Page tree
Skip to end of metadata
Go to start of metadata

Read intro below, and/or the preparation work at the bottom of this page.  Then look at actual draft spec working page.

The Automotive Virtualization Platform

To allow further success of hypervisor environments in automotive it is essential that all vendors are able to provide compelling guest runtime environments that make the usual automotive I/O devices available. The essential devices could be defined and agreed upon by the industry.  If such a set of devices and device features is defined a document can be crafted that defines these devices and their features that in their combination create a virtual platform.

PORTABLE

  • The virtual platform definition would allow the development of virtual machine guests that like appliances in the enterprise world, could be moved among different hypervisor systems without (minimal) modification.

SPECIFIED

  • A clear and detailed specification is required to achieve true portability, and to give real support to vendors. 

Hand-waving about some virtual architecture is not our goal.  The specification shall enable efficient reuse and collaborative progress among companies in this industry.  

REUSE

  • Specifications like these are best developed as open specifications and just like open-source code, it should stand on the shoulders of previous work. 

It should be built upon already published work, where VIRTIO is the most prominent we have found, but it is not enough.  It is clear that the automotive industry needs to do more and that it should (will) one way or another make sure that different initiatives combine into common agreements. 

The project group has discussed and found consensus on how that automotive virtual platform specification should be done:

For an automotive virtualized platform, VIRTIO is primarily a starting point because the defined devices enable running only basic system functionality but lack the multimedia and other devices often found in automotive environments.

Instead of developing additional domain or even vendor specific device frameworks and models a collaborative development could greatly reduce the individual development of project efforts and thus spur the adoption of hypervisor based environments in the automotive sector.

Additional efforts definitely need to be spent in the area of audio virtualization and the in virtualization of co-processors like DSPs, image processors and codec accelerators. Developing a standard model for the diverse buses found in cars like CAN or ethernet AVB could also be additional fields of investigation.

In the work done here (see table below) we categorize device types and ensure that each need has been thoroughly analyzed so that as an industry we can truly define the right specification, to be applicable in >95% of the typical situations.

About VIRTIO

The VIRTIO standard aims to provide a standardized interface and device models for device para-virtualization in hypervisor environments.

With development going back until 20XX the virtio device model was first introduced for the educational "lguest" hypervisor and became the default I/O virtulaization method used with qemu/KVM and recently the default model used by many cloud providers. The virtio devices have been partially standardized by the OASIS standardisation body in 2015 with the VIRTIO 1.0 specification which describes the transport layer and a limited set of device models.

The currently standardized device models are: network, block, console, entropy, memory ballooning and SCSI devices. Additionally to the formally standardized devices several additional devices exists as "implemented" devices such as GPU, input, 9pfs, vsock and crypto. Some of which are currently in the process of standardization.

Virtio relies on a dma-like memory model meaning that all allocations are handled by the "driver" part of the device and the "device" implementation can directly access the allocated buffers for reading or writing. This enables a resource saving and fast operating mode. Metadata is transported using so called virt-queues that resemble ring-buffers. Depending on the architecture used, different transport and device discovery modes are supported: PCI for x86, mmio for ARM and channel-IO for s390. These transports are geared toward the most efficient implementation per CPU architecture and allow for efficient implementations depending on the environment.

In recent years some hardware devices, like network controllers and NVMe based storage systems have evolved to be similar or compatible with the VIRTIO protocol, to allow hardware assisted I/O virtualization using para-virtualized device models.

VIRTIO Benefits

VIRTIO's main benefit for the automotive industry lies in it's sheer existence and operating system support. The fact that standardized device models exists allows for multiple compatible implementations of both driver and server parts of the system. The ubiquity of VIRTIO in cloud and enterprise virtualization makes drivers readily available in all major operating systems which keeps driver maintenance effort to a minimum.

Due to the driver defined memory allocation model, vendors can choose to limit the resource usage and define safety properties to their own requirements without impeding the standardized model and stay interoperable with existing device implementations.

The DMA-like nature of the devices allows for high-performant implementations that the easily compete with hardware assisted I/O virtualization models while still providing relative ease of implementation.

Many vendors of automotive grade hypervisors have already adopted virtio based devices into their system offerings due to the above mentioned reasons and the benefit of building upon these open source technologies has greatly improved the availability of commodity devices like network and block storage.

GENIVI and AGLs Role

(Opensynergy's proposal)

Vendor neutral industry bodies like GENVI and AGL can act a forum between hypervisor vendors, users and the hardware manufactures in a form that allows open collaboration in development and maintaining a standard platform definition.

The regular events can be used as occasions for interoperability testing and standard steering. The participation in the OASIS-VIRTIO committee can be delegated to such organization to voice the automotive industries concerns and advice in the technical committee.

GENIVI has in the past maintained domain specific APIs and standards can easily act as a body that makes sure the standard is not only maintained but also advanced as new technologies evolve.

AGL can provide the necessary collaboration with the upstream kernel project and the Linux Foundation which in-turn opens up for collaboration with the key industry players in cloud and enterprise computing.



Evaluation Process for existing specifications (e.g. VIRTIO)

For each topic:

    1. Discuss and write down the automotive requirements
    2. Read VIRTIO chapter
    3. Decide if VIRTIO is appropriate and complete for requirements (Gap Analysis)
    4. Write down what the industry needs to do to close the gap

Consider topics not yet listed (e.g. unique automotive requirements)

Note also:  Bottom of this page has a "brainstorm" list of criteria to consider.




Virtual Device Categories

The key challenge for defining a shared virtual platform definition is to first identify the various device driver types such a platform must provide, and to evaluate if existing work so far (e.g. VIRTIO) covers what the automotive industry needs: 


(Virtual) DeviceExplanation

Champion
+ interested people

Completeness / Applicability evaluation

(tick) (warning) (error)

Comments and discussion
Block StorageFlash/Disk/persistent storage

Kai


Possibly definition not enough since it is very generic. What happens below when the challenges

Kai working on evaluating these details. Let's investigate practical implementation of TRIM command, as example.


Network

Access to (shared) physical ethernet

and guest-to-guest communication

Nikola (AvB only)

(warning) Still open slot for general networking.



ConsoleText terminal inputGunnar(warning) WIP

Done initial browsing of the specification. Opinions still pending.

cryptoAccess to cryptographic services
(hardware accelerated)
Sang-bum

GPUGraphics hardwareMatti/DmitrySee GPU Summary, VIRTIO GPU Operation Highlights pagesSee GPU Summary page
InputTraditionally keyboard/mouse/etc
- for automotive = expanded?
Matti

Oct 2018: Input standard being discussed for inclusion in VIRTIO but accepted in principle. Let's wait a few weeks and look at proposal once in VIRTIO.

Mouse/touch events may need to remap coordinates in combined virtual systems but interface may still not be affected by this.

vsockCommunication between guest (VM)
and host (hypervisor)



9pfs9P = protocol to expose host (hypervisor)
file systems to the guest. FS=filesystem.
Gunnar

Use-cases: ?

Completeness: Protocol: (tick), VIRTIO spec: (warning) (see comment)

Need in Embedded/Automotive: (error) None? Can we find a use-case?

Applicability: (tick)(warning)(error) For what it does, seems ok. But might not be really needed and therefore "not applicable". Is there something else/more needed?

(warning) WIP

Essentially used for "shared folder" capability between host & guest, as in desktop (or maybe some server) usage. Its applicability to embedded hypervisor usage, in which the "host" is not really being used by itself) seems questionable. What's the use-case?

In VIRTIO spec: A PCI type device can indicate that it is going to use the 9P protocol. The specification also has 9P as a specific separate device type. Other than that, I found no further description of such a device type. The protocol is specified elsewhere and complemented by scattered information regarding the specific implementations (Xen, KVM, QEMU, ...)

The protocol seems proven and supposedly OK for what it does. Possibly more security features needed, depending on use-case. VIRTIO however seems to defer the definition completely to "somewhere else"? At least a reference to a canonical specification would seem appropriate.

It is a minimalistic network file-system protocol. It seems apt for the task. Other network protocols like NFS, SMB/SAMBA etc. would be too heavy. It feels a bit esoteric, and while "reinventing" is bad, in this simple case would not be the worst ever, if VIRTIO had defined something else. Flexibility and security features seem somewhat glossed over. There's basically only "fixed user" or "pass-through" for mapping ownership on files in guest/host.

Links: Virtio 1.0 spec : {PCI-9P, 9P device type}.
Kernel support: Xen/Linux 4.12+ FE driver
Xen implementation details

A note on its documentation/definition not being very precise

A set of man pages seemingly defining P9? intro, others

QEMU instruction how to set up a VirtFS (P9).
Example/info how to natively mount a 9P network filesystem,
Source code for 9pfs FUSE driver


Example how to set up in Linux system


vIOMMU

IOMMU coordinates of DMA devices'
connection to memory.

Dmitry

See IOMMU Summary page

Applicability:

(tick) Limit guest devices' scope to access system memory during DMA

(tick) Enable scatter-gather accesses due to remapping (DMA buffers do not need to be physically-contiguous)

(question) Nested virtualization. Any use-cases for automotive?
Group conclusion: Not needed, generally speaking
→ see comments

ARM is actively working on the specification, more features are coming.

Nested virtualization? The use of Linux Containers inside a VM was mentioned. That in itself is not really nested virtualization. Namespace-based containers, is just a kernel feature providing separation independent of a hypervisor. However, Kata Containers is an approach to tie Linux containers into a hypervisor layer, making them "fully" virtualized. A theoretical situation arises that involves the use of Kata Containers on a Linux system that itself already runs in a VM. That might constitute an example of nested virtualization, but it was decided that this is not a mainstream idea, possibly not supported or feasible, and in each case likely more trouble than it is worth. "Flattening" the virtualization approach so that all units still run on one hypervisor is a likely outcome. Further research into partitioning methods is likely but for now this falls outside of a mainstream automotive virtual platform definition. We highlighted that Linux containers in their normal namespace based implementation are already a very useful system partition tool and it can be trivially applied also if the Linux kernel runs in a VM.

Audio
Matti


Some info on Linux/Xen code here:

HVWS: Xen input and experience on Audio, Display, Input and TEE

Sensors

Automotive sensors:


Artem

Automotive sensors? Radar/LiDAR/? (or are they separate ECUs?)

Standard embedded sensor (ambient light...)

Some OS have requirements - eg. Android requires orientation sensor.

CPUs/SoCs have "internal" sensors too. Relating to temperature and power mgmt. Some internal control tweaks for power management (core frequency / voltage) are like tiny internal actuators. Virtual access to those? Same or different APIs?

Artem proposed that Systems Control Management Interface (SCMI) protocol is flexible and an appropriate abstraction for sensors. It is also appropriate for controlling power-management and related things. The hardware access implementation is according to ARM offloaded to a "Systems Control Processor" but this is a virtual concept. It could be a dedicated core in some cases, perhaps in others not. EPAM/Xen tried out putting code in ARM-TF, to act as this SCP.

SCMI destined (?) to become a ARM-wide standard in a currently fragmented reality.

Presentation attached (PDF)

Upper protocol defined, but could imagine different lower transport. One mailbox-style transport is kind-of defined by ARM spec? Discussion if VIRTIO transport would be appropriate. A "SCMI device" type added to VIRTIO?

Challenges:

  • Current situation in ARM is fragmented with many overlapping unique APIs across chip vendors.
  • Is this doable also on x86, and is it likely to be adopted?
  • Discuss applicability beyond "sensors" and where boundaries are drawn.

Reference:

What about PINCTRL, and handling the many multiplexed pins in a modern SoC. Any remaining need for lower-level protocols for accessing/virtualizing hardware?

Media Acceleration (VPUP, IPU, CODEC)Hardware support for codec/processing
Abstraction of SoC specifics
DSPs
Tensor processors
ArtemVPU = "AI" CPU optimized for visual recognition
USB
Franz
Example Assigning Host USB device to a Guest VM in KVM, here:

https://www.linux-kvm.org/page/USB_Host_Device_Assigned_to_Guest

Which use cases do we want to address?

•USB 2.0 (EHCI controller)
•USB 3.0 (xHCI controllers will replace ECHI)
•USB C
•Host only
•Device Classes:

  • Mass storage. Enable use of USB device with volume provider
  • Communications (e.g. serial, Ethernet)
  • Human interface (e.g. keyboard, mouse)

•On-The-Go (system can function as both USB host and USB device)
•Hot-plug (partial support):

  • Static configuration of device “tree”. A device can be plugged into a port. Dynamically detect device type.
  • Device tree cannot grow dynamically, i.e. cannot plug in a hub

Other Serial devices?

(Where does LIN, etc. fit in?)




LIN-bus:

  • Source code for linux-lin driver (for Linux, not necessarily virtual environment):
  • Paper by Czech Technical University & Volkswagen Group Research:
    LIN based on SocketCAN → 1. OSADL article, 2. paper (PDF).
    • The paper concludes that LIN data frames are similar enough to CAN frames that it can reuse CAN software infrastructure (such as the SocketCAN standard). LIN is a serial bus, implemented with a UARTs, and therefore standard UART device drivers would be used. For virtual environments, we can rely on the same conclusions, and therefore refer to the answer given for CAN.
    • On the other hand, LIN is most popular for its simplicity / low cost (even lower than CAN) and used in very simple ECUs or to/from input devices like switches, knobs and buttons. On the larger CPU it is likely to be run by a separate dedicated microcontroller, or at least small on-chip CPU core. Therefore it can often be considered out-of-scope for the CPU that implements virtualization.
CAN
Franz

virtio-can: VIRTIO-based CAN driver



Ethernet (incl. AvB/TSN)
NikolaThe required features are not present in the network virtio devices as of virtio 1.0.

Must have requirements:

  • IEEE 802.1AS compatible egress and ingress timestamps on ethernet frames available in the virtio consumer OS

Good to have:

General architectural considerations:

  • What if there is more than one consumer of the IEEE 802.1AS defined network timebase on the same system?
Bluetooth

Sang-bum

+OpenSynergy with BT experience

Is it possible? Needed?
Memory Balloon Device

Applicable?

Random Number Generator




Although used also for crypto. Look at what VIRTIO proposes. (Discussion should cover hardware-supported random generator, of course).
Watchdog

Very important for embedded systems... Let's see what is there and what we need to do.





VIRTIO-defined devices

The VIRTIO 1.0 specification is organized a bit differently, and more generic than our detailed list above.  Here is a much abbreviated table of contents for VIRTIO 1.0, just to give an overview on the most important parts.  Consider, especially, the limited types of devices.  All defined devices are under these categories only for the 1.0 version.

2 Basic Facilities of a Virtio Device
2.4 Virtqueues
3.1 Device Initialization
3.2 Device Operation
3.3 Device Cleanup
4 Virtio Transport Options
4.1 Virtio Over PCI Bus
4.2 Virtio Over MMIO
4.3 Virtio Over Channel I/O
5 Device Types
   5.1 Network Device
   5.2 Block Device
   5.3 Console Device
   5.4 Entropy Device
   5.5 Traditional Memory Balloon Device
   5.6 SCSI Host Device

7 Conformance
7.2 Driver Conformance
7.3 Device Conformance
7.4 Legacy Interface: Transitional Device and Transitional Driver Conformance

B Creating New Device Types



When looking at a particular proposed device standard, evaluate characteristics/criteria:

Criteria (brainstorm):

- Availability?
- Is there a proposal for standard (specification)?
- Is it accepted in VIRTIO?
- Is it a de-facto standard?
- Implementation status
- ...In QEMU/Linux kernel?
- ...FOSS in a GitHub Repo?
- ...Commercial/closed-source implementations?
- ... Number of implementations?

- Complexity estimation?
- ... e.g. CAN Device class, vs GPU (need to consider large User-space
library, complex HALs,

- Performance?
- ...mostly implementation dependent?  Are the technical requirements hindering implementing it efficiently, for some reason?  Does it matter?

- Code Maturity?

-- ...implementation dependent, but evaluate that which exists

- Evaluate: Security aspects
- Evaluate: Functional Safety aspects

Importance for automotive use-case
- ...Is it generally applicable for many use cases or for a special case?



  • No labels

1 Comment

  1. Matti, 

    Good summary of what VirtIO is.

    I think some of the facets of what you have mentioned could be represented as a block diagram mapping the various parts. For e.g. the memory allocation by the "driver" being separated from "device". It would be nice to know what goes in what domain.

    I get that  VirtIO is paravirtualization of devices. Does this mean that it is predominantly applicable to Type 2 hypervisors?  Can it also be used in bare-metal hypervisors? If so, does it favor micro-kernel / monolithic? If it can be used as a standard interface even for bare-metal hypervisors, I am also wondering how complex it will be as that will have a direct bearing on safety/security. Note that some of the recent automotive applications are around consolidating safety-critical ad safety-agnostic partitions. Of course, it will be a huge benefit, if we can define and deliver a solution for this.