Re: [Qemu-devel] [PATCH 0/4] add failover feature for assigned network devices

[Michael S. Tsirkin]

On Tue, Jun 11, 2019 at 11:42:54AM -0400, Laine Stump wrote:
> On 5/17/19 8:58 AM, Jens Freimann wrote:
> > This is another attempt at implementing the host side of the
> > net_failover concept
> > (https://www.kernel.org/doc/html/latest/networking/net_failover.html)
> > 
> > Changes since last RFC:
> > - work around circular dependency of commandline options. Just add
> >    failover=on to the virtio-net standby options and reference it from
> >    primary (vfio-pci) device with standby=<id>
> > - add patch 3/4 to allow migration of vfio-pci device when it is part of a
> >    failover pair, still disallow for all other devices
> > - add patch 4/4 to allow unplug of device during migrationm, make an
> >    exception for failover primary devices. I'd like feedback on how to
> >    solve this more elegant. I added a boolean to DeviceState, have it
> >    default to false for all devices except for primary devices.
> > - not tested yet with surprise removal
> > - I don't expect this to go in as it is, still needs more testing but
> >    I'd like to get feedback on above mentioned changes.
> > 
> > The general idea is that we have a pair of devices, a vfio-pci and a
> > emulated device. Before migration the vfio device is unplugged and data
> > flows to the emulated device, on the target side another vfio-pci device
> > is plugged in to take over the data-path. In the guest the net_failover
> > module will pair net devices with the same MAC address.
> > 
> > * In the first patch the infrastructure for hiding the device is added
> >    for the qbus and qdev APIs.
> > 
> > * In the second patch the virtio-net uses the API to defer adding the vfio
> >    device until the VIRTIO_NET_F_STANDBY feature is acked.
> > 
> > Previous discussion:
> >    RFC v1 https://patchwork.ozlabs.org/cover/989098/
> >    RFC v2 https://www.mail-archive.com/address@hidden/msg606906.html
> > 
> > To summarize concerns/feedback from previous discussion:
> > 1.- guest OS can reject or worse _delay_ unplug by any amount of time.
> >    Migration might get stuck for unpredictable time with unclear reason.
> >    This approach combines two tricky things, hot/unplug and migration.
> >    -> We can surprise-remove the PCI device and in QEMU we can do all
> >       necessary rollbacks transparent to management software. Will it be
> >       easy, probably not.
> > 2. PCI devices are a precious ressource. The primary device should never
> >    be added to QEMU if it won't be used by guest instead of hiding it in
> >    QEMU.
> >    -> We only hotplug the device when the standby feature bit was
> >       negotiated. We save the device cmdline options until we need it for
> >       qdev_device_add()
> >       Hiding a device can be a useful concept to model. For example a
> >       pci device in a powered-off slot could be marked as hidden until the 
> > slot is
> >       powered on (mst).
> > 3. Management layer software should handle this. Open Stack already has
> >    components/code to handle unplug/replug VFIO devices and metadata to
> >    provide to the guest for detecting which devices should be paired.
> >    -> An approach that includes all software from firmware to
> >       higher-level management software wasn't tried in the last years. This 
> > is
> >       an attempt to keep it simple and contained in QEMU as much as 
> > possible.
> > 4. Hotplugging a device and then making it part of a failover setup is
> >     not possible
> >    -> addressed by extending qdev hotplug functions to check for hidden
> >       attribute, so e.g. device_add can be used to plug a device.
> > 
> > 
> > I have tested this with a mlx5 NIC and was able to migrate the VM with
> > above mentioned workarounds for open problems.
> > 
> > Command line example:
> > 
> > qemu-system-x86_64 -enable-kvm -m 3072 -smp 3 \
> >          -machine q35,kernel-irqchip=split -cpu host   \
> >          -k fr   \
> >          -serial stdio   \
> >          -net none \
> >          -qmp unix:/tmp/qmp.socket,server,nowait \
> >          -monitor telnet:127.0.0.1:5555,server,nowait \
> >          -device 
> > pcie-root-port,id=root0,multifunction=on,chassis=0,addr=0xa \
> >          -device pcie-root-port,id=root1,bus=pcie.0,chassis=1 \
> >          -device pcie-root-port,id=root2,bus=pcie.0,chassis=2 \
> >          -netdev 
> > tap,script=/root/bin/bridge.sh,downscript=no,id=hostnet1,vhost=on \
> >          -device 
> > virtio-net-pci,netdev=hostnet1,id=net1,mac=52:54:00:6f:55:cc,bus=root2,failover=on
> >  \
> >          /root/rhel-guest-image-8.0-1781.x86_64.qcow2
> > 
> > Then the primary device can be hotplugged via
> >   (qemu) device_add vfio-pci,host=5e:00.2,id=hostdev0,bus=root1,standby=net1
> 
> 
> I guess this is the commandline on the migration destination, and as far as
> I understand from this example, on the destination we (meaning libvirt or
> higher level management application) must *not* include the assigned device
> on the qemu commandline, but must instead hotplug the device later after the
> guest CPUs have been restarted on the destination.
> 
> So if I'm understanding correctly, the idea is that on the migration source,
> the device may have been hotplugged, or may have been included when qemu was
> originally started. Then qemu automatically handles the unplug of the device
> on the source, but it seems qemu does nothing on the destination, leaving
> that up to libvirt or a higher layer to implement.

Good point. I don't see why it would not work just as well
with device present straight away.

Did I miss something?

I think Jens was just testing local machine migration
and of course you can only assign a device to 1 VM at a time.

> Then in order for this to work, libvirt (or OpenStack or oVirt or whoever)
> needs to understand that the device in the libvirt config (it will still be
> in the libvirt config, since from libvirt's POV it hasn't been unplugged):
> 
> 1) shouldn't be included in the qemu commandline on the destination,
> 
> 2) will almost surely need to be replaced with a different device on the
> destination (since it's almost certain that the destination won't have an
> available device at the same PCI address)
> 
> 3) will probably need to be unbinded from the VF net driver (does this need
> to happen before migration is finished? If we want to lower the probability
> of a failure after we're already committed to the migration, then I think we
> must, but libvirt isn't set up for that in any way).
> 
> 4) will need to be hotplugged after the migration has finished *and* after
> the guest CPUs have been restarted on the destination.
> 
> 
> While it will be possible to assure that there is a destination device, and
> to replace the old device with new in the config (and maybe, either with
> some major reworking of device assignment code, or offloading the
> responsibility to the management application(s), possible to re-bind the
> device to the vfio-pci driver), prior to marking the migration as
> "successful" (thus committing to running it on the destination), we can't
> say as much for actually assigning the device. So if the assignment fails,
> then what happens?
> 
> 
> So a few issues I see that will need to be solved by [someone] (apparently
> either libvirt or management):
> 
> a) there isn't anything in libvirt's XML grammar that allows us to signify a
> device that is "present in the config but shouldn't be included in the
> commandline"
> 
> b) someone will need to replace the device from the source with an
> equivalent device on the destination in the libvirt XML. There are other
> cases of management modifying the XML during migration (I think), but this
> does point out that putting the "auto-unplug code into qemu isn't turning
> this into a trivial
> 
> c) there is nothing in libvirt's migration logic that can cause a device to
> be re-binded to vfio-pci prior to completion of a migration. Unless this is
> added to libvirt (or the re-bind operation is passed off to the management
> application), we will need to live with the possibility that hotplugging the
> device will fail due to failed re-bind *after* we've committed to the
> migration.
> 
> d) once the guest CPUs are restarted on the destination, [someone] (libvirt
> or management) needs to hotplug the new device on the destination. (I'm
> guessing that a hotplug can only be done while the guest CPUs are running;
> correct me if this is wrong!)
> 
> This sounds like a lot of complexity for something that was supposed to be
> handled completely/transparently by qemu :-P.