examples/distributor/main.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2017 Intel Corporation
   3  */
   4
   5 #include <stdint.h>
   6 #include <inttypes.h>
   7 #include <unistd.h>
   8 #include <signal.h>
   9 #include <getopt.h>
  10
  11 #include <rte_eal.h>
  12 #include <rte_ethdev.h>
  13 #include <rte_cycles.h>
  14 #include <rte_malloc.h>
  15 #include <rte_debug.h>
  16 #include <rte_prefetch.h>
  17 #include <rte_distributor.h>
  18 #include <rte_pause.h>
  19 #include <rte_power.h>
  20
  21 #define RX_RING_SIZE 1024
  22 #define TX_RING_SIZE 1024
  23 #define NUM_MBUFS ((64*1024)-1)
  24 #define MBUF_CACHE_SIZE 128
  25 #define BURST_SIZE 64
  26 #define SCHED_RX_RING_SZ 8192
  27 #define SCHED_TX_RING_SZ 65536
  28 #define BURST_SIZE_TX 32
  29
  30 #define RTE_LOGTYPE_DISTRAPP RTE_LOGTYPE_USER1
  31
  32 #define ANSI_COLOR_RED     "\x1b[31m"
  33 #define ANSI_COLOR_RESET   "\x1b[0m"
  34
  35 /* mask of enabled ports */
  36 static uint32_t enabled_port_mask;
  37 volatile uint8_t quit_signal;
  38 volatile uint8_t quit_signal_rx;
  39 volatile uint8_t quit_signal_dist;
  40 volatile uint8_t quit_signal_work;
  41 unsigned int power_lib_initialised;
  42
  43 static volatile struct app_stats {
  44         struct {
  45                 uint64_t rx_pkts;
  46                 uint64_t returned_pkts;
  47                 uint64_t enqueued_pkts;
  48                 uint64_t enqdrop_pkts;
  49         } rx __rte_cache_aligned;
  50         int pad1 __rte_cache_aligned;
  51
  52         struct {
  53                 uint64_t in_pkts;
  54                 uint64_t ret_pkts;
  55                 uint64_t sent_pkts;
  56                 uint64_t enqdrop_pkts;
  57         } dist __rte_cache_aligned;
  58         int pad2 __rte_cache_aligned;
  59
  60         struct {
  61                 uint64_t dequeue_pkts;
  62                 uint64_t tx_pkts;
  63                 uint64_t enqdrop_pkts;
  64         } tx __rte_cache_aligned;
  65         int pad3 __rte_cache_aligned;
  66
  67         uint64_t worker_pkts[64] __rte_cache_aligned;
  68
  69         int pad4 __rte_cache_aligned;
  70
  71         uint64_t worker_bursts[64][8] __rte_cache_aligned;
  72
  73         int pad5 __rte_cache_aligned;
  74
  75         uint64_t port_rx_pkts[64] __rte_cache_aligned;
  76         uint64_t port_tx_pkts[64] __rte_cache_aligned;
  77 } app_stats;
  78
  79 struct app_stats prev_app_stats;
  80
  81 static const struct rte_eth_conf port_conf_default = {
  82         .rxmode = {
  83                 .mq_mode = ETH_MQ_RX_RSS,
  84                 .max_rx_pkt_len = RTE_ETHER_MAX_LEN,
  85         },
  86         .txmode = {
  87                 .mq_mode = ETH_MQ_TX_NONE,
  88         },
  89         .rx_adv_conf = {
  90                 .rss_conf = {
  91                         .rss_hf = ETH_RSS_IP | ETH_RSS_UDP |
  92                                 ETH_RSS_TCP | ETH_RSS_SCTP,
  93                 }
  94         },
  95 };
  96
  97 struct output_buffer {
  98         unsigned count;
  99         struct rte_mbuf *mbufs[BURST_SIZE];
 100 };
 101
 102 static void print_stats(void);
 103
 104 /*
 105  * Initialises a given port using global settings and with the rx buffers
 106  * coming from the mbuf_pool passed as parameter
 107  */
 108 static inline int
 109 port_init(uint16_t port, struct rte_mempool *mbuf_pool)
 110 {
 111         struct rte_eth_conf port_conf = port_conf_default;
 112         const uint16_t rxRings = 1, txRings = rte_lcore_count() - 1;
 113         int retval;
 114         uint16_t q;
 115         uint16_t nb_rxd = RX_RING_SIZE;
 116         uint16_t nb_txd = TX_RING_SIZE;
 117         struct rte_eth_dev_info dev_info;
 118         struct rte_eth_txconf txconf;
 119
 120         if (!rte_eth_dev_is_valid_port(port))
 121                 return -1;
 122
 123         retval = rte_eth_dev_info_get(port, &dev_info);
 124         if (retval != 0) {
 125                 printf("Error during getting device (port %u) info: %s\n",
 126                                 port, strerror(-retval));
 127                 return retval;
 128         }
 129
 130         if (dev_info.tx_offload_capa & DEV_TX_OFFLOAD_MBUF_FAST_FREE)
 131                 port_conf.txmode.offloads |=
 132                         DEV_TX_OFFLOAD_MBUF_FAST_FREE;
 133
 134         port_conf.rx_adv_conf.rss_conf.rss_hf &=
 135                 dev_info.flow_type_rss_offloads;
 136         if (port_conf.rx_adv_conf.rss_conf.rss_hf !=
 137                         port_conf_default.rx_adv_conf.rss_conf.rss_hf) {
 138                 printf("Port %u modified RSS hash function based on hardware support,"
 139                         "requested:%#"PRIx64" configured:%#"PRIx64"\n",
 140                         port,
 141                         port_conf_default.rx_adv_conf.rss_conf.rss_hf,
 142                         port_conf.rx_adv_conf.rss_conf.rss_hf);
 143         }
 144
 145         retval = rte_eth_dev_configure(port, rxRings, txRings, &port_conf);
 146         if (retval != 0)
 147                 return retval;
 148
 149         retval = rte_eth_dev_adjust_nb_rx_tx_desc(port, &nb_rxd, &nb_txd);
 150         if (retval != 0)
 151                 return retval;
 152
 153         for (q = 0; q < rxRings; q++) {
 154                 retval = rte_eth_rx_queue_setup(port, q, nb_rxd,
 155                                                 rte_eth_dev_socket_id(port),
 156                                                 NULL, mbuf_pool);
 157                 if (retval < 0)
 158                         return retval;
 159         }
 160
 161         txconf = dev_info.default_txconf;
 162         txconf.offloads = port_conf.txmode.offloads;
 163         for (q = 0; q < txRings; q++) {
 164                 retval = rte_eth_tx_queue_setup(port, q, nb_txd,
 165                                                 rte_eth_dev_socket_id(port),
 166                                                 &txconf);
 167                 if (retval < 0)
 168                         return retval;
 169         }
 170
 171         retval = rte_eth_dev_start(port);
 172         if (retval < 0)
 173                 return retval;
 174
 175         struct rte_eth_link link;
 176         rte_eth_link_get_nowait(port, &link);
 177         while (!link.link_status) {
 178                 printf("Waiting for Link up on port %"PRIu16"\n", port);
 179                 sleep(1);
 180                 rte_eth_link_get_nowait(port, &link);
 181         }
 182
 183         if (!link.link_status) {
 184                 printf("Link down on port %"PRIu16"\n", port);
 185                 return 0;
 186         }
 187
 188         struct rte_ether_addr addr;
 189         rte_eth_macaddr_get(port, &addr);
 190         printf("Port %u MAC: %02"PRIx8" %02"PRIx8" %02"PRIx8
 191                         " %02"PRIx8" %02"PRIx8" %02"PRIx8"\n",
 192                         port,
 193                         addr.addr_bytes[0], addr.addr_bytes[1],
 194                         addr.addr_bytes[2], addr.addr_bytes[3],
 195                         addr.addr_bytes[4], addr.addr_bytes[5]);
 196
 197         rte_eth_promiscuous_enable(port);
 198
 199         return 0;
 200 }
 201
 202 struct lcore_params {
 203         unsigned worker_id;
 204         struct rte_distributor *d;
 205         struct rte_ring *rx_dist_ring;
 206         struct rte_ring *dist_tx_ring;
 207         struct rte_mempool *mem_pool;
 208 };
 209
 210 static int
 211 lcore_rx(struct lcore_params *p)
 212 {
 213         const uint16_t nb_ports = rte_eth_dev_count_avail();
 214         const int socket_id = rte_socket_id();
 215         uint16_t port;
 216         struct rte_mbuf *bufs[BURST_SIZE*2];
 217
 218         RTE_ETH_FOREACH_DEV(port) {
 219                 /* skip ports that are not enabled */
 220                 if ((enabled_port_mask & (1 << port)) == 0)
 221                         continue;
 222
 223                 if (rte_eth_dev_socket_id(port) > 0 &&
 224                                 rte_eth_dev_socket_id(port) != socket_id)
 225                         printf("WARNING, port %u is on remote NUMA node to "
 226                                         "RX thread.\n\tPerformance will not "
 227                                         "be optimal.\n", port);
 228         }
 229
 230         printf("\nCore %u doing packet RX.\n", rte_lcore_id());
 231         port = 0;
 232         while (!quit_signal_rx) {
 233
 234                 /* skip ports that are not enabled */
 235                 if ((enabled_port_mask & (1 << port)) == 0) {
 236                         if (++port == nb_ports)
 237                                 port = 0;
 238                         continue;
 239                 }
 240                 const uint16_t nb_rx = rte_eth_rx_burst(port, 0, bufs,
 241                                 BURST_SIZE);
 242                 if (unlikely(nb_rx == 0)) {
 243                         if (++port == nb_ports)
 244                                 port = 0;
 245                         continue;
 246                 }
 247                 app_stats.rx.rx_pkts += nb_rx;
 248
 249 /*
 250  * You can run the distributor on the rx core with this code. Returned
 251  * packets are then send straight to the tx core.
 252  */
 253 #if 0
 254         rte_distributor_process(d, bufs, nb_rx);
 255         const uint16_t nb_ret = rte_distributor_returned_pktsd,
 256                         bufs, BURST_SIZE*2);
 257
 258                 app_stats.rx.returned_pkts += nb_ret;
 259                 if (unlikely(nb_ret == 0)) {
 260                         if (++port == nb_ports)
 261                                 port = 0;
 262                         continue;
 263                 }
 264
 265                 struct rte_ring *tx_ring = p->dist_tx_ring;
 266                 uint16_t sent = rte_ring_enqueue_burst(tx_ring,
 267                                 (void *)bufs, nb_ret, NULL);
 268 #else
 269                 uint16_t nb_ret = nb_rx;
 270                 /*
 271                  * Swap the following two lines if you want the rx traffic
 272                  * to go directly to tx, no distribution.
 273                  */
 274                 struct rte_ring *out_ring = p->rx_dist_ring;
 275                 /* struct rte_ring *out_ring = p->dist_tx_ring; */
 276
 277                 uint16_t sent = rte_ring_enqueue_burst(out_ring,
 278                                 (void *)bufs, nb_ret, NULL);
 279 #endif
 280
 281                 app_stats.rx.enqueued_pkts += sent;
 282                 if (unlikely(sent < nb_ret)) {
 283                         app_stats.rx.enqdrop_pkts +=  nb_ret - sent;
 284                         RTE_LOG_DP(DEBUG, DISTRAPP,
 285                                 "%s:Packet loss due to full ring\n", __func__);
 286                         while (sent < nb_ret)
 287                                 rte_pktmbuf_free(bufs[sent++]);
 288                 }
 289                 if (++port == nb_ports)
 290                         port = 0;
 291         }
 292         if (power_lib_initialised)
 293                 rte_power_exit(rte_lcore_id());
 294         /* set worker & tx threads quit flag */
 295         printf("\nCore %u exiting rx task.\n", rte_lcore_id());
 296         quit_signal = 1;
 297         return 0;
 298 }
 299
 300 static inline void
 301 flush_one_port(struct output_buffer *outbuf, uint8_t outp)
 302 {
 303         unsigned int nb_tx = rte_eth_tx_burst(outp, 0,
 304                         outbuf->mbufs, outbuf->count);
 305         app_stats.tx.tx_pkts += outbuf->count;
 306
 307         if (unlikely(nb_tx < outbuf->count)) {
 308                 app_stats.tx.enqdrop_pkts +=  outbuf->count - nb_tx;
 309                 do {
 310                         rte_pktmbuf_free(outbuf->mbufs[nb_tx]);
 311                 } while (++nb_tx < outbuf->count);
 312         }
 313         outbuf->count = 0;
 314 }
 315
 316 static inline void
 317 flush_all_ports(struct output_buffer *tx_buffers)
 318 {
 319         uint16_t outp;
 320
 321         RTE_ETH_FOREACH_DEV(outp) {
 322                 /* skip ports that are not enabled */
 323                 if ((enabled_port_mask & (1 << outp)) == 0)
 324                         continue;
 325
 326                 if (tx_buffers[outp].count == 0)
 327                         continue;
 328
 329                 flush_one_port(&tx_buffers[outp], outp);
 330         }
 331 }
 332
 333
 334
 335 static int
 336 lcore_distributor(struct lcore_params *p)
 337 {
 338         struct rte_ring *in_r = p->rx_dist_ring;
 339         struct rte_ring *out_r = p->dist_tx_ring;
 340         struct rte_mbuf *bufs[BURST_SIZE * 4];
 341         struct rte_distributor *d = p->d;
 342
 343         printf("\nCore %u acting as distributor core.\n", rte_lcore_id());
 344         while (!quit_signal_dist) {
 345                 const uint16_t nb_rx = rte_ring_dequeue_burst(in_r,
 346                                 (void *)bufs, BURST_SIZE*1, NULL);
 347                 if (nb_rx) {
 348                         app_stats.dist.in_pkts += nb_rx;
 349
 350                         /* Distribute the packets */
 351                         rte_distributor_process(d, bufs, nb_rx);
 352                         /* Handle Returns */
 353                         const uint16_t nb_ret =
 354                                 rte_distributor_returned_pkts(d,
 355                                         bufs, BURST_SIZE*2);
 356
 357                         if (unlikely(nb_ret == 0))
 358                                 continue;
 359                         app_stats.dist.ret_pkts += nb_ret;
 360
 361                         uint16_t sent = rte_ring_enqueue_burst(out_r,
 362                                         (void *)bufs, nb_ret, NULL);
 363                         app_stats.dist.sent_pkts += sent;
 364                         if (unlikely(sent < nb_ret)) {
 365                                 app_stats.dist.enqdrop_pkts += nb_ret - sent;
 366                                 RTE_LOG(DEBUG, DISTRAPP,
 367                                         "%s:Packet loss due to full out ring\n",
 368                                         __func__);
 369                                 while (sent < nb_ret)
 370                                         rte_pktmbuf_free(bufs[sent++]);
 371                         }
 372                 }
 373         }
 374         printf("\nCore %u exiting distributor task.\n", rte_lcore_id());
 375         quit_signal_work = 1;
 376         if (power_lib_initialised)
 377                 rte_power_exit(rte_lcore_id());
 378         rte_distributor_flush(d);
 379         /* Unblock any returns so workers can exit */
 380         rte_distributor_clear_returns(d);
 381         quit_signal_rx = 1;
 382         return 0;
 383 }
 384
 385
 386 static int
 387 lcore_tx(struct rte_ring *in_r)
 388 {
 389         static struct output_buffer tx_buffers[RTE_MAX_ETHPORTS];
 390         const int socket_id = rte_socket_id();
 391         uint16_t port;
 392
 393         RTE_ETH_FOREACH_DEV(port) {
 394                 /* skip ports that are not enabled */
 395                 if ((enabled_port_mask & (1 << port)) == 0)
 396                         continue;
 397
 398                 if (rte_eth_dev_socket_id(port) > 0 &&
 399                                 rte_eth_dev_socket_id(port) != socket_id)
 400                         printf("WARNING, port %u is on remote NUMA node to "
 401                                         "TX thread.\n\tPerformance will not "
 402                                         "be optimal.\n", port);
 403         }
 404
 405         printf("\nCore %u doing packet TX.\n", rte_lcore_id());
 406         while (!quit_signal) {
 407
 408                 RTE_ETH_FOREACH_DEV(port) {
 409                         /* skip ports that are not enabled */
 410                         if ((enabled_port_mask & (1 << port)) == 0)
 411                                 continue;
 412
 413                         struct rte_mbuf *bufs[BURST_SIZE_TX];
 414                         const uint16_t nb_rx = rte_ring_dequeue_burst(in_r,
 415                                         (void *)bufs, BURST_SIZE_TX, NULL);
 416                         app_stats.tx.dequeue_pkts += nb_rx;
 417
 418                         /* if we get no traffic, flush anything we have */
 419                         if (unlikely(nb_rx == 0)) {
 420                                 flush_all_ports(tx_buffers);
 421                                 continue;
 422                         }
 423
 424                         /* for traffic we receive, queue it up for transmit */
 425                         uint16_t i;
 426                         rte_prefetch_non_temporal((void *)bufs[0]);
 427                         rte_prefetch_non_temporal((void *)bufs[1]);
 428                         rte_prefetch_non_temporal((void *)bufs[2]);
 429                         for (i = 0; i < nb_rx; i++) {
 430                                 struct output_buffer *outbuf;
 431                                 uint8_t outp;
 432                                 rte_prefetch_non_temporal((void *)bufs[i + 3]);
 433                                 /*
 434                                  * workers should update in_port to hold the
 435                                  * output port value
 436                                  */
 437                                 outp = bufs[i]->port;
 438                                 /* skip ports that are not enabled */
 439                                 if ((enabled_port_mask & (1 << outp)) == 0)
 440                                         continue;
 441
 442                                 outbuf = &tx_buffers[outp];
 443                                 outbuf->mbufs[outbuf->count++] = bufs[i];
 444                                 if (outbuf->count == BURST_SIZE_TX)
 445                                         flush_one_port(outbuf, outp);
 446                         }
 447                 }
 448         }
 449         if (power_lib_initialised)
 450                 rte_power_exit(rte_lcore_id());
 451         printf("\nCore %u exiting tx task.\n", rte_lcore_id());
 452         return 0;
 453 }
 454
 455 static void
 456 int_handler(int sig_num)
 457 {
 458         printf("Exiting on signal %d\n", sig_num);
 459         /* set quit flag for rx thread to exit */
 460         quit_signal_dist = 1;
 461 }
 462
 463 static void
 464 print_stats(void)
 465 {
 466         struct rte_eth_stats eth_stats;
 467         unsigned int i, j;
 468         const unsigned int num_workers = rte_lcore_count() - 4;
 469
 470         RTE_ETH_FOREACH_DEV(i) {
 471                 rte_eth_stats_get(i, &eth_stats);
 472                 app_stats.port_rx_pkts[i] = eth_stats.ipackets;
 473                 app_stats.port_tx_pkts[i] = eth_stats.opackets;
 474         }
 475
 476         printf("\n\nRX Thread:\n");
 477         RTE_ETH_FOREACH_DEV(i) {
 478                 printf("Port %u Pktsin : %5.2f\n", i,
 479                                 (app_stats.port_rx_pkts[i] -
 480                                 prev_app_stats.port_rx_pkts[i])/1000000.0);
 481                 prev_app_stats.port_rx_pkts[i] = app_stats.port_rx_pkts[i];
 482         }
 483         printf(" - Received:    %5.2f\n",
 484                         (app_stats.rx.rx_pkts -
 485                         prev_app_stats.rx.rx_pkts)/1000000.0);
 486         printf(" - Returned:    %5.2f\n",
 487                         (app_stats.rx.returned_pkts -
 488                         prev_app_stats.rx.returned_pkts)/1000000.0);
 489         printf(" - Enqueued:    %5.2f\n",
 490                         (app_stats.rx.enqueued_pkts -
 491                         prev_app_stats.rx.enqueued_pkts)/1000000.0);
 492         printf(" - Dropped:     %s%5.2f%s\n", ANSI_COLOR_RED,
 493                         (app_stats.rx.enqdrop_pkts -
 494                         prev_app_stats.rx.enqdrop_pkts)/1000000.0,
 495                         ANSI_COLOR_RESET);
 496
 497         printf("Distributor thread:\n");
 498         printf(" - In:          %5.2f\n",
 499                         (app_stats.dist.in_pkts -
 500                         prev_app_stats.dist.in_pkts)/1000000.0);
 501         printf(" - Returned:    %5.2f\n",
 502                         (app_stats.dist.ret_pkts -
 503                         prev_app_stats.dist.ret_pkts)/1000000.0);
 504         printf(" - Sent:        %5.2f\n",
 505                         (app_stats.dist.sent_pkts -
 506                         prev_app_stats.dist.sent_pkts)/1000000.0);
 507         printf(" - Dropped      %s%5.2f%s\n", ANSI_COLOR_RED,
 508                         (app_stats.dist.enqdrop_pkts -
 509                         prev_app_stats.dist.enqdrop_pkts)/1000000.0,
 510                         ANSI_COLOR_RESET);
 511
 512         printf("TX thread:\n");
 513         printf(" - Dequeued:    %5.2f\n",
 514                         (app_stats.tx.dequeue_pkts -
 515                         prev_app_stats.tx.dequeue_pkts)/1000000.0);
 516         RTE_ETH_FOREACH_DEV(i) {
 517                 printf("Port %u Pktsout: %5.2f\n",
 518                                 i, (app_stats.port_tx_pkts[i] -
 519                                 prev_app_stats.port_tx_pkts[i])/1000000.0);
 520                 prev_app_stats.port_tx_pkts[i] = app_stats.port_tx_pkts[i];
 521         }
 522         printf(" - Transmitted: %5.2f\n",
 523                         (app_stats.tx.tx_pkts -
 524                         prev_app_stats.tx.tx_pkts)/1000000.0);
 525         printf(" - Dropped:     %s%5.2f%s\n", ANSI_COLOR_RED,
 526                         (app_stats.tx.enqdrop_pkts -
 527                         prev_app_stats.tx.enqdrop_pkts)/1000000.0,
 528                         ANSI_COLOR_RESET);
 529
 530         prev_app_stats.rx.rx_pkts = app_stats.rx.rx_pkts;
 531         prev_app_stats.rx.returned_pkts = app_stats.rx.returned_pkts;
 532         prev_app_stats.rx.enqueued_pkts = app_stats.rx.enqueued_pkts;
 533         prev_app_stats.rx.enqdrop_pkts = app_stats.rx.enqdrop_pkts;
 534         prev_app_stats.dist.in_pkts = app_stats.dist.in_pkts;
 535         prev_app_stats.dist.ret_pkts = app_stats.dist.ret_pkts;
 536         prev_app_stats.dist.sent_pkts = app_stats.dist.sent_pkts;
 537         prev_app_stats.dist.enqdrop_pkts = app_stats.dist.enqdrop_pkts;
 538         prev_app_stats.tx.dequeue_pkts = app_stats.tx.dequeue_pkts;
 539         prev_app_stats.tx.tx_pkts = app_stats.tx.tx_pkts;
 540         prev_app_stats.tx.enqdrop_pkts = app_stats.tx.enqdrop_pkts;
 541
 542         for (i = 0; i < num_workers; i++) {
 543                 printf("Worker %02u Pkts: %5.2f. Bursts(1-8): ", i,
 544                                 (app_stats.worker_pkts[i] -
 545                                 prev_app_stats.worker_pkts[i])/1000000.0);
 546                 for (j = 0; j < 8; j++) {
 547                         printf("%"PRIu64" ", app_stats.worker_bursts[i][j]);
 548                         app_stats.worker_bursts[i][j] = 0;
 549                 }
 550                 printf("\n");
 551                 prev_app_stats.worker_pkts[i] = app_stats.worker_pkts[i];
 552         }
 553 }
 554
 555 static int
 556 lcore_worker(struct lcore_params *p)
 557 {
 558         struct rte_distributor *d = p->d;
 559         const unsigned id = p->worker_id;
 560         unsigned int num = 0;
 561         unsigned int i;
 562
 563         /*
 564          * for single port, xor_val will be zero so we won't modify the output
 565          * port, otherwise we send traffic from 0 to 1, 2 to 3, and vice versa
 566          */
 567         const unsigned xor_val = (rte_eth_dev_count_avail() > 1);
 568         struct rte_mbuf *buf[8] __rte_cache_aligned;
 569
 570         for (i = 0; i < 8; i++)
 571                 buf[i] = NULL;
 572
 573         app_stats.worker_pkts[p->worker_id] = 1;
 574
 575         printf("\nCore %u acting as worker core.\n", rte_lcore_id());
 576         while (!quit_signal_work) {
 577                 num = rte_distributor_get_pkt(d, id, buf, buf, num);
 578                 /* Do a little bit of work for each packet */
 579                 for (i = 0; i < num; i++) {
 580                         uint64_t t = rte_rdtsc()+100;
 581
 582                         while (rte_rdtsc() < t)
 583                                 rte_pause();
 584                         buf[i]->port ^= xor_val;
 585                 }
 586
 587                 app_stats.worker_pkts[p->worker_id] += num;
 588                 if (num > 0)
 589                         app_stats.worker_bursts[p->worker_id][num-1]++;
 590         }
 591         if (power_lib_initialised)
 592                 rte_power_exit(rte_lcore_id());
 593         rte_free(p);
 594         return 0;
 595 }
 596
 597 static int
 598 init_power_library(void)
 599 {
 600         int ret = 0, lcore_id;
 601         RTE_LCORE_FOREACH_SLAVE(lcore_id) {
 602                 /* init power management library */
 603                 ret = rte_power_init(lcore_id);
 604                 if (ret) {
 605                         RTE_LOG(ERR, POWER,
 606                                 "Library initialization failed on core %u\n",
 607                                 lcore_id);
 608                         /*
 609                          * Return on first failure, we'll fall back
 610                          * to non-power operation
 611                          */
 612                         return ret;
 613                 }
 614         }
 615         return ret;
 616 }
 617
 618 /* display usage */
 619 static void
 620 print_usage(const char *prgname)
 621 {
 622         printf("%s [EAL options] -- -p PORTMASK\n"
 623                         "  -p PORTMASK: hexadecimal bitmask of ports to configure\n",
 624                         prgname);
 625 }
 626
 627 static int
 628 parse_portmask(const char *portmask)
 629 {
 630         char *end = NULL;
 631         unsigned long pm;
 632
 633         /* parse hexadecimal string */
 634         pm = strtoul(portmask, &end, 16);
 635         if ((portmask[0] == '\0') || (end == NULL) || (*end != '\0'))
 636                 return -1;
 637
 638         if (pm == 0)
 639                 return -1;
 640
 641         return pm;
 642 }
 643
 644 /* Parse the argument given in the command line of the application */
 645 static int
 646 parse_args(int argc, char **argv)
 647 {
 648         int opt;
 649         char **argvopt;
 650         int option_index;
 651         char *prgname = argv[0];
 652         static struct option lgopts[] = {
 653                 {NULL, 0, 0, 0}
 654         };
 655
 656         argvopt = argv;
 657
 658         while ((opt = getopt_long(argc, argvopt, "p:",
 659                         lgopts, &option_index)) != EOF) {
 660
 661                 switch (opt) {
 662                 /* portmask */
 663                 case 'p':
 664                         enabled_port_mask = parse_portmask(optarg);
 665                         if (enabled_port_mask == 0) {
 666                                 printf("invalid portmask\n");
 667                                 print_usage(prgname);
 668                                 return -1;
 669                         }
 670                         break;
 671
 672                 default:
 673                         print_usage(prgname);
 674                         return -1;
 675                 }
 676         }
 677
 678         if (optind <= 1) {
 679                 print_usage(prgname);
 680                 return -1;
 681         }
 682
 683         argv[optind-1] = prgname;
 684
 685         optind = 1; /* reset getopt lib */
 686         return 0;
 687 }
 688
 689 /* Main function, does initialization and calls the per-lcore functions */
 690 int
 691 main(int argc, char *argv[])
 692 {
 693         struct rte_mempool *mbuf_pool;
 694         struct rte_distributor *d;
 695         struct rte_ring *dist_tx_ring;
 696         struct rte_ring *rx_dist_ring;
 697         struct rte_power_core_capabilities lcore_cap;
 698         unsigned int lcore_id, worker_id = 0;
 699         int distr_core_id = -1, rx_core_id = -1, tx_core_id = -1;
 700         unsigned nb_ports;
 701         uint16_t portid;
 702         uint16_t nb_ports_available;
 703         uint64_t t, freq;
 704
 705         /* catch ctrl-c so we can print on exit */
 706         signal(SIGINT, int_handler);
 707
 708         /* init EAL */
 709         int ret = rte_eal_init(argc, argv);
 710         if (ret < 0)
 711                 rte_exit(EXIT_FAILURE, "Error with EAL initialization\n");
 712         argc -= ret;
 713         argv += ret;
 714
 715         /* parse application arguments (after the EAL ones) */
 716         ret = parse_args(argc, argv);
 717         if (ret < 0)
 718                 rte_exit(EXIT_FAILURE, "Invalid distributor parameters\n");
 719
 720         if (rte_lcore_count() < 5)
 721                 rte_exit(EXIT_FAILURE, "Error, This application needs at "
 722                                 "least 5 logical cores to run:\n"
 723                                 "1 lcore for stats (can be core 0)\n"
 724                                 "1 lcore for packet RX\n"
 725                                 "1 lcore for distribution\n"
 726                                 "1 lcore for packet TX\n"
 727                                 "and at least 1 lcore for worker threads\n");
 728
 729         if (init_power_library() == 0)
 730                 power_lib_initialised = 1;
 731
 732         nb_ports = rte_eth_dev_count_avail();
 733         if (nb_ports == 0)
 734                 rte_exit(EXIT_FAILURE, "Error: no ethernet ports detected\n");
 735         if (nb_ports != 1 && (nb_ports & 1))
 736                 rte_exit(EXIT_FAILURE, "Error: number of ports must be even, except "
 737                                 "when using a single port\n");
 738
 739         mbuf_pool = rte_pktmbuf_pool_create("MBUF_POOL",
 740                 NUM_MBUFS * nb_ports, MBUF_CACHE_SIZE, 0,
 741                 RTE_MBUF_DEFAULT_BUF_SIZE, rte_socket_id());
 742         if (mbuf_pool == NULL)
 743                 rte_exit(EXIT_FAILURE, "Cannot create mbuf pool\n");
 744         nb_ports_available = nb_ports;
 745
 746         /* initialize all ports */
 747         RTE_ETH_FOREACH_DEV(portid) {
 748                 /* skip ports that are not enabled */
 749                 if ((enabled_port_mask & (1 << portid)) == 0) {
 750                         printf("\nSkipping disabled port %d\n", portid);
 751                         nb_ports_available--;
 752                         continue;
 753                 }
 754                 /* init port */
 755                 printf("Initializing port %u... done\n", portid);
 756
 757                 if (port_init(portid, mbuf_pool) != 0)
 758                         rte_exit(EXIT_FAILURE, "Cannot initialize port %u\n",
 759                                         portid);
 760         }
 761
 762         if (!nb_ports_available) {
 763                 rte_exit(EXIT_FAILURE,
 764                                 "All available ports are disabled. Please set portmask.\n");
 765         }
 766
 767         d = rte_distributor_create("PKT_DIST", rte_socket_id(),
 768                         rte_lcore_count() - 4,
 769                         RTE_DIST_ALG_BURST);
 770         if (d == NULL)
 771                 rte_exit(EXIT_FAILURE, "Cannot create distributor\n");
 772
 773         /*
 774          * scheduler ring is read by the transmitter core, and written to
 775          * by scheduler core
 776          */
 777         dist_tx_ring = rte_ring_create("Output_ring", SCHED_TX_RING_SZ,
 778                         rte_socket_id(), RING_F_SC_DEQ | RING_F_SP_ENQ);
 779         if (dist_tx_ring == NULL)
 780                 rte_exit(EXIT_FAILURE, "Cannot create output ring\n");
 781
 782         rx_dist_ring = rte_ring_create("Input_ring", SCHED_RX_RING_SZ,
 783                         rte_socket_id(), RING_F_SC_DEQ | RING_F_SP_ENQ);
 784         if (rx_dist_ring == NULL)
 785                 rte_exit(EXIT_FAILURE, "Cannot create output ring\n");
 786
 787         if (power_lib_initialised) {
 788                 /*
 789                  * Here we'll pre-assign lcore ids to the rx, tx and
 790                  * distributor workloads if there's higher frequency
 791                  * on those cores e.g. if Turbo Boost is enabled.
 792                  * It's also worth mentioning that it will assign cores in a
 793                  * specific order, so that if there's less than three
 794                  * available, the higher frequency cores will go to the
 795                  * distributor first, then rx, then tx.
 796                  */
 797                 RTE_LCORE_FOREACH_SLAVE(lcore_id) {
 798
 799                         rte_power_get_capabilities(lcore_id, &lcore_cap);
 800
 801                         if (lcore_cap.priority != 1)
 802                                 continue;
 803
 804                         if (distr_core_id < 0) {
 805                                 distr_core_id = lcore_id;
 806                                 printf("Distributor on priority core %d\n",
 807                                         lcore_id);
 808                                 continue;
 809                         }
 810                         if (rx_core_id < 0) {
 811                                 rx_core_id = lcore_id;
 812                                 printf("Rx on priority core %d\n",
 813                                         lcore_id);
 814                                 continue;
 815                         }
 816                         if (tx_core_id < 0) {
 817                                 tx_core_id = lcore_id;
 818                                 printf("Tx on priority core %d\n",
 819                                         lcore_id);
 820                                 continue;
 821                         }
 822                 }
 823         }
 824
 825         /*
 826          * If there's any of the key workloads left without an lcore_id
 827          * after the high performing core assignment above, pre-assign
 828          * them here.
 829          */
 830         RTE_LCORE_FOREACH_SLAVE(lcore_id) {
 831                 if (lcore_id == (unsigned int)distr_core_id ||
 832                                 lcore_id == (unsigned int)rx_core_id ||
 833                                 lcore_id == (unsigned int)tx_core_id)
 834                         continue;
 835                 if (distr_core_id < 0) {
 836                         distr_core_id = lcore_id;
 837                         printf("Distributor on core %d\n", lcore_id);
 838                         continue;
 839                 }
 840                 if (rx_core_id < 0) {
 841                         rx_core_id = lcore_id;
 842                         printf("Rx on core %d\n", lcore_id);
 843                         continue;
 844                 }
 845                 if (tx_core_id < 0) {
 846                         tx_core_id = lcore_id;
 847                         printf("Tx on core %d\n", lcore_id);
 848                         continue;
 849                 }
 850         }
 851
 852         printf(" tx id %d, dist id %d, rx id %d\n",
 853                         tx_core_id,
 854                         distr_core_id,
 855                         rx_core_id);
 856
 857         /*
 858          * Kick off all the worker threads first, avoiding the pre-assigned
 859          * lcore_ids for tx, rx and distributor workloads.
 860          */
 861         RTE_LCORE_FOREACH_SLAVE(lcore_id) {
 862                 if (lcore_id == (unsigned int)distr_core_id ||
 863                                 lcore_id == (unsigned int)rx_core_id ||
 864                                 lcore_id == (unsigned int)tx_core_id)
 865                         continue;
 866                 printf("Starting thread %d as worker, lcore_id %d\n",
 867                                 worker_id, lcore_id);
 868                 struct lcore_params *p =
 869                         rte_malloc(NULL, sizeof(*p), 0);
 870                 if (!p)
 871                         rte_panic("malloc failure\n");
 872                 *p = (struct lcore_params){worker_id++, d, rx_dist_ring,
 873                         dist_tx_ring, mbuf_pool};
 874
 875                 rte_eal_remote_launch((lcore_function_t *)lcore_worker,
 876                                 p, lcore_id);
 877         }
 878
 879         /* Start tx core */
 880         rte_eal_remote_launch((lcore_function_t *)lcore_tx,
 881                         dist_tx_ring, tx_core_id);
 882
 883         /* Start distributor core */
 884         struct lcore_params *pd =
 885                 rte_malloc(NULL, sizeof(*pd), 0);
 886         if (!pd)
 887                 rte_panic("malloc failure\n");
 888         *pd = (struct lcore_params){worker_id++, d,
 889                 rx_dist_ring, dist_tx_ring, mbuf_pool};
 890         rte_eal_remote_launch(
 891                         (lcore_function_t *)lcore_distributor,
 892                         pd, distr_core_id);
 893
 894         /* Start rx core */
 895         struct lcore_params *pr =
 896                 rte_malloc(NULL, sizeof(*pr), 0);
 897         if (!pr)
 898                 rte_panic("malloc failure\n");
 899         *pr = (struct lcore_params){worker_id++, d, rx_dist_ring,
 900                 dist_tx_ring, mbuf_pool};
 901         rte_eal_remote_launch((lcore_function_t *)lcore_rx,
 902                         pr, rx_core_id);
 903
 904         freq = rte_get_timer_hz();
 905         t = rte_rdtsc() + freq;
 906         while (!quit_signal_dist) {
 907                 if (t < rte_rdtsc()) {
 908                         print_stats();
 909                         t = rte_rdtsc() + freq;
 910                 }
 911                 usleep(1000);
 912         }
 913
 914         RTE_LCORE_FOREACH_SLAVE(lcore_id) {
 915                 if (rte_eal_wait_lcore(lcore_id) < 0)
 916                         return -1;
 917         }
 918
 919         print_stats();
 920
 921         rte_free(pd);
 922         rte_free(pr);
 923
 924         return 0;
 925 }