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