Skip to content

Latest commit

 

History

History
443 lines (374 loc) · 19.4 KB

vm-configuration.md

File metadata and controls

443 lines (374 loc) · 19.4 KB

Virtual Machine Configuration

The configuration of virtual machines is one of the core tasks to accomplish with kubevirt. At the public kubernetes API layer, this will involve the user creating manifests that follow the published schema for one or more third party resource types. Internally, kubevirt will be using libvirt for managing QEMU, so there is a need to convert between the public resource descriptions and the libvirt XML document formats.

This document will only consider the configuration associated with a specific instance of a virtual machine. It will ignore the topic of 'templating' by which a higher level generalised config can be used as a cookie cutter for creating many guests.

Frontend / backend split

The libvirt domain XML schema defines configuration for two related, but separate concepts. The frontend configuration elements describe the hardware to expose for the virtual machine. This covers aspects like peripheral device models, PCI addresses, RAM/CPU setup, BIOS metadata and more. The backend configuration elements describe how the frontends are integrated with the host operating system services. This covers aspects like which file path is used for the virtual disk storage, which network interface is connected to the virtual interface, and performance tuning metrics such as CPU scheduler policies, disk and NIC I/O limits, or NUMA placement.

Generally the choice of backend config is independent of the choice of frontend config. For example, a guest can be given multiple virtual NUMA nodes, even if the host only has a single physical NUMA node. Some combinations may result in sub-optimal performance, but the choice is typically left to the administrator or management apps. In certain places, however, there are strict relationships. For example, the vhostuser network backend in QEMU requires uses of a virtio-net frontend device.

The frontend configuration is what is seen by the guest, and for some guests it is important that this does not change after initial provisioning. For example, if the PCI topology changes, Windows may no longer be able to determine which disks to use, or it may require license key reactivation. Changing the backend configuration is totally transparent to the guest (aside from performance differences inherent in different backend choices).

VM configuration diagram

Portability issues

The backend configuration information in the libvirt XML schema for configuring a virtual machine is a representation that is focused on deployment on a specific host. It is explicitly not a format that is guaranteed portable across all hosts in a clusters.

Similarly, the frontend configuration information is focused on deployment of a specific guest instance. It is not a format that can be directly used as a template for creating many guests, without making certain changes, such as MAC addresses in NICs, BIOS UUIDs, disk serial numbers and so on.

If a guest is deployed on a particular host, and then needs moving to a different host, there is a non-negligible chance that changes will be required in backend config of the XML configuration. For example, the ports currently used for VNC may not be available on the new host, or the storage devices that need to be used may require a different /dev/sdNN device path.

There are ways to minimise these differences, to increase portability between hosts. For example, instead of hardcoding a VNC port, allow libvirt to perform automatic allocation, or instead of using an unstable /dev/sdNN path, Linux provides /dev/disk/by-path/XXX symlinks which might be stable across hosts.

Security issues

There are many security features in libvirt, however, they are primarily focused around protection from malicious or misbehaving guests. For example, sVirt will stop guests attacking each other, or the host. NWFilter will stop guests doing IP/MAC/ARP spoofing on the network.

There is a fine grained access control mechanism for the host facing management API, however, the checks are performed against the (subject, object, action) tuple. So while you can block a user or application from being able to invoke the 'virDomainDefineXML' action, if you allow this API, you cannot then control what information is passed in its parameters. For example, when configuring disks in the domain XML, arbitrary host paths can be specified as the backing storage for the virtual disk. Libvirt will happily let /etc/passwd be specified as the disk backing store. When libvirt starts the QEMU instance, it will then tell sVirt to grant access to the /etc/passwd file.

IOW, any application which can invoke the virDomainDefineXML method on the libvirtd daemon has a privilege level which is equivalent to that of libvirtd itself, which is root. Thus applications calling virDomainDefineXML (and many other APIs) which accept configuration input from untrusted users, to feed into an XML document, must be very careful in their data handling.

Config states

Naively one might expect a virtual machine to have a single document describing its configuration. In practice, this is unreasonably restrictive.

The configuration provided by a tenant or management application will typically not express every single frontend or backend setting. When the XML document is provided to libvirt, it will fill in any important missing pieces of information in an attempt to fully represent the virtual hardware configuration and backend integration. For example, if the tenant did not supply PCI addresses, libvirt will assign them prior to starting QEMU. Even when libvirt fills in missing data this is only ever libvirt's best attempt at a complete virtual hardware configuration. For example, for ISA serial ports, libvirt does not record the hardware I/O port associated with the serial port, since that is not currently a configurable option & thus can be assumed stable without being explicitly represented. A future libvirt or QEMU, may allow I/O port configuration, and thus when upgrading to newer libvirt, the XML document previously considered to be complete may have yet more defaults filled in.

Applications that wish to be able to start the guest multiple times, while guaranteeing a stable hardware ABI, need to capture the augmented configuration and use it on subsequent boot attempts. Depending on which libvirt APIs are used, the augmented configuration can be obtained without starting the guest, or at the same time as starting a guest.

When a guest is started, some aspects of backend configuration may be dynamically filled in. For example, an automatically chosen VNC port will be reported in the XML document. This is information that an application will require for interaction with the guest for as long as it is running, but once shutoff the information can be discarded.

Running guests can have a variety of configuration changes made to them, in particular, hotplugging & unplugging of virtual devices such as disks, NICs, etc. These operations are not all guaranteed to be synchronous, nor are they even guaranteed to complete at all. For example, unplugging a PCI device requires co-operation from the guest OS. A guest that is still running the BIOS, or early boot up, may never honour the unplug request. Malicious or buggy guests can ignore a request forever too, or take an unexpectedly long time to complete it. There is thus a distinction between the configuration that a running guests currently has, vs what it is desired to have, which is only resolved when the unplug operation completes.

If disks and memory snapshots are taken of a guest, it is necessary to record the configuration associated with the guest at that time. When the snapshot is later activated, the guest must be set to use this original configuration, which may have different frontend device hardware from the current state. If the devices have changed since the snapshot was taken, the management application may not be able to honour the original virtual hardware configuration if the tenant user no longer has permission to connect the original devices to suitable backend services on the host. For example, if a NIC was unplugged since the snapshot was taken, even if the snapshot configuration has kept track of the existence of this NIC, the tenant may not be permitted to plug it into any host network. A snapshot can be taken regardless of whether a guest is running or inactive. The latter case, the snapshot would only cover disks, while in the format case, the snapshot would cover disks, and optionally memory too.

Bearing all this in mind there can be multiple configurations associated with a single instance of a virtual machine.

  1. Initial minimal tenant provided configuration
  2. Expanded configuration for inactive VM
  3. Expanded configuration for running VM reflecting desired state
  4. Expanded configuration for running VM reflecting current state
  5. Zero or many snapshots of the config for either inactive or running VM

An application may choose not to track all of these configurations explicitly. For example, by converting the minimal configuration, to the expanded inactive configuration, an application is fixing the virtual machine to the best practice at a given point in time. This is good if a stable hardware ABI is desired, but is not required. Some applications may instead prefer to always have the latest best practice applied to a guest each time it boots, and let the guest adapt to changes dynamically. Similarly an application tracking the inactive configuration of the guest, has a choice about how to handle updates to reflect hotplug/unplug operations. An unplug operation could be applied to the inactive configuration unconditionally, without waiting for completion on the running VM. This would ensure that the device goes away on the next cold boot, even if the guest OS ignores the hotunplug.

Additional document formats

The libvirt domain XML format is just one of many XML formats managed via the libvirt APIs. Many of the additional XML formats are used in conjunction with the domain XML, when managing individual virtual machines.

The secret APIs in libvirt and their associated secret XML format provide a way to push security sensitive credentials into libvirt. This covers items such as network service client login passwords to be used by QEMU, credentials to be required for network servers run by QEMU (such as SPICE) and disk encryption keys or passphrases. These are not included in the domain XML document since these documents are widely shared and logged without concern for security sensitive data. Thus any part of the domain XML configuration that requires security sensitive data will instead provide the UUID of a libvirt secret object. The sensitive data will be provided to the secret object ahead of time via a dedicated API.

The nwfilter APIs in libvirt and their associated XML format provide a way to define firewall rules to filter traffic that flows over guest virtual network interfaces. These can be used for a wide variety of tasks, but most common is their use to prevent IP, MAC and ARP spoofing attacks where one guest tries to impersonate another guest to captures its traffic or inflict a denial of service.

The node device APIs in libvirt and their associated XML format provide a way to dynamically create new host devices for use by virtual machines. For example they allow to create virtual GPU devices, or NPIV virtual SCSI host adapters which can then be assigned to the guest.

Design implications

The key take away from the analysis above is that kubevirt needs to maintain a strict separation between the libvirt XML document format, and the format used in the public APIs. The tenant users who are specifying the virtual machine configuration cannot be permitted to have direct control over most aspects of backend configuration. This is because they are both untrusted users, and do not have knowledge about the compute hosts which is required in order to make the correct choices.

The public k8s virtual machine configuration must allow users to fully specify the frontend virtual hardware, in order to enable a stable ABI to be attained.

For the backend configuration, some parts can be automatically determined by kubevirt without requiring any user specification. For example, there is no need for the user to specify what VNC/SPICE port to use, or where to find the UEFI BIOS files. It is sufficient to know that VNC or UEFI must be enabled. Other areas of backend configuration will require the user to specify resources in an abstracted manner. For example, instead of providing a file path for a disk, the tenant would provide the name of a PersistentVolumeClaim resource, or alternative k8s API resource. Further areas of backend configuration will require users to specify desired goals, which kubevirt then translates into specific configuration information. For example, the tenant may request exclusive use of CPUs, and kubevirt has to decide which host CPUs to bind virtual CPUs to, potentially impacting NUMA node placement of the guest as a result.

Given a k8s API resource describing the virtual machine, it will be possible to generate a libvirt domain XML document from scratch. The particular API usage scenario, however, may affect what information is provided in the domain XML. The domain XML generated at the time a guest is started, will be the most fully specified XML in terms of both frontend and backend configuration. If domain XML is generated in isolation, however, while the frontend can be fully specified, the backend configuration may remain partially specified. For example, it would be possible to expand a PVC to a backend disk config, but without knowledge of a specific host to run the guest, the domain XML cannot contain information about CPU pinning.

Going in the reverse direction is more difficult in general. The conversion from the high level k8s resource to the domain XML format is a lossy process. So given a domain XML document in isolation, it is not possible to generate a full k8s resource description of the virtual machine. At most it can contain frontend information. Translating the backend information in the domain XML, to the backend information required in the k8s resource is impractical or impossible in many cases. For example, there's no practical way to determine that a disk path on a particular host corresponds to a specific PVC. There is none the less a need for a conversion from domain XML to the k8s resource, in order to preserve the virtual hardware information for subsequent boots. Rather than trying to perform this conversion in isolation, however, it is more practical to take an initial k8s resource configuration (which contains all the high level backend information) and then augment the frontend configuration data with any missing pieces.

A further complication is that a single k8s API resource, may end up triggering the generation of multiple libvirt XML documents, the domain XML and associated documents for secrets, network filters and node devices. So there is not a strict 1-1 mapping between the k8s API resource representation and the libvirt low level representation.

Mapping rules for mapping libvirt XML to the KubeVirt API

In order to support libvirt features, we tend to expose the libvirt API as much as possible in a 1:1 fashion. However, there are some differences which need to be taken into account:

  1. Our API is json/yaml and not xml.
  2. Host specifics are not exposed on the API.
  3. We want to make use of the golang xml/yaml/json mappers as much as possible.
  4. Ordering of devices can be important (e.g. for pci address assignment).
  5. KubeVirt also delivers specialized device types (mostly for network and storage), which need to be fit into the existing structure.

Devices

All devices need to be ordered the same way libvirt internally orders them

Libvirt sorts devices into buckets based on their types. This means that if no PCI addresses are specified by hand, libvirt will always assign lower PCI addresses to disks than to video devices. No matter in which order they appear in the XML. The order within devices of the same type is kept.

The buckets are arranged in the following order:

emulator
disk
controller
lease
filesystem
interface
smartcard
serial
parallel
console
channel
input
tpm
graphics
sound
video
hostdev
redirdev
redirfilter
hub
watchdog
memballoon
rng
nvram
panic
shmem
memory
iommu

In order to avoid surprises we need to keep that order also on the API level.

Use optional structs to represent different sources or targets

By having an inlined wrapper-struct, containing optional structs corresponding to different types, we get a reasonable usable API and can maintain explicit mapping withing the code.

Given a Disk which can have different DiskSources

type Disk struct {
    Name string `json:"name"`
    DiskSource `json:"diskSource,inline"`
}

type DiskSource struct {
   ISCSI *ISCSIDiskSource `json"iscsi,omitempty"`
   NoCloud *NoCloudDiskSource `json"noCloud,omitempty"`
}

it will map to

disk:
  name: mydisk
  diskSource:
    iscsi:
      blub: bla
disk:
  name: mydisk
  diskSource:
    noCloud:
      blub: bla

Device Naming

Every device has a mandatory name field. Amongst other benefits, this allows easy tracking of the devices for management applications and users, simplifies declarative hotplug and allows VirtualMachinePresets to detect conflicts based on device names.

Specification changes

Only the user changes the specification spec of a VirtualMachine. There exists no back-propagation of any defaults or devices which are added by libvirt on the node. A specification reflects the requirement from the user and will only be changed by the system by the apiserver on creation time or by initializers. After the VirtualMachine is accepted by the system and was processed by all initializers, controllers will ever only change status and metadata fields. VirtualMachinePresets are an example of an initializer which can manipulate the spec, before the VirtualMachine is fully accepted by the system.

Devices inherited from the machine type

They are normally not visible on the VirtualMachine specification. If there will be the need to represent them, then they need to be represented in the status.

Mapping decisions for specific device types

Disks

disk, lun, floppy and cdrom are different structs within devices.disks array. They soley contain the guest side information of the disk. Each devices.disks entry continas one of them and volumeName attribute, which references a named entry in spec.volumes:

spec:
  domain:
    resources:
      initial:
        memory: 8Mi
    devices:
      disks:
      - name: disk0
        disk:
          dev: vda
        volumeName: volume0
      - name: cdrom0
        cdrom:
          dev: vdb
          readonly: true
          tray: open
        volumeName: volume1
      - name: floppy0
        floppy:
          dev: vdc
          readonly: true
          tray: open
        volumeName: volume2
      - name: lun0
        lun:
          dev: vdd
          readonly: true
        volumeName: volume3
  volumes:
  - name: volume0
    registryDisk:
      image: test/image
  - name: volume1
    cloudInitNoCloud:
      secretRef:
        name: testsecret
  - name: volume2
    iscsi:
      targetPortal: '1234'
      iqn: ''
      lun: 0
      secretRef:
        name: testsecret
  - name: volume3
    persistentVolumeClaim:
      claimName: testclaim