summaryrefslogtreecommitdiffstats
path: root/src/plugins/af_xdp/af_xdp_doc.md
blob: 7d83d7129188ebe9a953314757ccbe6883d939a4 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
# AF_XDP Ethernet driver {#af_xdp_doc}

This driver relies on Linux AF_XDP socket to rx/tx Ethernet packets.

## Maturity level
Under development: it should work, but has not been thoroughly tested.

## Features
 - copy and zero-copy mode
 - multiqueue
 - API
 - custom eBPF program
 - polling, interrupt and adaptive mode

## Limitations
Because of AF_XDP restrictions, the MTU is limited to below PAGE_SIZE
(4096-bytes on most systems) minus 256-bytes, and they are additional
limitations depending upon specific Linux device drivers.
As a rule of thumb, a MTU of 3000-bytes or less should be safe.
Furthermore, upon UMEM creation, the kernel allocates a
physically-contiguous structure, whose size is proportional to the number
of 4KB pages contained in the UMEM. That allocation might fail when
the number of buffers allocated by VPP is too high. That number can be
controlled with the `buffers { buffers-per-numa }` configuration option.
Finally, note that because of this limitation, this plugin is unlikely
to be compatible with the use of 1GB hugepages.

## Requirements
The Linux kernel interface must be up and have enough queues before
creating the VPP AF_XDP interface, otherwise Linux will deny creating
the AF_XDP socket.
The AF_XDP interface will claim NIC RX queue starting from 0, up to the
requested number of RX queues (only 1 by default). It means all packets
destined to NIC RX queue `[0, num_rx_queues[` will be received by the
AF_XDP interface, and only them. Depending on your configuration, there
will usually be several RX queues (typically 1 per core) and packets are
spread accross queues by RSS. In order to receive consistent traffic,
you **must** program the NIC dispatching accordingly. The simplest way
to get all the packets is to specify `num-rx-queues all` to grab all
available queues or to reconfigure the Linux kernel driver to use only
`num_rx_queues` RX queues (ie all NIC queues will be associated with
the AF_XDP socket):
```
~# ethtool -L <iface> combined <num_rx_queues>
```
Additionally, the VPP AF_XDP interface will use a MAC address generated at
creation time instead of the Linux kernel interface MAC. As Linux kernel
interface are not in promiscuous mode by default (see below) this will
results in a useless configuration where the VPP AF_XDP interface only
receives packets destined to the Linux kernel interface MAC just to drop
them because the destination MAC does not match VPP AF_XDP interface MAC.
If you want to use the Linux interface MAC for the VPP AF_XDP interface,
you can change it afterwards in VPP:
```
~# vppctl set int mac address <iface> <mac>
```
Finally, if you wish to receive all packets and not only the packets
destined to the Linux kernel interface MAC you need to set the Linux
kernel interface in promiscuous mode:
```
~# ip link set dev <iface> promisc on
```

## Security considerations
When creating an AF_XDP interface, it will receive all packets arriving
to the NIC RX queue `[0, num_rx_queues[`. You need to configure the Linux
kernel NIC driver properly to ensure that only intented packets will
arrive in this queue. There is no way to filter the packets after-the-fact
using eg. netfilter or eBPF.

## Quickstart
1. Put the Linux kernel interface up and in promiscuous mode:
```
~# ip l set dev enp216s0f0 promisc on up
```
2. Create the AF_XDP interface:
```
~# vppctl create int af_xdp host-if enp216s0f0 num-rx-queues all
```
3. Use the interface as usual, eg.:
```
~# vppctl set int ip addr enp216s0f0/0 1.1.1.1/24
~# vppctl set int st enp216s0f0/0 up
~# vppctl ping 1.1.1.100`
```

## Custom eBPF XDP program
This driver relies on libbpf and as such relies on the `xsks_map` eBPF
map.  The default behavior is to use the XDP program already attached
to the interface if any, otherwise load the default one.
You can request to load a custom XDP program with the `prog` option when
creating the interface in VPP:
```
~# vppctl create int af_xdp host-if enp216s0f0 num-rx-queues 4 prog extras/bpf/af_xdp.bpf.o
```
In that case it will replace any previously attached program.  A custom
XDP program example is provided in `extras/bpf/`.

## Performance consideration
AF_XDP relies on the Linux kernel NIC driver to rx/tx packets. To reach
high-performance (10's MPPS), the Linux kernel NIC driver must support
zero-copy mode and its RX path must run on a dedicated core in the NUMA
where the NIC is physically connected.