| 1 |
.. SPDX-License-Identifier: GPL-2.0 |
| 2 |
|
| 3 |
============================== |
| 4 |
The QorIQ DPAA Ethernet Driver |
| 5 |
============================== |
| 6 |
|
| 7 |
Authors: |
| 8 |
- Madalin Bucur <[email protected]> |
| 9 |
- Camelia Groza <[email protected]> |
| 10 |
|
| 11 |
.. Contents |
| 12 |
|
| 13 |
- DPAA Ethernet Overview |
| 14 |
- DPAA Ethernet Supported SoCs |
| 15 |
- Configuring DPAA Ethernet in your kernel |
| 16 |
- DPAA Ethernet Frame Processing |
| 17 |
- DPAA Ethernet Features |
| 18 |
- DPAA IRQ Affinity and Receive Side Scaling |
| 19 |
- Debugging |
| 20 |
|
| 21 |
DPAA Ethernet Overview |
| 22 |
====================== |
| 23 |
|
| 24 |
DPAA stands for Data Path Acceleration Architecture and it is a |
| 25 |
set of networking acceleration IPs that are available on several |
| 26 |
generations of SoCs, both on PowerPC and ARM64. |
| 27 |
|
| 28 |
The Freescale DPAA architecture consists of a series of hardware blocks |
| 29 |
that support Ethernet connectivity. The Ethernet driver depends upon the |
| 30 |
following drivers in the Linux kernel: |
| 31 |
|
| 32 |
- Peripheral Access Memory Unit (PAMU) (* needed only for PPC platforms) |
| 33 |
drivers/iommu/fsl_* |
| 34 |
- Frame Manager (FMan) |
| 35 |
drivers/net/ethernet/freescale/fman |
| 36 |
- Queue Manager (QMan), Buffer Manager (BMan) |
| 37 |
drivers/soc/fsl/qbman |
| 38 |
|
| 39 |
A simplified view of the dpaa_eth interfaces mapped to FMan MACs:: |
| 40 |
|
| 41 |
dpaa_eth /eth0\ ... /ethN\ |
| 42 |
driver | | | | |
| 43 |
------------- ---- ----------- ---- ------------- |
| 44 |
-Ports / Tx Rx \ ... / Tx Rx \ |
| 45 |
FMan | | | | |
| 46 |
-MACs | MAC0 | | MACN | |
| 47 |
/ dtsec0 \ ... / dtsecN \ (or tgec) |
| 48 |
/ \ / \(or memac) |
| 49 |
--------- -------------- --- -------------- --------- |
| 50 |
FMan, FMan Port, FMan SP, FMan MURAM drivers |
| 51 |
--------------------------------------------------------- |
| 52 |
FMan HW blocks: MURAM, MACs, Ports, SP |
| 53 |
--------------------------------------------------------- |
| 54 |
|
| 55 |
The dpaa_eth relation to the QMan, BMan and FMan:: |
| 56 |
|
| 57 |
________________________________ |
| 58 |
dpaa_eth / eth0 \ |
| 59 |
driver / \ |
| 60 |
--------- -^- -^- -^- --- --------- |
| 61 |
QMan driver / \ / \ / \ \ / | BMan | |
| 62 |
|Rx | |Rx | |Tx | |Tx | | driver | |
| 63 |
--------- |Dfl| |Err| |Cnf| |FQs| | | |
| 64 |
QMan HW |FQ | |FQ | |FQs| | | | | |
| 65 |
/ \ / \ / \ \ / | | |
| 66 |
--------- --- --- --- -v- --------- |
| 67 |
| FMan QMI | | |
| 68 |
| FMan HW FMan BMI | BMan HW | |
| 69 |
----------------------- -------- |
| 70 |
|
| 71 |
where the acronyms used above (and in the code) are: |
| 72 |
|
| 73 |
=============== =========================================================== |
| 74 |
DPAA Data Path Acceleration Architecture |
| 75 |
FMan DPAA Frame Manager |
| 76 |
QMan DPAA Queue Manager |
| 77 |
BMan DPAA Buffers Manager |
| 78 |
QMI QMan interface in FMan |
| 79 |
BMI BMan interface in FMan |
| 80 |
FMan SP FMan Storage Profiles |
| 81 |
MURAM Multi-user RAM in FMan |
| 82 |
FQ QMan Frame Queue |
| 83 |
Rx Dfl FQ default reception FQ |
| 84 |
Rx Err FQ Rx error frames FQ |
| 85 |
Tx Cnf FQ Tx confirmation FQs |
| 86 |
Tx FQs transmission frame queues |
| 87 |
dtsec datapath three speed Ethernet controller (10/100/1000 Mbps) |
| 88 |
tgec ten gigabit Ethernet controller (10 Gbps) |
| 89 |
memac multirate Ethernet MAC (10/100/1000/10000) |
| 90 |
=============== =========================================================== |
| 91 |
|
| 92 |
DPAA Ethernet Supported SoCs |
| 93 |
============================ |
| 94 |
|
| 95 |
The DPAA drivers enable the Ethernet controllers present on the following SoCs: |
| 96 |
|
| 97 |
PPC |
| 98 |
- P1023 |
| 99 |
- P2041 |
| 100 |
- P3041 |
| 101 |
- P4080 |
| 102 |
- P5020 |
| 103 |
- P5040 |
| 104 |
- T1023 |
| 105 |
- T1024 |
| 106 |
- T1040 |
| 107 |
- T1042 |
| 108 |
- T2080 |
| 109 |
- T4240 |
| 110 |
- B4860 |
| 111 |
|
| 112 |
ARM |
| 113 |
- LS1043A |
| 114 |
- LS1046A |
| 115 |
|
| 116 |
Configuring DPAA Ethernet in your kernel |
| 117 |
======================================== |
| 118 |
|
| 119 |
To enable the DPAA Ethernet driver, the following Kconfig options are required:: |
| 120 |
|
| 121 |
# common for arch/arm64 and arch/powerpc platforms |
| 122 |
CONFIG_FSL_DPAA=y |
| 123 |
CONFIG_FSL_FMAN=y |
| 124 |
CONFIG_FSL_DPAA_ETH=y |
| 125 |
CONFIG_FSL_XGMAC_MDIO=y |
| 126 |
|
| 127 |
# for arch/powerpc only |
| 128 |
CONFIG_FSL_PAMU=y |
| 129 |
|
| 130 |
# common options needed for the PHYs used on the RDBs |
| 131 |
CONFIG_VITESSE_PHY=y |
| 132 |
CONFIG_REALTEK_PHY=y |
| 133 |
CONFIG_AQUANTIA_PHY=y |
| 134 |
|
| 135 |
DPAA Ethernet Frame Processing |
| 136 |
============================== |
| 137 |
|
| 138 |
On Rx, buffers for the incoming frames are retrieved from the buffers found |
| 139 |
in the dedicated interface buffer pool. The driver initializes and seeds these |
| 140 |
with one page buffers. |
| 141 |
|
| 142 |
On Tx, all transmitted frames are returned to the driver through Tx |
| 143 |
confirmation frame queues. The driver is then responsible for freeing the |
| 144 |
buffers. In order to do this properly, a backpointer is added to the buffer |
| 145 |
before transmission that points to the skb. When the buffer returns to the |
| 146 |
driver on a confirmation FQ, the skb can be correctly consumed. |
| 147 |
|
| 148 |
DPAA Ethernet Features |
| 149 |
====================== |
| 150 |
|
| 151 |
Currently the DPAA Ethernet driver enables the basic features required for |
| 152 |
a Linux Ethernet driver. The support for advanced features will be added |
| 153 |
gradually. |
| 154 |
|
| 155 |
The driver has Rx and Tx checksum offloading for UDP and TCP. Currently the Rx |
| 156 |
checksum offload feature is enabled by default and cannot be controlled through |
| 157 |
ethtool. Also, rx-flow-hash and rx-hashing was added. The addition of RSS |
| 158 |
provides a big performance boost for the forwarding scenarios, allowing |
| 159 |
different traffic flows received by one interface to be processed by different |
| 160 |
CPUs in parallel. |
| 161 |
|
| 162 |
The driver has support for multiple prioritized Tx traffic classes. Priorities |
| 163 |
range from 0 (lowest) to 3 (highest). These are mapped to HW workqueues with |
| 164 |
strict priority levels. Each traffic class contains NR_CPU TX queues. By |
| 165 |
default, only one traffic class is enabled and the lowest priority Tx queues |
| 166 |
are used. Higher priority traffic classes can be enabled with the mqprio |
| 167 |
qdisc. For example, all four traffic classes are enabled on an interface with |
| 168 |
the following command. Furthermore, skb priority levels are mapped to traffic |
| 169 |
classes as follows: |
| 170 |
|
| 171 |
* priorities 0 to 3 - traffic class 0 (low priority) |
| 172 |
* priorities 4 to 7 - traffic class 1 (medium-low priority) |
| 173 |
* priorities 8 to 11 - traffic class 2 (medium-high priority) |
| 174 |
* priorities 12 to 15 - traffic class 3 (high priority) |
| 175 |
|
| 176 |
:: |
| 177 |
|
| 178 |
tc qdisc add dev <int> root handle 1: \ |
| 179 |
mqprio num_tc 4 map 0 0 0 0 1 1 1 1 2 2 2 2 3 3 3 3 hw 1 |
| 180 |
|
| 181 |
DPAA IRQ Affinity and Receive Side Scaling |
| 182 |
========================================== |
| 183 |
|
| 184 |
Traffic coming on the DPAA Rx queues or on the DPAA Tx confirmation |
| 185 |
queues is seen by the CPU as ingress traffic on a certain portal. |
| 186 |
The DPAA QMan portal interrupts are affined each to a certain CPU. |
| 187 |
The same portal interrupt services all the QMan portal consumers. |
| 188 |
|
| 189 |
By default the DPAA Ethernet driver enables RSS, making use of the |
| 190 |
DPAA FMan Parser and Keygen blocks to distribute traffic on 128 |
| 191 |
hardware frame queues using a hash on IP v4/v6 source and destination |
| 192 |
and L4 source and destination ports, in present in the received frame. |
| 193 |
When RSS is disabled, all traffic received by a certain interface is |
| 194 |
received on the default Rx frame queue. The default DPAA Rx frame |
| 195 |
queues are configured to put the received traffic into a pool channel |
| 196 |
that allows any available CPU portal to dequeue the ingress traffic. |
| 197 |
The default frame queues have the HOLDACTIVE option set, ensuring that |
| 198 |
traffic bursts from a certain queue are serviced by the same CPU. |
| 199 |
This ensures a very low rate of frame reordering. A drawback of this |
| 200 |
is that only one CPU at a time can service the traffic received by a |
| 201 |
certain interface when RSS is not enabled. |
| 202 |
|
| 203 |
To implement RSS, the DPAA Ethernet driver allocates an extra set of |
| 204 |
128 Rx frame queues that are configured to dedicated channels, in a |
| 205 |
round-robin manner. The mapping of the frame queues to CPUs is now |
| 206 |
hardcoded, there is no indirection table to move traffic for a certain |
| 207 |
FQ (hash result) to another CPU. The ingress traffic arriving on one |
| 208 |
of these frame queues will arrive at the same portal and will always |
| 209 |
be processed by the same CPU. This ensures intra-flow order preservation |
| 210 |
and workload distribution for multiple traffic flows. |
| 211 |
|
| 212 |
RSS can be turned off for a certain interface using ethtool, i.e.:: |
| 213 |
|
| 214 |
# ethtool -N fm1-mac9 rx-flow-hash tcp4 "" |
| 215 |
|
| 216 |
To turn it back on, one needs to set rx-flow-hash for tcp4/6 or udp4/6:: |
| 217 |
|
| 218 |
# ethtool -N fm1-mac9 rx-flow-hash udp4 sfdn |
| 219 |
|
| 220 |
There is no independent control for individual protocols, any command |
| 221 |
run for one of tcp4|udp4|ah4|esp4|sctp4|tcp6|udp6|ah6|esp6|sctp6 is |
| 222 |
going to control the rx-flow-hashing for all protocols on that interface. |
| 223 |
|
| 224 |
Besides using the FMan Keygen computed hash for spreading traffic on the |
| 225 |
128 Rx FQs, the DPAA Ethernet driver also sets the skb hash value when |
| 226 |
the NETIF_F_RXHASH feature is on (active by default). This can be turned |
| 227 |
on or off through ethtool, i.e.:: |
| 228 |
|
| 229 |
# ethtool -K fm1-mac9 rx-hashing off |
| 230 |
# ethtool -k fm1-mac9 | grep hash |
| 231 |
receive-hashing: off |
| 232 |
# ethtool -K fm1-mac9 rx-hashing on |
| 233 |
Actual changes: |
| 234 |
receive-hashing: on |
| 235 |
# ethtool -k fm1-mac9 | grep hash |
| 236 |
receive-hashing: on |
| 237 |
|
| 238 |
Please note that Rx hashing depends upon the rx-flow-hashing being on |
| 239 |
for that interface - turning off rx-flow-hashing will also disable the |
| 240 |
rx-hashing (without ethtool reporting it as off as that depends on the |
| 241 |
NETIF_F_RXHASH feature flag). |
| 242 |
|
| 243 |
Debugging |
| 244 |
========= |
| 245 |
|
| 246 |
The following statistics are exported for each interface through ethtool: |
| 247 |
|
| 248 |
- interrupt count per CPU |
| 249 |
- Rx packets count per CPU |
| 250 |
- Tx packets count per CPU |
| 251 |
- Tx confirmed packets count per CPU |
| 252 |
- Tx S/G frames count per CPU |
| 253 |
- Tx error count per CPU |
| 254 |
- Rx error count per CPU |
| 255 |
- Rx error count per type |
| 256 |
- congestion related statistics: |
| 257 |
|
| 258 |
- congestion status |
| 259 |
- time spent in congestion |
| 260 |
- number of time the device entered congestion |
| 261 |
- dropped packets count per cause |
| 262 |
|
| 263 |
The driver also exports the following information in sysfs: |
| 264 |
|
| 265 |
- the FQ IDs for each FQ type |
| 266 |
/sys/devices/platform/soc/<addr>.fman/<addr>.ethernet/dpaa-ethernet.<id>/net/fm<nr>-mac<nr>/fqids |
| 267 |
|
| 268 |
- the ID of the buffer pool in use |
| 269 |
/sys/devices/platform/soc/<addr>.fman/<addr>.ethernet/dpaa-ethernet.<id>/net/fm<nr>-mac<nr>/bpids |