[dpdk-users] Intel 10G X550T NIC takes a long time(25s+) after DPDK initialization to passing traffic correctly.

[Jun Gan <geminijun@gmail.com>]

Hi,



We have an intel 10G X550T NIC, after bind it to igb_uio and launched our
dpdk application.



Although I don’t see any error reported while calling “rte_eth_tx_burst”, I
cannot see any packets on peer side, until at least 25 seconds later.
Everything finally became normal.



I don’t see this issue with my Mellanox ConnectX-5 NIC, using same code.



My dpdk init code, nothing special I think:



Status InitDpdk() {

  Status status = SpdkHelper::Init();

  if (status != kStatusOk) {

    return GeneralError::EC_FAILURE;

  }



  rc = rte_openlog_stream(stderr);

  assert(rc == 0);



  int dev_socket_id = rte_eth_dev_socket_id(eth_port_);



  rte_eth_dev_info_get(eth_port_, &dev_info_);

  if ((dev_info_.tx_offload_capa & DEV_TX_OFFLOAD_MULTI_SEGS) == 0) {

    return GeneralError::EC_FAILURE;

  }



  int size_pktpool =

      2048 //dev_info_.rx_desc_lim.nb_max + dev_info_.tx_desc_lim.nb_max

          + 512 /* PKTPOOL_EXTRA_SIZE */;



  // Packet mem pool for tx

  tx_pkt_pool_ = rte_pktmbuf_pool_create("pkt_pool_tx",

                                         size_pktpool,

                                         32 /* per lcore cache size*/,

                                         0,

                                         RTE_MBUF_DEFAULT_BUF_SIZE,

                                         dev_socket_id);

  assert(tx_pkt_pool_ != NULL);



  // Packet mem pool for rx

  rx_pkt_pool_ = rte_pktmbuf_pool_create("pkt_pool_rx",

                                         size_pktpool,

                                         32 /* per lcore cache size*/,

                                         0,

                                         RTE_MBUF_DEFAULT_BUF_SIZE,

                                         dev_socket_id);

  assert(rx_pkt_pool_ != NULL);



  struct rte_eth_conf cfg_port;

  memset(&cfg_port, 0, sizeof(cfg_port));

  cfg_port.txmode.mq_mode = ETH_MQ_TX_NONE;

  cfg_port.txmode.offloads |= DEV_TX_OFFLOAD_MULTI_SEGS;



  // Reference for how to offload:

  // https://doc.dpdk.org/guides/prog_guide/mbuf_lib.html

  // Section "Meta Information".

  // Also rte_mbuf.h and examples.



  // Offload Tx checksums if possible

  std::string offloads;

  if (dev_info_.tx_offload_capa & DEV_TX_OFFLOAD_IPV4_CKSUM) {

    offloads += " IPv4 Tx cksum";

    cfg_port.txmode.offloads |= DEV_TX_OFFLOAD_IPV4_CKSUM;

  }

  if (dev_info_.tx_offload_capa & DEV_TX_OFFLOAD_UDP_CKSUM) {

    offloads += " UDP Tx cksum";

    cfg_port.txmode.offloads |= DEV_TX_OFFLOAD_UDP_CKSUM;

  }

  // Offload Rx checksums if possible

  if (dev_info_.rx_offload_capa & DEV_RX_OFFLOAD_IPV4_CKSUM) {

    offloads += " IPv4 Rx cksum";

    cfg_port.rxmode.offloads |= DEV_RX_OFFLOAD_IPV4_CKSUM;

  }

  if (dev_info_.rx_offload_capa & DEV_RX_OFFLOAD_UDP_CKSUM) {

    offloads += " UDP Rx cksum";

    cfg_port.rxmode.offloads |= DEV_RX_OFFLOAD_UDP_CKSUM;

  }



  rc = rte_eth_dev_configure(eth_port_, 1, 1, &cfg_port);

  assert(rc == 0);

  if (rc) {

    return GeneralError::EC_FAILURE;

  }



  uint16_t n_rx_desc = 1024;

  uint16_t n_tx_desc = 1024;

  rc = rte_eth_dev_adjust_nb_rx_tx_desc(eth_port_, &n_rx_desc, &n_tx_desc);

  assert(rc == 0);

  if (rc) {

    return GeneralError::EC_FAILURE;

  }



  struct rte_eth_rxconf rxconf;

  rxconf = dev_info_.default_rxconf;

  rc = rte_eth_rx_queue_setup(eth_port_, 0, n_rx_desc,

                              dev_socket_id, &rxconf, rx_pkt_pool_);

  assert(rc == 0);

  if (rc) {

    return GeneralError::EC_FAILURE;

  }



  struct rte_eth_txconf txconf;

  txconf = dev_info_.default_txconf;

  rc = rte_eth_tx_queue_setup(eth_port_, 0, n_tx_desc,

                              dev_socket_id, &txconf);

  assert(rc == 0);

  if (rc) {

    return GeneralError::EC_FAILURE;

  }



  rc = rte_eth_dev_start(eth_port_);

  assert(rc == 0);

  if (rc) {

    return GeneralError::EC_FAILURE;

  }



  // RTE Ring for our TX packets, with multiple producers but single
consumer

  tx_ring_ = rte_ring_create("TX Ring", tx_ring_size_,

                             dev_socket_id, RING_F_SC_DEQ);

  assert(tx_ring_ != nullptr);

  if (tx_ring_ == nullptr) {

    return GeneralError::EC_FAILURE;

  }

  return kStatusOk;

}



Really appreciate it if someone can help!

-- 
Jun Gan