virtio-paper¶

abstract¶

virtio: a series of efficient, well-maintained Linux drivers
- can be adapted for various hypervisor using a shim layer.
vring: ring buffer transport implementation
provide an implementation
- presents (1) vring transport (2) device config as a PCI device
- means:
  - guest OS merely need a new PCI driver
  - hypervisor only need add vring support to virtual device (they implement)
my idea
- 通用的 (generic) IO PV protocol, 高相容性
  - 向下相容: 不同的 hypervisor 可以使用同一套 IO PV protocol 跟 guest driver 溝通. 如同不同 hypervisor 都使用相同硬體介面.
  - 向上相容: [?] 跟原生的 kernel driver 儘量相似, 希望只要改掉硬體介面即可 porting 成 virtio driver.
- protocol 組成
  - 實作必須具有 IO PV 的必備要素, 比如說 transport buffer, vring 便用於此
  - [?] virtio 同時是 protocol, 也是 API layer. 這讓我想到 remote procedure call
my idea2
- design 目標
- API (config + transport)
- 應用的 driver, 需要特殊處理的應用: PCI driver
- 效能分析
- 目前有使用的實作
- 未來的擴展

1. intro
1. virtio: the three goals
1. virtio: a Linux internal abstraction API
- Virtqueues: a transport abstraction
1. virtio_ring: a transport implementation for virtio
- 4.1. a note on Zero-Copy and religion of page flipping
1. current virtio drivers
- 5.1. virtio block driver
- 5.2. virtio network driver
1. virtio_pci: a PCI implementation of vring and virtio
1. performance
1. adoption
1. future work
1. conclusions

通用的 IO Para-Virt Protocol, 跨 hypervisor/guest VM.

3 goals
1. driver unification
2. uniformity to provide common ABI for (1) general publication (2) use of buffer.
3. device probing and config
如果 developer 熟悉 Linux, 他們可能會直接 map Linux API 在 (virtual IO mechanism?) ABI 上面.
cross-device: common ABI for (1) general publication (2) use of buffer.
provide a two complete ABI implementations
- using (1) virtio_ring infrastructure (2) Linux API for virtual IO device.
- implement final part of virtual IO device: device probe and config
explicit seperate (1) driver (2) transport (3) config
- 反例: 如果其他 hypervisor 要用 Xen Linux network driver, 必須 support Xenbus probing and config system

4 parts of config op
1. RW feature bits
2. RW config space
3. RW status bits
4. device reset
feature bit
- device feature: e.g. VIRTIO_NET_F_CSUM is checksum offload feature of net device
- device specific
config space
- when device feature VIRTIO_NET_F_MAC is set
- MAC address of device is in config space
RW 8 bits status word
- indicate device probe
- VIRTIO_CONFIG_S_DRIVER_OK is set means that host knows what feature it understands and wants to use.

Transport Abstraction: Virtqueue

virtio-blk has one queue
virtio-net/console has two queues, input/output both use one.
each buffer is a scatter-gather array
virtqueue op struct: 5 ops
- add_buf
- kick
- get_buf
- enable/disable_cb
data exchange flow: add_buf() => kick host() => host update data => get_buf()
5 ops
- disable_cb is a hint that guest doesn’t want to know “when the buffer is used” => same as disable interrupt.

virtio ring consist of 3 parts
- descriptor array: (addr, length) pair chain.
- avail ring: guest indicate. chain is ready to use.
- used ring: host indicate. chain is used.
descriptor array:
- (addr, length)
- optional next
- flags: 2 bits, 1 for RW, 1 for next option
used ring 有故意跟 available ring/descriptor array 放在不同 page. 這樣 cache 的表現會比較好.
interrupt suppression flag
- available ring 跟 used ring 都有.
- for optimization (virtqueue kick (vmexit/trap) and interrupt completion)
- available ring 是 guest 用來通知 host 不用送 completion interrupt (disable_cb)
- used ring 是 host 用來通知 guest 不用 virtqueue kick host.
- optimization example?

efficient IO need 2 things
1. the number of notification per op. => by virtio_ring interrupt suppression flag.
2. amount of cache-cold data which is accessed.
Zero-Copy
page flipping