26. HNS3 Poll Mode Driver

The hns3 PMD (librte_net_hns3) provides poll mode driver support for the inbuilt HiSilicon Network Subsystem(HNS) network engine found in the HiSilicon Kunpeng 920 SoC and Kunpeng 930 SoC .

26.1. Features

Features of the HNS3 PMD are:

  • Multiple queues for TX and RX
  • Receive Side Scaling (RSS)
  • Packet type information
  • Checksum offload
  • TSO offload
  • LRO offload
  • Promiscuous mode
  • Multicast mode
  • Port hardware statistics
  • Jumbo frames
  • Link state information
  • Interrupt mode for RX
  • VLAN stripping and inserting
  • QinQ inserting
  • DCB
  • Scattered and gather for TX and RX
  • Vector Poll mode driver
  • Dump register
  • SR-IOV VF
  • Multi-process
  • MAC/VLAN filter
  • MTU update
  • NUMA support
  • Generic flow API
  • IEEE1588/802.1AS timestamping

26.2. Prerequisites

26.3. Configuration

26.3.1. Compilation Options

The following options can be modified in the config/rte_config.h file.

  • RTE_LIBRTE_HNS3_MAX_TQP_NUM_PER_PF (default 256)

    Number of MAX queues reserved for PF.

26.3.2. Runtime Configuration

  • rx_func_hint (default none)

    Used to select Rx burst function, supported value are vec, sve, simple, common. vec, if supported use the vec Rx function which indicates the default vector algorithm, neon for Kunpeng Arm platform. sve, if supported use the sve Rx function which indicates the sve algorithm. simple, if supported use the simple Rx function which indicates the scalar simple algorithm. common, if supported use the common Rx function which indicates the scalar scattered algorithm.

    When provided parameter is not supported, vec usage condition will be first checked, if meets, use the vec. Then, simple, at last common.

    For example:: -a 0000:7d:00.0,rx_func_hint=simple

  • tx_func_hint (default none)

    Used to select Tx burst function, supported value are vec, sve, simple, common. vec, if supported use the vec Tx function which indicates the default vector algorithm, neon for Kunpeng Arm platform. sve, if supported use the sve Tx function which indicates the sve algorithm. simple, if supported use the simple Tx function which indicates the scalar simple algorithm. common, if supported use the common Tx function which indicates the scalar algorithm.

    When provided parameter is not supported, vec usage condition will be first checked, if meets, use the vec. Then, simple, at last common.

    For example:: -a 0000:7d:00.0,tx_func_hint=common

  • dev_caps_mask (default 0)

    Used to mask the capability which queried from firmware. This args take hexadecimal bitmask where each bit represents whether mask corresponding capability. eg. If the capability is 0xFFFF queried from firmware, and the args value is 0xF which means the bit0~bit3 should be masked off, then the capability will be 0xFFF0. Its main purpose is to debug and avoid problems.

    For example:: -a 0000:7d:00.0,dev_caps_mask=0xF

  • mbx_time_limit_ms (default 500)

    Used to define the mailbox time limit by user. Current, the max waiting time for MBX response is 500ms, but in some scenarios, it is not enough. Since it depends on the response of the kernel mode driver, and its response time is related to the scheduling of the system. In this special scenario, most of the cores are isolated, and only a few cores are used for system scheduling. When a large number of services are started, the scheduling of the system will be very busy, and the reply of the mbx message will time out, which will cause our PMD initialization to fail. So provide access to set mailbox time limit for user.

    For example:: -a 0000:7d:00.0,mbx_time_limit_ms=600

26.4. Driver compilation and testing

Refer to the document compiling and testing a PMD for a NIC for details.

26.5. Sample Application Notes

26.5.1. VLAN filter

VLAN filter only works when Promiscuous mode is off.

To start testpmd, and add VLAN 10 to port 0:

./<build_dir>/app/dpdk-testpmd -l 0-15 -n 4 -- -i --forward-mode=mac
...

testpmd> set promisc 0 off
testpmd> vlan set filter on 0
testpmd> rx_vlan add 10 0

26.5.2. Flow Director

The Flow Director works in receive mode to identify specific flows or sets of flows and route them to specific queues. The Flow Director filters can match the different fields for different type of packet: flow type, specific input set per flow type.

Start testpmd:

./<build_dir>/app/dpdk-testpmd -l 0-15 -n 4 -- -i --rxq=8 --txq=8 \
                               --nb-cores=8 --nb-ports=1

Add a rule to direct ipv4-udp packet whose dst_ip=2.2.2.5, src_ip=2.2.2.3, src_port=32, dst_port=32 to queue 1:

testpmd> flow create 0 ingress pattern eth / ipv4 src is 2.2.2.3 \
         dst is 2.2.2.5 / udp src is 32 dst is 32 / end \
         actions mark id 1 / queue index 1 / end

26.5.3. Generic flow API

  • RSS Flow

    RSS Flow supports to set hash input set, hash function, enable hash and configure queues. For example: Configure queues as queue 0, 1, 2, 3.

    testpmd> flow create 0 ingress pattern end actions rss types end \
  queues 0 1 2 3 end / end

    Enable hash and set input set for IPv4-TCP.

    testpmd> flow create 0 ingress pattern eth / ipv4 / tcp / end \
  actions rss types ipv4-tcp l3-src-only end queues end / end

    Set symmetric hash enable for flow type IPv4-TCP.

    testpmd> flow create 0 ingress pattern eth / ipv4 / tcp / end \
  actions rss types ipv4-tcp end queues end func symmetric_toeplitz / end

    Set hash function as simple xor.

    testpmd> flow create 0 ingress pattern end actions rss types end \
  queues end func simple_xor / end

26.6. Statistics

HNS3 supports various methods to report statistics:

Port statistics can be queried using rte_eth_stats_get(). The number of packets received or sent successfully by the PMD. While the received and sent packet bytes are through SW only. The imissed counter is the amount of packets that could not be delivered to SW because a queue was full. The oerror counter is the amount of packets that are dropped by HW in Tx.

Extended statistics can be queried using rte_eth_xstats_get(). The extended statistics expose a wider set of counters counted by the device. The extended port statistics contains packets statistics per queue, Mac statistics, HW reset count and IO error count.

Finally per-flow statistics can by queried using rte_flow_query when attaching a count action for specific flow. The flow counter counts the number of packets received successfully by the port and match the specific flow.

26.7. Performance tuning

26.7.1. Hardware configuration

32 GB DIMMs is used to ensure that each channel is fully configured. Dynamic CPU Tuning is disabled.

26.7.2. Queue depth configuration

According to the actual test, the performance is best when the queue depth ranges from 1024 to 2048.

26.7.3. IO burst configuration

According to the actual test, the performance is best when IO burst is set to 64. IO burst is the number of packets per burst.

26.7.4. Queue number configuration

When the number of port queues corresponds to the number of CPU cores, the performance will be better.

26.7.5. Hugepage configuration

For 4K systems, 1 GB hugepages are recommended. For 64 KB systems, 512 MB hugepages are recommended.

26.7.6. CPU core isolation

To reduce the possibility of context switching, kernel isolation parameter should be provided to avoid scheduling the CPU core used by DPDK application threads for other tasks. Before starting the Linux OS, add the kernel isolation boot parameter. For example, “isolcpus=1-18 nohz_full=1-18 rcu_nocbs=1-18”.

26.8. Limitations or Known issues

Currently, we only support VF device driven by DPDK driver when PF is driven by kernel mode hns3 ethdev driver. VF is not supported when PF is driven by DPDK driver.

For sake of Rx/Tx performance, IEEE 1588 is not supported when using vec or sve burst function. When enabling IEEE 1588, Rx/Tx burst mode should be simple or common. It is recommended that enable IEEE 1588 before ethdev start. In this way, the correct Rx/Tx burst function can be selected.

Build with ICC is not supported yet. X86-32, Power8, ARMv7 and BSD are not supported yet.