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/qemu-devel@nongnu.org/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.
