projects/linux·blob master · refs

174 lines · 7.2 KiB

1.. SPDX-License-Identifier: GPL-2.0
2.. include:: <isonum.txt>
3
4===============
5Multi-PF Netdev
6===============
7
8Contents
9========
10
11- `Background`_
12- `Overview`_
13- `mlx5 implementation`_
14- `Channels distribution`_
15- `Observability`_
16- `Steering`_
17- `Mutually exclusive features`_
18
19Background
20==========
21
22The Multi-PF NIC technology enables several CPUs within a multi-socket server to connect directly to
23the network, each through its own dedicated PCIe interface. Through either a connection harness that
24splits the PCIe lanes between two cards or by bifurcating a PCIe slot for a single card. This
25results in eliminating the network traffic traversing over the internal bus between the sockets,
26significantly reducing overhead and latency, in addition to reducing CPU utilization and increasing
27network throughput.
28
29Overview
30========
31
32The feature adds support for combining multiple PFs of the same port in a Multi-PF environment under
33one netdev instance. It is implemented in the netdev layer. Lower-layer instances like pci func,
34sysfs entry, and devlink are kept separate.
35Passing traffic through different devices belonging to different NUMA sockets saves cross-NUMA
36traffic and allows apps running on the same netdev from different NUMAs to still feel a sense of
37proximity to the device and achieve improved performance.
38
39mlx5 implementation
40===================
41
42Multi-PF or Socket-direct in mlx5 is achieved by grouping PFs together which belong to the same
43NIC and has the socket-direct property enabled, once all PFs are probed, we create a single netdev
44to represent all of them, symmetrically, we destroy the netdev whenever any of the PFs is removed.
45
46The netdev network channels are distributed between all devices, a proper configuration would utilize
47the correct close NUMA node when working on a certain app/CPU.
48
49We pick one PF to be a primary (leader), and it fills a special role. The other devices
50(secondaries) are disconnected from the network at the chip level (set to silent mode). In silent
51mode, no south <-> north traffic flowing directly through a secondary PF. It needs the assistance of
52the leader PF (east <-> west traffic) to function. All Rx/Tx traffic is steered through the primary
53to/from the secondaries.
54
55Currently, we limit the support to PFs only, and up to two PFs (sockets).
56
57Channels distribution
58=====================
59
60We distribute the channels between the different PFs to achieve local NUMA node performance
61on multiple NUMA nodes.
62
63Each combined channel works against one specific PF, creating all its datapath queues against it. We
64distribute channels to PFs in a round-robin policy.
65
66::
67
68 Example for 2 PFs and 5 channels:
69 +--------+--------+
70 | ch idx | PF idx |
71 +--------+--------+
72 | 0 | 0 |
73 | 1 | 1 |
74 | 2 | 0 |
75 | 3 | 1 |
76 | 4 | 0 |
77 +--------+--------+
78
79
80The reason we prefer round-robin is, it is less influenced by changes in the number of channels. The
81mapping between a channel index and a PF is fixed, no matter how many channels the user configures.
82As the channel stats are persistent across channel's closure, changing the mapping every single time
83would turn the accumulative stats less representing of the channel's history.
84
85This is achieved by using the correct core device instance (mdev) in each channel, instead of them
86all using the same instance under "priv->mdev".
87
88Observability
89=============
90The relation between PF, irq, napi, and queue can be observed via netlink spec::
91
92 $ ./tools/net/ynl/pyynl/cli.py --spec Documentation/netlink/specs/netdev.yaml --dump queue-get --json='{"ifindex": 13}'
93 [{'id': 0, 'ifindex': 13, 'napi-id': 539, 'type': 'rx'},
94 {'id': 1, 'ifindex': 13, 'napi-id': 540, 'type': 'rx'},
95 {'id': 2, 'ifindex': 13, 'napi-id': 541, 'type': 'rx'},
96 {'id': 3, 'ifindex': 13, 'napi-id': 542, 'type': 'rx'},
97 {'id': 4, 'ifindex': 13, 'napi-id': 543, 'type': 'rx'},
98 {'id': 0, 'ifindex': 13, 'napi-id': 539, 'type': 'tx'},
99 {'id': 1, 'ifindex': 13, 'napi-id': 540, 'type': 'tx'},
100 {'id': 2, 'ifindex': 13, 'napi-id': 541, 'type': 'tx'},
101 {'id': 3, 'ifindex': 13, 'napi-id': 542, 'type': 'tx'},
102 {'id': 4, 'ifindex': 13, 'napi-id': 543, 'type': 'tx'}]
103
104 $ ./tools/net/ynl/pyynl/cli.py --spec Documentation/netlink/specs/netdev.yaml --dump napi-get --json='{"ifindex": 13}'
105 [{'id': 543, 'ifindex': 13, 'irq': 42},
106 {'id': 542, 'ifindex': 13, 'irq': 41},
107 {'id': 541, 'ifindex': 13, 'irq': 40},
108 {'id': 540, 'ifindex': 13, 'irq': 39},
109 {'id': 539, 'ifindex': 13, 'irq': 36}]
110
111Here you can clearly observe our channels distribution policy::
112
113 $ ls /proc/irq/{36,39,40,41,42}/mlx5* -d -1
114 /proc/irq/36/mlx5_comp0@pci:0000:08:00.0
115 /proc/irq/39/mlx5_comp0@pci:0000:09:00.0
116 /proc/irq/40/mlx5_comp1@pci:0000:08:00.0
117 /proc/irq/41/mlx5_comp1@pci:0000:09:00.0
118 /proc/irq/42/mlx5_comp2@pci:0000:08:00.0
119
120Steering
121========
122Secondary PFs are set to "silent" mode, meaning they are disconnected from the network.
123
124In Rx, the steering tables belong to the primary PF only, and it is its role to distribute incoming
125traffic to other PFs, via cross-vhca steering capabilities. Still maintain a single default RSS table,
126that is capable of pointing to the receive queues of a different PF.
127
128In Tx, the primary PF creates a new Tx flow table, which is aliased by the secondaries, so they can
129go out to the network through it.
130
131In addition, we set default XPS configuration that, based on the CPU, selects an SQ belonging to the
132PF on the same node as the CPU.
133
134XPS default config example:
135
136NUMA node(s): 2
137NUMA node0 CPU(s): 0-11
138NUMA node1 CPU(s): 12-23
139
140PF0 on node0, PF1 on node1.
141
142- /sys/class/net/eth2/queues/tx-0/xps_cpus:000001
143- /sys/class/net/eth2/queues/tx-1/xps_cpus:001000
144- /sys/class/net/eth2/queues/tx-2/xps_cpus:000002
145- /sys/class/net/eth2/queues/tx-3/xps_cpus:002000
146- /sys/class/net/eth2/queues/tx-4/xps_cpus:000004
147- /sys/class/net/eth2/queues/tx-5/xps_cpus:004000
148- /sys/class/net/eth2/queues/tx-6/xps_cpus:000008
149- /sys/class/net/eth2/queues/tx-7/xps_cpus:008000
150- /sys/class/net/eth2/queues/tx-8/xps_cpus:000010
151- /sys/class/net/eth2/queues/tx-9/xps_cpus:010000
152- /sys/class/net/eth2/queues/tx-10/xps_cpus:000020
153- /sys/class/net/eth2/queues/tx-11/xps_cpus:020000
154- /sys/class/net/eth2/queues/tx-12/xps_cpus:000040
155- /sys/class/net/eth2/queues/tx-13/xps_cpus:040000
156- /sys/class/net/eth2/queues/tx-14/xps_cpus:000080
157- /sys/class/net/eth2/queues/tx-15/xps_cpus:080000
158- /sys/class/net/eth2/queues/tx-16/xps_cpus:000100
159- /sys/class/net/eth2/queues/tx-17/xps_cpus:100000
160- /sys/class/net/eth2/queues/tx-18/xps_cpus:000200
161- /sys/class/net/eth2/queues/tx-19/xps_cpus:200000
162- /sys/class/net/eth2/queues/tx-20/xps_cpus:000400
163- /sys/class/net/eth2/queues/tx-21/xps_cpus:400000
164- /sys/class/net/eth2/queues/tx-22/xps_cpus:000800
165- /sys/class/net/eth2/queues/tx-23/xps_cpus:800000
166
167Mutually exclusive features
168===========================
169
170The nature of Multi-PF, where different channels work with different PFs, conflicts with
171stateful features where the state is maintained in one of the PFs.
172For example, in the TLS device-offload feature, special context objects are created per connection
173and maintained in the PF. Transitioning between different RQs/SQs would break the feature. Hence,
174we disable this combination for now.