projects / linux-block.git / blame
| Commit | Line | Data |
|---|---|---|
| a5050c61 | 1 | Hyper-V network driver |
| 2 | ====================== | |
| 3 | ||
| 4 | Compatibility | |
| 5 | ============= | |
| 6 | ||
| 7 | This driver is compatible with Windows Server 2012 R2, 2016 and | |
| 8 | Windows 10. | |
| 9 | ||
| 10 | Features | |
| 11 | ======== | |
| 12 | ||
| 13 | Checksum offload | |
| 14 | ---------------- | |
| 15 | The netvsc driver supports checksum offload as long as the | |
| 16 | Hyper-V host version does. Windows Server 2016 and Azure | |
| 17 | support checksum offload for TCP and UDP for both IPv4 and | |
| 18 | IPv6. Windows Server 2012 only supports checksum offload for TCP. | |
| 19 | ||
| 20 | Receive Side Scaling | |
| 21 | -------------------- | |
| 78005d91 HZ |
22 | Hyper-V supports receive side scaling. For TCP & UDP, packets can |
| 23 | be distributed among available queues based on IP address and port | |
| 3b0c3458 HZ |
24 | number. |
| 25 | ||
| 78005d91 HZ |
26 | For TCP & UDP, we can switch hash level between L3 and L4 by ethtool |
| 27 | command. TCP/UDP over IPv4 and v6 can be set differently. The default | |
| 3b0c3458 HZ |
28 | hash level is L4. We currently only allow switching TX hash level |
| 29 | from within the guests. | |
| 30 | ||
| 31 | On Azure, fragmented UDP packets have high loss rate with L4 | |
| 32 | hashing. Using L3 hashing is recommended in this case. | |
| 33 | ||
| 34 | For example, for UDP over IPv4 on eth0: | |
| d35d6e92 | 35 | To include UDP port numbers in hashing: |
| 3b0c3458 | 36 | ethtool -N eth0 rx-flow-hash udp4 sdfn |
| d35d6e92 | 37 | To exclude UDP port numbers in hashing: |
| 3b0c3458 HZ |
38 | ethtool -N eth0 rx-flow-hash udp4 sd |
| 39 | To show UDP hash level: | |
| 40 | ethtool -n eth0 rx-flow-hash udp4 | |
| a5050c61 | 41 | |
| 42 | Generic Receive Offload, aka GRO | |
| 43 | -------------------------------- | |
| 44 | The driver supports GRO and it is enabled by default. GRO coalesces | |
| 45 | like packets and significantly reduces CPU usage under heavy Rx | |
| 46 | load. | |
| 47 | ||
| f1951c22 HZ |
48 | Large Receive Offload (LRO), or Receive Side Coalescing (RSC) |
| 49 | ------------------------------------------------------------- | |
| 50 | The driver supports LRO/RSC in the vSwitch feature. It reduces the per packet | |
| 51 | processing overhead by coalescing multiple TCP segments when possible. The | |
| 52 | feature is enabled by default on VMs running on Windows Server 2019 and | |
| 53 | later. It may be changed by ethtool command: | |
| 54 | ethtool -K eth0 lro on | |
| 55 | ethtool -K eth0 lro off | |
| 56 | ||
| a5050c61 | 57 | SR-IOV support |
| 58 | -------------- | |
| 59 | Hyper-V supports SR-IOV as a hardware acceleration option. If SR-IOV | |
| 60 | is enabled in both the vSwitch and the guest configuration, then the | |
| 61 | Virtual Function (VF) device is passed to the guest as a PCI | |
| 62 | device. In this case, both a synthetic (netvsc) and VF device are | |
| 63 | visible in the guest OS and both NIC's have the same MAC address. | |
| 64 | ||
| 65 | The VF is enslaved by netvsc device. The netvsc driver will transparently | |
| 66 | switch the data path to the VF when it is available and up. | |
| 67 | Network state (addresses, firewall, etc) should be applied only to the | |
| 68 | netvsc device; the slave device should not be accessed directly in | |
| 69 | most cases. The exceptions are if some special queue discipline or | |
| 70 | flow direction is desired, these should be applied directly to the | |
| 71 | VF slave device. | |
| 72 | ||
| 73 | Receive Buffer | |
| 74 | -------------- | |
| 75 | Packets are received into a receive area which is created when device | |
| 76 | is probed. The receive area is broken into MTU sized chunks and each may | |
| 77 | contain one or more packets. The number of receive sections may be changed | |
| 78 | via ethtool Rx ring parameters. | |
| 79 | ||
| 80 | There is a similar send buffer which is used to aggregate packets for sending. | |
| 81 | The send area is broken into chunks of 6144 bytes, each of section may | |
| 82 | contain one or more packets. The send buffer is an optimization, the driver | |
| 83 | will use slower method to handle very large packets or if the send buffer | |
| 84 | area is exhausted. |