactoring becomes a bit lower... Paolo > I.e., with the proposed refactoring we'd have on the virtio-bus: > > - VirtioDevice > + VirtioBlockDevice > + VirtioSCSIDevice - has-a scsi-bus > ... > > In turn that means that every VirtioDevice to be exposed as PCI device > to the guest needs it own PCIDevice exposing a private virtio-bus. > > - PCIDevice > - VirtioPCIDevice - has-a virtio-bus > + virtio-blk-pci - has-a VirtioBlockDevice on its virtio-bus > + virtio-scsi-pci - has-a VirtioSCSIDevice on its virtio-bus > ... > > This also happens to solve most of the migration compatibility pretty > nicely because the wapping PCI devices would be used almost as before, > some state may need to be forwarded to the VirtioDevice. > > Finally supplying a public device_initialize() or so would be helpful > for the latter since VMState cannot cross pointers IIRC. I'll look into > that part since inlining the old qdev functions cripples my Tegra work. > > Andreas >
