On Wed, Jun 04, 2025 at 03:48:40PM +0200, Cédric Le Goater wrote: > > I don't see any advantage to making this a class attribute. I looked for > > examples > > of using such attributes for vfio to configure pci, and found very little. > > It > > sounds like overkill since vfio already sets and gets PCIDevice members > > directly > > in many places. > > > > I defined skip_reset_on_cpr based on this existing example: > > > > vfio_instance_init() > > pci_dev->cap_present |= QEMU_PCI_CAP_EXPRESS > > pci_dev->cap_present can be modified at realize time. skip_reset_on_cpr > is a constant, for which a class attribute are more appropriate. > This is minor. > > Michael, > > Are you ok with the 'skip_reset_on_cpr' bool ?
Generally yes, but maybe cap_present bit is even cleaner? vfio already pokes at it, and we have history of encoding quirks there, see QEMU_PCIE_LNKSTA_DLLLA_BITNR for example. > > > I wonder if the resettable interface, and more specifically the > > > RESET_TYPE_SNAPSHOT_LOAD type, might be useful. Have you explored > > > this alternative ? > > > > RESET_TYPE_SNAPSHOT_LOAD (or a new type such as RESET_TYPE_CPR) would skip > > reset for all devices, but we only skip for vfio_pci. All other devices > > (including virtio) save and restore state using standard migration vmstate, > > and must call reset. > OK. > > C.
