DPDK l2fwd源码学习
1. 主函数分析
1 /* 命令行解析 2 * 参数输入 ./l2fwd -c 0x3 -n 4 -- -p 3 -q 1 3 * -c 为十六进制的分配的逻辑内核数量 4 * -n 为十进制的内存通道数量,EAL参数和程序参数用--分开 5 * -q 为分配给每个核心的收发队列数量(端口数量) 6 * -p为十六进制的分配的端口数 7 * -t 为可选默认10s打印时间间隔参数 8 */ 9 int main(int argc, char **argv) 10 { 11 struct lcore_queue_conf *qconf; 12 int ret; 13 uint16_t nb_ports; 14 uint16_t nb_ports_available = 0; 15 uint16_t portid, last_port; 16 unsigned lcore_id, rx_lcore_id; 17 unsigned nb_ports_in_mask = 0; 18 unsigned int nb_lcores = 0; 19 unsigned int nb_mbufs; 20 21 /* init EAL */ 22 /* 初始化EAL参数,并解析参数,系统函数getopt以及getopt_long, 23 * 这些处理命令行参数的函数,处理到“--”时就会停止,分割参 24 */ 25 ret = rte_eal_init(argc, argv); 26 if (ret < 0) 27 rte_exit(EXIT_FAILURE, "Invalid EAL arguments\n"); 28 //argc减去EAL参数的同时,argv加上EAL的参数,保证解析程序参数的时候已经跳过了EAL参数 29 argc -= ret; 30 argv += ret; 31 32 force_quit = false; 33 signal(SIGINT, signal_handler); 34 signal(SIGTERM, signal_handler); 35 36 /* parse application arguments (after the EAL ones) */ 37 //解析l2fwd程序参数 38 ret = l2fwd_parse_args(argc, argv); 39 if (ret < 0) 40 rte_exit(EXIT_FAILURE, "Invalid L2FWD arguments\n"); 41 42 printf("MAC updating %s\n", mac_updating ? "enabled" : "disabled"); 43 44 /* convert to number of cycles */ 45 //-t参数,打印时间间隔 46 timer_period *= rte_get_timer_hz(); 47 48 nb_ports = rte_eth_dev_count_avail(); 49 if (nb_ports == 0) 50 rte_exit(EXIT_FAILURE, "No Ethernet ports - bye\n"); 51 52 /* check port mask to possible port mask */ 53 /* 54 * DPDK运行时创建的大页内存中,创建报文内存池, 55 * 其中socket不是套接字,是numa框架中的socket, 56 * 每个socket都有数个node,每个node右包括数个core。 57 * 每个socket都有自己的内存,每个socket里的处理器访问自己内存的速度最快, 58 * 访问其他socket的内存则较慢。 59 */ 60 if (l2fwd_enabled_port_mask & ~((1 << nb_ports) - 1)) 61 rte_exit(EXIT_FAILURE, "Invalid portmask; possible (0x%x)\n", 62 (1 << nb_ports) - 1); 63 64 /* reset l2fwd_dst_ports */ 65 //设置二层转发目的端口 66 for (portid = 0; portid < RTE_MAX_ETHPORTS; portid++) 67 l2fwd_dst_ports[portid] = 0; 68 //初始化所有的目的端口为0 69 last_port = 0; 70 71 /* 72 * Each logical core is assigned a dedicated TX queue on each port. 73 */ 74 RTE_ETH_FOREACH_DEV(portid) { 75 /* skip ports that are not enabled */ 76 /* l2fwd_enabled_port_mask 可用端口位掩码 77 * 跳过未分配或是不可用端口。 78 * 可用端口位掩码表示,左数第n位如果为1,表示端口n可用,如果左数第n位如果为0,表示端口n不可用。 79 * 要得到第x位为1还是0,我们的方法是将1左移x位,得到一个只在x位为1,其他位都为0的数,再与位掩码相与。 80 * 结果为1,那么第x位为1,结果位0,那么第x位为0. 81 */ 82 if ((l2fwd_enabled_port_mask & (1 << portid)) == 0) 83 continue; 84 //此处,当输入端口数,即nb_ports为1时,dst_port[0] = 0; 85 //此处,当输入端口数,即nb_ports为2时,dst_port[0] = 0,dst_port[2] = 1,dst_port[1] = 2; 86 //此处,当输入端口数,即nb_ports为3时,dst_port[0] = 0,dst_port[2] = 1,dst_port[1] = 2; 87 //此处,当输入端口数,即nb_ports为4时,....dst_port[4] = 3,dst_port[3] = 4; 88 89 if (nb_ports_in_mask % 2) { 90 l2fwd_dst_ports[portid] = last_port; 91 l2fwd_dst_ports[last_port] = portid; 92 } 93 else 94 last_port = portid; 95 96 nb_ports_in_mask++; 97 } 98 if (nb_ports_in_mask % 2) { 99 printf("Notice: odd number of ports in portmask.\n"); 100 l2fwd_dst_ports[last_port] = last_port; 101 } 102 103 rx_lcore_id = 0; 104 qconf = NULL; 105 106 /* Initialize the port/queue configuration of each logical core */ 107 RTE_ETH_FOREACH_DEV(portid) { 108 /* skip ports that are not enabled */ 109 if ((l2fwd_enabled_port_mask & (1 << portid)) == 0) 110 continue; 111 112 /* get the lcore_id for this port */ 113 //l2fwd_rx_queue_per_lcore即参数-q 114 while (rte_lcore_is_enabled(rx_lcore_id) == 0 || 115 lcore_queue_conf[rx_lcore_id].n_rx_port == 116 l2fwd_rx_queue_per_lcore) { 117 rx_lcore_id++; 118 if (rx_lcore_id >= RTE_MAX_LCORE) 119 rte_exit(EXIT_FAILURE, "Not enough cores\n"); 120 } 121 122 if (qconf != &lcore_queue_conf[rx_lcore_id]) { 123 /* Assigned a new logical core in the loop above. */ 124 qconf = &lcore_queue_conf[rx_lcore_id]; 125 nb_lcores++; 126 } 127 128 qconf->rx_port_list[qconf->n_rx_port] = portid; 129 qconf->n_rx_port++; 130 printf("Lcore %u: RX port %u\n", rx_lcore_id, portid); 131 } 132 133 nb_mbufs = RTE_MAX(nb_ports * (nb_rxd + nb_txd + MAX_PKT_BURST + 134 nb_lcores * MEMPOOL_CACHE_SIZE), 8192U); 135 136 /* create the mbuf pool */ 137 l2fwd_pktmbuf_pool = rte_pktmbuf_pool_create("mbuf_pool", nb_mbufs, 138 MEMPOOL_CACHE_SIZE, 0, RTE_MBUF_DEFAULT_BUF_SIZE, 139 rte_socket_id()); 140 if (l2fwd_pktmbuf_pool == NULL) 141 rte_exit(EXIT_FAILURE, "Cannot init mbuf pool\n"); 142 143 /* Initialise each port */ 144 RTE_ETH_FOREACH_DEV(portid) { 145 struct rte_eth_rxconf rxq_conf; 146 struct rte_eth_txconf txq_conf; 147 struct rte_eth_conf local_port_conf = port_conf; 148 struct rte_eth_dev_info dev_info; 149 150 /* skip ports that are not enabled */ 151 if ((l2fwd_enabled_port_mask & (1 << portid)) == 0) { 152 printf("Skipping disabled port %u\n", portid); 153 continue; 154 } 155 nb_ports_available++; 156 157 /* init port */ 158 printf("Initializing port %u... ", portid); 159 //清除读写缓冲区 160 fflush(stdout); 161 162 //配置端口,将一些配置写进设备dev的一些字段,以及检查设备支持什么类型的中断、支持的包大小 163 ret = rte_eth_dev_info_get(portid, &dev_info); 164 if (ret != 0) 165 rte_exit(EXIT_FAILURE, 166 "Error during getting device (port %u) info: %s\n", 167 portid, strerror(-ret)); 168 169 if (dev_info.tx_offload_capa & DEV_TX_OFFLOAD_MBUF_FAST_FREE) 170 local_port_conf.txmode.offloads |= 171 DEV_TX_OFFLOAD_MBUF_FAST_FREE; 172 ret = rte_eth_dev_configure(portid, 1, 1, &local_port_conf); 173 if (ret < 0) 174 rte_exit(EXIT_FAILURE, "Cannot configure device: err=%d, port=%u\n", 175 ret, portid); 176 177 ret = rte_eth_dev_adjust_nb_rx_tx_desc(portid, &nb_rxd, 178 &nb_txd); 179 if (ret < 0) 180 rte_exit(EXIT_FAILURE, 181 "Cannot adjust number of descriptors: err=%d, port=%u\n", 182 ret, portid); 183 184 //获取设备的MAC地址,存入l2fwd_ports_eth_addr[]数组,后续打印MAC地址 185 ret = rte_eth_macaddr_get(portid, 186 &l2fwd_ports_eth_addr[portid]); 187 if (ret < 0) 188 rte_exit(EXIT_FAILURE, 189 "Cannot get MAC address: err=%d, port=%u\n", 190 ret, portid); 191 192 /* init one RX queue */ 193 //清除读写缓冲区 194 fflush(stdout); 195 rxq_conf = dev_info.default_rxconf; 196 rxq_conf.offloads = local_port_conf.rxmode.offloads; 197 //设置接收队列,nb_rxd指收取队列的大小,最大能够存储mbuf的数量 198 ret = rte_eth_rx_queue_setup(portid, 0, nb_rxd, 199 rte_eth_dev_socket_id(portid), 200 &rxq_conf, 201 l2fwd_pktmbuf_pool); 202 if (ret < 0) 203 rte_exit(EXIT_FAILURE, "rte_eth_rx_queue_setup:err=%d, port=%u\n", 204 ret, portid); 205 206 /* init one TX queue on each port */ 207 fflush(stdout); 208 txq_conf = dev_info.default_txconf; 209 txq_conf.offloads = local_port_conf.txmode.offloads; 210 //初始化一个发送队列,nb_txd指发送队列的大小,最大能够存储mbuf的数量 211 ret = rte_eth_tx_queue_setup(portid, 0, nb_txd, 212 rte_eth_dev_socket_id(portid), 213 &txq_conf); 214 if (ret < 0) 215 rte_exit(EXIT_FAILURE, "rte_eth_tx_queue_setup:err=%d, port=%u\n", 216 ret, portid); 217 218 /* Initialize TX buffers */ 219 //为每个端口分配接收缓冲区,根据numa架构的socket就近分配 220 tx_buffer[portid] = rte_zmalloc_socket("tx_buffer", 221 RTE_ETH_TX_BUFFER_SIZE(MAX_PKT_BURST), 0, 222 rte_eth_dev_socket_id(portid)); 223 if (tx_buffer[portid] == NULL) 224 rte_exit(EXIT_FAILURE, "Cannot allocate buffer for tx on port %u\n", 225 portid); 226 227 rte_eth_tx_buffer_init(tx_buffer[portid], MAX_PKT_BURST); 228 229 ret = rte_eth_tx_buffer_set_err_callback(tx_buffer[portid], 230 rte_eth_tx_buffer_count_callback, 231 &port_statistics[portid].dropped); 232 if (ret < 0) 233 rte_exit(EXIT_FAILURE, 234 "Cannot set error callback for tx buffer on port %u\n", 235 portid); 236 237 ret = rte_eth_dev_set_ptypes(portid, RTE_PTYPE_UNKNOWN, NULL, 238 0); 239 if (ret < 0) 240 printf("Port %u, Failed to disable Ptype parsing\n", 241 portid); 242 /* Start device */ 243 //启动端口 244 ret = rte_eth_dev_start(portid); 245 if (ret < 0) 246 rte_exit(EXIT_FAILURE, "rte_eth_dev_start:err=%d, port=%u\n", 247 ret, portid); 248 249 printf("done: \n"); 250 251 ret = rte_eth_promiscuous_enable(portid); 252 if (ret != 0) 253 rte_exit(EXIT_FAILURE, 254 "rte_eth_promiscuous_enable:err=%s, port=%u\n", 255 rte_strerror(-ret), portid); 256 257 printf("Port %u, MAC address: %02X:%02X:%02X:%02X:%02X:%02X\n\n", 258 portid, 259 l2fwd_ports_eth_addr[portid].addr_bytes[0], 260 l2fwd_ports_eth_addr[portid].addr_bytes[1], 261 l2fwd_ports_eth_addr[portid].addr_bytes[2], 262 l2fwd_ports_eth_addr[portid].addr_bytes[3], 263 l2fwd_ports_eth_addr[portid].addr_bytes[4], 264 l2fwd_ports_eth_addr[portid].addr_bytes[5]); 265 266 /* initialize port stats */ 267 //初始化端口数据,就是后面要打印的,接收、发送、drop的包数 268 memset(&port_statistics, 0, sizeof(port_statistics)); 269 } 270 271 if (!nb_ports_available) { 272 rte_exit(EXIT_FAILURE, 273 "All available ports are disabled. Please set portmask.\n"); 274 } 275 276 277 //检查每个端口的连接状态 278 check_all_ports_link_status(l2fwd_enabled_port_mask); 279 280 ret = 0; 281 /* launch per-lcore init on every lcore */ 282 //在每个逻辑内核上启动线程,开始转发,l2fwd_launch_one_lcore实际上运行的是l2fwd_main_loop 283 rte_eal_mp_remote_launch(l2fwd_launch_one_lcore, NULL, CALL_MASTER); 284 RTE_LCORE_FOREACH_SLAVE(lcore_id) { 285 if (rte_eal_wait_lcore(lcore_id) < 0) { 286 ret = -1; 287 break; 288 } 289 } 290 291 RTE_ETH_FOREACH_DEV(portid) { 292 if ((l2fwd_enabled_port_mask & (1 << portid)) == 0) 293 continue; 294 printf("Closing port %d...", portid); 295 rte_eth_dev_stop(portid); 296 rte_eth_dev_close(portid); 297 printf(" Done\n"); 298 } 299 printf("Bye...\n"); 300 301 return ret; 302 }
程序的主要流程如下:
二. 二层转发和普通的端口转发区别:
特点 | L2fwd | basicfwd |
---|---|---|
端口数量 | 两者都用端口掩码来指定,L2fwd支持奇数个 | 只能是偶数个 |
lcore数量 | 多个,每个lcore负责一个port | 一个lcore,执行类似repeater的程序 |
转发逻辑 | 转发时会改写MAC地址 | 只能是 0<-->1,2<-->3 这样的 pair 互相转发 |
Tx_buffer | 有发包缓存队列,收的包会缓存到发包队列里,一段时间后或者队列满后才会转发 | 没有发包缓存,Rx收到包后直接Tx出去 |
三. 任务分发
每个逻辑核在任务分发后会执行如下的循环,直到退出:
40 /* 41 * Check that every SLAVE lcores are in WAIT state, then call 42 * rte_eal_remote_launch() for all of them. If call_master is true 43 * (set to CALL_MASTER), also call the function on the master lcore. 44 */ 45 int 46 rte_eal_mp_remote_launch(int (*f)(void *), void *arg, 47 enum rte_rmt_call_master_t call_master) 48 { 49 int lcore_id; 50 int master = rte_get_master_lcore(); 51 52 /* check state of lcores */ 53 RTE_LCORE_FOREACH_SLAVE(lcore_id) { 54 if (lcore_config[lcore_id].state != WAIT) 55 return -EBUSY; 56 } 57 58 /* send messages to cores */ 59 RTE_LCORE_FOREACH_SLAVE(lcore_id) { 60 rte_eal_remote_launch(f, arg, lcore_id); 61 } 62 63 if (call_master == CALL_MASTER) { 64 lcore_config[master].ret = f(arg); 65 lcore_config[master].state = FINISHED; 66 } 67 68 return 0; 69 }
rte_eal_mp_remote_launch(l2fwd_launch_one_lcore, NULL, CALL_MASTER)
283 static int 284 l2fwd_launch_one_lcore(__attribute__((unused)) void *dummy) 285 { 286 l2fwd_main_loop(); 287 return 0; 288 }
1 /* main processing loop */ 2 static void 3 l2fwd_main_loop(void) 4 { 5 struct rte_mbuf *pkts_burst[MAX_PKT_BURST]; 6 struct rte_mbuf *m; 7 int sent; 8 unsigned lcore_id; 9 uint64_t prev_tsc, diff_tsc, cur_tsc, timer_tsc; 10 unsigned i, j, portid, nb_rx; 11 struct lcore_queue_conf *qconf; 12 const uint64_t drain_tsc = (rte_get_tsc_hz() + US_PER_S - 1) / US_PER_S * 13 BURST_TX_DRAIN_US; 14 struct rte_eth_dev_tx_buffer *buffer; 15 16 prev_tsc = 0; 17 timer_tsc = 0; 18 19 //获取自己的lcore_id 20 lcore_id = rte_lcore_id(); 21 qconf = &lcore_queue_conf[lcore_id]; 22 23 //分配后多余的lcore,无事可做,orz 24 if (qconf->n_rx_port == 0) { 25 RTE_LOG(INFO, L2FWD, "lcore %u has nothing to do\n", lcore_id); 26 return; 27 } 28 29 //有事做的核,很开心的进入了主循环~ 30 RTE_LOG(INFO, L2FWD, "entering main loop on lcore %u\n", lcore_id); 31 32 for (i = 0; i < qconf->n_rx_port; i++) { 33 34 portid = qconf->rx_port_list[i]; 35 RTE_LOG(INFO, L2FWD, " -- lcoreid=%u portid=%u\n", lcore_id, 36 portid); 37 38 } 39 40 //直到发生了强制退出,在这里就是ctrl+c或者kill了这个进程 41 while (!force_quit) { 42 43 cur_tsc = rte_rdtsc(); 44 45 /* 46 * TX burst queue drain 47 */ 48 //计算时间片 49 diff_tsc = cur_tsc - prev_tsc; 50 //过了100us,把发送buffer里的报文发出去 51 if (unlikely(diff_tsc > drain_tsc)) { 52 53 for (i = 0; i < qconf->n_rx_port; i++) { 54 55 portid = l2fwd_dst_ports[qconf->rx_port_list[i]]; 56 buffer = tx_buffer[portid]; 57 58 sent = rte_eth_tx_buffer_flush(portid, 0, buffer); 59 if (sent) 60 port_statistics[portid].tx += sent; 61 62 } 63 64 /* if timer is enabled */ 65 //到了时间片了打印各端口的数据 66 if (timer_period > 0) { 67 68 /* advance the timer */ 69 timer_tsc += diff_tsc; 70 71 /* if timer has reached its timeout */ 72 if (unlikely(timer_tsc >= timer_period)) { 73 74 /* do this only on master core */ 75 if (lcore_id == rte_get_master_lcore()) { 76 //打印让master主线程来做 77 print_stats(); 78 /* reset the timer */ 79 timer_tsc = 0; 80 } 81 } 82 } 83 84 prev_tsc = cur_tsc; 85 } 86 87 /* 88 * Read packet from RX queues 89 */ 90 //没有到发送时间片的话,读接收队列里的报文 91 for (i = 0; i < qconf->n_rx_port; i++) { 92 93 portid = qconf->rx_port_list[i]; 94 nb_rx = rte_eth_rx_burst(portid, 0, 95 pkts_burst, MAX_PKT_BURST); 96 97 //计数,收到的报文数 98 port_statistics[portid].rx += nb_rx; 99 100 for (j = 0; j < nb_rx; j++) { 101 m = pkts_burst[j]; 102 rte_prefetch0(rte_pktmbuf_mtod(m, void *)); 103 //updating mac地址以及目的端口发送buffer满了的话,尝试发送 104 l2fwd_simple_forward(m, portid); 105 } 106 } 107 } 108 }
流程图:
四. 测试实验
* 参数输入 ./l2fwd -c 0x3 -n 4 -- -p 3 -q 1 * -c 为十六进制的分配的逻辑内核数量 * -n 为十进制的内存通道数量,EAL参数和程序参数用--分开 * -q 为分配给每个核心的收发队列数量(端口数量) * -p为十六进制的分配的端口数 * -t 为可选默认10s打印时间间隔参数