On Thu, Sep 08, 2022 at 15:16:56 +0800, Yalan Zhang wrote:
> Hi Laine,
>
> As for the hot-unplug behavior, I have one more question about it, could
> you please help to confirm?
>
> unplugging a PCI device properly requires cooperation from the guest OS.
> > If the guest OS isn't running yet, the unplug won't complete, so qemu
> > (and libvirt) still show the device as plugged into the guest.
> >
> > virsh reports success on the unplug because unplugging a device is done
> > asynchronously - the "success" means "libvirt successfully told qemu to
> > unplug the device, qemu has told the virtual machine to unplug the
> > device, and is waiting for acknowledgment from the virtual machine that
> > the guest has completed removal". At some later time the guest OS may
> > complete its part of the unplug; when that happens, qemu will get a
> > notification and will send an event to libvirt - at that time the device
> > will be removed from libvirt's list of devices.
> >
> > tl;dr - this is all expected.
> >
>
> The question is that, when I unplug it during boot, the virsh cmd will
> succeed but the interface still exists, which is expected.
> After the vm boot successfully, the guest OS will *not* complete this
> removal. When I tried to detach it again, it reported that the device was
> in the process of unplugging.
> Is this acceptable?
>
> # virsh detach-interface rhel_new network 52:54:00:36:a8:d4
> Interface detached successfully
> # virsh domiflist rhel_new
>  Interface   Type      Source    Model    MAC
> -------------------------------------------------------------
>  vnet4       network   default   virtio   52:54:00:36:a8:d4
>
> # virsh detach-interface rhel_new network 52:54:00:36:a8:d4
> error: Failed to detach interface
> error: internal error: unable to execute QEMU command 'device_del': Device
> net0 is already in the process of unplug

The same problem was already reported for disks:

https://bugzilla.redhat.com/show_bug.cgi?id=2087047

https://gitlab.com/libvirt/libvirt/-/issues/309

The main problem is that qemu doesn't re-send the request to unplug the
device and rather reprots an eroror. At the same time the guest OS
doesn't notice it any more, so the unplug can't be finished until the VM
is rebooted.