examples/load_balancer/runtime.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2012 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
  35 #include <stdio.h>
  36 #include <stdlib.h>
  37 #include <stdint.h>
  38 #include <inttypes.h>
  39 #include <sys/types.h>
  40 #include <string.h>
  41 #include <sys/queue.h>
  42 #include <stdarg.h>
  43 #include <errno.h>
  44 #include <getopt.h>
  45
  46 #include <rte_common.h>
  47 #include <rte_byteorder.h>
  48 #include <rte_log.h>
  49 #include <rte_memory.h>
  50 #include <rte_memcpy.h>
  51 #include <rte_memzone.h>
  52 #include <rte_tailq.h>
  53 #include <rte_eal.h>
  54 #include <rte_per_lcore.h>
  55 #include <rte_launch.h>
  56 #include <rte_atomic.h>
  57 #include <rte_cycles.h>
  58 #include <rte_prefetch.h>
  59 #include <rte_lcore.h>
  60 #include <rte_per_lcore.h>
  61 #include <rte_branch_prediction.h>
  62 #include <rte_interrupts.h>
  63 #include <rte_pci.h>
  64 #include <rte_random.h>
  65 #include <rte_debug.h>
  66 #include <rte_ether.h>
  67 #include <rte_ethdev.h>
  68 #include <rte_ring.h>
  69 #include <rte_mempool.h>
  70 #include <rte_mbuf.h>
  71 #include <rte_ip.h>
  72 #include <rte_tcp.h>
  73 #include <rte_lpm.h>
  74
  75 #include "main.h"
  76
  77 #ifndef APP_LCORE_IO_FLUSH
  78 #define APP_LCORE_IO_FLUSH           1000000
  79 #endif
  80
  81 #ifndef APP_LCORE_WORKER_FLUSH
  82 #define APP_LCORE_WORKER_FLUSH       1000000
  83 #endif
  84
  85 #ifndef APP_STATS
  86 #define APP_STATS                    1000000
  87 #endif
  88
  89 #define APP_IO_RX_DROP_ALL_PACKETS   0
  90 #define APP_WORKER_DROP_ALL_PACKETS  0
  91 #define APP_IO_TX_DROP_ALL_PACKETS   0
  92
  93 #ifndef APP_IO_RX_PREFETCH_ENABLE
  94 #define APP_IO_RX_PREFETCH_ENABLE    1
  95 #endif
  96
  97 #ifndef APP_WORKER_PREFETCH_ENABLE
  98 #define APP_WORKER_PREFETCH_ENABLE   1
  99 #endif
 100
 101 #ifndef APP_IO_TX_PREFETCH_ENABLE
 102 #define APP_IO_TX_PREFETCH_ENABLE    1
 103 #endif
 104
 105 #if APP_IO_RX_PREFETCH_ENABLE
 106 #define APP_IO_RX_PREFETCH0(p)       rte_prefetch0(p)
 107 #define APP_IO_RX_PREFETCH1(p)       rte_prefetch1(p)
 108 #else
 109 #define APP_IO_RX_PREFETCH0(p)
 110 #define APP_IO_RX_PREFETCH1(p)
 111 #endif
 112
 113 #if APP_WORKER_PREFETCH_ENABLE
 114 #define APP_WORKER_PREFETCH0(p)      rte_prefetch0(p)
 115 #define APP_WORKER_PREFETCH1(p)      rte_prefetch1(p)
 116 #else
 117 #define APP_WORKER_PREFETCH0(p)
 118 #define APP_WORKER_PREFETCH1(p)
 119 #endif
 120
 121 #if APP_IO_TX_PREFETCH_ENABLE
 122 #define APP_IO_TX_PREFETCH0(p)       rte_prefetch0(p)
 123 #define APP_IO_TX_PREFETCH1(p)       rte_prefetch1(p)
 124 #else
 125 #define APP_IO_TX_PREFETCH0(p)
 126 #define APP_IO_TX_PREFETCH1(p)
 127 #endif
 128
 129 static inline void
 130 app_lcore_io_rx_buffer_to_send (
 131         struct app_lcore_params_io *lp,
 132         uint32_t worker,
 133         struct rte_mbuf *mbuf,
 134         uint32_t bsz)
 135 {
 136         uint32_t pos;
 137         int ret;
 138
 139         pos = lp->rx.mbuf_out[worker].n_mbufs;
 140         lp->rx.mbuf_out[worker].array[pos ++] = mbuf;
 141         if (likely(pos < bsz)) {
 142                 lp->rx.mbuf_out[worker].n_mbufs = pos;
 143                 return;
 144         }
 145
 146         ret = rte_ring_sp_enqueue_bulk(
 147                 lp->rx.rings[worker],
 148                 (void **) lp->rx.mbuf_out[worker].array,
 149                 bsz);
 150
 151         if (unlikely(ret == -ENOBUFS)) {
 152                 uint32_t k;
 153                 for (k = 0; k < bsz; k ++) {
 154                         struct rte_mbuf *m = lp->rx.mbuf_out[worker].array[k];
 155                         rte_pktmbuf_free(m);
 156                 }
 157         }
 158
 159         lp->rx.mbuf_out[worker].n_mbufs = 0;
 160         lp->rx.mbuf_out_flush[worker] = 0;
 161
 162 #if APP_STATS
 163         lp->rx.rings_iters[worker] ++;
 164         if (likely(ret == 0)) {
 165                 lp->rx.rings_count[worker] ++;
 166         }
 167         if (unlikely(lp->rx.rings_iters[worker] == APP_STATS)) {
 168                 uint32_t lcore = rte_lcore_id();
 169
 170                 printf("\tI/O RX %u out (worker %u): enq success rate = %.2f\n",
 171                         lcore,
 172                         worker,
 173                         ((double) lp->rx.rings_count[worker]) / ((double) lp->rx.rings_iters[worker]));
 174                 lp->rx.rings_iters[worker] = 0;
 175                 lp->rx.rings_count[worker] = 0;
 176         }
 177 #endif
 178 }
 179
 180 static inline void
 181 app_lcore_io_rx(
 182         struct app_lcore_params_io *lp,
 183         uint32_t n_workers,
 184         uint32_t bsz_rd,
 185         uint32_t bsz_wr,
 186         uint8_t pos_lb)
 187 {
 188         struct rte_mbuf *mbuf_1_0, *mbuf_1_1, *mbuf_2_0, *mbuf_2_1;
 189         uint8_t *data_1_0, *data_1_1;
 190         uint32_t i;
 191
 192         for (i = 0; i < lp->rx.n_nic_queues; i ++) {
 193                 uint8_t port = lp->rx.nic_queues[i].port;
 194                 uint8_t queue = lp->rx.nic_queues[i].queue;
 195                 uint32_t n_mbufs, j;
 196
 197                 n_mbufs = rte_eth_rx_burst(
 198                         port,
 199                         queue,
 200                         lp->rx.mbuf_in.array,
 201                         (uint16_t) bsz_rd);
 202
 203                 if (unlikely(n_mbufs == 0)) {
 204                         continue;
 205                 }
 206
 207 #if APP_STATS
 208                 lp->rx.nic_queues_iters[i] ++;
 209                 lp->rx.nic_queues_count[i] += n_mbufs;
 210                 if (unlikely(lp->rx.nic_queues_iters[i] == APP_STATS)) {
 211                         struct rte_eth_stats stats;
 212                         uint32_t lcore = rte_lcore_id();
 213
 214                         rte_eth_stats_get(port, &stats);
 215
 216                         printf("I/O RX %u in (NIC port %u): NIC drop ratio = %.2f avg burst size = %.2f\n",
 217                                 lcore,
 218                                 (uint32_t) port,
 219                                 (double) stats.ierrors / (double) (stats.ierrors + stats.ipackets),
 220                                 ((double) lp->rx.nic_queues_count[i]) / ((double) lp->rx.nic_queues_iters[i]));
 221                         lp->rx.nic_queues_iters[i] = 0;
 222                         lp->rx.nic_queues_count[i] = 0;
 223                 }
 224 #endif
 225
 226 #if APP_IO_RX_DROP_ALL_PACKETS
 227                 for (j = 0; j < n_mbufs; j ++) {
 228                         struct rte_mbuf *pkt = lp->rx.mbuf_in.array[j];
 229                         rte_pktmbuf_free(pkt);
 230                 }
 231
 232                 continue;
 233 #endif
 234
 235                 mbuf_1_0 = lp->rx.mbuf_in.array[0];
 236                 mbuf_1_1 = lp->rx.mbuf_in.array[1];
 237                 data_1_0 = rte_pktmbuf_mtod(mbuf_1_0, uint8_t *);
 238                 if (likely(n_mbufs > 1)) {
 239                         data_1_1 = rte_pktmbuf_mtod(mbuf_1_1, uint8_t *);
 240                 }
 241
 242                 mbuf_2_0 = lp->rx.mbuf_in.array[2];
 243                 mbuf_2_1 = lp->rx.mbuf_in.array[3];
 244                 APP_IO_RX_PREFETCH0(mbuf_2_0);
 245                 APP_IO_RX_PREFETCH0(mbuf_2_1);
 246
 247                 for (j = 0; j + 3 < n_mbufs; j += 2) {
 248                         struct rte_mbuf *mbuf_0_0, *mbuf_0_1;
 249                         uint8_t *data_0_0, *data_0_1;
 250                         uint32_t worker_0, worker_1;
 251
 252                         mbuf_0_0 = mbuf_1_0;
 253                         mbuf_0_1 = mbuf_1_1;
 254                         data_0_0 = data_1_0;
 255                         data_0_1 = data_1_1;
 256
 257                         mbuf_1_0 = mbuf_2_0;
 258                         mbuf_1_1 = mbuf_2_1;
 259                         data_1_0 = rte_pktmbuf_mtod(mbuf_2_0, uint8_t *);
 260                         data_1_1 = rte_pktmbuf_mtod(mbuf_2_1, uint8_t *);
 261                         APP_IO_RX_PREFETCH0(data_1_0);
 262                         APP_IO_RX_PREFETCH0(data_1_1);
 263
 264                         mbuf_2_0 = lp->rx.mbuf_in.array[j+4];
 265                         mbuf_2_1 = lp->rx.mbuf_in.array[j+5];
 266                         APP_IO_RX_PREFETCH0(mbuf_2_0);
 267                         APP_IO_RX_PREFETCH0(mbuf_2_1);
 268
 269                         worker_0 = data_0_0[pos_lb] & (n_workers - 1);
 270                         worker_1 = data_0_1[pos_lb] & (n_workers - 1);
 271
 272                         app_lcore_io_rx_buffer_to_send(lp, worker_0, mbuf_0_0, bsz_wr);
 273                         app_lcore_io_rx_buffer_to_send(lp, worker_1, mbuf_0_1, bsz_wr);
 274                 }
 275
 276                 /* Handle the last 1, 2 (when n_mbufs is even) or 3 (when n_mbufs is odd) packets  */
 277                 for ( ; j < n_mbufs; j += 1) {
 278                         struct rte_mbuf *mbuf;
 279                         uint8_t *data;
 280                         uint32_t worker;
 281
 282                         mbuf = mbuf_1_0;
 283                         mbuf_1_0 = mbuf_1_1;
 284                         mbuf_1_1 = mbuf_2_0;
 285                         mbuf_2_0 = mbuf_2_1;
 286
 287                         data = rte_pktmbuf_mtod(mbuf, uint8_t *);
 288
 289                         APP_IO_RX_PREFETCH0(mbuf_1_0);
 290
 291                         worker = data[pos_lb] & (n_workers - 1);
 292
 293                         app_lcore_io_rx_buffer_to_send(lp, worker, mbuf, bsz_wr);
 294                 }
 295         }
 296 }
 297
 298 static inline void
 299 app_lcore_io_rx_flush(struct app_lcore_params_io *lp, uint32_t n_workers)
 300 {
 301         uint32_t worker;
 302
 303         for (worker = 0; worker < n_workers; worker ++) {
 304                 int ret;
 305
 306                 if (likely((lp->rx.mbuf_out_flush[worker] == 0) ||
 307                            (lp->rx.mbuf_out[worker].n_mbufs == 0))) {
 308                         lp->rx.mbuf_out_flush[worker] = 1;
 309                         continue;
 310                 }
 311
 312                 ret = rte_ring_sp_enqueue_bulk(
 313                         lp->rx.rings[worker],
 314                         (void **) lp->rx.mbuf_out[worker].array,
 315                         lp->rx.mbuf_out[worker].n_mbufs);
 316
 317                 if (unlikely(ret < 0)) {
 318                         uint32_t k;
 319                         for (k = 0; k < lp->rx.mbuf_out[worker].n_mbufs; k ++) {
 320                                 struct rte_mbuf *pkt_to_free = lp->rx.mbuf_out[worker].array[k];
 321                                 rte_pktmbuf_free(pkt_to_free);
 322                         }
 323                 }
 324
 325                 lp->rx.mbuf_out[worker].n_mbufs = 0;
 326                 lp->rx.mbuf_out_flush[worker] = 1;
 327         }
 328 }
 329
 330 static inline void
 331 app_lcore_io_tx(
 332         struct app_lcore_params_io *lp,
 333         uint32_t n_workers,
 334         uint32_t bsz_rd,
 335         uint32_t bsz_wr)
 336 {
 337         uint32_t worker;
 338
 339         for (worker = 0; worker < n_workers; worker ++) {
 340                 uint32_t i;
 341
 342                 for (i = 0; i < lp->tx.n_nic_ports; i ++) {
 343                         uint8_t port = lp->tx.nic_ports[i];
 344                         struct rte_ring *ring = lp->tx.rings[port][worker];
 345                         uint32_t n_mbufs, n_pkts;
 346                         int ret;
 347
 348                         n_mbufs = lp->tx.mbuf_out[port].n_mbufs;
 349                         ret = rte_ring_sc_dequeue_bulk(
 350                                 ring,
 351                                 (void **) &lp->tx.mbuf_out[port].array[n_mbufs],
 352                                 bsz_rd);
 353
 354                         if (unlikely(ret == -ENOENT)) {
 355                                 continue;
 356                         }
 357
 358                         n_mbufs += bsz_rd;
 359
 360 #if APP_IO_TX_DROP_ALL_PACKETS
 361                         {
 362                                 uint32_t j;
 363                                 APP_IO_TX_PREFETCH0(lp->tx.mbuf_out[port].array[0]);
 364                                 APP_IO_TX_PREFETCH0(lp->tx.mbuf_out[port].array[1]);
 365
 366                                 for (j = 0; j < n_mbufs; j ++) {
 367                                         if (likely(j < n_mbufs - 2)) {
 368                                                 APP_IO_TX_PREFETCH0(lp->tx.mbuf_out[port].array[j + 2]);
 369                                         }
 370
 371                                         rte_pktmbuf_free(lp->tx.mbuf_out[port].array[j]);
 372                                 }
 373
 374                                 lp->tx.mbuf_out[port].n_mbufs = 0;
 375
 376                                 continue;
 377                         }
 378 #endif
 379
 380                         if (unlikely(n_mbufs < bsz_wr)) {
 381                                 lp->tx.mbuf_out[port].n_mbufs = n_mbufs;
 382                                 continue;
 383                         }
 384
 385                         n_pkts = rte_eth_tx_burst(
 386                                 port,
 387                                 0,
 388                                 lp->tx.mbuf_out[port].array,
 389                                 (uint16_t) n_mbufs);
 390
 391 #if APP_STATS
 392                         lp->tx.nic_ports_iters[port] ++;
 393                         lp->tx.nic_ports_count[port] += n_pkts;
 394                         if (unlikely(lp->tx.nic_ports_iters[port] == APP_STATS)) {
 395                                 uint32_t lcore = rte_lcore_id();
 396
 397                                 printf("\t\t\tI/O TX %u out (port %u): avg burst size = %.2f\n",
 398                                         lcore,
 399                                         (uint32_t) port,
 400                                         ((double) lp->tx.nic_ports_count[port]) / ((double) lp->tx.nic_ports_iters[port]));
 401                                 lp->tx.nic_ports_iters[port] = 0;
 402                                 lp->tx.nic_ports_count[port] = 0;
 403                         }
 404 #endif
 405
 406                         if (unlikely(n_pkts < n_mbufs)) {
 407                                 uint32_t k;
 408                                 for (k = n_pkts; k < n_mbufs; k ++) {
 409                                         struct rte_mbuf *pkt_to_free = lp->tx.mbuf_out[port].array[k];
 410                                         rte_pktmbuf_free(pkt_to_free);
 411                                 }
 412                         }
 413                         lp->tx.mbuf_out[port].n_mbufs = 0;
 414                         lp->tx.mbuf_out_flush[port] = 0;
 415                 }
 416         }
 417 }
 418
 419 static inline void
 420 app_lcore_io_tx_flush(struct app_lcore_params_io *lp)
 421 {
 422         uint8_t port;
 423
 424         for (port = 0; port < lp->tx.n_nic_ports; port ++) {
 425                 uint32_t n_pkts;
 426
 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
 510                 if (unlikely(ret == -ENOENT)) {
 511                         continue;
 512                 }
 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                         uint8_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 = (struct ipv4_hdr *)(rte_pktmbuf_mtod(pkt, unsigned char *) + sizeof(struct ether_hdr));
 541                         ipv4_dst = rte_be_to_cpu_32(ipv4_hdr->dst_addr);
 542
 543                         if (unlikely(rte_lpm_lookup(lp->lpm_table, ipv4_dst, &port) != 0)) {
 544                                 port = pkt->pkt.in_port;
 545                         }
 546
 547                         pos = lp->mbuf_out[port].n_mbufs;
 548
 549                         lp->mbuf_out[port].array[pos ++] = pkt;
 550                         if (likely(pos < bsz_wr)) {
 551                                 lp->mbuf_out[port].n_mbufs = pos;
 552                                 continue;
 553                         }
 554
 555                         ret = rte_ring_sp_enqueue_bulk(
 556                                 lp->rings_out[port],
 557                                 (void **) lp->mbuf_out[port].array,
 558                                 bsz_wr);
 559
 560 #if APP_STATS
 561                         lp->rings_out_iters[port] ++;
 562                         if (ret == 0) {
 563                                 lp->rings_out_count[port] += 1;
 564                         }
 565                         if (lp->rings_out_iters[port] == APP_STATS){
 566                                 printf("\t\tWorker %u out (NIC port %u): enq success rate = %.2f\n",
 567                                         lp->worker_id,
 568                                         (uint32_t) port,
 569                                         ((double) lp->rings_out_count[port]) / ((double) lp->rings_out_iters[port]));
 570                                 lp->rings_out_iters[port] = 0;
 571                                 lp->rings_out_count[port] = 0;
 572                         }
 573 #endif
 574
 575                         if (unlikely(ret == -ENOBUFS)) {
 576                                 uint32_t k;
 577                                 for (k = 0; k < bsz_wr; k ++) {
 578                                         struct rte_mbuf *pkt_to_free = lp->mbuf_out[port].array[k];
 579                                         rte_pktmbuf_free(pkt_to_free);
 580                                 }
 581                         }
 582
 583                         lp->mbuf_out[port].n_mbufs = 0;
 584                         lp->mbuf_out_flush[port] = 0;
 585                 }
 586         }
 587 }
 588
 589 static inline void
 590 app_lcore_worker_flush(struct app_lcore_params_worker *lp)
 591 {
 592         uint32_t port;
 593
 594         for (port = 0; port < APP_MAX_NIC_PORTS; port ++) {
 595                 int ret;
 596
 597                 if (unlikely(lp->rings_out[port] == NULL)) {
 598                         continue;
 599                 }
 600
 601                 if (likely((lp->mbuf_out_flush[port] == 0) ||
 602                            (lp->mbuf_out[port].n_mbufs == 0))) {
 603                         lp->mbuf_out_flush[port] = 1;
 604                         continue;
 605                 }
 606
 607                 ret = rte_ring_sp_enqueue_bulk(
 608                         lp->rings_out[port],
 609                         (void **) lp->mbuf_out[port].array,
 610                         lp->mbuf_out[port].n_mbufs);
 611
 612                 if (unlikely(ret < 0)) {
 613                         uint32_t k;
 614                         for (k = 0; k < lp->mbuf_out[port].n_mbufs; k ++) {
 615                                 struct rte_mbuf *pkt_to_free = lp->mbuf_out[port].array[k];
 616                                 rte_pktmbuf_free(pkt_to_free);
 617                         }
 618                 }
 619
 620                 lp->mbuf_out[port].n_mbufs = 0;
 621                 lp->mbuf_out_flush[port] = 1;
 622         }
 623 }
 624
 625 static void
 626 app_lcore_main_loop_worker(void) {
 627         uint32_t lcore = rte_lcore_id();
 628         struct app_lcore_params_worker *lp = &app.lcore_params[lcore].worker;
 629         uint64_t i = 0;
 630
 631         uint32_t bsz_rd = app.burst_size_worker_read;
 632         uint32_t bsz_wr = app.burst_size_worker_write;
 633
 634         for ( ; ; ) {
 635                 if (APP_LCORE_WORKER_FLUSH && (unlikely(i == APP_LCORE_WORKER_FLUSH))) {
 636                         app_lcore_worker_flush(lp);
 637                         i = 0;
 638                 }
 639
 640                 app_lcore_worker(lp, bsz_rd, bsz_wr);
 641
 642                 i ++;
 643         }
 644 }
 645
 646 int
 647 app_lcore_main_loop(__attribute__((unused)) void *arg)
 648 {
 649         struct app_lcore_params *lp;
 650         uint32_t lcore;
 651
 652         lcore = rte_lcore_id();
 653         lp = &app.lcore_params[lcore];
 654
 655         if (lp->type == e_APP_LCORE_IO) {
 656                 printf("Logical core %u (I/O) main loop.\n", lcore);
 657                 app_lcore_main_loop_io();
 658         }
 659
 660         if (lp->type == e_APP_LCORE_WORKER) {
 661                 printf("Logical core %u (worker %u) main loop.\n",
 662                         lcore,
 663                         lp->worker.worker_id);
 664                 app_lcore_main_loop_worker();
 665         }
 666
 667         return 0;
 668 }