Libvirt provides management of both physical and virtual host devices (historically also referred to as node devices) like USB, PCI, SCSI, and network devices. This also includes various virtualization capabilities which the aforementioned devices provide for utilization, for example SR-IOV, NPIV, MDEV, DRM, etc.
The node device driver provides means to list and show details about host
devices (virsh nodedev-list
, virsh nodedev-info
,
and virsh nodedev-dumpxml
), which are generic and can be used
with all devices. It also provides the means to manage virtual devices.
Persistently-defined virtual devices are only supported for mediated
devices, while transient devices are supported by both mediated devices
and NPIV (more
info about NPIV)).
Persistent virtual devices are managed with
virsh nodedev-define
and virsh nodedev-undefine
.
Persistent devices can be configured to start manually or automatically
using virsh nodedev-autostart
. Inactive devices can be made
active with virsh nodedev-start
.
Transient virtual devices are started and stopped with the commands
virsh nodedev-create
and virsh nodedev-destroy
.
Devices on the host system are arranged in a tree-like hierarchy, with
the root node being called computer
. The node device driver
supports udev backend (HAL backend was removed in 6.8.0
).
Details of the XML format of a host device can be found here. Of particular interest is the
capability
element, which describes features supported by
the device. Some specific device types are addressed in more detail
below.
<device> <name>pci_0000_00_17_0</name> <path>/sys/devices/pci0000:00/0000:00:17.0</path> <parent>computer</parent> <driver> <name>ahci</name> </driver> <capability type='pci'> ... </capability> </device>
capability
type
attribute are pci
and
phys_function
(see SR-IOV below).
<device> <name>pci_0000_04_00_1</name> <path>/sys/devices/pci0000:00/0000:00:06.0/0000:04:00.1</path> <parent>pci_0000_00_06_0</parent> <driver> <name>igb</name> </driver> <capability type='pci'> <domain>0</domain> <bus>4</bus> <slot>0</slot> <function>1</function> <product id='0x10c9'>82576 Gigabit Network Connection</product> <vendor id='0x8086'>Intel Corporation</vendor> <iommuGroup number='15'> <address domain='0x0000' bus='0x04' slot='0x00' function='0x1'/> </iommuGroup> <numa node='0'/> <pci-express> <link validity='cap' port='1' speed='2.5' width='2'/> <link validity='sta' speed='2.5' width='2'/> </pci-express> </capability> </device>
The XML format for a PCI device stays the same for any further
capabilities it supports, a single nested <capability>
element will be included for each capability the device supports.
Single root input/output virtualization (SR-IOV) allows sharing of the PCIe resources by multiple virtual environments. That is achieved by slicing up a single full-featured physical resource called physical function (PF) into multiple devices called virtual functions (VFs) sharing their configuration with the underlying PF. Despite the SR-IOV specification, the amount of VFs that can be created on a PF varies among manufacturers.
Suppose the NIC above was also SR-IOV capable, it would
also include a nested
<capability>
element enumerating all virtual
functions available on the physical device (physical port) like in the
example below.
<capability type='pci'> ... <capability type='virt_functions' maxCount='7'> <address domain='0x0000' bus='0x04' slot='0x10' function='0x1'/> <address domain='0x0000' bus='0x04' slot='0x10' function='0x3'/> <address domain='0x0000' bus='0x04' slot='0x10' function='0x5'/> <address domain='0x0000' bus='0x04' slot='0x10' function='0x7'/> <address domain='0x0000' bus='0x04' slot='0x11' function='0x1'/> <address domain='0x0000' bus='0x04' slot='0x11' function='0x3'/> <address domain='0x0000' bus='0x04' slot='0x11' function='0x5'/> </capability> ... </capability>
A SR-IOV child device on the other hand, would then report its top level
capability type as a phys_function
instead:
<device> ... <capability type='phys_function'> <address domain='0x0000' bus='0x04' slot='0x00' function='0x0'/> </capability> ... </device>
A device capable of creating mediated devices will include a nested
capability mdev_types
which enumerates all supported mdev
types on the physical device, along with the type attributes available
through sysfs. A detailed description of the XML format for the
mdev_types
capability can be found
here.
The following example shows how we might represent an NVIDIA GPU device that supports mediated devices. See below for more information about mediated devices.
<device> ... <driver> <name>nvidia</name> </driver> <capability type='pci'> ... <capability type='mdev_types'> <type id='nvidia-11'> <name>GRID M60-0B</name> <deviceAPI>vfio-pci</deviceAPI> <availableInstances>16</availableInstances> </type> <!-- Here would come the rest of the available mdev types --> </capability> ... </capability> </device>
A device that exposes a PCI/PCIe VPD capability will include a nested
capability vpd
which presents data stored in the Vital Product
Data (VPD). VPD provides a device name and a number of other standard-defined
read-only fields (change level, manufacture id, part number, serial number) and
vendor-specific read-only fields. Additionally, if a device supports it,
read-write fields (asset tag, vendor-specific fields or system fields) may
also be present. The VPD capability is optional for PCI/PCIe devices and the
set of exposed fields may vary depending on a device. The XML format follows
the binary format described in "I.3. VPD Definitions" in PCI Local Bus (2.2+)
and the identical format in PCIe 4.0+. At the time of writing, the support for
exposing this capability is only present on Linux-based systems (kernel version
v2.6.26 is the first one to expose VPD via sysfs which Libvirt relies on).
Reading the VPD contents requires root privileges, therefore,
virsh nodedev-dumpxml
must be executed accordingly.
A description of the XML format for the vpd
capability can
be found here.
The following example shows a VPD representation for a device that exposes the VPD capability with read-only and read-write fields. Among other things, the VPD of this particular device includes a unique board serial number.
<device> <name>pci_0000_42_00_0</name> <capability type='pci'> <class>0x020000</class> <domain>0</domain> <bus>66</bus> <slot>0</slot> <function>0</function> <product id='0xa2d6'>MT42822 BlueField-2 integrated ConnectX-6 Dx network controller</product> <vendor id='0x15b3'>Mellanox Technologies</vendor> <capability type='virt_functions' maxCount='16'/> <capability type='vpd'> <name>BlueField-2 DPU 25GbE Dual-Port SFP56, Crypto Enabled, 16GB on-board DDR, 1GbE OOB management, Tall Bracket</name> <fields access='readonly'> <change_level>B1</change_level> <manufacture_id>foobar</manufacture_id> <part_number>MBF2H332A-AEEOT</part_number> <serial_number>MT2113X00000</serial_number> <vendor_field index='0'>PCIeGen4 x8</vendor_field> <vendor_field index='2'>MBF2H332A-AEEOT</vendor_field> <vendor_field index='3'>3c53d07eec484d8aab34dabd24fe575aa</vendor_field> <vendor_field index='A'>MLX:MN=MLNX:CSKU=V2:UUID=V3:PCI=V0:MODL=BF2H332A</vendor_field> </fields> <fields access='readwrite'> <asset_tag>fooasset</asset_tag> <vendor_field index='0'>vendorfield0</vendor_field> <vendor_field index='2'>vendorfield2</vendor_field> <vendor_field index='A'>vendorfieldA</vendor_field> <system_field index='B'>systemfieldB</system_field> <system_field index='0'>systemfield0</system_field> </fields> </capability> <iommuGroup number='65'> <address domain='0x0000' bus='0x42' slot='0x00' function='0x0'/> </iommuGroup> <numa node='0'/> <pci-express> <link validity='cap' port='0' speed='16' width='8'/> <link validity='sta' speed='8' width='8'/> </pci-express> </capability> </device>
Mediated devices (Since 3.2.0) are software
devices defining resource allocation on the backing physical device which
in turn allows the parent physical device's resources to be divided into
several mediated devices, thus sharing the physical device's performance
among multiple guests. Unlike SR-IOV however, where a PCIe device appears
as multiple separate PCIe devices on the host's PCI bus, mediated devices
only appear on the mdev virtual bus. Therefore, no detach/reattach
procedure from/to the host driver procedure is involved even though
mediated devices are used in a direct device assignment manner. A
detailed description of the XML format for the mdev
capability can be found here.
<device> <name>mdev_4b20d080_1b54_4048_85b3_a6a62d165c01</name> <path>/sys/devices/pci0000:00/0000:00:02.0/4b20d080-1b54-4048-85b3-a6a62d165c01</path> <parent>pci_0000_06_00_0</parent> <driver> <name>vfio_mdev</name> </driver> <capability type='mdev'> <type id='nvidia-11'/> <uuid>4b20d080-1b54-4048-85b3-a6a62d165c01</uuid> <iommuGroup number='12'/> </capability> </device>
The support of mediated device's framework in libvirt's node device driver covers the following features:
Because mediated devices are instantiated from vendor specific templates, simply called 'types', information describing these types is contained within the parent device's capabilities (see the example in PCI host devices). To list all devices capable of creating mediated devices, the following command can be used.
$ virsh nodedev-list --cap mdev_types
To see the supported mediated device types on a specific physical device use the following:
$ virsh nodedev-dumpxml <device>
Before creating a mediated device, unbind the device from the respective device driver, eg. subchannel I/O driver for a CCW device. Then bind the device to the respective VFIO driver. For a CCW device, also unbind the corresponding subchannel of the CCW device from the subchannel I/O driver and then bind the subchannel (instead of the CCW device) to the vfio_ccw driver. The below example shows the unbinding and binding steps for a CCW device.
device="0.0.1234" subchannel="0.0.0123" echo $device > /sys/bus/ccw/devices/$device/driver/unbind echo $subchannel > /sys/bus/css/devices/$subchannel/driver/unbind echo $subchannel > /sys/bus/css/drivers/vfio_ccw/bind
To instantiate a transient mediated device, create an XML file representing the device. See above for information about the mediated device xml format.
$ virsh nodedev-create <xml-file> Node device '<device-name>' created from '<xml-file>'
If you would like to persistently define the device so that it will be
maintained across host reboots, use virsh nodedev-define
instead of nodedev-create
:
$ virsh nodedev-define <xml-file> Node device '<device-name>' defined from '<xml-file>'
To start an instance of this device definition, use the following command:
$ virsh nodedev-start <device-name>
Active mediated device instances can be stopped using virsh
nodedev-destroy
, and persistent device definitions can be removed
using virsh nodedev-undefine
.
If a mediated device is defined persistently, it can also be set to be automatically started whenever the host reboots or when the parent device becomes available. In order to autostart a mediated device, use the following command:
$ virsh nodedev-autostart <device-name>