examples/load_balancer/init.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2013 Intel Corporation. All rights reserved.
   5  *   All rights reserved.
   6  *
   7  *   Redistribution and use in source and binary forms, with or without
   8  *   modification, are permitted provided that the following conditions
   9  *   are met:
  10  *
  11  *     * Redistributions of source code must retain the above copyright
  12  *       notice, this list of conditions and the following disclaimer.
  13  *     * Redistributions in binary form must reproduce the above copyright
  14  *       notice, this list of conditions and the following disclaimer in
  15  *       the documentation and/or other materials provided with the
  16  *       distribution.
  17  *     * Neither the name of Intel Corporation nor the names of its
  18  *       contributors may be used to endorse or promote products derived
  19  *       from this software without specific prior written permission.
  20  *
  21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  32  */
  33
  34 #include <stdio.h>
  35 #include <stdlib.h>
  36 #include <stdint.h>
  37 #include <inttypes.h>
  38 #include <sys/types.h>
  39 #include <string.h>
  40 #include <sys/queue.h>
  41 #include <stdarg.h>
  42 #include <errno.h>
  43 #include <getopt.h>
  44
  45 #include <rte_common.h>
  46 #include <rte_byteorder.h>
  47 #include <rte_log.h>
  48 #include <rte_memory.h>
  49 #include <rte_memcpy.h>
  50 #include <rte_memzone.h>
  51 #include <rte_tailq.h>
  52 #include <rte_eal.h>
  53 #include <rte_per_lcore.h>
  54 #include <rte_launch.h>
  55 #include <rte_atomic.h>
  56 #include <rte_cycles.h>
  57 #include <rte_prefetch.h>
  58 #include <rte_lcore.h>
  59 #include <rte_per_lcore.h>
  60 #include <rte_branch_prediction.h>
  61 #include <rte_interrupts.h>
  62 #include <rte_pci.h>
  63 #include <rte_random.h>
  64 #include <rte_debug.h>
  65 #include <rte_ether.h>
  66 #include <rte_ethdev.h>
  67 #include <rte_ring.h>
  68 #include <rte_mempool.h>
  69 #include <rte_mbuf.h>
  70 #include <rte_string_fns.h>
  71 #include <rte_ip.h>
  72 #include <rte_tcp.h>
  73 #include <rte_lpm.h>
  74
  75 #include "main.h"
  76
  77 static struct rte_eth_conf port_conf = {
  78         .rxmode = {
  79                 .split_hdr_size = 0,
  80                 .header_split   = 0, /**< Header Split disabled */
  81                 .hw_ip_checksum = 1, /**< IP checksum offload enabled */
  82                 .hw_vlan_filter = 0, /**< VLAN filtering disabled */
  83                 .jumbo_frame    = 0, /**< Jumbo Frame Support disabled */
  84                 .hw_strip_crc   = 0, /**< CRC stripped by hardware */
  85         },
  86         .rx_adv_conf = {
  87                 .rss_conf = {
  88                         .rss_key = NULL,
  89                         .rss_hf = ETH_RSS_IPV4 | ETH_RSS_IPV6,
  90                 },
  91         },
  92         .txmode = {
  93                 .mq_mode = ETH_MQ_TX_NONE,
  94         },
  95 };
  96
  97 static struct rte_eth_rxconf rx_conf = {
  98         .rx_thresh = {
  99                 .pthresh = APP_DEFAULT_NIC_RX_PTHRESH,
 100                 .hthresh = APP_DEFAULT_NIC_RX_HTHRESH,
 101                 .wthresh = APP_DEFAULT_NIC_RX_WTHRESH,
 102         },
 103         .rx_free_thresh = APP_DEFAULT_NIC_RX_FREE_THRESH,
 104         .rx_drop_en = APP_DEFAULT_NIC_RX_DROP_EN,
 105 };
 106
 107 static struct rte_eth_txconf tx_conf = {
 108         .tx_thresh = {
 109                 .pthresh = APP_DEFAULT_NIC_TX_PTHRESH,
 110                 .hthresh = APP_DEFAULT_NIC_TX_HTHRESH,
 111                 .wthresh = APP_DEFAULT_NIC_TX_WTHRESH,
 112         },
 113         .tx_free_thresh = APP_DEFAULT_NIC_TX_FREE_THRESH,
 114         .tx_rs_thresh = APP_DEFAULT_NIC_TX_RS_THRESH,
 115 };
 116
 117 static void
 118 app_assign_worker_ids(void)
 119 {
 120         uint32_t lcore, worker_id;
 121
 122         /* Assign ID for each worker */
 123         worker_id = 0;
 124         for (lcore = 0; lcore < APP_MAX_LCORES; lcore ++) {
 125                 struct app_lcore_params_worker *lp_worker = &app.lcore_params[lcore].worker;
 126
 127                 if (app.lcore_params[lcore].type != e_APP_LCORE_WORKER) {
 128                         continue;
 129                 }
 130
 131                 lp_worker->worker_id = worker_id;
 132                 worker_id ++;
 133         }
 134 }
 135
 136 static void
 137 app_init_mbuf_pools(void)
 138 {
 139         unsigned socket, lcore;
 140
 141         /* Init the buffer pools */
 142         for (socket = 0; socket < APP_MAX_SOCKETS; socket ++) {
 143                 char name[32];
 144                 if (app_is_socket_used(socket) == 0) {
 145                         continue;
 146                 }
 147
 148                 rte_snprintf(name, sizeof(name), "mbuf_pool_%u", socket);
 149                 printf("Creating the mbuf pool for socket %u ...\n", socket);
 150                 app.pools[socket] = rte_mempool_create(
 151                         name,
 152                         APP_DEFAULT_MEMPOOL_BUFFERS,
 153                         APP_DEFAULT_MBUF_SIZE,
 154                         APP_DEFAULT_MEMPOOL_CACHE_SIZE,
 155                         sizeof(struct rte_pktmbuf_pool_private),
 156                         rte_pktmbuf_pool_init, NULL,
 157                         rte_pktmbuf_init, NULL,
 158                         socket,
 159                         0);
 160                 if (app.pools[socket] == NULL) {
 161                         rte_panic("Cannot create mbuf pool on socket %u\n", socket);
 162                 }
 163         }
 164
 165         for (lcore = 0; lcore < APP_MAX_LCORES; lcore ++) {
 166                 if (app.lcore_params[lcore].type == e_APP_LCORE_DISABLED) {
 167                         continue;
 168                 }
 169
 170                 socket = rte_lcore_to_socket_id(lcore);
 171                 app.lcore_params[lcore].pool = app.pools[socket];
 172         }
 173 }
 174
 175 static void
 176 app_init_lpm_tables(void)
 177 {
 178         unsigned socket, lcore;
 179
 180         /* Init the LPM tables */
 181         for (socket = 0; socket < APP_MAX_SOCKETS; socket ++) {
 182                 char name[32];
 183                 uint32_t rule;
 184
 185                 if (app_is_socket_used(socket) == 0) {
 186                         continue;
 187                 }
 188
 189                 rte_snprintf(name, sizeof(name), "lpm_table_%u", socket);
 190                 printf("Creating the LPM table for socket %u ...\n", socket);
 191                 app.lpm_tables[socket] = rte_lpm_create(
 192                         name,
 193                         socket,
 194                         APP_MAX_LPM_RULES,
 195                         0);
 196                 if (app.lpm_tables[socket] == NULL) {
 197                         rte_panic("Unable to create LPM table on socket %u\n", socket);
 198                 }
 199
 200                 for (rule = 0; rule < app.n_lpm_rules; rule ++) {
 201                         int ret;
 202
 203                         ret = rte_lpm_add(app.lpm_tables[socket],
 204                                 app.lpm_rules[rule].ip,
 205                                 app.lpm_rules[rule].depth,
 206                                 app.lpm_rules[rule].if_out);
 207
 208                         if (ret < 0) {
 209                                 rte_panic("Unable to add entry %u (%x/%u => %u) to the LPM table on socket %u (%d)\n",
 210                                         (unsigned) rule,
 211                                         (unsigned) app.lpm_rules[rule].ip,
 212                                         (unsigned) app.lpm_rules[rule].depth,
 213                                         (unsigned) app.lpm_rules[rule].if_out,
 214                                         socket,
 215                                         ret);
 216                         }
 217                 }
 218
 219         }
 220
 221         for (lcore = 0; lcore < APP_MAX_LCORES; lcore ++) {
 222                 if (app.lcore_params[lcore].type != e_APP_LCORE_WORKER) {
 223                         continue;
 224                 }
 225
 226                 socket = rte_lcore_to_socket_id(lcore);
 227                 app.lcore_params[lcore].worker.lpm_table = app.lpm_tables[socket];
 228         }
 229 }
 230
 231 static void
 232 app_init_rings_rx(void)
 233 {
 234         unsigned lcore;
 235
 236         /* Initialize the rings for the RX side */
 237         for (lcore = 0; lcore < APP_MAX_LCORES; lcore ++) {
 238                 struct app_lcore_params_io *lp_io = &app.lcore_params[lcore].io;
 239                 unsigned socket_io, lcore_worker;
 240
 241                 if ((app.lcore_params[lcore].type != e_APP_LCORE_IO) ||
 242                     (lp_io->rx.n_nic_queues == 0)) {
 243                         continue;
 244                 }
 245
 246                 socket_io = rte_lcore_to_socket_id(lcore);
 247
 248                 for (lcore_worker = 0; lcore_worker < APP_MAX_LCORES; lcore_worker ++) {
 249                         char name[32];
 250                         struct app_lcore_params_worker *lp_worker = &app.lcore_params[lcore_worker].worker;
 251                         struct rte_ring *ring = NULL;
 252
 253                         if (app.lcore_params[lcore_worker].type != e_APP_LCORE_WORKER) {
 254                                 continue;
 255                         }
 256
 257                         printf("Creating ring to connect I/O lcore %u (socket %u) with worker lcore %u ...\n",
 258                                 lcore,
 259                                 socket_io,
 260                                 lcore_worker);
 261                         rte_snprintf(name, sizeof(name), "app_ring_rx_s%u_io%u_w%u",
 262                                 socket_io,
 263                                 lcore,
 264                                 lcore_worker);
 265                         ring = rte_ring_create(
 266                                 name,
 267                                 app.ring_rx_size,
 268                                 socket_io,
 269                                 RING_F_SP_ENQ | RING_F_SC_DEQ);
 270                         if (ring == NULL) {
 271                                 rte_panic("Cannot create ring to connect I/O core %u with worker core %u\n",
 272                                         lcore,
 273                                         lcore_worker);
 274                         }
 275
 276                         lp_io->rx.rings[lp_io->rx.n_rings] = ring;
 277                         lp_io->rx.n_rings ++;
 278
 279                         lp_worker->rings_in[lp_worker->n_rings_in] = ring;
 280                         lp_worker->n_rings_in ++;
 281                 }
 282         }
 283
 284         for (lcore = 0; lcore < APP_MAX_LCORES; lcore ++) {
 285                 struct app_lcore_params_io *lp_io = &app.lcore_params[lcore].io;
 286
 287                 if ((app.lcore_params[lcore].type != e_APP_LCORE_IO) ||
 288                     (lp_io->rx.n_nic_queues == 0)) {
 289                         continue;
 290                 }
 291
 292                 if (lp_io->rx.n_rings != app_get_lcores_worker()) {
 293                         rte_panic("Algorithmic error (I/O RX rings)\n");
 294                 }
 295         }
 296
 297         for (lcore = 0; lcore < APP_MAX_LCORES; lcore ++) {
 298                 struct app_lcore_params_worker *lp_worker = &app.lcore_params[lcore].worker;
 299
 300                 if (app.lcore_params[lcore].type != e_APP_LCORE_WORKER) {
 301                         continue;
 302                 }
 303
 304                 if (lp_worker->n_rings_in != app_get_lcores_io_rx()) {
 305                         rte_panic("Algorithmic error (worker input rings)\n");
 306                 }
 307         }
 308 }
 309
 310 static void
 311 app_init_rings_tx(void)
 312 {
 313         unsigned lcore;
 314
 315         /* Initialize the rings for the TX side */
 316         for (lcore = 0; lcore < APP_MAX_LCORES; lcore ++) {
 317                 struct app_lcore_params_worker *lp_worker = &app.lcore_params[lcore].worker;
 318                 unsigned port;
 319
 320                 if (app.lcore_params[lcore].type != e_APP_LCORE_WORKER) {
 321                         continue;
 322                 }
 323
 324                 for (port = 0; port < APP_MAX_NIC_PORTS; port ++) {
 325                         char name[32];
 326                         struct app_lcore_params_io *lp_io = NULL;
 327                         struct rte_ring *ring;
 328                         uint32_t socket_io, lcore_io;
 329
 330                         if (app.nic_tx_port_mask[port] == 0) {
 331                                 continue;
 332                         }
 333
 334                         if (app_get_lcore_for_nic_tx((uint8_t) port, &lcore_io) < 0) {
 335                                 rte_panic("Algorithmic error (no I/O core to handle TX of port %u)\n",
 336                                         port);
 337                         }
 338
 339                         lp_io = &app.lcore_params[lcore_io].io;
 340                         socket_io = rte_lcore_to_socket_id(lcore_io);
 341
 342                         printf("Creating ring to connect worker lcore %u with TX port %u (through I/O lcore %u) (socket %u) ...\n",
 343                                 lcore, port, (unsigned)lcore_io, (unsigned)socket_io);
 344                         rte_snprintf(name, sizeof(name), "app_ring_tx_s%u_w%u_p%u", socket_io, lcore, port);
 345                         ring = rte_ring_create(
 346                                 name,
 347                                 app.ring_tx_size,
 348                                 socket_io,
 349                                 RING_F_SP_ENQ | RING_F_SC_DEQ);
 350                         if (ring == NULL) {
 351                                 rte_panic("Cannot create ring to connect worker core %u with TX port %u\n",
 352                                         lcore,
 353                                         port);
 354                         }
 355
 356                         lp_worker->rings_out[port] = ring;
 357                         lp_io->tx.rings[port][lp_worker->worker_id] = ring;
 358                 }
 359         }
 360
 361         for (lcore = 0; lcore < APP_MAX_LCORES; lcore ++) {
 362                 struct app_lcore_params_io *lp_io = &app.lcore_params[lcore].io;
 363                 unsigned i;
 364
 365                 if ((app.lcore_params[lcore].type != e_APP_LCORE_IO) ||
 366                     (lp_io->tx.n_nic_ports == 0)) {
 367                         continue;
 368                 }
 369
 370                 for (i = 0; i < lp_io->tx.n_nic_ports; i ++){
 371                         unsigned port, j;
 372
 373                         port = lp_io->tx.nic_ports[i];
 374                         for (j = 0; j < app_get_lcores_worker(); j ++) {
 375                                 if (lp_io->tx.rings[port][j] == NULL) {
 376                                         rte_panic("Algorithmic error (I/O TX rings)\n");
 377                                 }
 378                         }
 379                 }
 380         }
 381 }
 382
 383 /* Check the link status of all ports in up to 9s, and print them finally */
 384 static void
 385 check_all_ports_link_status(uint8_t port_num, uint32_t port_mask)
 386 {
 387 #define CHECK_INTERVAL 100 /* 100ms */
 388 #define MAX_CHECK_TIME 90 /* 9s (90 * 100ms) in total */
 389         uint8_t portid, count, all_ports_up, print_flag = 0;
 390         struct rte_eth_link link;
 391         uint32_t n_rx_queues, n_tx_queues;
 392
 393         printf("\nChecking link status");
 394         fflush(stdout);
 395         for (count = 0; count <= MAX_CHECK_TIME; count++) {
 396                 all_ports_up = 1;
 397                 for (portid = 0; portid < port_num; portid++) {
 398                         if ((port_mask & (1 << portid)) == 0)
 399                                 continue;
 400                         n_rx_queues = app_get_nic_rx_queues_per_port(portid);
 401                         n_tx_queues = app.nic_tx_port_mask[portid];
 402                         if ((n_rx_queues == 0) && (n_tx_queues == 0))
 403                                 continue;
 404                         memset(&link, 0, sizeof(link));
 405                         rte_eth_link_get_nowait(portid, &link);
 406                         /* print link status if flag set */
 407                         if (print_flag == 1) {
 408                                 if (link.link_status)
 409                                         printf("Port %d Link Up - speed %u "
 410                                                 "Mbps - %s\n", (uint8_t)portid,
 411                                                 (unsigned)link.link_speed,
 412                                 (link.link_duplex == ETH_LINK_FULL_DUPLEX) ?
 413                                         ("full-duplex") : ("half-duplex\n"));
 414                                 else
 415                                         printf("Port %d Link Down\n",
 416                                                         (uint8_t)portid);
 417                                 continue;
 418                         }
 419                         /* clear all_ports_up flag if any link down */
 420                         if (link.link_status == 0) {
 421                                 all_ports_up = 0;
 422                                 break;
 423                         }
 424                 }
 425                 /* after finally printing all link status, get out */
 426                 if (print_flag == 1)
 427                         break;
 428
 429                 if (all_ports_up == 0) {
 430                         printf(".");
 431                         fflush(stdout);
 432                         rte_delay_ms(CHECK_INTERVAL);
 433                 }
 434
 435                 /* set the print_flag if all ports up or timeout */
 436                 if (all_ports_up == 1 || count == (MAX_CHECK_TIME - 1)) {
 437                         print_flag = 1;
 438                         printf("done\n");
 439                 }
 440         }
 441 }
 442
 443 static void
 444 app_init_nics(void)
 445 {
 446         unsigned socket;
 447         uint32_t lcore;
 448         uint8_t port, queue;
 449         int ret;
 450         uint32_t n_rx_queues, n_tx_queues;
 451
 452         /* Init driver */
 453         printf("Initializing the PMD driver ...\n");
 454         if (rte_pmd_init_all() < 0) {
 455                 rte_panic("Cannot init PMD\n");
 456         }
 457
 458         if (rte_eal_pci_probe() < 0) {
 459                 rte_panic("Cannot probe PCI\n");
 460         }
 461
 462         /* Init NIC ports and queues, then start the ports */
 463         for (port = 0; port < APP_MAX_NIC_PORTS; port ++) {
 464                 struct rte_mempool *pool;
 465
 466                 n_rx_queues = app_get_nic_rx_queues_per_port(port);
 467                 n_tx_queues = app.nic_tx_port_mask[port];
 468
 469                 if ((n_rx_queues == 0) && (n_tx_queues == 0)) {
 470                         continue;
 471                 }
 472
 473                 /* Init port */
 474                 printf("Initializing NIC port %u ...\n", (unsigned) port);
 475                 ret = rte_eth_dev_configure(
 476                         port,
 477                         (uint8_t) n_rx_queues,
 478                         (uint8_t) n_tx_queues,
 479                         &port_conf);
 480                 if (ret < 0) {
 481                         rte_panic("Cannot init NIC port %u (%d)\n", (unsigned) port, ret);
 482                 }
 483                 rte_eth_promiscuous_enable(port);
 484
 485                 /* Init RX queues */
 486                 for (queue = 0; queue < APP_MAX_RX_QUEUES_PER_NIC_PORT; queue ++) {
 487                         if (app.nic_rx_queue_mask[port][queue] == 0) {
 488                                 continue;
 489                         }
 490
 491                         app_get_lcore_for_nic_rx(port, queue, &lcore);
 492                         socket = rte_lcore_to_socket_id(lcore);
 493                         pool = app.lcore_params[lcore].pool;
 494
 495                         printf("Initializing NIC port %u RX queue %u ...\n",
 496                                 (unsigned) port,
 497                                 (unsigned) queue);
 498                         ret = rte_eth_rx_queue_setup(
 499                                 port,
 500                                 queue,
 501                                 (uint16_t) app.nic_rx_ring_size,
 502                                 socket,
 503                                 &rx_conf,
 504                                 pool);
 505                         if (ret < 0) {
 506                                 rte_panic("Cannot init RX queue %u for port %u (%d)\n",
 507                                         (unsigned) queue,
 508                                         (unsigned) port,
 509                                         ret);
 510                         }
 511                 }
 512
 513                 /* Init TX queues */
 514                 if (app.nic_tx_port_mask[port] == 1) {
 515                         app_get_lcore_for_nic_tx(port, &lcore);
 516                         socket = rte_lcore_to_socket_id(lcore);
 517                         printf("Initializing NIC port %u TX queue 0 ...\n",
 518                                 (unsigned) port);
 519                         ret = rte_eth_tx_queue_setup(
 520                                 port,
 521                                 0,
 522                                 (uint16_t) app.nic_tx_ring_size,
 523                                 socket,
 524                                 &tx_conf);
 525                         if (ret < 0) {
 526                                 rte_panic("Cannot init TX queue 0 for port %d (%d)\n",
 527                                         port,
 528                                         ret);
 529                         }
 530                 }
 531
 532                 /* Start port */
 533                 ret = rte_eth_dev_start(port);
 534                 if (ret < 0) {
 535                         rte_panic("Cannot start port %d (%d)\n", port, ret);
 536                 }
 537         }
 538
 539         check_all_ports_link_status(APP_MAX_NIC_PORTS, (~0x0));
 540 }
 541
 542 void
 543 app_init(void)
 544 {
 545         app_assign_worker_ids();
 546         app_init_mbuf_pools();
 547         app_init_lpm_tables();
 548         app_init_rings_rx();
 549         app_init_rings_tx();
 550         app_init_nics();
 551
 552         printf("Initialization completed.\n");
 553 }