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_eal.h>
  52 #include <rte_per_lcore.h>
  53 #include <rte_launch.h>
  54 #include <rte_atomic.h>
  55 #include <rte_cycles.h>
  56 #include <rte_prefetch.h>
  57 #include <rte_lcore.h>
  58 #include <rte_per_lcore.h>
  59 #include <rte_branch_prediction.h>
  60 #include <rte_interrupts.h>
  61 #include <rte_pci.h>
  62 #include <rte_random.h>
  63 #include <rte_debug.h>
  64 #include <rte_ether.h>
  65 #include <rte_ethdev.h>
  66 #include <rte_ring.h>
  67 #include <rte_mempool.h>
  68 #include <rte_mbuf.h>
  69 #include <rte_string_fns.h>
  70 #include <rte_ip.h>
  71 #include <rte_tcp.h>
  72 #include <rte_lpm.h>
  73
  74 #include "main.h"
  75
  76 static struct rte_eth_conf port_conf = {
  77         .rxmode = {
  78                 .mq_mode        = ETH_MQ_RX_RSS,
  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_IP,
  90                 },
  91         },
  92         .txmode = {
  93                 .mq_mode = ETH_MQ_TX_NONE,
  94         },
  95 };
  96
  97 static void
  98 app_assign_worker_ids(void)
  99 {
 100         uint32_t lcore, worker_id;
 101
 102         /* Assign ID for each worker */
 103         worker_id = 0;
 104         for (lcore = 0; lcore < APP_MAX_LCORES; lcore ++) {
 105                 struct app_lcore_params_worker *lp_worker = &app.lcore_params[lcore].worker;
 106
 107                 if (app.lcore_params[lcore].type != e_APP_LCORE_WORKER) {
 108                         continue;
 109                 }
 110
 111                 lp_worker->worker_id = worker_id;
 112                 worker_id ++;
 113         }
 114 }
 115
 116 static void
 117 app_init_mbuf_pools(void)
 118 {
 119         unsigned socket, lcore;
 120
 121         /* Init the buffer pools */
 122         for (socket = 0; socket < APP_MAX_SOCKETS; socket ++) {
 123                 char name[32];
 124                 if (app_is_socket_used(socket) == 0) {
 125                         continue;
 126                 }
 127
 128                 snprintf(name, sizeof(name), "mbuf_pool_%u", socket);
 129                 printf("Creating the mbuf pool for socket %u ...\n", socket);
 130                 app.pools[socket] = rte_pktmbuf_pool_create(
 131                         name, APP_DEFAULT_MEMPOOL_BUFFERS,
 132                         APP_DEFAULT_MEMPOOL_CACHE_SIZE,
 133                         0, APP_DEFAULT_MBUF_DATA_SIZE, socket);
 134                 if (app.pools[socket] == NULL) {
 135                         rte_panic("Cannot create mbuf pool on socket %u\n", socket);
 136                 }
 137         }
 138
 139         for (lcore = 0; lcore < APP_MAX_LCORES; lcore ++) {
 140                 if (app.lcore_params[lcore].type == e_APP_LCORE_DISABLED) {
 141                         continue;
 142                 }
 143
 144                 socket = rte_lcore_to_socket_id(lcore);
 145                 app.lcore_params[lcore].pool = app.pools[socket];
 146         }
 147 }
 148
 149 static void
 150 app_init_lpm_tables(void)
 151 {
 152         unsigned socket, lcore;
 153
 154         /* Init the LPM tables */
 155         for (socket = 0; socket < APP_MAX_SOCKETS; socket ++) {
 156                 char name[32];
 157                 uint32_t rule;
 158
 159                 if (app_is_socket_used(socket) == 0) {
 160                         continue;
 161                 }
 162
 163                 snprintf(name, sizeof(name), "lpm_table_%u", socket);
 164                 printf("Creating the LPM table for socket %u ...\n", socket);
 165                 app.lpm_tables[socket] = rte_lpm_create(
 166                         name,
 167                         socket,
 168                         APP_MAX_LPM_RULES,
 169                         0);
 170                 if (app.lpm_tables[socket] == NULL) {
 171                         rte_panic("Unable to create LPM table on socket %u\n", socket);
 172                 }
 173
 174                 for (rule = 0; rule < app.n_lpm_rules; rule ++) {
 175                         int ret;
 176
 177                         ret = rte_lpm_add(app.lpm_tables[socket],
 178                                 app.lpm_rules[rule].ip,
 179                                 app.lpm_rules[rule].depth,
 180                                 app.lpm_rules[rule].if_out);
 181
 182                         if (ret < 0) {
 183                                 rte_panic("Unable to add entry %u (%x/%u => %u) to the LPM table on socket %u (%d)\n",
 184                                         (unsigned) rule,
 185                                         (unsigned) app.lpm_rules[rule].ip,
 186                                         (unsigned) app.lpm_rules[rule].depth,
 187                                         (unsigned) app.lpm_rules[rule].if_out,
 188                                         socket,
 189                                         ret);
 190                         }
 191                 }
 192
 193         }
 194
 195         for (lcore = 0; lcore < APP_MAX_LCORES; lcore ++) {
 196                 if (app.lcore_params[lcore].type != e_APP_LCORE_WORKER) {
 197                         continue;
 198                 }
 199
 200                 socket = rte_lcore_to_socket_id(lcore);
 201                 app.lcore_params[lcore].worker.lpm_table = app.lpm_tables[socket];
 202         }
 203 }
 204
 205 static void
 206 app_init_rings_rx(void)
 207 {
 208         unsigned lcore;
 209
 210         /* Initialize the rings for the RX side */
 211         for (lcore = 0; lcore < APP_MAX_LCORES; lcore ++) {
 212                 struct app_lcore_params_io *lp_io = &app.lcore_params[lcore].io;
 213                 unsigned socket_io, lcore_worker;
 214
 215                 if ((app.lcore_params[lcore].type != e_APP_LCORE_IO) ||
 216                     (lp_io->rx.n_nic_queues == 0)) {
 217                         continue;
 218                 }
 219
 220                 socket_io = rte_lcore_to_socket_id(lcore);
 221
 222                 for (lcore_worker = 0; lcore_worker < APP_MAX_LCORES; lcore_worker ++) {
 223                         char name[32];
 224                         struct app_lcore_params_worker *lp_worker = &app.lcore_params[lcore_worker].worker;
 225                         struct rte_ring *ring = NULL;
 226
 227                         if (app.lcore_params[lcore_worker].type != e_APP_LCORE_WORKER) {
 228                                 continue;
 229                         }
 230
 231                         printf("Creating ring to connect I/O lcore %u (socket %u) with worker lcore %u ...\n",
 232                                 lcore,
 233                                 socket_io,
 234                                 lcore_worker);
 235                         snprintf(name, sizeof(name), "app_ring_rx_s%u_io%u_w%u",
 236                                 socket_io,
 237                                 lcore,
 238                                 lcore_worker);
 239                         ring = rte_ring_create(
 240                                 name,
 241                                 app.ring_rx_size,
 242                                 socket_io,
 243                                 RING_F_SP_ENQ | RING_F_SC_DEQ);
 244                         if (ring == NULL) {
 245                                 rte_panic("Cannot create ring to connect I/O core %u with worker core %u\n",
 246                                         lcore,
 247                                         lcore_worker);
 248                         }
 249
 250                         lp_io->rx.rings[lp_io->rx.n_rings] = ring;
 251                         lp_io->rx.n_rings ++;
 252
 253                         lp_worker->rings_in[lp_worker->n_rings_in] = ring;
 254                         lp_worker->n_rings_in ++;
 255                 }
 256         }
 257
 258         for (lcore = 0; lcore < APP_MAX_LCORES; lcore ++) {
 259                 struct app_lcore_params_io *lp_io = &app.lcore_params[lcore].io;
 260
 261                 if ((app.lcore_params[lcore].type != e_APP_LCORE_IO) ||
 262                     (lp_io->rx.n_nic_queues == 0)) {
 263                         continue;
 264                 }
 265
 266                 if (lp_io->rx.n_rings != app_get_lcores_worker()) {
 267                         rte_panic("Algorithmic error (I/O RX rings)\n");
 268                 }
 269         }
 270
 271         for (lcore = 0; lcore < APP_MAX_LCORES; lcore ++) {
 272                 struct app_lcore_params_worker *lp_worker = &app.lcore_params[lcore].worker;
 273
 274                 if (app.lcore_params[lcore].type != e_APP_LCORE_WORKER) {
 275                         continue;
 276                 }
 277
 278                 if (lp_worker->n_rings_in != app_get_lcores_io_rx()) {
 279                         rte_panic("Algorithmic error (worker input rings)\n");
 280                 }
 281         }
 282 }
 283
 284 static void
 285 app_init_rings_tx(void)
 286 {
 287         unsigned lcore;
 288
 289         /* Initialize the rings for the TX side */
 290         for (lcore = 0; lcore < APP_MAX_LCORES; lcore ++) {
 291                 struct app_lcore_params_worker *lp_worker = &app.lcore_params[lcore].worker;
 292                 unsigned port;
 293
 294                 if (app.lcore_params[lcore].type != e_APP_LCORE_WORKER) {
 295                         continue;
 296                 }
 297
 298                 for (port = 0; port < APP_MAX_NIC_PORTS; port ++) {
 299                         char name[32];
 300                         struct app_lcore_params_io *lp_io = NULL;
 301                         struct rte_ring *ring;
 302                         uint32_t socket_io, lcore_io;
 303
 304                         if (app.nic_tx_port_mask[port] == 0) {
 305                                 continue;
 306                         }
 307
 308                         if (app_get_lcore_for_nic_tx((uint8_t) port, &lcore_io) < 0) {
 309                                 rte_panic("Algorithmic error (no I/O core to handle TX of port %u)\n",
 310                                         port);
 311                         }
 312
 313                         lp_io = &app.lcore_params[lcore_io].io;
 314                         socket_io = rte_lcore_to_socket_id(lcore_io);
 315
 316                         printf("Creating ring to connect worker lcore %u with TX port %u (through I/O lcore %u) (socket %u) ...\n",
 317                                 lcore, port, (unsigned)lcore_io, (unsigned)socket_io);
 318                         snprintf(name, sizeof(name), "app_ring_tx_s%u_w%u_p%u", socket_io, lcore, port);
 319                         ring = rte_ring_create(
 320                                 name,
 321                                 app.ring_tx_size,
 322                                 socket_io,
 323                                 RING_F_SP_ENQ | RING_F_SC_DEQ);
 324                         if (ring == NULL) {
 325                                 rte_panic("Cannot create ring to connect worker core %u with TX port %u\n",
 326                                         lcore,
 327                                         port);
 328                         }
 329
 330                         lp_worker->rings_out[port] = ring;
 331                         lp_io->tx.rings[port][lp_worker->worker_id] = ring;
 332                 }
 333         }
 334
 335         for (lcore = 0; lcore < APP_MAX_LCORES; lcore ++) {
 336                 struct app_lcore_params_io *lp_io = &app.lcore_params[lcore].io;
 337                 unsigned i;
 338
 339                 if ((app.lcore_params[lcore].type != e_APP_LCORE_IO) ||
 340                     (lp_io->tx.n_nic_ports == 0)) {
 341                         continue;
 342                 }
 343
 344                 for (i = 0; i < lp_io->tx.n_nic_ports; i ++){
 345                         unsigned port, j;
 346
 347                         port = lp_io->tx.nic_ports[i];
 348                         for (j = 0; j < app_get_lcores_worker(); j ++) {
 349                                 if (lp_io->tx.rings[port][j] == NULL) {
 350                                         rte_panic("Algorithmic error (I/O TX rings)\n");
 351                                 }
 352                         }
 353                 }
 354         }
 355 }
 356
 357 /* Check the link status of all ports in up to 9s, and print them finally */
 358 static void
 359 check_all_ports_link_status(uint8_t port_num, uint32_t port_mask)
 360 {
 361 #define CHECK_INTERVAL 100 /* 100ms */
 362 #define MAX_CHECK_TIME 90 /* 9s (90 * 100ms) in total */
 363         uint8_t portid, count, all_ports_up, print_flag = 0;
 364         struct rte_eth_link link;
 365         uint32_t n_rx_queues, n_tx_queues;
 366
 367         printf("\nChecking link status");
 368         fflush(stdout);
 369         for (count = 0; count <= MAX_CHECK_TIME; count++) {
 370                 all_ports_up = 1;
 371                 for (portid = 0; portid < port_num; portid++) {
 372                         if ((port_mask & (1 << portid)) == 0)
 373                                 continue;
 374                         n_rx_queues = app_get_nic_rx_queues_per_port(portid);
 375                         n_tx_queues = app.nic_tx_port_mask[portid];
 376                         if ((n_rx_queues == 0) && (n_tx_queues == 0))
 377                                 continue;
 378                         memset(&link, 0, sizeof(link));
 379                         rte_eth_link_get_nowait(portid, &link);
 380                         /* print link status if flag set */
 381                         if (print_flag == 1) {
 382                                 if (link.link_status)
 383                                         printf("Port %d Link Up - speed %u "
 384                                                 "Mbps - %s\n", (uint8_t)portid,
 385                                                 (unsigned)link.link_speed,
 386                                 (link.link_duplex == ETH_LINK_FULL_DUPLEX) ?
 387                                         ("full-duplex") : ("half-duplex\n"));
 388                                 else
 389                                         printf("Port %d Link Down\n",
 390                                                         (uint8_t)portid);
 391                                 continue;
 392                         }
 393                         /* clear all_ports_up flag if any link down */
 394                         if (link.link_status == 0) {
 395                                 all_ports_up = 0;
 396                                 break;
 397                         }
 398                 }
 399                 /* after finally printing all link status, get out */
 400                 if (print_flag == 1)
 401                         break;
 402
 403                 if (all_ports_up == 0) {
 404                         printf(".");
 405                         fflush(stdout);
 406                         rte_delay_ms(CHECK_INTERVAL);
 407                 }
 408
 409                 /* set the print_flag if all ports up or timeout */
 410                 if (all_ports_up == 1 || count == (MAX_CHECK_TIME - 1)) {
 411                         print_flag = 1;
 412                         printf("done\n");
 413                 }
 414         }
 415 }
 416
 417 static void
 418 app_init_nics(void)
 419 {
 420         unsigned socket;
 421         uint32_t lcore;
 422         uint8_t port, queue;
 423         int ret;
 424         uint32_t n_rx_queues, n_tx_queues;
 425
 426         /* Init NIC ports and queues, then start the ports */
 427         for (port = 0; port < APP_MAX_NIC_PORTS; port ++) {
 428                 struct rte_mempool *pool;
 429
 430                 n_rx_queues = app_get_nic_rx_queues_per_port(port);
 431                 n_tx_queues = app.nic_tx_port_mask[port];
 432
 433                 if ((n_rx_queues == 0) && (n_tx_queues == 0)) {
 434                         continue;
 435                 }
 436
 437                 /* Init port */
 438                 printf("Initializing NIC port %u ...\n", (unsigned) port);
 439                 ret = rte_eth_dev_configure(
 440                         port,
 441                         (uint8_t) n_rx_queues,
 442                         (uint8_t) n_tx_queues,
 443                         &port_conf);
 444                 if (ret < 0) {
 445                         rte_panic("Cannot init NIC port %u (%d)\n", (unsigned) port, ret);
 446                 }
 447                 rte_eth_promiscuous_enable(port);
 448
 449                 /* Init RX queues */
 450                 for (queue = 0; queue < APP_MAX_RX_QUEUES_PER_NIC_PORT; queue ++) {
 451                         if (app.nic_rx_queue_mask[port][queue] == 0) {
 452                                 continue;
 453                         }
 454
 455                         app_get_lcore_for_nic_rx(port, queue, &lcore);
 456                         socket = rte_lcore_to_socket_id(lcore);
 457                         pool = app.lcore_params[lcore].pool;
 458
 459                         printf("Initializing NIC port %u RX queue %u ...\n",
 460                                 (unsigned) port,
 461                                 (unsigned) queue);
 462                         ret = rte_eth_rx_queue_setup(
 463                                 port,
 464                                 queue,
 465                                 (uint16_t) app.nic_rx_ring_size,
 466                                 socket,
 467                                 NULL,
 468                                 pool);
 469                         if (ret < 0) {
 470                                 rte_panic("Cannot init RX queue %u for port %u (%d)\n",
 471                                         (unsigned) queue,
 472                                         (unsigned) port,
 473                                         ret);
 474                         }
 475                 }
 476
 477                 /* Init TX queues */
 478                 if (app.nic_tx_port_mask[port] == 1) {
 479                         app_get_lcore_for_nic_tx(port, &lcore);
 480                         socket = rte_lcore_to_socket_id(lcore);
 481                         printf("Initializing NIC port %u TX queue 0 ...\n",
 482                                 (unsigned) port);
 483                         ret = rte_eth_tx_queue_setup(
 484                                 port,
 485                                 0,
 486                                 (uint16_t) app.nic_tx_ring_size,
 487                                 socket,
 488                                 NULL);
 489                         if (ret < 0) {
 490                                 rte_panic("Cannot init TX queue 0 for port %d (%d)\n",
 491                                         port,
 492                                         ret);
 493                         }
 494                 }
 495
 496                 /* Start port */
 497                 ret = rte_eth_dev_start(port);
 498                 if (ret < 0) {
 499                         rte_panic("Cannot start port %d (%d)\n", port, ret);
 500                 }
 501         }
 502
 503         check_all_ports_link_status(APP_MAX_NIC_PORTS, (~0x0));
 504 }
 505
 506 void
 507 app_init(void)
 508 {
 509         app_assign_worker_ids();
 510         app_init_mbuf_pools();
 511         app_init_lpm_tables();
 512         app_init_rings_rx();
 513         app_init_rings_tx();
 514         app_init_nics();
 515
 516         printf("Initialization completed.\n");
 517 }