Madalin Bucur | 5ee6842 | 2017-05-29 09:22:06 +0300 | [diff] [blame] | 1 | The QorIQ DPAA Ethernet Driver |
| 2 | ============================== |
| 3 | |
| 4 | Authors: |
| 5 | Madalin Bucur <madalin.bucur@nxp.com> |
| 6 | Camelia Groza <camelia.groza@nxp.com> |
| 7 | |
| 8 | Contents |
| 9 | ======== |
| 10 | |
| 11 | - DPAA Ethernet Overview |
| 12 | - DPAA Ethernet Supported SoCs |
| 13 | - Configuring DPAA Ethernet in your kernel |
| 14 | - DPAA Ethernet Frame Processing |
| 15 | - DPAA Ethernet Features |
Madalin Bucur | 0659191 | 2017-08-27 16:13:42 +0300 | [diff] [blame] | 16 | - DPAA IRQ Affinity and Receive Side Scaling |
Madalin Bucur | 5ee6842 | 2017-05-29 09:22:06 +0300 | [diff] [blame] | 17 | - Debugging |
| 18 | |
| 19 | DPAA Ethernet Overview |
| 20 | ====================== |
| 21 | |
| 22 | DPAA stands for Data Path Acceleration Architecture and it is a |
| 23 | set of networking acceleration IPs that are available on several |
| 24 | generations of SoCs, both on PowerPC and ARM64. |
| 25 | |
| 26 | The Freescale DPAA architecture consists of a series of hardware blocks |
| 27 | that support Ethernet connectivity. The Ethernet driver depends upon the |
| 28 | following drivers in the Linux kernel: |
| 29 | |
| 30 | - Peripheral Access Memory Unit (PAMU) (* needed only for PPC platforms) |
| 31 | drivers/iommu/fsl_* |
| 32 | - Frame Manager (FMan) |
| 33 | drivers/net/ethernet/freescale/fman |
| 34 | - Queue Manager (QMan), Buffer Manager (BMan) |
| 35 | drivers/soc/fsl/qbman |
| 36 | |
| 37 | A simplified view of the dpaa_eth interfaces mapped to FMan MACs: |
| 38 | |
| 39 | dpaa_eth /eth0\ ... /ethN\ |
| 40 | driver | | | | |
| 41 | ------------- ---- ----------- ---- ------------- |
| 42 | -Ports / Tx Rx \ ... / Tx Rx \ |
| 43 | FMan | | | | |
| 44 | -MACs | MAC0 | | MACN | |
| 45 | / dtsec0 \ ... / dtsecN \ (or tgec) |
| 46 | / \ / \(or memac) |
| 47 | --------- -------------- --- -------------- --------- |
| 48 | FMan, FMan Port, FMan SP, FMan MURAM drivers |
| 49 | --------------------------------------------------------- |
| 50 | FMan HW blocks: MURAM, MACs, Ports, SP |
| 51 | --------------------------------------------------------- |
| 52 | |
| 53 | The dpaa_eth relation to the QMan, BMan and FMan: |
| 54 | ________________________________ |
| 55 | dpaa_eth / eth0 \ |
| 56 | driver / \ |
| 57 | --------- -^- -^- -^- --- --------- |
| 58 | QMan driver / \ / \ / \ \ / | BMan | |
| 59 | |Rx | |Rx | |Tx | |Tx | | driver | |
| 60 | --------- |Dfl| |Err| |Cnf| |FQs| | | |
| 61 | QMan HW |FQ | |FQ | |FQs| | | | | |
| 62 | / \ / \ / \ \ / | | |
| 63 | --------- --- --- --- -v- --------- |
| 64 | | FMan QMI | | |
| 65 | | FMan HW FMan BMI | BMan HW | |
| 66 | ----------------------- -------- |
| 67 | |
| 68 | where the acronyms used above (and in the code) are: |
| 69 | DPAA = Data Path Acceleration Architecture |
| 70 | FMan = DPAA Frame Manager |
| 71 | QMan = DPAA Queue Manager |
| 72 | BMan = DPAA Buffers Manager |
| 73 | QMI = QMan interface in FMan |
| 74 | BMI = BMan interface in FMan |
| 75 | FMan SP = FMan Storage Profiles |
| 76 | MURAM = Multi-user RAM in FMan |
| 77 | FQ = QMan Frame Queue |
| 78 | Rx Dfl FQ = default reception FQ |
| 79 | Rx Err FQ = Rx error frames FQ |
| 80 | Tx Cnf FQ = Tx confirmation FQs |
| 81 | Tx FQs = transmission frame queues |
| 82 | dtsec = datapath three speed Ethernet controller (10/100/1000 Mbps) |
| 83 | tgec = ten gigabit Ethernet controller (10 Gbps) |
| 84 | memac = multirate Ethernet MAC (10/100/1000/10000) |
| 85 | |
| 86 | DPAA Ethernet Supported SoCs |
| 87 | ============================ |
| 88 | |
| 89 | The DPAA drivers enable the Ethernet controllers present on the following SoCs: |
| 90 | |
| 91 | # PPC |
| 92 | P1023 |
| 93 | P2041 |
| 94 | P3041 |
| 95 | P4080 |
| 96 | P5020 |
| 97 | P5040 |
| 98 | T1023 |
| 99 | T1024 |
| 100 | T1040 |
| 101 | T1042 |
| 102 | T2080 |
| 103 | T4240 |
| 104 | B4860 |
| 105 | |
| 106 | # ARM |
| 107 | LS1043A |
| 108 | LS1046A |
| 109 | |
| 110 | Configuring DPAA Ethernet in your kernel |
| 111 | ======================================== |
| 112 | |
| 113 | To enable the DPAA Ethernet driver, the following Kconfig options are required: |
| 114 | |
| 115 | # common for arch/arm64 and arch/powerpc platforms |
| 116 | CONFIG_FSL_DPAA=y |
| 117 | CONFIG_FSL_FMAN=y |
| 118 | CONFIG_FSL_DPAA_ETH=y |
| 119 | CONFIG_FSL_XGMAC_MDIO=y |
| 120 | |
| 121 | # for arch/powerpc only |
| 122 | CONFIG_FSL_PAMU=y |
| 123 | |
| 124 | # common options needed for the PHYs used on the RDBs |
| 125 | CONFIG_VITESSE_PHY=y |
| 126 | CONFIG_REALTEK_PHY=y |
| 127 | CONFIG_AQUANTIA_PHY=y |
| 128 | |
| 129 | DPAA Ethernet Frame Processing |
| 130 | ============================== |
| 131 | |
| 132 | On Rx, buffers for the incoming frames are retrieved from one of the three |
| 133 | existing buffers pools. The driver initializes and seeds these, each with |
| 134 | buffers of different sizes: 1KB, 2KB and 4KB. |
| 135 | |
| 136 | On Tx, all transmitted frames are returned to the driver through Tx |
| 137 | confirmation frame queues. The driver is then responsible for freeing the |
| 138 | buffers. In order to do this properly, a backpointer is added to the buffer |
| 139 | before transmission that points to the skb. When the buffer returns to the |
| 140 | driver on a confirmation FQ, the skb can be correctly consumed. |
| 141 | |
| 142 | DPAA Ethernet Features |
| 143 | ====================== |
| 144 | |
| 145 | Currently the DPAA Ethernet driver enables the basic features required for |
| 146 | a Linux Ethernet driver. The support for advanced features will be added |
| 147 | gradually. |
| 148 | |
| 149 | The driver has Rx and Tx checksum offloading for UDP and TCP. Currently the Rx |
| 150 | checksum offload feature is enabled by default and cannot be controlled through |
Madalin Bucur | 0659191 | 2017-08-27 16:13:42 +0300 | [diff] [blame] | 151 | ethtool. Also, rx-flow-hash and rx-hashing was added. The addition of RSS |
| 152 | provides a big performance boost for the forwarding scenarios, allowing |
| 153 | different traffic flows received by one interface to be processed by different |
| 154 | CPUs in parallel. |
Madalin Bucur | 5ee6842 | 2017-05-29 09:22:06 +0300 | [diff] [blame] | 155 | |
| 156 | The driver has support for multiple prioritized Tx traffic classes. Priorities |
| 157 | range from 0 (lowest) to 3 (highest). These are mapped to HW workqueues with |
| 158 | strict priority levels. Each traffic class contains NR_CPU TX queues. By |
| 159 | default, only one traffic class is enabled and the lowest priority Tx queues |
| 160 | are used. Higher priority traffic classes can be enabled with the mqprio |
| 161 | qdisc. For example, all four traffic classes are enabled on an interface with |
| 162 | the following command. Furthermore, skb priority levels are mapped to traffic |
| 163 | classes as follows: |
| 164 | |
| 165 | * priorities 0 to 3 - traffic class 0 (low priority) |
| 166 | * priorities 4 to 7 - traffic class 1 (medium-low priority) |
| 167 | * priorities 8 to 11 - traffic class 2 (medium-high priority) |
| 168 | * priorities 12 to 15 - traffic class 3 (high priority) |
| 169 | |
| 170 | tc qdisc add dev <int> root handle 1: \ |
| 171 | mqprio num_tc 4 map 0 0 0 0 1 1 1 1 2 2 2 2 3 3 3 3 hw 1 |
| 172 | |
Madalin Bucur | 0659191 | 2017-08-27 16:13:42 +0300 | [diff] [blame] | 173 | DPAA IRQ Affinity and Receive Side Scaling |
| 174 | ========================================== |
| 175 | |
| 176 | Traffic coming on the DPAA Rx queues or on the DPAA Tx confirmation |
| 177 | queues is seen by the CPU as ingress traffic on a certain portal. |
| 178 | The DPAA QMan portal interrupts are affined each to a certain CPU. |
| 179 | The same portal interrupt services all the QMan portal consumers. |
| 180 | |
| 181 | By default the DPAA Ethernet driver enables RSS, making use of the |
| 182 | DPAA FMan Parser and Keygen blocks to distribute traffic on 128 |
| 183 | hardware frame queues using a hash on IP v4/v6 source and destination |
| 184 | and L4 source and destination ports, in present in the received frame. |
| 185 | When RSS is disabled, all traffic received by a certain interface is |
| 186 | received on the default Rx frame queue. The default DPAA Rx frame |
| 187 | queues are configured to put the received traffic into a pool channel |
| 188 | that allows any available CPU portal to dequeue the ingress traffic. |
| 189 | The default frame queues have the HOLDACTIVE option set, ensuring that |
| 190 | traffic bursts from a certain queue are serviced by the same CPU. |
| 191 | This ensures a very low rate of frame reordering. A drawback of this |
| 192 | is that only one CPU at a time can service the traffic received by a |
| 193 | certain interface when RSS is not enabled. |
| 194 | |
| 195 | To implement RSS, the DPAA Ethernet driver allocates an extra set of |
| 196 | 128 Rx frame queues that are configured to dedicated channels, in a |
| 197 | round-robin manner. The mapping of the frame queues to CPUs is now |
| 198 | hardcoded, there is no indirection table to move traffic for a certain |
| 199 | FQ (hash result) to another CPU. The ingress traffic arriving on one |
| 200 | of these frame queues will arrive at the same portal and will always |
| 201 | be processed by the same CPU. This ensures intra-flow order preservation |
| 202 | and workload distribution for multiple traffic flows. |
| 203 | |
| 204 | RSS can be turned off for a certain interface using ethtool, i.e. |
| 205 | |
| 206 | # ethtool -N fm1-mac9 rx-flow-hash tcp4 "" |
| 207 | |
| 208 | To turn it back on, one needs to set rx-flow-hash for tcp4/6 or udp4/6: |
| 209 | |
| 210 | # ethtool -N fm1-mac9 rx-flow-hash udp4 sfdn |
| 211 | |
| 212 | There is no independent control for individual protocols, any command |
| 213 | run for one of tcp4|udp4|ah4|esp4|sctp4|tcp6|udp6|ah6|esp6|sctp6 is |
| 214 | going to control the rx-flow-hashing for all protocols on that interface. |
| 215 | |
| 216 | Besides using the FMan Keygen computed hash for spreading traffic on the |
| 217 | 128 Rx FQs, the DPAA Ethernet driver also sets the skb hash value when |
| 218 | the NETIF_F_RXHASH feature is on (active by default). This can be turned |
| 219 | on or off through ethtool, i.e.: |
| 220 | |
| 221 | # ethtool -K fm1-mac9 rx-hashing off |
| 222 | # ethtool -k fm1-mac9 | grep hash |
| 223 | receive-hashing: off |
| 224 | # ethtool -K fm1-mac9 rx-hashing on |
| 225 | Actual changes: |
| 226 | receive-hashing: on |
| 227 | # ethtool -k fm1-mac9 | grep hash |
| 228 | receive-hashing: on |
| 229 | |
| 230 | Please note that Rx hashing depends upon the rx-flow-hashing being on |
| 231 | for that interface - turning off rx-flow-hashing will also disable the |
| 232 | rx-hashing (without ethtool reporting it as off as that depends on the |
| 233 | NETIF_F_RXHASH feature flag). |
| 234 | |
Madalin Bucur | 5ee6842 | 2017-05-29 09:22:06 +0300 | [diff] [blame] | 235 | Debugging |
| 236 | ========= |
| 237 | |
| 238 | The following statistics are exported for each interface through ethtool: |
| 239 | |
| 240 | - interrupt count per CPU |
| 241 | - Rx packets count per CPU |
| 242 | - Tx packets count per CPU |
| 243 | - Tx confirmed packets count per CPU |
| 244 | - Tx S/G frames count per CPU |
| 245 | - Tx error count per CPU |
| 246 | - Rx error count per CPU |
| 247 | - Rx error count per type |
| 248 | - congestion related statistics: |
| 249 | - congestion status |
| 250 | - time spent in congestion |
| 251 | - number of time the device entered congestion |
| 252 | - dropped packets count per cause |
| 253 | |
| 254 | The driver also exports the following information in sysfs: |
| 255 | |
| 256 | - the FQ IDs for each FQ type |
| 257 | /sys/devices/platform/dpaa-ethernet.0/net/<int>/fqids |
| 258 | |
| 259 | - the IDs of the buffer pools in use |
| 260 | /sys/devices/platform/dpaa-ethernet.0/net/<int>/bpids |