examples/load_balancer/runtime.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_eal.h>
  51 #include <rte_launch.h>
  52 #include <rte_atomic.h>
  53 #include <rte_cycles.h>
  54 #include <rte_prefetch.h>
  55 #include <rte_lcore.h>
  56 #include <rte_per_lcore.h>
  57 #include <rte_branch_prediction.h>
  58 #include <rte_interrupts.h>
  59 #include <rte_pci.h>
  60 #include <rte_random.h>
  61 #include <rte_debug.h>
  62 #include <rte_ether.h>
  63 #include <rte_ethdev.h>
  64 #include <rte_ring.h>
  65 #include <rte_mempool.h>
  66 #include <rte_mbuf.h>
  67 #include <rte_ip.h>
  68 #include <rte_tcp.h>
  69 #include <rte_lpm.h>
  70
  71 #include "main.h"
  72
  73 #ifndef APP_LCORE_IO_FLUSH
  74 #define APP_LCORE_IO_FLUSH           1000000
  75 #endif
  76
  77 #ifndef APP_LCORE_WORKER_FLUSH
  78 #define APP_LCORE_WORKER_FLUSH       1000000
  79 #endif
  80
  81 #ifndef APP_STATS
  82 #define APP_STATS                    1000000
  83 #endif
  84
  85 #define APP_IO_RX_DROP_ALL_PACKETS   0
  86 #define APP_WORKER_DROP_ALL_PACKETS  0
  87 #define APP_IO_TX_DROP_ALL_PACKETS   0
  88
  89 #ifndef APP_IO_RX_PREFETCH_ENABLE
  90 #define APP_IO_RX_PREFETCH_ENABLE    1
  91 #endif
  92
  93 #ifndef APP_WORKER_PREFETCH_ENABLE
  94 #define APP_WORKER_PREFETCH_ENABLE   1
  95 #endif
  96
  97 #ifndef APP_IO_TX_PREFETCH_ENABLE
  98 #define APP_IO_TX_PREFETCH_ENABLE    1
  99 #endif
 100
 101 #if APP_IO_RX_PREFETCH_ENABLE
 102 #define APP_IO_RX_PREFETCH0(p)       rte_prefetch0(p)
 103 #define APP_IO_RX_PREFETCH1(p)       rte_prefetch1(p)
 104 #else
 105 #define APP_IO_RX_PREFETCH0(p)
 106 #define APP_IO_RX_PREFETCH1(p)
 107 #endif
 108
 109 #if APP_WORKER_PREFETCH_ENABLE
 110 #define APP_WORKER_PREFETCH0(p)      rte_prefetch0(p)
 111 #define APP_WORKER_PREFETCH1(p)      rte_prefetch1(p)
 112 #else
 113 #define APP_WORKER_PREFETCH0(p)
 114 #define APP_WORKER_PREFETCH1(p)
 115 #endif
 116
 117 #if APP_IO_TX_PREFETCH_ENABLE
 118 #define APP_IO_TX_PREFETCH0(p)       rte_prefetch0(p)
 119 #define APP_IO_TX_PREFETCH1(p)       rte_prefetch1(p)
 120 #else
 121 #define APP_IO_TX_PREFETCH0(p)
 122 #define APP_IO_TX_PREFETCH1(p)
 123 #endif
 124
 125 static inline void
 126 app_lcore_io_rx_buffer_to_send (
 127         struct app_lcore_params_io *lp,
 128         uint32_t worker,
 129         struct rte_mbuf *mbuf,
 130         uint32_t bsz)
 131 {
 132         uint32_t pos;
 133         int ret;
 134
 135         pos = lp->rx.mbuf_out[worker].n_mbufs;
 136         lp->rx.mbuf_out[worker].array[pos ++] = mbuf;
 137         if (likely(pos < bsz)) {
 138                 lp->rx.mbuf_out[worker].n_mbufs = pos;
 139                 return;
 140         }
 141
 142         ret = rte_ring_sp_enqueue_bulk(
 143                 lp->rx.rings[worker],
 144                 (void **) lp->rx.mbuf_out[worker].array,
 145                 bsz,
 146                 NULL);
 147
 148         if (unlikely(ret == 0)) {
 149                 uint32_t k;
 150                 for (k = 0; k < bsz; k ++) {
 151                         struct rte_mbuf *m = lp->rx.mbuf_out[worker].array[k];
 152                         rte_pktmbuf_free(m);
 153                 }
 154         }
 155
 156         lp->rx.mbuf_out[worker].n_mbufs = 0;
 157         lp->rx.mbuf_out_flush[worker] = 0;
 158
 159 #if APP_STATS
 160         lp->rx.rings_iters[worker] ++;
 161         if (likely(ret == 0)) {
 162                 lp->rx.rings_count[worker] ++;
 163         }
 164         if (unlikely(lp->rx.rings_iters[worker] == APP_STATS)) {
 165                 unsigned lcore = rte_lcore_id();
 166
 167                 printf("\tI/O RX %u out (worker %u): enq success rate = %.2f\n",
 168                         lcore,
 169                         (unsigned)worker,
 170                         ((double) lp->rx.rings_count[worker]) / ((double) lp->rx.rings_iters[worker]));
 171                 lp->rx.rings_iters[worker] = 0;
 172                 lp->rx.rings_count[worker] = 0;
 173         }
 174 #endif
 175 }
 176
 177 static inline void
 178 app_lcore_io_rx(
 179         struct app_lcore_params_io *lp,
 180         uint32_t n_workers,
 181         uint32_t bsz_rd,
 182         uint32_t bsz_wr,
 183         uint8_t pos_lb)
 184 {
 185         struct rte_mbuf *mbuf_1_0, *mbuf_1_1, *mbuf_2_0, *mbuf_2_1;
 186         uint8_t *data_1_0, *data_1_1 = NULL;
 187         uint32_t i;
 188
 189         for (i = 0; i < lp->rx.n_nic_queues; i ++) {
 190                 uint16_t port = lp->rx.nic_queues[i].port;
 191                 uint8_t queue = lp->rx.nic_queues[i].queue;
 192                 uint32_t n_mbufs, j;
 193
 194                 n_mbufs = rte_eth_rx_burst(
 195                         port,
 196                         queue,
 197                         lp->rx.mbuf_in.array,
 198                         (uint16_t) bsz_rd);
 199
 200                 if (unlikely(n_mbufs == 0)) {
 201                         continue;
 202                 }
 203
 204 #if APP_STATS
 205                 lp->rx.nic_queues_iters[i] ++;
 206                 lp->rx.nic_queues_count[i] += n_mbufs;
 207                 if (unlikely(lp->rx.nic_queues_iters[i] == APP_STATS)) {
 208                         struct rte_eth_stats stats;
 209                         unsigned lcore = rte_lcore_id();
 210
 211                         rte_eth_stats_get(port, &stats);
 212
 213                         printf("I/O RX %u in (NIC port %u): NIC drop ratio = %.2f avg burst size = %.2f\n",
 214                                 lcore,
 215                                 port,
 216                                 (double) stats.imissed / (double) (stats.imissed + stats.ipackets),
 217                                 ((double) lp->rx.nic_queues_count[i]) / ((double) lp->rx.nic_queues_iters[i]));
 218                         lp->rx.nic_queues_iters[i] = 0;
 219                         lp->rx.nic_queues_count[i] = 0;
 220                 }
 221 #endif
 222
 223 #if APP_IO_RX_DROP_ALL_PACKETS
 224                 for (j = 0; j < n_mbufs; j ++) {
 225                         struct rte_mbuf *pkt = lp->rx.mbuf_in.array[j];
 226                         rte_pktmbuf_free(pkt);
 227                 }
 228
 229                 continue;
 230 #endif
 231
 232                 mbuf_1_0 = lp->rx.mbuf_in.array[0];
 233                 mbuf_1_1 = lp->rx.mbuf_in.array[1];
 234                 data_1_0 = rte_pktmbuf_mtod(mbuf_1_0, uint8_t *);
 235                 if (likely(n_mbufs > 1)) {
 236                         data_1_1 = rte_pktmbuf_mtod(mbuf_1_1, uint8_t *);
 237                 }
 238
 239                 mbuf_2_0 = lp->rx.mbuf_in.array[2];
 240                 mbuf_2_1 = lp->rx.mbuf_in.array[3];
 241                 APP_IO_RX_PREFETCH0(mbuf_2_0);
 242                 APP_IO_RX_PREFETCH0(mbuf_2_1);
 243
 244                 for (j = 0; j + 3 < n_mbufs; j += 2) {
 245                         struct rte_mbuf *mbuf_0_0, *mbuf_0_1;
 246                         uint8_t *data_0_0, *data_0_1;
 247                         uint32_t worker_0, worker_1;
 248
 249                         mbuf_0_0 = mbuf_1_0;
 250                         mbuf_0_1 = mbuf_1_1;
 251                         data_0_0 = data_1_0;
 252                         data_0_1 = data_1_1;
 253
 254                         mbuf_1_0 = mbuf_2_0;
 255                         mbuf_1_1 = mbuf_2_1;
 256                         data_1_0 = rte_pktmbuf_mtod(mbuf_2_0, uint8_t *);
 257                         data_1_1 = rte_pktmbuf_mtod(mbuf_2_1, uint8_t *);
 258                         APP_IO_RX_PREFETCH0(data_1_0);
 259                         APP_IO_RX_PREFETCH0(data_1_1);
 260
 261                         mbuf_2_0 = lp->rx.mbuf_in.array[j+4];
 262                         mbuf_2_1 = lp->rx.mbuf_in.array[j+5];
 263                         APP_IO_RX_PREFETCH0(mbuf_2_0);
 264                         APP_IO_RX_PREFETCH0(mbuf_2_1);
 265
 266                         worker_0 = data_0_0[pos_lb] & (n_workers - 1);
 267                         worker_1 = data_0_1[pos_lb] & (n_workers - 1);
 268
 269                         app_lcore_io_rx_buffer_to_send(lp, worker_0, mbuf_0_0, bsz_wr);
 270                         app_lcore_io_rx_buffer_to_send(lp, worker_1, mbuf_0_1, bsz_wr);
 271                 }
 272
 273                 /* Handle the last 1, 2 (when n_mbufs is even) or 3 (when n_mbufs is odd) packets  */
 274                 for ( ; j < n_mbufs; j += 1) {
 275                         struct rte_mbuf *mbuf;
 276                         uint8_t *data;
 277                         uint32_t worker;
 278
 279                         mbuf = mbuf_1_0;
 280                         mbuf_1_0 = mbuf_1_1;
 281                         mbuf_1_1 = mbuf_2_0;
 282                         mbuf_2_0 = mbuf_2_1;
 283
 284                         data = rte_pktmbuf_mtod(mbuf, uint8_t *);
 285
 286                         APP_IO_RX_PREFETCH0(mbuf_1_0);
 287
 288                         worker = data[pos_lb] & (n_workers - 1);
 289
 290                         app_lcore_io_rx_buffer_to_send(lp, worker, mbuf, bsz_wr);
 291                 }
 292         }
 293 }
 294
 295 static inline void
 296 app_lcore_io_rx_flush(struct app_lcore_params_io *lp, uint32_t n_workers)
 297 {
 298         uint32_t worker;
 299
 300         for (worker = 0; worker < n_workers; worker ++) {
 301                 int ret;
 302
 303                 if (likely((lp->rx.mbuf_out_flush[worker] == 0) ||
 304                            (lp->rx.mbuf_out[worker].n_mbufs == 0))) {
 305                         lp->rx.mbuf_out_flush[worker] = 1;
 306                         continue;
 307                 }
 308
 309                 ret = rte_ring_sp_enqueue_bulk(
 310                         lp->rx.rings[worker],
 311                         (void **) lp->rx.mbuf_out[worker].array,
 312                         lp->rx.mbuf_out[worker].n_mbufs,
 313                         NULL);
 314
 315                 if (unlikely(ret == 0)) {
 316                         uint32_t k;
 317                         for (k = 0; k < lp->rx.mbuf_out[worker].n_mbufs; k ++) {
 318                                 struct rte_mbuf *pkt_to_free = lp->rx.mbuf_out[worker].array[k];
 319                                 rte_pktmbuf_free(pkt_to_free);
 320                         }
 321                 }
 322
 323                 lp->rx.mbuf_out[worker].n_mbufs = 0;
 324                 lp->rx.mbuf_out_flush[worker] = 1;
 325         }
 326 }
 327
 328 static inline void
 329 app_lcore_io_tx(
 330         struct app_lcore_params_io *lp,
 331         uint32_t n_workers,
 332         uint32_t bsz_rd,
 333         uint32_t bsz_wr)
 334 {
 335         uint32_t worker;
 336
 337         for (worker = 0; worker < n_workers; worker ++) {
 338                 uint32_t i;
 339
 340                 for (i = 0; i < lp->tx.n_nic_ports; i ++) {
 341                         uint16_t port = lp->tx.nic_ports[i];
 342                         struct rte_ring *ring = lp->tx.rings[port][worker];
 343                         uint32_t n_mbufs, n_pkts;
 344                         int ret;
 345
 346                         n_mbufs = lp->tx.mbuf_out[port].n_mbufs;
 347                         ret = rte_ring_sc_dequeue_bulk(
 348                                 ring,
 349                                 (void **) &lp->tx.mbuf_out[port].array[n_mbufs],
 350                                 bsz_rd,
 351                                 NULL);
 352
 353                         if (unlikely(ret == 0))
 354                                 continue;
 355
 356                         n_mbufs += bsz_rd;
 357
 358 #if APP_IO_TX_DROP_ALL_PACKETS
 359                         {
 360                                 uint32_t j;
 361                                 APP_IO_TX_PREFETCH0(lp->tx.mbuf_out[port].array[0]);
 362                                 APP_IO_TX_PREFETCH0(lp->tx.mbuf_out[port].array[1]);
 363
 364                                 for (j = 0; j < n_mbufs; j ++) {
 365                                         if (likely(j < n_mbufs - 2)) {
 366                                                 APP_IO_TX_PREFETCH0(lp->tx.mbuf_out[port].array[j + 2]);
 367                                         }
 368
 369                                         rte_pktmbuf_free(lp->tx.mbuf_out[port].array[j]);
 370                                 }
 371
 372                                 lp->tx.mbuf_out[port].n_mbufs = 0;
 373
 374                                 continue;
 375                         }
 376 #endif
 377
 378                         if (unlikely(n_mbufs < bsz_wr)) {
 379                                 lp->tx.mbuf_out[port].n_mbufs = n_mbufs;
 380                                 continue;
 381                         }
 382
 383                         n_pkts = rte_eth_tx_burst(
 384                                 port,
 385                                 0,
 386                                 lp->tx.mbuf_out[port].array,
 387                                 (uint16_t) n_mbufs);
 388
 389 #if APP_STATS
 390                         lp->tx.nic_ports_iters[port] ++;
 391                         lp->tx.nic_ports_count[port] += n_pkts;
 392                         if (unlikely(lp->tx.nic_ports_iters[port] == APP_STATS)) {
 393                                 unsigned lcore = rte_lcore_id();
 394
 395                                 printf("\t\t\tI/O TX %u out (port %u): avg burst size = %.2f\n",
 396                                         lcore,
 397                                         port,
 398                                         ((double) lp->tx.nic_ports_count[port]) / ((double) lp->tx.nic_ports_iters[port]));
 399                                 lp->tx.nic_ports_iters[port] = 0;
 400                                 lp->tx.nic_ports_count[port] = 0;
 401                         }
 402 #endif
 403
 404                         if (unlikely(n_pkts < n_mbufs)) {
 405                                 uint32_t k;
 406                                 for (k = n_pkts; k < n_mbufs; k ++) {
 407                                         struct rte_mbuf *pkt_to_free = lp->tx.mbuf_out[port].array[k];
 408                                         rte_pktmbuf_free(pkt_to_free);
 409                                 }
 410                         }
 411                         lp->tx.mbuf_out[port].n_mbufs = 0;
 412                         lp->tx.mbuf_out_flush[port] = 0;
 413                 }
 414         }
 415 }
 416
 417 static inline void
 418 app_lcore_io_tx_flush(struct app_lcore_params_io *lp)
 419 {
 420         uint16_t port;
 421         uint32_t i;
 422
 423         for (i = 0; i < lp->tx.n_nic_ports; i++) {
 424                 uint32_t n_pkts;
 425
 426                 port = lp->tx.nic_ports[i];
 427                 if (likely((lp->tx.mbuf_out_flush[port] == 0) ||
 428                            (lp->tx.mbuf_out[port].n_mbufs == 0))) {
 429                         lp->tx.mbuf_out_flush[port] = 1;
 430                         continue;
 431                 }
 432
 433                 n_pkts = rte_eth_tx_burst(
 434                         port,
 435                         0,
 436                         lp->tx.mbuf_out[port].array,
 437                         (uint16_t) lp->tx.mbuf_out[port].n_mbufs);
 438
 439                 if (unlikely(n_pkts < lp->tx.mbuf_out[port].n_mbufs)) {
 440                         uint32_t k;
 441                         for (k = n_pkts; k < lp->tx.mbuf_out[port].n_mbufs; k ++) {
 442                                 struct rte_mbuf *pkt_to_free = lp->tx.mbuf_out[port].array[k];
 443                                 rte_pktmbuf_free(pkt_to_free);
 444                         }
 445                 }
 446
 447                 lp->tx.mbuf_out[port].n_mbufs = 0;
 448                 lp->tx.mbuf_out_flush[port] = 1;
 449         }
 450 }
 451
 452 static void
 453 app_lcore_main_loop_io(void)
 454 {
 455         uint32_t lcore = rte_lcore_id();
 456         struct app_lcore_params_io *lp = &app.lcore_params[lcore].io;
 457         uint32_t n_workers = app_get_lcores_worker();
 458         uint64_t i = 0;
 459
 460         uint32_t bsz_rx_rd = app.burst_size_io_rx_read;
 461         uint32_t bsz_rx_wr = app.burst_size_io_rx_write;
 462         uint32_t bsz_tx_rd = app.burst_size_io_tx_read;
 463         uint32_t bsz_tx_wr = app.burst_size_io_tx_write;
 464
 465         uint8_t pos_lb = app.pos_lb;
 466
 467         for ( ; ; ) {
 468                 if (APP_LCORE_IO_FLUSH && (unlikely(i == APP_LCORE_IO_FLUSH))) {
 469                         if (likely(lp->rx.n_nic_queues > 0)) {
 470                                 app_lcore_io_rx_flush(lp, n_workers);
 471                         }
 472
 473                         if (likely(lp->tx.n_nic_ports > 0)) {
 474                                 app_lcore_io_tx_flush(lp);
 475                         }
 476
 477                         i = 0;
 478                 }
 479
 480                 if (likely(lp->rx.n_nic_queues > 0)) {
 481                         app_lcore_io_rx(lp, n_workers, bsz_rx_rd, bsz_rx_wr, pos_lb);
 482                 }
 483
 484                 if (likely(lp->tx.n_nic_ports > 0)) {
 485                         app_lcore_io_tx(lp, n_workers, bsz_tx_rd, bsz_tx_wr);
 486                 }
 487
 488                 i ++;
 489         }
 490 }
 491
 492 static inline void
 493 app_lcore_worker(
 494         struct app_lcore_params_worker *lp,
 495         uint32_t bsz_rd,
 496         uint32_t bsz_wr)
 497 {
 498         uint32_t i;
 499
 500         for (i = 0; i < lp->n_rings_in; i ++) {
 501                 struct rte_ring *ring_in = lp->rings_in[i];
 502                 uint32_t j;
 503                 int ret;
 504
 505                 ret = rte_ring_sc_dequeue_bulk(
 506                         ring_in,
 507                         (void **) lp->mbuf_in.array,
 508                         bsz_rd,
 509                         NULL);
 510
 511                 if (unlikely(ret == 0))
 512                         continue;
 513
 514 #if APP_WORKER_DROP_ALL_PACKETS
 515                 for (j = 0; j < bsz_rd; j ++) {
 516                         struct rte_mbuf *pkt = lp->mbuf_in.array[j];
 517                         rte_pktmbuf_free(pkt);
 518                 }
 519
 520                 continue;
 521 #endif
 522
 523                 APP_WORKER_PREFETCH1(rte_pktmbuf_mtod(lp->mbuf_in.array[0], unsigned char *));
 524                 APP_WORKER_PREFETCH0(lp->mbuf_in.array[1]);
 525
 526                 for (j = 0; j < bsz_rd; j ++) {
 527                         struct rte_mbuf *pkt;
 528                         struct ipv4_hdr *ipv4_hdr;
 529                         uint32_t ipv4_dst, pos;
 530                         uint32_t port;
 531
 532                         if (likely(j < bsz_rd - 1)) {
 533                                 APP_WORKER_PREFETCH1(rte_pktmbuf_mtod(lp->mbuf_in.array[j+1], unsigned char *));
 534                         }
 535                         if (likely(j < bsz_rd - 2)) {
 536                                 APP_WORKER_PREFETCH0(lp->mbuf_in.array[j+2]);
 537                         }
 538
 539                         pkt = lp->mbuf_in.array[j];
 540                         ipv4_hdr = rte_pktmbuf_mtod_offset(pkt,
 541                                                            struct ipv4_hdr *,
 542                                                            sizeof(struct ether_hdr));
 543                         ipv4_dst = rte_be_to_cpu_32(ipv4_hdr->dst_addr);
 544
 545                         if (unlikely(rte_lpm_lookup(lp->lpm_table, ipv4_dst, &port) != 0)) {
 546                                 port = pkt->port;
 547                         }
 548
 549                         pos = lp->mbuf_out[port].n_mbufs;
 550
 551                         lp->mbuf_out[port].array[pos ++] = pkt;
 552                         if (likely(pos < bsz_wr)) {
 553                                 lp->mbuf_out[port].n_mbufs = pos;
 554                                 continue;
 555                         }
 556
 557                         ret = rte_ring_sp_enqueue_bulk(
 558                                 lp->rings_out[port],
 559                                 (void **) lp->mbuf_out[port].array,
 560                                 bsz_wr,
 561                                 NULL);
 562
 563 #if APP_STATS
 564                         lp->rings_out_iters[port] ++;
 565                         if (ret > 0) {
 566                                 lp->rings_out_count[port] += 1;
 567                         }
 568                         if (lp->rings_out_iters[port] == APP_STATS){
 569                                 printf("\t\tWorker %u out (NIC port %u): enq success rate = %.2f\n",
 570                                         (unsigned) lp->worker_id,
 571                                         port,
 572                                         ((double) lp->rings_out_count[port]) / ((double) lp->rings_out_iters[port]));
 573                                 lp->rings_out_iters[port] = 0;
 574                                 lp->rings_out_count[port] = 0;
 575                         }
 576 #endif
 577
 578                         if (unlikely(ret == 0)) {
 579                                 uint32_t k;
 580                                 for (k = 0; k < bsz_wr; k ++) {
 581                                         struct rte_mbuf *pkt_to_free = lp->mbuf_out[port].array[k];
 582                                         rte_pktmbuf_free(pkt_to_free);
 583                                 }
 584                         }
 585
 586                         lp->mbuf_out[port].n_mbufs = 0;
 587                         lp->mbuf_out_flush[port] = 0;
 588                 }
 589         }
 590 }
 591
 592 static inline void
 593 app_lcore_worker_flush(struct app_lcore_params_worker *lp)
 594 {
 595         uint32_t port;
 596
 597         for (port = 0; port < APP_MAX_NIC_PORTS; port ++) {
 598                 int ret;
 599
 600                 if (unlikely(lp->rings_out[port] == NULL)) {
 601                         continue;
 602                 }
 603
 604                 if (likely((lp->mbuf_out_flush[port] == 0) ||
 605                            (lp->mbuf_out[port].n_mbufs == 0))) {
 606                         lp->mbuf_out_flush[port] = 1;
 607                         continue;
 608                 }
 609
 610                 ret = rte_ring_sp_enqueue_bulk(
 611                         lp->rings_out[port],
 612                         (void **) lp->mbuf_out[port].array,
 613                         lp->mbuf_out[port].n_mbufs,
 614                         NULL);
 615
 616                 if (unlikely(ret == 0)) {
 617                         uint32_t k;
 618                         for (k = 0; k < lp->mbuf_out[port].n_mbufs; k ++) {
 619                                 struct rte_mbuf *pkt_to_free = lp->mbuf_out[port].array[k];
 620                                 rte_pktmbuf_free(pkt_to_free);
 621                         }
 622                 }
 623
 624                 lp->mbuf_out[port].n_mbufs = 0;
 625                 lp->mbuf_out_flush[port] = 1;
 626         }
 627 }
 628
 629 static void
 630 app_lcore_main_loop_worker(void) {
 631         uint32_t lcore = rte_lcore_id();
 632         struct app_lcore_params_worker *lp = &app.lcore_params[lcore].worker;
 633         uint64_t i = 0;
 634
 635         uint32_t bsz_rd = app.burst_size_worker_read;
 636         uint32_t bsz_wr = app.burst_size_worker_write;
 637
 638         for ( ; ; ) {
 639                 if (APP_LCORE_WORKER_FLUSH && (unlikely(i == APP_LCORE_WORKER_FLUSH))) {
 640                         app_lcore_worker_flush(lp);
 641                         i = 0;
 642                 }
 643
 644                 app_lcore_worker(lp, bsz_rd, bsz_wr);
 645
 646                 i ++;
 647         }
 648 }
 649
 650 int
 651 app_lcore_main_loop(__attribute__((unused)) void *arg)
 652 {
 653         struct app_lcore_params *lp;
 654         unsigned lcore;
 655
 656         lcore = rte_lcore_id();
 657         lp = &app.lcore_params[lcore];
 658
 659         if (lp->type == e_APP_LCORE_IO) {
 660                 printf("Logical core %u (I/O) main loop.\n", lcore);
 661                 app_lcore_main_loop_io();
 662         }
 663
 664         if (lp->type == e_APP_LCORE_WORKER) {
 665                 printf("Logical core %u (worker %u) main loop.\n",
 666                         lcore,
 667                         (unsigned) lp->worker.worker_id);
 668                 app_lcore_main_loop_worker();
 669         }
 670
 671         return 0;
 672 }