PCI Passthrough base
PCI Passthrough is used to use pci devices in VMs without any virtualization.
This blueprint will provide base integration of PCI passthrough with OpenStack based on PCI labels.
It will include:
1) Mechanism to specify available PCI devices on host through nova.conf
2) Mechanism to store PCI devices state in DB
3) Specific InstanceTypeExt
4) Scheduling based on PCI labels
PCI label is high level abstract name of device, all operations such as:
a) scheduling,
b) specifying required by instance_types devices
c) attaching pci device to instance # I mean work in DB, not virt layer
are make with PCI lables, not PCI addresses.
Using PCI lables makes this approach very flexible.
Why labels and not device address?
For example if you have 10 nodes, on each SR-IOV then each node will have 7 pci addresses.
So if all these addresses on all nodes have one logical name, for example "with_eth".
Then it is pretty simple to organize:
1) scheduling // get all nodes that have available devices with this label
2) specifying what devices you need
3) attaching devices
How to use it (base use case):
1) Specify on all compute nodes available for passthrough PCI devices:
add to nova.conf:
pci_passthrough
# This will create/update db records for PCI devices on host (when we run nova compue)
2) Add to any InstanceType extra specs {'pci_passthrou
3) Create Instance with this InstanceType, it will create VM with device with address "xxxx:xx:xx.x"
Blueprint information
- Status:
- Complete
- Approver:
- Russell Bryant
- Priority:
- Low
- Drafter:
- Boris Pavlovic
- Direction:
- Approved
- Assignee:
- jiang, yunhong
- Definition:
- Approved
- Series goal:
- Accepted for havana
- Implementation:
-
Implemented
- Milestone target:
-
2013.2
- Started by
- Boris Pavlovic
- Completed by
- Russell Bryant
Related branches
Related bugs
Sprints
Whiteboard
UML diagrams of changes + Code will be soon. -- boris-42
----
Does this blueprint answers the case where PCI device is intended to be used as Network interface (connect to certain virtual network)? How it guarantees that chosen device has access to provider network that Virtual Network is defined on?
----
Nope, there is no fine grade work here. This is base for PCI passthrough, that allows only:
1) specifying devices in nova.conf that are able for pci passthrough
2) abstract from the address to labels (to be able to make "the same" devices on different hostes)
3) guarantee that two VMs won't be able to get on device in the same time
4) scheduling by labels
5) specifying required labels (device with this labels should be passed through) instance_
----
Should the admin configure an entry {"label": "some_name", "address": "xxxx:xx:xx.x"} for each address? It can be tedious if one has more than a dozen. -- itzikb1
----
There will be two different ways to specify devices:
a) manually
b) auto-discovering
So you will be able to chose the method. -- boris-42
----
How will auto-discovering work? -- itzikb1
----
please check https:/
----
Gerrit topic: https:/
Addressed by: https:/
PCI passthrough DB Layer
Addressed by: https:/
PCI passthrough Conductor layer
Addressed by: https:/
PCI passthrough Utils layer
Addressed by: https:/
PCI passthrough Scheduler layer
Addressed by: https:/
PCI passthrough Compute layer
Gerrit topic: https:/
Addressed by: https:/
PCI passthrough DB Layer
Addressed by: https:/
PCI passthrough Utils layer
Gerrit topic: https:/
Gerrit topic: https:/
Addressed by: https:/
Add PCI device object support
Addressed by: https:/
Add scheduler support for PCI passthrough
Addressed by: https:/
Add PCI device filters support
Addressed by: https:/
PCI devices resource tracker
Addressed by: https:/
Add PCI device tracker to compute resource tracker
Addressed by: https:/
PCI device auto discover
Addressed by: https:/
PCI alias support
Addressed by: https:/
PCI utils and exceptions
Addressed by: https:/
Pci Device DB support
Addressed by: https:/
Add PCI stats
Work Items
Dependency tree
* Blueprints in grey have been implemented.
