Il 09/12/2013 22:15, Igor Mammedov ha scritto: >> > Check is idempotent, and can be called before realize makes any change >> > (it could also be called after the device is added to >> > /machine/unattached, it's not a big difference). >> > >> > Plug is called after realize. > PCIE case: "check" before realize will work since code that does check depends > only on hotplug device (i.e. PCIE slot) and do not access not yet realized > device at all. > > however > SHPC case: check code access pci_slot that is derived from PCIDevice.devfn, > which in turn could be initialized in realize() (see pci_qdev_init() devfn > auto allocation). So it's not possible to call check before realize() it > should be called from realize(). > > Perhaps other hotplug buses/devices have similar limitations, where it's not > fine to access device state from outside before calling it's realize(), so it > should be some post_realize() hook then to make it generic which leads to the > following: > if ->plug() called after realize() fails, all we need to do is to > fail "realize" property setter. That should cause > qdev_device_add() -> object_unparent() -> device_unparent() -> unrealize() > doing all necessary cleanup.
If you can make it work, that'd be great. Otherwise, let's do it later, but please make a wiki page with a todo list. We already have too many items on the same critical path (hotplug, memdev, NUMA,...). Paolo
