dpdk-18.11网卡多队列RSS设置

背景

最近在作将基于dpdk-16.11.1开发的程序,转移到基于dpdk-18.11版本下开发。遇到了网卡RSS配置的问题,在这里纪录一下。函数

问题

dpdk-16.11.1

在dpdk-16.11.1上的程序以下:测试

static uint8_t rss_intel_key[40] = { 0x6D, 0x5A, 0x6D, 0x5A, 0x6D, 0x5A, 0x6D, 0x5A, 0x6D, 0x5A, 0x6D, 0x5A, 0x6D, 0x5A, 0x6D, 0x5A, 0x6D, 0x5A, 0x6D, 0x5A, 0x6D, 0x5A, 0x6D, 0x5A, 0x6D, 0x5A, 0x6D, 0x5A, 0x6D, 0x5A, 0x6D, 0x5A, 0x6D, 0x5A, 0x6D, 0x5A, 0x6D, 0x5A, 0x6D, 0x5A }; static const struct rte_eth_conf port_conf = { .rxmode = { .mq_mode = ETH_MQ_RX_RSS, .split_hdr_size = 0, .header_split = 0, /**< Header Split disabled */ .hw_ip_checksum = 0, /**< IP checksum offload disabled */ .hw_vlan_filter = 0, /**< VLAN filtering disabled */ .jumbo_frame = 0, /**< Jumbo Frame Support disabled */ .hw_strip_crc = 0, /**< CRC stripped by hardware */ }, .rx_adv_conf = { .rss_conf = { .rss_key = rss_intel_key, .rss_hf = ETH_RSS_PROTO_MASK, }, }, .txmode = { .mq_mode = ETH_MQ_TX_NONE, }, };

rte_eth_dev_configure函数声明ui

int rte_eth_dev_configure(uint16_t port_id, uint16_t nb_rx_queue, uint16_t nb_tx_queue, const struct rte_eth_conf *eth_conf);

不管是千兆网卡仍是万兆网卡,在调用  rte_eth_dev_configure(port_id, nb_rx_queue, nb_tx_queue, &port_conf)  函数时均可以正常启动。spa

dpdk-18.11

dpdk-18.11中对 struct rte_eth_conf 结构体进行了修改,使用offloads代替了以前的几个变量标志位,不过这都影响不大。code

static uint8_t rss_intel_key[40] = { 0x6D, 0x5A, 0x6D, 0x5A, 0x6D, 0x5A, 0x6D, 0x5A, 0x6D, 0x5A, 0x6D, 0x5A, 0x6D, 0x5A, 0x6D, 0x5A, 0x6D, 0x5A, 0x6D, 0x5A, 0x6D, 0x5A, 0x6D, 0x5A, 0x6D, 0x5A, 0x6D, 0x5A, 0x6D, 0x5A, 0x6D, 0x5A, 0x6D, 0x5A, 0x6D, 0x5A, 0x6D, 0x5A, 0x6D, 0x5A }; static const struct rte_eth_conf port_conf = { .rxmode = { .mq_mode = ETH_MQ_RX_RSS, .max_rx_pkt_len = 0, .split_hdr_size = 0, .offloads = 0, }, .rx_adv_conf = { .rss_conf = { .rss_key = rss_intel_key, .rss_key_len = 40, .rss_hf = ETH_RSS_PROTO_MASK, }, }, .txmode = { .mq_mode = ETH_MQ_TX_NONE, }, };

在调用  rte_eth_dev_configure(port_id, nb_rx_queue, nb_tx_queue, &port_conf)  函数是和16.11.1上是一致的。blog

在编译完成后,进行运行时发现,报错没法运行。报错以下:ip

Ethdev port_id=0 invalid rss_hf: 0x3ffffc, valid value: 0x38d34

port_id 0是一个igb驱动的网卡,型号为I350。开发

报错的意思是在调用 rte_eth_dev_configure 函数时,传入的最后一个函数参数(也就是网卡配置)中的 rx_adv_conf.rss_conf.rss_hf 参数值有问题,这个值是无效的。有效值是 0x38d34。get

找到dpdk源码中打印错误信息的位置,先后代码以下:源码

int rte_eth_dev_rss_hash_update(uint16_t port_id, struct rte_eth_rss_conf *rss_conf) { struct rte_eth_dev *dev; struct rte_eth_dev_info dev_info = { .flow_type_rss_offloads = 0, }; RTE_ETH_VALID_PORTID_OR_ERR_RET(port_id, -ENODEV); dev = &rte_eth_devices[port_id]; rte_eth_dev_info_get(port_id, &dev_info); if ((dev_info.flow_type_rss_offloads | rss_conf->rss_hf) != dev_info.flow_type_rss_offloads) { RTE_ETHDEV_LOG(ERR, "Ethdev port_id=%u invalid rss_hf: 0x%"PRIx64", valid value: 0x%"PRIx64"\n", port_id, rss_conf->rss_hf, dev_info.flow_type_rss_offloads); return -EINVAL; } RTE_FUNC_PTR_OR_ERR_RET(*dev->dev_ops->rss_hash_update, -ENOTSUP); return eth_err(port_id, (*dev->dev_ops->rss_hash_update)(dev, rss_conf)); }

 

 dev_info.flow_type_rss_offloads 就是网卡支持的flow类型, rss_conf->rss_hf 就是咱们调用 rte_eth_dev_configure 函数的最后一个参考中的  rx_adv_conf.rss_conf.rss_hf 的值,

这里的if判断意思是配置的flow类型必须是网卡支持的flow类型,若是配置了网卡不支持的类型,就会报错。

通过计算器算出 0x38d34的二进制是111000110100110100,再配合在 rte_ethdev.h 中 ETH_RSS_开头的宏定义,得出111000110100110100就是下面全部宏定义的 或 值:

ETH_RSS_IPV4 | \ ETH_RSS_NONFRAG_IPV4_TCP| \ ETH_RSS_NONFRAG_IPV4_UDP| \ ETH_RSS_IPV6 | \ ETH_RSS_NONFRAG_IPV6_TCP | \ ETH_RSS_NONFRAG_IPV6_UDP | \ ETH_RSS_IPV6_EX | \ ETH_RSS_IPV6_TCP_EX | \ ETH_RSS_IPV6_UDP_EX

而咱们配置中的参数  ETH_RSS_PROTO_MASK 显然比上面的类型要多,也就是 ETH_RSS_PROTO_MASK 定义的一些类型,网卡不支持,故而报错。

咱们再看16.11.1中的源码

int rte_eth_dev_rss_hash_update(uint8_t port_id, struct rte_eth_rss_conf *rss_conf) { struct rte_eth_dev *dev; uint16_t rss_hash_protos; RTE_ETH_VALID_PORTID_OR_ERR_RET(port_id, -ENODEV); rss_hash_protos = rss_conf->rss_hf; if ((rss_hash_protos != 0) && ((rss_hash_protos & ETH_RSS_PROTO_MASK) == 0)) { RTE_PMD_DEBUG_TRACE("Invalid rss_hash_protos=0x%x\n", rss_hash_protos); return -EINVAL; } dev = &rte_eth_devices[port_id]; RTE_FUNC_PTR_OR_ERR_RET(*dev->dev_ops->rss_hash_update, -ENOTSUP); return (*dev->dev_ops->rss_hash_update)(dev, rss_conf); }

显然判断逻辑不一样的。这里的判断意思是,只要有配置,而且配置中的flow类型至少有1项是在  ETH_RSS_PROTO_MASK  中的就能够,除非你配置的类型都不在  ETH_RSS_PROTO_MASK 中才会报错,这里就能够看出,即便你配置了网卡不支持的特性,也不会报错。

至此问题已解决。

总结

在dpdk-18.11中,配置网卡rss,必须网卡支持这种特性才行,不然会报错。

下面是整理的 igb , ixgbe ,  i40e 驱动网卡所支持的flow类型。

igb支持的特性

测试使用I350及I211网卡,有效值为0x38d34,二进制为111000110100110100,对应宏定义为

#define ETH_RSS_E1000_IGB (\ ETH_RSS_IPV4 | \ ETH_RSS_NONFRAG_IPV4_TCP| \ ETH_RSS_NONFRAG_IPV4_UDP| \ ETH_RSS_IPV6 | \ ETH_RSS_NONFRAG_IPV6_TCP | \ ETH_RSS_NONFRAG_IPV6_UDP | \ ETH_RSS_IPV6_EX | \ ETH_RSS_IPV6_TCP_EX | \ ETH_RSS_IPV6_UDP_EX)

ixgbe支持的特性

测试使用82599ES网卡,有效值和igb相同

#define ETH_RSS_IXGBE ETH_RSS_E1000_IGB

i40e支持的特性

测试使用X710网卡,有效值为0x7ef8,二进制为111111011111000,对应宏定义为

#define ETH_RSS_I40E (\ ETH_RSS_FRAG_IPV4 | \ ETH_RSS_NONFRAG_IPV4_TCP | \ ETH_RSS_NONFRAG_IPV4_UDP | \ ETH_RSS_NONFRAG_IPV4_SCTP | \ ETH_RSS_NONFRAG_IPV4_OTHER | \ ETH_RSS_FRAG_IPV6 | \ ETH_RSS_NONFRAG_IPV6_TCP | \ ETH_RSS_NONFRAG_IPV6_UDP | \ ETH_RSS_NONFRAG_IPV6_SCTP | \ ETH_RSS_NONFRAG_IPV6_OTHER | \ ETH_RSS_L2_PAYLOAD)