Re: [Qemu-devel] [PATCH v4 04/14] pc: prepare for multi stage hotplug handlers

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

On Wed, Jun 13, 2018 at 09:37:54PM +0200, David Hildenbrand wrote:
> >>>       [ ... specific to machine_foo wiring ...]
> >>>
> >>>       virtio_mem_plug() {
> >>>          [... some machine specific wiring ...]
> >>>
> >>>          bus_hotplug_ctrl = qdev_get_bus_hotplug_handler()
> >>>          bus_hotplug_ctrl->plug(bus_hotplug_ctrl, dev)
> >>>
> >>>          [... some more machine specific wiring ...]
> >>>       }
> >>>
> >>>       [ ... specific to machine_foo wiring ...]
> >>>
> >>> i.e. device itself doesn't participate in attaching to external entities,
> >>> those entities (machine or bus controller virtio device is attached to)
> >>> do wiring on their own within their own domain.
> >>
> >> I am fine with this, but Michael asked if this ("virtio_mem_plug()")
> >> could go via new DeviceClass functions. Michael, would that also work
> >> for your use case?
> > 
> > It's not virtio specifically, I'm interested in how this will work for
> > PCI generally.  Right now we call do_pci_register_device which
> > immediately makes it guest visible.
> 
> So you're telling me that a PCI device exposes itself to the system in
> pci_qdev_realize() instead of letting a hotplug handler handle that? My
> assumption is that the PCI bridge hotplug handler handles this.

Well given how things work in qemu that's not exactly
the case. See below.

> What am
> I missing?
> 
> I can see that e.g. for a virtio device the realize call chain is
> 
> pci_qdev_realize() -> virtio_pci_realize() -> virtio_XXX__pci_realize ->
> virtio_XXX_realize()
> 
> If any realization in pci_qdev_realize() fails, we do a
> do_pci_unregister_device().
> 
> So if it is true what you're saying than we're already exposing
> partially realized (and possibly unrealized again) devices via PCI. I
> *guess* because we're holding the iothread mutex this is okay and
> actually not visible.

For now but we need ability to have separate new commands for
realize and plug, so we will drop the mutex.

> And we only seem to be sending events in the PCI
> bridge hotplug handlers, so my assumption is that this is fine.

For core PCI, it's mostly just this line:

    bus->devices[devfn] = pci_dev;

which makes it accessible to pci config cycles.

But failover also cares about vfio, which seems to set up
e.g. irqfs on realize.




> > 
> > 
> >>
> >> -- 
> >>
> >> Thanks,
> >>
> >> David / dhildenb
> 
> 
> -- 
> 
> Thanks,
> 
> David / dhildenb