KVM Device Failover
Support for transparent failover between an assigned and an emulated device, allows enabling the migration and overcommit dynamically, while still gaining the performance benefits of device assignment and without disrupting the guest operation.
- Name: Michael S. Tsirkin
- Email: email@example.com
- Name: Gal Hammer
- Email: firstname.lastname@example.org
- Name: Cole Robinson
- Email: email@example.com
- Name: Laine Stump
- Email: firstname.lastname@example.org
- Targeted release: Fedora 19
- Last updated: Jan 29, 2013
- Percentage of completion: 50%
For virtual machines, device assignment is the best option for performance. However, when a device is assigned to a VM, both migration and memory overcommit are currently disabled.
This feature aims at removing the performance/features tradeoff, by switching to an emulated device in a way that is almost transparent to users, for configurations where both host and guest are Fedora.
Fedora should detect that the emulated device serves as a failover for the assigned device. When requested by the hypervisor, it will stop and eject the assigned device, switching to failover. After this point, migration and memory overcommit are possible, while device configuration is preserved. Once e.g. migration completes, the reverse switch can take place.
Thus the device is controlled by: - before migration: device specific driver loaded in guest - during migration: driver loaded in host, virtio or emulated
device driver loaded in guest
- after migration: device specific driver loaded in guest
At the kernel level, for networking, this can be done by and creating a bond in a failover configuration, and for storage, using multipath, on top of both the assigned and the emulated device.
Benefit to Fedora
Complex virt setups now have less operational caveats, which makes things simpler for users.
Work left to do:
- kvm needs to be extended to notify the guest that the two devices are setup in a fallback configuration
- For networking, udev and/or network manager needs to be extended to detect this and setup bonding
- For storage, need to setup device-mapper-multipath to autodetect this configuration
- libvirt has to be extended to specify this configuration
- libvirt has to be extended to request failover, and ack on guest ack of the failover
- above covers linux guests
if possible, guest agent for windows should be extended to add this support in windows guests as well
How To Test
Two systems with device assignment (IOMMU) support are required to test this feature. To test the feature, specify an assigned device, start guest and migrate.
XXX: Explicit test steps here for test day
User will see that they can specify an assigned network or storage device and still migrate the guest seamlessly.
For networking, https://fedoraproject.org/wiki/Features/NetworkManagerBonding
None necessary, revert to previous release behaviour.
Links to related upstream documentation:
http://www.linux-kvm.org/page/Hotadd_pci_devices https://access.redhat.com/knowledge/docs/en-US/Red_Hat_Enterprise_Linux/6/html-single/DM_Multipath/index.html http://unixfoo.blogspot.com/2007/10/yet-to-add.html
- KVM guests with assigned host devices can now be migrated across hosts. The assigned device will be replaced during migration with an emulated device in a transparent manner.