| 1 |
.. SPDX-License-Identifier: GPL-2.0 |
| 2 |
|
| 3 |
====================== |
| 4 |
Hyper-V network driver |
| 5 |
====================== |
| 6 |
|
| 7 |
Compatibility |
| 8 |
============= |
| 9 |
|
| 10 |
This driver is compatible with Windows Server 2012 R2, 2016 and |
| 11 |
Windows 10. |
| 12 |
|
| 13 |
Features |
| 14 |
======== |
| 15 |
|
| 16 |
Checksum offload |
| 17 |
---------------- |
| 18 |
The netvsc driver supports checksum offload as long as the |
| 19 |
Hyper-V host version does. Windows Server 2016 and Azure |
| 20 |
support checksum offload for TCP and UDP for both IPv4 and |
| 21 |
IPv6. Windows Server 2012 only supports checksum offload for TCP. |
| 22 |
|
| 23 |
Receive Side Scaling |
| 24 |
-------------------- |
| 25 |
Hyper-V supports receive side scaling. For TCP & UDP, packets can |
| 26 |
be distributed among available queues based on IP address and port |
| 27 |
number. |
| 28 |
|
| 29 |
For TCP & UDP, we can switch hash level between L3 and L4 by ethtool |
| 30 |
command. TCP/UDP over IPv4 and v6 can be set differently. The default |
| 31 |
hash level is L4. We currently only allow switching TX hash level |
| 32 |
from within the guests. |
| 33 |
|
| 34 |
On Azure, fragmented UDP packets have high loss rate with L4 |
| 35 |
hashing. Using L3 hashing is recommended in this case. |
| 36 |
|
| 37 |
For example, for UDP over IPv4 on eth0: |
| 38 |
|
| 39 |
To include UDP port numbers in hashing:: |
| 40 |
|
| 41 |
ethtool -N eth0 rx-flow-hash udp4 sdfn |
| 42 |
|
| 43 |
To exclude UDP port numbers in hashing:: |
| 44 |
|
| 45 |
ethtool -N eth0 rx-flow-hash udp4 sd |
| 46 |
|
| 47 |
To show UDP hash level:: |
| 48 |
|
| 49 |
ethtool -n eth0 rx-flow-hash udp4 |
| 50 |
|
| 51 |
Generic Receive Offload, aka GRO |
| 52 |
-------------------------------- |
| 53 |
The driver supports GRO and it is enabled by default. GRO coalesces |
| 54 |
like packets and significantly reduces CPU usage under heavy Rx |
| 55 |
load. |
| 56 |
|
| 57 |
Large Receive Offload (LRO), or Receive Side Coalescing (RSC) |
| 58 |
------------------------------------------------------------- |
| 59 |
The driver supports LRO/RSC in the vSwitch feature. It reduces the per packet |
| 60 |
processing overhead by coalescing multiple TCP segments when possible. The |
| 61 |
feature is enabled by default on VMs running on Windows Server 2019 and |
| 62 |
later. It may be changed by ethtool command:: |
| 63 |
|
| 64 |
ethtool -K eth0 lro on |
| 65 |
ethtool -K eth0 lro off |
| 66 |
|
| 67 |
SR-IOV support |
| 68 |
-------------- |
| 69 |
Hyper-V supports SR-IOV as a hardware acceleration option. If SR-IOV |
| 70 |
is enabled in both the vSwitch and the guest configuration, then the |
| 71 |
Virtual Function (VF) device is passed to the guest as a PCI |
| 72 |
device. In this case, both a synthetic (netvsc) and VF device are |
| 73 |
visible in the guest OS and both NIC's have the same MAC address. |
| 74 |
|
| 75 |
The VF is enslaved by netvsc device. The netvsc driver will transparently |
| 76 |
switch the data path to the VF when it is available and up. |
| 77 |
Network state (addresses, firewall, etc) should be applied only to the |
| 78 |
netvsc device; the slave device should not be accessed directly in |
| 79 |
most cases. The exceptions are if some special queue discipline or |
| 80 |
flow direction is desired, these should be applied directly to the |
| 81 |
VF slave device. |
| 82 |
|
| 83 |
Receive Buffer |
| 84 |
-------------- |
| 85 |
Packets are received into a receive area which is created when device |
| 86 |
is probed. The receive area is broken into MTU sized chunks and each may |
| 87 |
contain one or more packets. The number of receive sections may be changed |
| 88 |
via ethtool Rx ring parameters. |
| 89 |
|
| 90 |
There is a similar send buffer which is used to aggregate packets |
| 91 |
for sending. The send area is broken into chunks, typically of 6144 |
| 92 |
bytes, each of section may contain one or more packets. Small |
| 93 |
packets are usually transmitted via copy to the send buffer. However, |
| 94 |
if the buffer is temporarily exhausted, or the packet to be transmitted is |
| 95 |
an LSO packet, the driver will provide the host with pointers to the data |
| 96 |
from the SKB. This attempts to achieve a balance between the overhead of |
| 97 |
data copy and the impact of remapping VM memory to be accessible by the |
| 98 |
host. |
| 99 |
|
| 100 |
XDP support |
| 101 |
----------- |
| 102 |
XDP (eXpress Data Path) is a feature that runs eBPF bytecode at the early |
| 103 |
stage when packets arrive at a NIC card. The goal is to increase performance |
| 104 |
for packet processing, reducing the overhead of SKB allocation and other |
| 105 |
upper network layers. |
| 106 |
|
| 107 |
hv_netvsc supports XDP in native mode, and transparently sets the XDP |
| 108 |
program on the associated VF NIC as well. |
| 109 |
|
| 110 |
Setting / unsetting XDP program on synthetic NIC (netvsc) propagates to |
| 111 |
VF NIC automatically. Setting / unsetting XDP program on VF NIC directly |
| 112 |
is not recommended, also not propagated to synthetic NIC, and may be |
| 113 |
overwritten by setting of synthetic NIC. |
| 114 |
|
| 115 |
XDP program cannot run with LRO (RSC) enabled, so you need to disable LRO |
| 116 |
before running XDP:: |
| 117 |
|
| 118 |
ethtool -K eth0 lro off |
| 119 |
|
| 120 |
XDP_REDIRECT action is not yet supported. |