examples/load_balancer/init.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2014 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                 .mq_mode        = ETH_MQ_RX_RSS,
  80                 .split_hdr_size = 0,
  81                 .header_split   = 0, /**< Header Split disabled */
  82                 .hw_ip_checksum = 1, /**< IP checksum offload enabled */
  83                 .hw_vlan_filter = 0, /**< VLAN filtering disabled */
  84                 .jumbo_frame    = 0, /**< Jumbo Frame Support disabled */
  85                 .hw_strip_crc   = 0, /**< CRC stripped by hardware */
  86         },
  87         .rx_adv_conf = {
  88                 .rss_conf = {
  89                         .rss_key = NULL,
  90                         .rss_hf = ETH_RSS_IP,
  91                 },
  92         },
  93         .txmode = {
  94                 .mq_mode = ETH_MQ_TX_NONE,
  95         },
  96 };
  97
  98 static void
  99 app_assign_worker_ids(void)
 100 {
 101         uint32_t lcore, worker_id;
 102
 103         /* Assign ID for each worker */
 104         worker_id = 0;
 105         for (lcore = 0; lcore < APP_MAX_LCORES; lcore ++) {
 106                 struct app_lcore_params_worker *lp_worker = &app.lcore_params[lcore].worker;
 107
 108                 if (app.lcore_params[lcore].type != e_APP_LCORE_WORKER) {
 109                         continue;
 110                 }
 111
 112                 lp_worker->worker_id = worker_id;
 113                 worker_id ++;
 114         }
 115 }
 116
 117 static void
 118 app_init_mbuf_pools(void)
 119 {
 120         unsigned socket, lcore;
 121
 122         /* Init the buffer pools */
 123         for (socket = 0; socket < APP_MAX_SOCKETS; socket ++) {
 124                 char name[32];
 125                 if (app_is_socket_used(socket) == 0) {
 126                         continue;
 127                 }
 128
 129                 snprintf(name, sizeof(name), "mbuf_pool_%u", socket);
 130                 printf("Creating the mbuf pool for socket %u ...\n", socket);
 131                 app.pools[socket] = rte_mempool_create(
 132                         name,
 133                         APP_DEFAULT_MEMPOOL_BUFFERS,
 134                         APP_DEFAULT_MBUF_SIZE,
 135                         APP_DEFAULT_MEMPOOL_CACHE_SIZE,
 136                         sizeof(struct rte_pktmbuf_pool_private),
 137                         rte_pktmbuf_pool_init, NULL,
 138                         rte_pktmbuf_init, NULL,
 139                         socket,
 140                         0);
 141                 if (app.pools[socket] == NULL) {
 142                         rte_panic("Cannot create mbuf pool on socket %u\n", socket);
 143                 }
 144         }
 145
 146         for (lcore = 0; lcore < APP_MAX_LCORES; lcore ++) {
 147                 if (app.lcore_params[lcore].type == e_APP_LCORE_DISABLED) {
 148                         continue;
 149                 }
 150
 151                 socket = rte_lcore_to_socket_id(lcore);
 152                 app.lcore_params[lcore].pool = app.pools[socket];
 153         }
 154 }
 155
 156 static void
 157 app_init_lpm_tables(void)
 158 {
 159         unsigned socket, lcore;
 160
 161         /* Init the LPM tables */
 162         for (socket = 0; socket < APP_MAX_SOCKETS; socket ++) {
 163                 char name[32];
 164                 uint32_t rule;
 165
 166                 if (app_is_socket_used(socket) == 0) {
 167                         continue;
 168                 }
 169
 170                 snprintf(name, sizeof(name), "lpm_table_%u", socket);
 171                 printf("Creating the LPM table for socket %u ...\n", socket);
 172                 app.lpm_tables[socket] = rte_lpm_create(
 173                         name,
 174                         socket,
 175                         APP_MAX_LPM_RULES,
 176                         0);
 177                 if (app.lpm_tables[socket] == NULL) {
 178                         rte_panic("Unable to create LPM table on socket %u\n", socket);
 179                 }
 180
 181                 for (rule = 0; rule < app.n_lpm_rules; rule ++) {
 182                         int ret;
 183
 184                         ret = rte_lpm_add(app.lpm_tables[socket],
 185                                 app.lpm_rules[rule].ip,
 186                                 app.lpm_rules[rule].depth,
 187                                 app.lpm_rules[rule].if_out);
 188
 189                         if (ret < 0) {
 190                                 rte_panic("Unable to add entry %u (%x/%u => %u) to the LPM table on socket %u (%d)\n",
 191                                         (unsigned) rule,
 192                                         (unsigned) app.lpm_rules[rule].ip,
 193                                         (unsigned) app.lpm_rules[rule].depth,
 194                                         (unsigned) app.lpm_rules[rule].if_out,
 195                                         socket,
 196                                         ret);
 197                         }
 198                 }
 199
 200         }
 201
 202         for (lcore = 0; lcore < APP_MAX_LCORES; lcore ++) {
 203                 if (app.lcore_params[lcore].type != e_APP_LCORE_WORKER) {
 204                         continue;
 205                 }
 206
 207                 socket = rte_lcore_to_socket_id(lcore);
 208                 app.lcore_params[lcore].worker.lpm_table = app.lpm_tables[socket];
 209         }
 210 }
 211
 212 static void
 213 app_init_rings_rx(void)
 214 {
 215         unsigned lcore;
 216
 217         /* Initialize the rings for the RX side */
 218         for (lcore = 0; lcore < APP_MAX_LCORES; lcore ++) {
 219                 struct app_lcore_params_io *lp_io = &app.lcore_params[lcore].io;
 220                 unsigned socket_io, lcore_worker;
 221
 222                 if ((app.lcore_params[lcore].type != e_APP_LCORE_IO) ||
 223                     (lp_io->rx.n_nic_queues == 0)) {
 224                         continue;
 225                 }
 226
 227                 socket_io = rte_lcore_to_socket_id(lcore);
 228
 229                 for (lcore_worker = 0; lcore_worker < APP_MAX_LCORES; lcore_worker ++) {
 230                         char name[32];
 231                         struct app_lcore_params_worker *lp_worker = &app.lcore_params[lcore_worker].worker;
 232                         struct rte_ring *ring = NULL;
 233
 234                         if (app.lcore_params[lcore_worker].type != e_APP_LCORE_WORKER) {
 235                                 continue;
 236                         }
 237
 238                         printf("Creating ring to connect I/O lcore %u (socket %u) with worker lcore %u ...\n",
 239                                 lcore,
 240                                 socket_io,
 241                                 lcore_worker);
 242                         snprintf(name, sizeof(name), "app_ring_rx_s%u_io%u_w%u",
 243                                 socket_io,
 244                                 lcore,
 245                                 lcore_worker);
 246                         ring = rte_ring_create(
 247                                 name,
 248                                 app.ring_rx_size,
 249                                 socket_io,
 250                                 RING_F_SP_ENQ | RING_F_SC_DEQ);
 251                         if (ring == NULL) {
 252                                 rte_panic("Cannot create ring to connect I/O core %u with worker core %u\n",
 253                                         lcore,
 254                                         lcore_worker);
 255                         }
 256
 257                         lp_io->rx.rings[lp_io->rx.n_rings] = ring;
 258                         lp_io->rx.n_rings ++;
 259
 260                         lp_worker->rings_in[lp_worker->n_rings_in] = ring;
 261                         lp_worker->n_rings_in ++;
 262                 }
 263         }
 264
 265         for (lcore = 0; lcore < APP_MAX_LCORES; lcore ++) {
 266                 struct app_lcore_params_io *lp_io = &app.lcore_params[lcore].io;
 267
 268                 if ((app.lcore_params[lcore].type != e_APP_LCORE_IO) ||
 269                     (lp_io->rx.n_nic_queues == 0)) {
 270                         continue;
 271                 }
 272
 273                 if (lp_io->rx.n_rings != app_get_lcores_worker()) {
 274                         rte_panic("Algorithmic error (I/O RX rings)\n");
 275                 }
 276         }
 277
 278         for (lcore = 0; lcore < APP_MAX_LCORES; lcore ++) {
 279                 struct app_lcore_params_worker *lp_worker = &app.lcore_params[lcore].worker;
 280
 281                 if (app.lcore_params[lcore].type != e_APP_LCORE_WORKER) {
 282                         continue;
 283                 }
 284
 285                 if (lp_worker->n_rings_in != app_get_lcores_io_rx()) {
 286                         rte_panic("Algorithmic error (worker input rings)\n");
 287                 }
 288         }
 289 }
 290
 291 static void
 292 app_init_rings_tx(void)
 293 {
 294         unsigned lcore;
 295
 296         /* Initialize the rings for the TX side */
 297         for (lcore = 0; lcore < APP_MAX_LCORES; lcore ++) {
 298                 struct app_lcore_params_worker *lp_worker = &app.lcore_params[lcore].worker;
 299                 unsigned port;
 300
 301                 if (app.lcore_params[lcore].type != e_APP_LCORE_WORKER) {
 302                         continue;
 303                 }
 304
 305                 for (port = 0; port < APP_MAX_NIC_PORTS; port ++) {
 306                         char name[32];
 307                         struct app_lcore_params_io *lp_io = NULL;
 308                         struct rte_ring *ring;
 309                         uint32_t socket_io, lcore_io;
 310
 311                         if (app.nic_tx_port_mask[port] == 0) {
 312                                 continue;
 313                         }
 314
 315                         if (app_get_lcore_for_nic_tx((uint8_t) port, &lcore_io) < 0) {
 316                                 rte_panic("Algorithmic error (no I/O core to handle TX of port %u)\n",
 317                                         port);
 318                         }
 319
 320                         lp_io = &app.lcore_params[lcore_io].io;
 321                         socket_io = rte_lcore_to_socket_id(lcore_io);
 322
 323                         printf("Creating ring to connect worker lcore %u with TX port %u (through I/O lcore %u) (socket %u) ...\n",
 324                                 lcore, port, (unsigned)lcore_io, (unsigned)socket_io);
 325                         snprintf(name, sizeof(name), "app_ring_tx_s%u_w%u_p%u", socket_io, lcore, port);
 326                         ring = rte_ring_create(
 327                                 name,
 328                                 app.ring_tx_size,
 329                                 socket_io,
 330                                 RING_F_SP_ENQ | RING_F_SC_DEQ);
 331                         if (ring == NULL) {
 332                                 rte_panic("Cannot create ring to connect worker core %u with TX port %u\n",
 333                                         lcore,
 334                                         port);
 335                         }
 336
 337                         lp_worker->rings_out[port] = ring;
 338                         lp_io->tx.rings[port][lp_worker->worker_id] = ring;
 339                 }
 340         }
 341
 342         for (lcore = 0; lcore < APP_MAX_LCORES; lcore ++) {
 343                 struct app_lcore_params_io *lp_io = &app.lcore_params[lcore].io;
 344                 unsigned i;
 345
 346                 if ((app.lcore_params[lcore].type != e_APP_LCORE_IO) ||
 347                     (lp_io->tx.n_nic_ports == 0)) {
 348                         continue;
 349                 }
 350
 351                 for (i = 0; i < lp_io->tx.n_nic_ports; i ++){
 352                         unsigned port, j;
 353
 354                         port = lp_io->tx.nic_ports[i];
 355                         for (j = 0; j < app_get_lcores_worker(); j ++) {
 356                                 if (lp_io->tx.rings[port][j] == NULL) {
 357                                         rte_panic("Algorithmic error (I/O TX rings)\n");
 358                                 }
 359                         }
 360                 }
 361         }
 362 }
 363
 364 /* Check the link status of all ports in up to 9s, and print them finally */
 365 static void
 366 check_all_ports_link_status(uint8_t port_num, uint32_t port_mask)
 367 {
 368 #define CHECK_INTERVAL 100 /* 100ms */
 369 #define MAX_CHECK_TIME 90 /* 9s (90 * 100ms) in total */
 370         uint8_t portid, count, all_ports_up, print_flag = 0;
 371         struct rte_eth_link link;
 372         uint32_t n_rx_queues, n_tx_queues;
 373
 374         printf("\nChecking link status");
 375         fflush(stdout);
 376         for (count = 0; count <= MAX_CHECK_TIME; count++) {
 377                 all_ports_up = 1;
 378                 for (portid = 0; portid < port_num; portid++) {
 379                         if ((port_mask & (1 << portid)) == 0)
 380                                 continue;
 381                         n_rx_queues = app_get_nic_rx_queues_per_port(portid);
 382                         n_tx_queues = app.nic_tx_port_mask[portid];
 383                         if ((n_rx_queues == 0) && (n_tx_queues == 0))
 384                                 continue;
 385                         memset(&link, 0, sizeof(link));
 386                         rte_eth_link_get_nowait(portid, &link);
 387                         /* print link status if flag set */
 388                         if (print_flag == 1) {
 389                                 if (link.link_status)
 390                                         printf("Port %d Link Up - speed %u "
 391                                                 "Mbps - %s\n", (uint8_t)portid,
 392                                                 (unsigned)link.link_speed,
 393                                 (link.link_duplex == ETH_LINK_FULL_DUPLEX) ?
 394                                         ("full-duplex") : ("half-duplex\n"));
 395                                 else
 396                                         printf("Port %d Link Down\n",
 397                                                         (uint8_t)portid);
 398                                 continue;
 399                         }
 400                         /* clear all_ports_up flag if any link down */
 401                         if (link.link_status == 0) {
 402                                 all_ports_up = 0;
 403                                 break;
 404                         }
 405                 }
 406                 /* after finally printing all link status, get out */
 407                 if (print_flag == 1)
 408                         break;
 409
 410                 if (all_ports_up == 0) {
 411                         printf(".");
 412                         fflush(stdout);
 413                         rte_delay_ms(CHECK_INTERVAL);
 414                 }
 415
 416                 /* set the print_flag if all ports up or timeout */
 417                 if (all_ports_up == 1 || count == (MAX_CHECK_TIME - 1)) {
 418                         print_flag = 1;
 419                         printf("done\n");
 420                 }
 421         }
 422 }
 423
 424 static void
 425 app_init_nics(void)
 426 {
 427         unsigned socket;
 428         uint32_t lcore;
 429         uint8_t port, queue;
 430         int ret;
 431         uint32_t n_rx_queues, n_tx_queues;
 432
 433         /* Init NIC ports and queues, then start the ports */
 434         for (port = 0; port < APP_MAX_NIC_PORTS; port ++) {
 435                 struct rte_mempool *pool;
 436
 437                 n_rx_queues = app_get_nic_rx_queues_per_port(port);
 438                 n_tx_queues = app.nic_tx_port_mask[port];
 439
 440                 if ((n_rx_queues == 0) && (n_tx_queues == 0)) {
 441                         continue;
 442                 }
 443
 444                 /* Init port */
 445                 printf("Initializing NIC port %u ...\n", (unsigned) port);
 446                 ret = rte_eth_dev_configure(
 447                         port,
 448                         (uint8_t) n_rx_queues,
 449                         (uint8_t) n_tx_queues,
 450                         &port_conf);
 451                 if (ret < 0) {
 452                         rte_panic("Cannot init NIC port %u (%d)\n", (unsigned) port, ret);
 453                 }
 454                 rte_eth_promiscuous_enable(port);
 455
 456                 /* Init RX queues */
 457                 for (queue = 0; queue < APP_MAX_RX_QUEUES_PER_NIC_PORT; queue ++) {
 458                         if (app.nic_rx_queue_mask[port][queue] == 0) {
 459                                 continue;
 460                         }
 461
 462                         app_get_lcore_for_nic_rx(port, queue, &lcore);
 463                         socket = rte_lcore_to_socket_id(lcore);
 464                         pool = app.lcore_params[lcore].pool;
 465
 466                         printf("Initializing NIC port %u RX queue %u ...\n",
 467                                 (unsigned) port,
 468                                 (unsigned) queue);
 469                         ret = rte_eth_rx_queue_setup(
 470                                 port,
 471                                 queue,
 472                                 (uint16_t) app.nic_rx_ring_size,
 473                                 socket,
 474                                 NULL,
 475                                 pool);
 476                         if (ret < 0) {
 477                                 rte_panic("Cannot init RX queue %u for port %u (%d)\n",
 478                                         (unsigned) queue,
 479                                         (unsigned) port,
 480                                         ret);
 481                         }
 482                 }
 483
 484                 /* Init TX queues */
 485                 if (app.nic_tx_port_mask[port] == 1) {
 486                         app_get_lcore_for_nic_tx(port, &lcore);
 487                         socket = rte_lcore_to_socket_id(lcore);
 488                         printf("Initializing NIC port %u TX queue 0 ...\n",
 489                                 (unsigned) port);
 490                         ret = rte_eth_tx_queue_setup(
 491                                 port,
 492                                 0,
 493                                 (uint16_t) app.nic_tx_ring_size,
 494                                 socket,
 495                                 NULL);
 496                         if (ret < 0) {
 497                                 rte_panic("Cannot init TX queue 0 for port %d (%d)\n",
 498                                         port,
 499                                         ret);
 500                         }
 501                 }
 502
 503                 /* Start port */
 504                 ret = rte_eth_dev_start(port);
 505                 if (ret < 0) {
 506                         rte_panic("Cannot start port %d (%d)\n", port, ret);
 507                 }
 508         }
 509
 510         check_all_ports_link_status(APP_MAX_NIC_PORTS, (~0x0));
 511 }
 512
 513 void
 514 app_init(void)
 515 {
 516         app_assign_worker_ids();
 517         app_init_mbuf_pools();
 518         app_init_lpm_tables();
 519         app_init_rings_rx();
 520         app_init_rings_tx();
 521         app_init_nics();
 522
 523         printf("Initialization completed.\n");
 524 }