app/test-flow-perf/main.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright 2020 Mellanox Technologies, Ltd
   3  *
   4  * This file contain the application main file
   5  * This application provides the user the ability to test the
   6  * insertion rate for specific rte_flow rule under stress state ~4M rule/
   7  *
   8  * Then it will also provide packet per second measurement after installing
   9  * all rules, the user may send traffic to test the PPS that match the rules
  10  * after all rules are installed, to check performance or functionality after
  11  * the stress.
  12  *
  13  * The flows insertion will go for all ports first, then it will print the
  14  * results, after that the application will go into forwarding packets mode
  15  * it will start receiving traffic if any and then forwarding it back and
  16  * gives packet per second measurement.
  17  */
  18
  19 #include <stdio.h>
  20 #include <stdlib.h>
  21 #include <string.h>
  22 #include <stdint.h>
  23 #include <inttypes.h>
  24 #include <stdarg.h>
  25 #include <errno.h>
  26 #include <getopt.h>
  27 #include <stdbool.h>
  28 #include <sys/time.h>
  29 #include <signal.h>
  30 #include <unistd.h>
  31
  32 #include <rte_malloc.h>
  33 #include <rte_mempool.h>
  34 #include <rte_mbuf.h>
  35 #include <rte_ethdev.h>
  36 #include <rte_flow.h>
  37
  38 #include "config.h"
  39 #include "flow_gen.h"
  40
  41 #define MAX_ITERATIONS             100
  42 #define DEFAULT_RULES_COUNT    4000000
  43 #define DEFAULT_ITERATION       100000
  44
  45 struct rte_flow *flow;
  46 static uint8_t flow_group;
  47
  48 static uint64_t flow_items[MAX_ITEMS_NUM];
  49 static uint64_t flow_actions[MAX_ACTIONS_NUM];
  50 static uint64_t flow_attrs[MAX_ATTRS_NUM];
  51 static uint8_t items_idx, actions_idx, attrs_idx;
  52
  53 static volatile bool force_quit;
  54 static bool dump_iterations;
  55 static bool delete_flag;
  56 static bool dump_socket_mem_flag;
  57 static bool enable_fwd;
  58
  59 static struct rte_mempool *mbuf_mp;
  60 static uint32_t nb_lcores;
  61 static uint32_t flows_count;
  62 static uint32_t iterations_number;
  63 static uint32_t hairpin_queues_num; /* total hairpin q number - default: 0 */
  64 static uint32_t nb_lcores;
  65
  66 #define MAX_PKT_BURST    32
  67 #define LCORE_MODE_PKT    1
  68 #define LCORE_MODE_STATS  2
  69 #define MAX_STREAMS      64
  70 #define MAX_LCORES       64
  71
  72 struct stream {
  73         int tx_port;
  74         int tx_queue;
  75         int rx_port;
  76         int rx_queue;
  77 };
  78
  79 struct lcore_info {
  80         int mode;
  81         int streams_nb;
  82         struct stream streams[MAX_STREAMS];
  83         /* stats */
  84         uint64_t tx_pkts;
  85         uint64_t tx_drops;
  86         uint64_t rx_pkts;
  87         struct rte_mbuf *pkts[MAX_PKT_BURST];
  88 } __rte_cache_aligned;
  89
  90 static struct lcore_info lcore_infos[MAX_LCORES];
  91
  92 static void
  93 usage(char *progname)
  94 {
  95         printf("\nusage: %s\n", progname);
  96         printf("\nControl configurations:\n");
  97         printf("  --flows-count=N: to set the number of needed"
  98                 " flows to insert, default is 4,000,000\n");
  99         printf("  --dump-iterations: To print rates for each"
 100                 " iteration\n");
 101         printf("  --deletion-rate: Enable deletion rate"
 102                 " calculations\n");
 103         printf("  --dump-socket-mem: To dump all socket memory\n");
 104         printf("  --enable-fwd: To enable packets forwarding"
 105                 " after insertion\n");
 106
 107         printf("To set flow attributes:\n");
 108         printf("  --ingress: set ingress attribute in flows\n");
 109         printf("  --egress: set egress attribute in flows\n");
 110         printf("  --transfer: set transfer attribute in flows\n");
 111         printf("  --group=N: set group for all flows,"
 112                 " default is 0\n");
 113
 114         printf("To set flow items:\n");
 115         printf("  --ether: add ether layer in flow items\n");
 116         printf("  --vlan: add vlan layer in flow items\n");
 117         printf("  --ipv4: add ipv4 layer in flow items\n");
 118         printf("  --ipv6: add ipv6 layer in flow items\n");
 119         printf("  --tcp: add tcp layer in flow items\n");
 120         printf("  --udp: add udp layer in flow items\n");
 121         printf("  --vxlan: add vxlan layer in flow items\n");
 122         printf("  --vxlan-gpe: add vxlan-gpe layer in flow items\n");
 123         printf("  --gre: add gre layer in flow items\n");
 124         printf("  --geneve: add geneve layer in flow items\n");
 125         printf("  --gtp: add gtp layer in flow items\n");
 126         printf("  --meta: add meta layer in flow items\n");
 127         printf("  --tag: add tag layer in flow items\n");
 128
 129         printf("To set flow actions:\n");
 130         printf("  --port-id: add port-id action in flow actions\n");
 131         printf("  --rss: add rss action in flow actions\n");
 132         printf("  --queue: add queue action in flow actions\n");
 133         printf("  --jump: add jump action in flow actions\n");
 134         printf("  --mark: add mark action in flow actions\n");
 135         printf("  --count: add count action in flow actions\n");
 136         printf("  --set-meta: add set meta action in flow actions\n");
 137         printf("  --set-tag: add set tag action in flow actions\n");
 138         printf("  --drop: add drop action in flow actions\n");
 139         printf("  --hairpin-queue=N: add hairpin-queue action in flow actions\n");
 140         printf("  --hairpin-rss=N: add hairpin-rss action in flow actions\n");
 141 }
 142
 143 static void
 144 args_parse(int argc, char **argv)
 145 {
 146         char **argvopt;
 147         int n, opt;
 148         int opt_idx;
 149         size_t i;
 150
 151         static const struct option_dict {
 152                 const char *str;
 153                 const uint64_t mask;
 154                 uint64_t *map;
 155                 uint8_t *map_idx;
 156
 157         } flow_options[] = {
 158                 {
 159                         .str = "ether",
 160                         .mask = FLOW_ITEM_MASK(RTE_FLOW_ITEM_TYPE_ETH),
 161                         .map = &flow_items[0],
 162                         .map_idx = &items_idx
 163                 },
 164                 {
 165                         .str = "ipv4",
 166                         .mask = FLOW_ITEM_MASK(RTE_FLOW_ITEM_TYPE_IPV4),
 167                         .map = &flow_items[0],
 168                         .map_idx = &items_idx
 169                 },
 170                 {
 171                         .str = "ipv6",
 172                         .mask = FLOW_ITEM_MASK(RTE_FLOW_ITEM_TYPE_IPV6),
 173                         .map = &flow_items[0],
 174                         .map_idx = &items_idx
 175                 },
 176                 {
 177                         .str = "vlan",
 178                         .mask = FLOW_ITEM_MASK(RTE_FLOW_ITEM_TYPE_VLAN),
 179                         .map = &flow_items[0],
 180                         .map_idx = &items_idx
 181                 },
 182                 {
 183                         .str = "tcp",
 184                         .mask = FLOW_ITEM_MASK(RTE_FLOW_ITEM_TYPE_TCP),
 185                         .map = &flow_items[0],
 186                         .map_idx = &items_idx
 187                 },
 188                 {
 189                         .str = "udp",
 190                         .mask = FLOW_ITEM_MASK(RTE_FLOW_ITEM_TYPE_UDP),
 191                         .map = &flow_items[0],
 192                         .map_idx = &items_idx
 193                 },
 194                 {
 195                         .str = "vxlan",
 196                         .mask = FLOW_ITEM_MASK(RTE_FLOW_ITEM_TYPE_VXLAN),
 197                         .map = &flow_items[0],
 198                         .map_idx = &items_idx
 199                 },
 200                 {
 201                         .str = "vxlan-gpe",
 202                         .mask = FLOW_ITEM_MASK(RTE_FLOW_ITEM_TYPE_VXLAN_GPE),
 203                         .map = &flow_items[0],
 204                         .map_idx = &items_idx
 205                 },
 206                 {
 207                         .str = "gre",
 208                         .mask = FLOW_ITEM_MASK(RTE_FLOW_ITEM_TYPE_GRE),
 209                         .map = &flow_items[0],
 210                         .map_idx = &items_idx
 211                 },
 212                 {
 213                         .str = "geneve",
 214                         .mask = FLOW_ITEM_MASK(RTE_FLOW_ITEM_TYPE_GENEVE),
 215                         .map = &flow_items[0],
 216                         .map_idx = &items_idx
 217                 },
 218                 {
 219                         .str = "gtp",
 220                         .mask = FLOW_ITEM_MASK(RTE_FLOW_ITEM_TYPE_GTP),
 221                         .map = &flow_items[0],
 222                         .map_idx = &items_idx
 223                 },
 224                 {
 225                         .str = "meta",
 226                         .mask = FLOW_ITEM_MASK(RTE_FLOW_ITEM_TYPE_META),
 227                         .map = &flow_items[0],
 228                         .map_idx = &items_idx
 229                 },
 230                 {
 231                         .str = "tag",
 232                         .mask = FLOW_ITEM_MASK(RTE_FLOW_ITEM_TYPE_TAG),
 233                         .map = &flow_items[0],
 234                         .map_idx = &items_idx
 235                 },
 236                 {
 237                         .str = "ingress",
 238                         .mask = INGRESS,
 239                         .map = &flow_attrs[0],
 240                         .map_idx = &attrs_idx
 241                 },
 242                 {
 243                         .str = "egress",
 244                         .mask = EGRESS,
 245                         .map = &flow_attrs[0],
 246                         .map_idx = &attrs_idx
 247                 },
 248                 {
 249                         .str = "transfer",
 250                         .mask = TRANSFER,
 251                         .map = &flow_attrs[0],
 252                         .map_idx = &attrs_idx
 253                 },
 254                 {
 255                         .str = "port-id",
 256                         .mask = FLOW_ACTION_MASK(RTE_FLOW_ACTION_TYPE_PORT_ID),
 257                         .map = &flow_actions[0],
 258                         .map_idx = &actions_idx
 259                 },
 260                 {
 261                         .str = "rss",
 262                         .mask = FLOW_ACTION_MASK(RTE_FLOW_ACTION_TYPE_RSS),
 263                         .map = &flow_actions[0],
 264                         .map_idx = &actions_idx
 265                 },
 266                 {
 267                         .str = "queue",
 268                         .mask = FLOW_ACTION_MASK(RTE_FLOW_ACTION_TYPE_QUEUE),
 269                         .map = &flow_actions[0],
 270                         .map_idx = &actions_idx
 271                 },
 272                 {
 273                         .str = "jump",
 274                         .mask = FLOW_ACTION_MASK(RTE_FLOW_ACTION_TYPE_JUMP),
 275                         .map = &flow_actions[0],
 276                         .map_idx = &actions_idx
 277                 },
 278                 {
 279                         .str = "mark",
 280                         .mask = FLOW_ACTION_MASK(RTE_FLOW_ACTION_TYPE_MARK),
 281                         .map = &flow_actions[0],
 282                         .map_idx = &actions_idx
 283                 },
 284                 {
 285                         .str = "count",
 286                         .mask = FLOW_ACTION_MASK(RTE_FLOW_ACTION_TYPE_COUNT),
 287                         .map = &flow_actions[0],
 288                         .map_idx = &actions_idx
 289                 },
 290                 {
 291                         .str = "set-meta",
 292                         .mask = FLOW_ACTION_MASK(RTE_FLOW_ACTION_TYPE_SET_META),
 293                         .map = &flow_actions[0],
 294                         .map_idx = &actions_idx
 295                 },
 296                 {
 297                         .str = "set-tag",
 298                         .mask = FLOW_ACTION_MASK(RTE_FLOW_ACTION_TYPE_SET_TAG),
 299                         .map = &flow_actions[0],
 300                         .map_idx = &actions_idx
 301                 },
 302                 {
 303                         .str = "drop",
 304                         .mask = FLOW_ACTION_MASK(RTE_FLOW_ACTION_TYPE_DROP),
 305                         .map = &flow_actions[0],
 306                         .map_idx = &actions_idx
 307                 }
 308         };
 309
 310         static const struct option lgopts[] = {
 311                 /* Control */
 312                 { "help",                       0, 0, 0 },
 313                 { "flows-count",                1, 0, 0 },
 314                 { "dump-iterations",            0, 0, 0 },
 315                 { "deletion-rate",              0, 0, 0 },
 316                 { "dump-socket-mem",            0, 0, 0 },
 317                 { "enable-fwd",                 0, 0, 0 },
 318                 /* Attributes */
 319                 { "ingress",                    0, 0, 0 },
 320                 { "egress",                     0, 0, 0 },
 321                 { "transfer",                   0, 0, 0 },
 322                 { "group",                      1, 0, 0 },
 323                 /* Items */
 324                 { "ether",                      0, 0, 0 },
 325                 { "vlan",                       0, 0, 0 },
 326                 { "ipv4",                       0, 0, 0 },
 327                 { "ipv6",                       0, 0, 0 },
 328                 { "tcp",                        0, 0, 0 },
 329                 { "udp",                        0, 0, 0 },
 330                 { "vxlan",                      0, 0, 0 },
 331                 { "vxlan-gpe",                  0, 0, 0 },
 332                 { "gre",                        0, 0, 0 },
 333                 { "geneve",                     0, 0, 0 },
 334                 { "gtp",                        0, 0, 0 },
 335                 { "meta",                       0, 0, 0 },
 336                 { "tag",                        0, 0, 0 },
 337                 /* Actions */
 338                 { "port-id",                    0, 0, 0 },
 339                 { "rss",                        0, 0, 0 },
 340                 { "queue",                      0, 0, 0 },
 341                 { "jump",                       0, 0, 0 },
 342                 { "mark",                       0, 0, 0 },
 343                 { "count",                      0, 0, 0 },
 344                 { "set-meta",                   0, 0, 0 },
 345                 { "set-tag",                    0, 0, 0 },
 346                 { "drop",                       0, 0, 0 },
 347                 { "hairpin-queue",              1, 0, 0 },
 348                 { "hairpin-rss",                1, 0, 0 },
 349         };
 350
 351         hairpin_queues_num = 0;
 352         argvopt = argv;
 353
 354         printf(":: Flow -> ");
 355         while ((opt = getopt_long(argc, argvopt, "",
 356                                 lgopts, &opt_idx)) != EOF) {
 357                 switch (opt) {
 358                 case 0:
 359                         if (strcmp(lgopts[opt_idx].name, "help") == 0) {
 360                                 usage(argv[0]);
 361                                 rte_exit(EXIT_SUCCESS, "Displayed help\n");
 362                         }
 363
 364                         if (strcmp(lgopts[opt_idx].name, "group") == 0) {
 365                                 n = atoi(optarg);
 366                                 if (n >= 0)
 367                                         flow_group = n;
 368                                 else
 369                                         rte_exit(EXIT_SUCCESS,
 370                                                 "flow group should be >= 0\n");
 371                                 printf("group %d ", flow_group);
 372                         }
 373
 374                         for (i = 0; i < RTE_DIM(flow_options); i++)
 375                                 if (strcmp(lgopts[opt_idx].name,
 376                                                 flow_options[i].str) == 0) {
 377                                         flow_options[i].map[
 378                                         (*flow_options[i].map_idx)++] =
 379                                                 flow_options[i].mask;
 380                                         printf("%s / ", flow_options[i].str);
 381                                 }
 382
 383                         if (strcmp(lgopts[opt_idx].name,
 384                                         "hairpin-rss") == 0) {
 385                                 n = atoi(optarg);
 386                                 if (n > 0)
 387                                         hairpin_queues_num = n;
 388                                 else
 389                                         rte_exit(EXIT_SUCCESS,
 390                                                 "Hairpin queues should be > 0\n");
 391
 392                                 flow_actions[actions_idx++] =
 393                                         HAIRPIN_RSS_ACTION;
 394                                 printf("hairpin-rss / ");
 395                         }
 396                         if (strcmp(lgopts[opt_idx].name,
 397                                         "hairpin-queue") == 0) {
 398                                 n = atoi(optarg);
 399                                 if (n > 0)
 400                                         hairpin_queues_num = n;
 401                                 else
 402                                         rte_exit(EXIT_SUCCESS,
 403                                                 "Hairpin queues should be > 0\n");
 404
 405                                 flow_actions[actions_idx++] =
 406                                         HAIRPIN_QUEUE_ACTION;
 407                                 printf("hairpin-queue / ");
 408                         }
 409
 410                         /* Control */
 411                         if (strcmp(lgopts[opt_idx].name,
 412                                         "flows-count") == 0) {
 413                                 n = atoi(optarg);
 414                                 if (n > (int) iterations_number)
 415                                         flows_count = n;
 416                                 else {
 417                                         printf("\n\nflows_count should be > %d\n",
 418                                                 iterations_number);
 419                                         rte_exit(EXIT_SUCCESS, " ");
 420                                 }
 421                         }
 422                         if (strcmp(lgopts[opt_idx].name,
 423                                         "dump-iterations") == 0)
 424                                 dump_iterations = true;
 425                         if (strcmp(lgopts[opt_idx].name,
 426                                         "deletion-rate") == 0)
 427                                 delete_flag = true;
 428                         if (strcmp(lgopts[opt_idx].name,
 429                                         "dump-socket-mem") == 0)
 430                                 dump_socket_mem_flag = true;
 431                         if (strcmp(lgopts[opt_idx].name,
 432                                         "enable-fwd") == 0)
 433                                 enable_fwd = true;
 434                         break;
 435                 default:
 436                         fprintf(stderr, "Invalid option: %s\n", argv[optind]);
 437                         usage(argv[0]);
 438                         rte_exit(EXIT_SUCCESS, "Invalid option\n");
 439                         break;
 440                 }
 441         }
 442         printf("end_flow\n");
 443 }
 444
 445 /* Dump the socket memory statistics on console */
 446 static size_t
 447 dump_socket_mem(FILE *f)
 448 {
 449         struct rte_malloc_socket_stats socket_stats;
 450         unsigned int i = 0;
 451         size_t total = 0;
 452         size_t alloc = 0;
 453         size_t free = 0;
 454         unsigned int n_alloc = 0;
 455         unsigned int n_free = 0;
 456         bool active_nodes = false;
 457
 458
 459         for (i = 0; i < RTE_MAX_NUMA_NODES; i++) {
 460                 if (rte_malloc_get_socket_stats(i, &socket_stats) ||
 461                     !socket_stats.heap_totalsz_bytes)
 462                         continue;
 463                 active_nodes = true;
 464                 total += socket_stats.heap_totalsz_bytes;
 465                 alloc += socket_stats.heap_allocsz_bytes;
 466                 free += socket_stats.heap_freesz_bytes;
 467                 n_alloc += socket_stats.alloc_count;
 468                 n_free += socket_stats.free_count;
 469                 if (dump_socket_mem_flag) {
 470                         fprintf(f, "::::::::::::::::::::::::::::::::::::::::");
 471                         fprintf(f,
 472                                 "\nSocket %u:\nsize(M) total: %.6lf\nalloc:"
 473                                 " %.6lf(%.3lf%%)\nfree: %.6lf"
 474                                 "\nmax: %.6lf"
 475                                 "\ncount alloc: %u\nfree: %u\n",
 476                                 i,
 477                                 socket_stats.heap_totalsz_bytes / 1.0e6,
 478                                 socket_stats.heap_allocsz_bytes / 1.0e6,
 479                                 (double)socket_stats.heap_allocsz_bytes * 100 /
 480                                 (double)socket_stats.heap_totalsz_bytes,
 481                                 socket_stats.heap_freesz_bytes / 1.0e6,
 482                                 socket_stats.greatest_free_size / 1.0e6,
 483                                 socket_stats.alloc_count,
 484                                 socket_stats.free_count);
 485                                 fprintf(f, "::::::::::::::::::::::::::::::::::::::::");
 486                 }
 487         }
 488         if (dump_socket_mem_flag && active_nodes) {
 489                 fprintf(f,
 490                         "\nTotal: size(M)\ntotal: %.6lf"
 491                         "\nalloc: %.6lf(%.3lf%%)\nfree: %.6lf"
 492                         "\ncount alloc: %u\nfree: %u\n",
 493                         total / 1.0e6, alloc / 1.0e6,
 494                         (double)alloc * 100 / (double)total, free / 1.0e6,
 495                         n_alloc, n_free);
 496                 fprintf(f, "::::::::::::::::::::::::::::::::::::::::\n");
 497         }
 498         return alloc;
 499 }
 500
 501 static void
 502 print_flow_error(struct rte_flow_error error)
 503 {
 504         printf("Flow can't be created %d message: %s\n",
 505                 error.type,
 506                 error.message ? error.message : "(no stated reason)");
 507 }
 508
 509 static inline void
 510 destroy_flows(int port_id, struct rte_flow **flow_list)
 511 {
 512         struct rte_flow_error error;
 513         clock_t start_iter, end_iter;
 514         double cpu_time_used = 0;
 515         double flows_rate;
 516         double cpu_time_per_iter[MAX_ITERATIONS];
 517         double delta;
 518         uint32_t i;
 519         int iter_id;
 520
 521         for (i = 0; i < MAX_ITERATIONS; i++)
 522                 cpu_time_per_iter[i] = -1;
 523
 524         if (iterations_number > flows_count)
 525                 iterations_number = flows_count;
 526
 527         /* Deletion Rate */
 528         printf("Flows Deletion on port = %d\n", port_id);
 529         start_iter = clock();
 530         for (i = 0; i < flows_count; i++) {
 531                 if (flow_list[i] == 0)
 532                         break;
 533
 534                 memset(&error, 0x33, sizeof(error));
 535                 if (rte_flow_destroy(port_id, flow_list[i], &error)) {
 536                         print_flow_error(error);
 537                         rte_exit(EXIT_FAILURE, "Error in deleting flow");
 538                 }
 539
 540                 if (i && !((i + 1) % iterations_number)) {
 541                         /* Save the deletion rate of each iter */
 542                         end_iter = clock();
 543                         delta = (double) (end_iter - start_iter);
 544                         iter_id = ((i + 1) / iterations_number) - 1;
 545                         cpu_time_per_iter[iter_id] =
 546                                 delta / CLOCKS_PER_SEC;
 547                         cpu_time_used += cpu_time_per_iter[iter_id];
 548                         start_iter = clock();
 549                 }
 550         }
 551
 552         /* Deletion rate per iteration */
 553         if (dump_iterations)
 554                 for (i = 0; i < MAX_ITERATIONS; i++) {
 555                         if (cpu_time_per_iter[i] == -1)
 556                                 continue;
 557                         delta = (double)(iterations_number /
 558                                 cpu_time_per_iter[i]);
 559                         flows_rate = delta / 1000;
 560                         printf(":: Iteration #%d: %d flows "
 561                                 "in %f sec[ Rate = %f K/Sec ]\n",
 562                                 i, iterations_number,
 563                                 cpu_time_per_iter[i], flows_rate);
 564                 }
 565
 566         /* Deletion rate for all flows */
 567         flows_rate = ((double) (flows_count / cpu_time_used) / 1000);
 568         printf("\n:: Total flow deletion rate -> %f K/Sec\n",
 569                 flows_rate);
 570         printf(":: The time for deleting %d in flows %f seconds\n",
 571                 flows_count, cpu_time_used);
 572 }
 573
 574 static inline void
 575 flows_handler(void)
 576 {
 577         struct rte_flow **flow_list;
 578         struct rte_flow_error error;
 579         clock_t start_iter, end_iter;
 580         double cpu_time_used;
 581         double flows_rate;
 582         double cpu_time_per_iter[MAX_ITERATIONS];
 583         double delta;
 584         uint16_t nr_ports;
 585         uint32_t i;
 586         int port_id;
 587         int iter_id;
 588         uint32_t flow_index;
 589         uint64_t global_items[MAX_ITEMS_NUM] = { 0 };
 590         uint64_t global_actions[MAX_ACTIONS_NUM] = { 0 };
 591
 592         global_items[0] = FLOW_ITEM_MASK(RTE_FLOW_ITEM_TYPE_ETH);
 593         global_actions[0] = FLOW_ITEM_MASK(RTE_FLOW_ACTION_TYPE_JUMP);
 594
 595         nr_ports = rte_eth_dev_count_avail();
 596
 597         for (i = 0; i < MAX_ITERATIONS; i++)
 598                 cpu_time_per_iter[i] = -1;
 599
 600         if (iterations_number > flows_count)
 601                 iterations_number = flows_count;
 602
 603         printf(":: Flows Count per port: %d\n", flows_count);
 604
 605         flow_list = rte_zmalloc("flow_list",
 606                 (sizeof(struct rte_flow *) * flows_count) + 1, 0);
 607         if (flow_list == NULL)
 608                 rte_exit(EXIT_FAILURE, "No Memory available!");
 609
 610         for (port_id = 0; port_id < nr_ports; port_id++) {
 611                 cpu_time_used = 0;
 612                 flow_index = 0;
 613                 if (flow_group > 0) {
 614                         /*
 615                          * Create global rule to jump into flow_group,
 616                          * this way the app will avoid the default rules.
 617                          *
 618                          * Global rule:
 619                          * group 0 eth / end actions jump group <flow_group>
 620                          *
 621                          */
 622                         flow = generate_flow(port_id, 0, flow_attrs,
 623                                 global_items, global_actions,
 624                                 flow_group, 0, 0, &error);
 625
 626                         if (flow == NULL) {
 627                                 print_flow_error(error);
 628                                 rte_exit(EXIT_FAILURE, "error in creating flow");
 629                         }
 630                         flow_list[flow_index++] = flow;
 631                 }
 632
 633                 /* Insertion Rate */
 634                 printf("Flows insertion on port = %d\n", port_id);
 635                 start_iter = clock();
 636                 for (i = 0; i < flows_count; i++) {
 637                         flow = generate_flow(port_id, flow_group,
 638                                 flow_attrs, flow_items, flow_actions,
 639                                 JUMP_ACTION_TABLE, i,
 640                                 hairpin_queues_num, &error);
 641
 642                         if (force_quit)
 643                                 i = flows_count;
 644
 645                         if (!flow) {
 646                                 print_flow_error(error);
 647                                 rte_exit(EXIT_FAILURE, "error in creating flow");
 648                         }
 649
 650                         flow_list[flow_index++] = flow;
 651
 652                         if (i && !((i + 1) % iterations_number)) {
 653                                 /* Save the insertion rate of each iter */
 654                                 end_iter = clock();
 655                                 delta = (double) (end_iter - start_iter);
 656                                 iter_id = ((i + 1) / iterations_number) - 1;
 657                                 cpu_time_per_iter[iter_id] =
 658                                         delta / CLOCKS_PER_SEC;
 659                                 cpu_time_used += cpu_time_per_iter[iter_id];
 660                                 start_iter = clock();
 661                         }
 662                 }
 663
 664                 /* Iteration rate per iteration */
 665                 if (dump_iterations)
 666                         for (i = 0; i < MAX_ITERATIONS; i++) {
 667                                 if (cpu_time_per_iter[i] == -1)
 668                                         continue;
 669                                 delta = (double)(iterations_number /
 670                                         cpu_time_per_iter[i]);
 671                                 flows_rate = delta / 1000;
 672                                 printf(":: Iteration #%d: %d flows "
 673                                         "in %f sec[ Rate = %f K/Sec ]\n",
 674                                         i, iterations_number,
 675                                         cpu_time_per_iter[i], flows_rate);
 676                         }
 677
 678                 /* Insertion rate for all flows */
 679                 flows_rate = ((double) (flows_count / cpu_time_used) / 1000);
 680                 printf("\n:: Total flow insertion rate -> %f K/Sec\n",
 681                                                 flows_rate);
 682                 printf(":: The time for creating %d in flows %f seconds\n",
 683                                                 flows_count, cpu_time_used);
 684
 685                 if (delete_flag)
 686                         destroy_flows(port_id, flow_list);
 687         }
 688 }
 689
 690 static void
 691 signal_handler(int signum)
 692 {
 693         if (signum == SIGINT || signum == SIGTERM) {
 694                 printf("\n\nSignal %d received, preparing to exit...\n",
 695                                         signum);
 696                 printf("Error: Stats are wrong due to sudden signal!\n\n");
 697                 force_quit = true;
 698         }
 699 }
 700
 701 static inline uint16_t
 702 do_rx(struct lcore_info *li, uint16_t rx_port, uint16_t rx_queue)
 703 {
 704         uint16_t cnt = 0;
 705         cnt = rte_eth_rx_burst(rx_port, rx_queue, li->pkts, MAX_PKT_BURST);
 706         li->rx_pkts += cnt;
 707         return cnt;
 708 }
 709
 710 static inline void
 711 do_tx(struct lcore_info *li, uint16_t cnt, uint16_t tx_port,
 712                         uint16_t tx_queue)
 713 {
 714         uint16_t nr_tx = 0;
 715         uint16_t i;
 716
 717         nr_tx = rte_eth_tx_burst(tx_port, tx_queue, li->pkts, cnt);
 718         li->tx_pkts  += nr_tx;
 719         li->tx_drops += cnt - nr_tx;
 720
 721         for (i = nr_tx; i < cnt; i++)
 722                 rte_pktmbuf_free(li->pkts[i]);
 723 }
 724
 725 /*
 726  * Method to convert numbers into pretty numbers that easy
 727  * to read. The design here is to add comma after each three
 728  * digits and set all of this inside buffer.
 729  *
 730  * For example if n = 1799321, the output will be
 731  * 1,799,321 after this method which is easier to read.
 732  */
 733 static char *
 734 pretty_number(uint64_t n, char *buf)
 735 {
 736         char p[6][4];
 737         int i = 0;
 738         int off = 0;
 739
 740         while (n > 1000) {
 741                 sprintf(p[i], "%03d", (int)(n % 1000));
 742                 n /= 1000;
 743                 i += 1;
 744         }
 745
 746         sprintf(p[i++], "%d", (int)n);
 747
 748         while (i--)
 749                 off += sprintf(buf + off, "%s,", p[i]);
 750         buf[strlen(buf) - 1] = '\0';
 751
 752         return buf;
 753 }
 754
 755 static void
 756 packet_per_second_stats(void)
 757 {
 758         struct lcore_info *old;
 759         struct lcore_info *li, *oli;
 760         int nr_lines = 0;
 761         int i;
 762
 763         old = rte_zmalloc("old",
 764                 sizeof(struct lcore_info) * MAX_LCORES, 0);
 765         if (old == NULL)
 766                 rte_exit(EXIT_FAILURE, "No Memory available!");
 767
 768         memcpy(old, lcore_infos,
 769                 sizeof(struct lcore_info) * MAX_LCORES);
 770
 771         while (!force_quit) {
 772                 uint64_t total_tx_pkts = 0;
 773                 uint64_t total_rx_pkts = 0;
 774                 uint64_t total_tx_drops = 0;
 775                 uint64_t tx_delta, rx_delta, drops_delta;
 776                 char buf[3][32];
 777                 int nr_valid_core = 0;
 778
 779                 sleep(1);
 780
 781                 if (nr_lines) {
 782                         char go_up_nr_lines[16];
 783
 784                         sprintf(go_up_nr_lines, "%c[%dA\r", 27, nr_lines);
 785                         printf("%s\r", go_up_nr_lines);
 786                 }
 787
 788                 printf("\n%6s %16s %16s %16s\n", "core", "tx", "tx drops", "rx");
 789                 printf("%6s %16s %16s %16s\n", "------", "----------------",
 790                         "----------------", "----------------");
 791                 nr_lines = 3;
 792                 for (i = 0; i < MAX_LCORES; i++) {
 793                         li  = &lcore_infos[i];
 794                         oli = &old[i];
 795                         if (li->mode != LCORE_MODE_PKT)
 796                                 continue;
 797
 798                         tx_delta    = li->tx_pkts  - oli->tx_pkts;
 799                         rx_delta    = li->rx_pkts  - oli->rx_pkts;
 800                         drops_delta = li->tx_drops - oli->tx_drops;
 801                         printf("%6d %16s %16s %16s\n", i,
 802                                 pretty_number(tx_delta,    buf[0]),
 803                                 pretty_number(drops_delta, buf[1]),
 804                                 pretty_number(rx_delta,    buf[2]));
 805
 806                         total_tx_pkts  += tx_delta;
 807                         total_rx_pkts  += rx_delta;
 808                         total_tx_drops += drops_delta;
 809
 810                         nr_valid_core++;
 811                         nr_lines += 1;
 812                 }
 813
 814                 if (nr_valid_core > 1) {
 815                         printf("%6s %16s %16s %16s\n", "total",
 816                                 pretty_number(total_tx_pkts,  buf[0]),
 817                                 pretty_number(total_tx_drops, buf[1]),
 818                                 pretty_number(total_rx_pkts,  buf[2]));
 819                         nr_lines += 1;
 820                 }
 821
 822                 memcpy(old, lcore_infos,
 823                         sizeof(struct lcore_info) * MAX_LCORES);
 824         }
 825 }
 826
 827 static int
 828 start_forwarding(void *data __rte_unused)
 829 {
 830         int lcore = rte_lcore_id();
 831         int stream_id;
 832         uint16_t cnt;
 833         struct lcore_info *li = &lcore_infos[lcore];
 834
 835         if (!li->mode)
 836                 return 0;
 837
 838         if (li->mode == LCORE_MODE_STATS) {
 839                 printf(":: started stats on lcore %u\n", lcore);
 840                 packet_per_second_stats();
 841                 return 0;
 842         }
 843
 844         while (!force_quit)
 845                 for (stream_id = 0; stream_id < MAX_STREAMS; stream_id++) {
 846                         if (li->streams[stream_id].rx_port == -1)
 847                                 continue;
 848
 849                         cnt = do_rx(li,
 850                                         li->streams[stream_id].rx_port,
 851                                         li->streams[stream_id].rx_queue);
 852                         if (cnt)
 853                                 do_tx(li, cnt,
 854                                         li->streams[stream_id].tx_port,
 855                                         li->streams[stream_id].tx_queue);
 856                 }
 857         return 0;
 858 }
 859
 860 static void
 861 init_lcore_info(void)
 862 {
 863         int i, j;
 864         unsigned int lcore;
 865         uint16_t nr_port;
 866         uint16_t queue;
 867         int port;
 868         int stream_id = 0;
 869         int streams_per_core;
 870         int unassigned_streams;
 871         int nb_fwd_streams;
 872         nr_port = rte_eth_dev_count_avail();
 873
 874         /* First logical core is reserved for stats printing */
 875         lcore = rte_get_next_lcore(-1, 0, 0);
 876         lcore_infos[lcore].mode = LCORE_MODE_STATS;
 877
 878         /*
 879          * Initialize all cores
 880          * All cores at first must have -1 value in all streams
 881          * This means that this stream is not used, or not set
 882          * yet.
 883          */
 884         for (i = 0; i < MAX_LCORES; i++)
 885                 for (j = 0; j < MAX_STREAMS; j++) {
 886                         lcore_infos[i].streams[j].tx_port = -1;
 887                         lcore_infos[i].streams[j].rx_port = -1;
 888                         lcore_infos[i].streams[j].tx_queue = -1;
 889                         lcore_infos[i].streams[j].rx_queue = -1;
 890                         lcore_infos[i].streams_nb = 0;
 891                 }
 892
 893         /*
 894          * Calculate the total streams count.
 895          * Also distribute those streams count between the available
 896          * logical cores except first core, since it's reserved for
 897          * stats prints.
 898          */
 899         nb_fwd_streams = nr_port * RXQ_NUM;
 900         if ((int)(nb_lcores - 1) >= nb_fwd_streams)
 901                 for (i = 0; i < (int)(nb_lcores - 1); i++) {
 902                         lcore = rte_get_next_lcore(lcore, 0, 0);
 903                         lcore_infos[lcore].streams_nb = 1;
 904                 }
 905         else {
 906                 streams_per_core = nb_fwd_streams / (nb_lcores - 1);
 907                 unassigned_streams = nb_fwd_streams % (nb_lcores - 1);
 908                 for (i = 0; i < (int)(nb_lcores - 1); i++) {
 909                         lcore = rte_get_next_lcore(lcore, 0, 0);
 910                         lcore_infos[lcore].streams_nb = streams_per_core;
 911                         if (unassigned_streams) {
 912                                 lcore_infos[lcore].streams_nb++;
 913                                 unassigned_streams--;
 914                         }
 915                 }
 916         }
 917
 918         /*
 919          * Set the streams for the cores according to each logical
 920          * core stream count.
 921          * The streams is built on the design of what received should
 922          * forward as well, this means that if you received packets on
 923          * port 0 queue 0 then the same queue should forward the
 924          * packets, using the same logical core.
 925          */
 926         lcore = rte_get_next_lcore(-1, 0, 0);
 927         for (port = 0; port < nr_port; port++) {
 928                 /* Create FWD stream */
 929                 for (queue = 0; queue < RXQ_NUM; queue++) {
 930                         if (!lcore_infos[lcore].streams_nb ||
 931                                 !(stream_id % lcore_infos[lcore].streams_nb)) {
 932                                 lcore = rte_get_next_lcore(lcore, 0, 0);
 933                                 lcore_infos[lcore].mode = LCORE_MODE_PKT;
 934                                 stream_id = 0;
 935                         }
 936                         lcore_infos[lcore].streams[stream_id].rx_queue = queue;
 937                         lcore_infos[lcore].streams[stream_id].tx_queue = queue;
 938                         lcore_infos[lcore].streams[stream_id].rx_port = port;
 939                         lcore_infos[lcore].streams[stream_id].tx_port = port;
 940                         stream_id++;
 941                 }
 942         }
 943
 944         /* Print all streams */
 945         printf(":: Stream -> core id[N]: (rx_port, rx_queue)->(tx_port, tx_queue)\n");
 946         for (i = 0; i < MAX_LCORES; i++)
 947                 for (j = 0; j < MAX_STREAMS; j++) {
 948                         /* No streams for this core */
 949                         if (lcore_infos[i].streams[j].tx_port == -1)
 950                                 break;
 951                         printf("Stream -> core id[%d]: (%d,%d)->(%d,%d)\n",
 952                                 i,
 953                                 lcore_infos[i].streams[j].rx_port,
 954                                 lcore_infos[i].streams[j].rx_queue,
 955                                 lcore_infos[i].streams[j].tx_port,
 956                                 lcore_infos[i].streams[j].tx_queue);
 957                 }
 958 }
 959
 960 static void
 961 init_port(void)
 962 {
 963         int ret;
 964         uint16_t std_queue;
 965         uint16_t hairpin_queue;
 966         uint16_t port_id;
 967         uint16_t nr_ports;
 968         uint16_t nr_queues;
 969         struct rte_eth_hairpin_conf hairpin_conf = {
 970                 .peer_count = 1,
 971         };
 972         struct rte_eth_conf port_conf = {
 973                 .rx_adv_conf = {
 974                         .rss_conf.rss_hf =
 975                                 GET_RSS_HF(),
 976                 }
 977         };
 978         struct rte_eth_txconf txq_conf;
 979         struct rte_eth_rxconf rxq_conf;
 980         struct rte_eth_dev_info dev_info;
 981
 982         nr_queues = RXQ_NUM;
 983         if (hairpin_queues_num != 0)
 984                 nr_queues = RXQ_NUM + hairpin_queues_num;
 985
 986         nr_ports = rte_eth_dev_count_avail();
 987         if (nr_ports == 0)
 988                 rte_exit(EXIT_FAILURE, "Error: no port detected\n");
 989
 990         mbuf_mp = rte_pktmbuf_pool_create("mbuf_pool",
 991                                         TOTAL_MBUF_NUM, MBUF_CACHE_SIZE,
 992                                         0, MBUF_SIZE,
 993                                         rte_socket_id());
 994         if (mbuf_mp == NULL)
 995                 rte_exit(EXIT_FAILURE, "Error: can't init mbuf pool\n");
 996
 997         for (port_id = 0; port_id < nr_ports; port_id++) {
 998                 ret = rte_eth_dev_info_get(port_id, &dev_info);
 999                 if (ret != 0)
1000                         rte_exit(EXIT_FAILURE,
1001                                 "Error during getting device"
1002                                 " (port %u) info: %s\n",
1003                                 port_id, strerror(-ret));
1004
1005                 port_conf.txmode.offloads &= dev_info.tx_offload_capa;
1006                 port_conf.rxmode.offloads &= dev_info.rx_offload_capa;
1007
1008                 printf(":: initializing port: %d\n", port_id);
1009
1010                 ret = rte_eth_dev_configure(port_id, nr_queues,
1011                                 nr_queues, &port_conf);
1012                 if (ret < 0)
1013                         rte_exit(EXIT_FAILURE,
1014                                 ":: cannot configure device: err=%d, port=%u\n",
1015                                 ret, port_id);
1016
1017                 rxq_conf = dev_info.default_rxconf;
1018                 for (std_queue = 0; std_queue < RXQ_NUM; std_queue++) {
1019                         ret = rte_eth_rx_queue_setup(port_id, std_queue, NR_RXD,
1020                                         rte_eth_dev_socket_id(port_id),
1021                                         &rxq_conf,
1022                                         mbuf_mp);
1023                         if (ret < 0)
1024                                 rte_exit(EXIT_FAILURE,
1025                                         ":: Rx queue setup failed: err=%d, port=%u\n",
1026                                         ret, port_id);
1027                 }
1028
1029                 txq_conf = dev_info.default_txconf;
1030                 for (std_queue = 0; std_queue < TXQ_NUM; std_queue++) {
1031                         ret = rte_eth_tx_queue_setup(port_id, std_queue, NR_TXD,
1032                                         rte_eth_dev_socket_id(port_id),
1033                                         &txq_conf);
1034                         if (ret < 0)
1035                                 rte_exit(EXIT_FAILURE,
1036                                         ":: Tx queue setup failed: err=%d, port=%u\n",
1037                                         ret, port_id);
1038                 }
1039
1040                 /* Catch all packets from traffic generator. */
1041                 ret = rte_eth_promiscuous_enable(port_id);
1042                 if (ret != 0)
1043                         rte_exit(EXIT_FAILURE,
1044                                 ":: promiscuous mode enable failed: err=%s, port=%u\n",
1045                                 rte_strerror(-ret), port_id);
1046
1047                 if (hairpin_queues_num != 0) {
1048                         /*
1049                          * Configure peer which represents hairpin Tx.
1050                          * Hairpin queue numbers start after standard queues
1051                          * (RXQ_NUM and TXQ_NUM).
1052                          */
1053                         for (hairpin_queue = RXQ_NUM, std_queue = 0;
1054                                         hairpin_queue < nr_queues;
1055                                         hairpin_queue++, std_queue++) {
1056                                 hairpin_conf.peers[0].port = port_id;
1057                                 hairpin_conf.peers[0].queue =
1058                                         std_queue + TXQ_NUM;
1059                                 ret = rte_eth_rx_hairpin_queue_setup(
1060                                                 port_id, hairpin_queue,
1061                                                 NR_RXD, &hairpin_conf);
1062                                 if (ret != 0)
1063                                         rte_exit(EXIT_FAILURE,
1064                                                 ":: Hairpin rx queue setup failed: err=%d, port=%u\n",
1065                                                 ret, port_id);
1066                         }
1067
1068                         for (hairpin_queue = TXQ_NUM, std_queue = 0;
1069                                         hairpin_queue < nr_queues;
1070                                         hairpin_queue++, std_queue++) {
1071                                 hairpin_conf.peers[0].port = port_id;
1072                                 hairpin_conf.peers[0].queue =
1073                                         std_queue + RXQ_NUM;
1074                                 ret = rte_eth_tx_hairpin_queue_setup(
1075                                                 port_id, hairpin_queue,
1076                                                 NR_TXD, &hairpin_conf);
1077                                 if (ret != 0)
1078                                         rte_exit(EXIT_FAILURE,
1079                                                 ":: Hairpin tx queue setup failed: err=%d, port=%u\n",
1080                                                 ret, port_id);
1081                         }
1082                 }
1083
1084                 ret = rte_eth_dev_start(port_id);
1085                 if (ret < 0)
1086                         rte_exit(EXIT_FAILURE,
1087                                 "rte_eth_dev_start:err=%d, port=%u\n",
1088                                 ret, port_id);
1089
1090                 printf(":: initializing port: %d done\n", port_id);
1091         }
1092 }
1093
1094 int
1095 main(int argc, char **argv)
1096 {
1097         int ret;
1098         uint16_t port;
1099         struct rte_flow_error error;
1100         int64_t alloc, last_alloc;
1101
1102         ret = rte_eal_init(argc, argv);
1103         if (ret < 0)
1104                 rte_exit(EXIT_FAILURE, "EAL init failed\n");
1105
1106         force_quit = false;
1107         dump_iterations = false;
1108         flows_count = DEFAULT_RULES_COUNT;
1109         iterations_number = DEFAULT_ITERATION;
1110         delete_flag = false;
1111         dump_socket_mem_flag = false;
1112         flow_group = 0;
1113
1114         signal(SIGINT, signal_handler);
1115         signal(SIGTERM, signal_handler);
1116
1117         argc -= ret;
1118         argv += ret;
1119         if (argc > 1)
1120                 args_parse(argc, argv);
1121
1122         init_port();
1123
1124         nb_lcores = rte_lcore_count();
1125         if (nb_lcores <= 1)
1126                 rte_exit(EXIT_FAILURE, "This app needs at least two cores\n");
1127
1128         last_alloc = (int64_t)dump_socket_mem(stdout);
1129         flows_handler();
1130         alloc = (int64_t)dump_socket_mem(stdout);
1131
1132         if (last_alloc)
1133                 fprintf(stdout, ":: Memory allocation change(M): %.6lf\n",
1134                 (alloc - last_alloc) / 1.0e6);
1135
1136         if (enable_fwd) {
1137                 init_lcore_info();
1138                 rte_eal_mp_remote_launch(start_forwarding, NULL, CALL_MASTER);
1139         }
1140
1141         RTE_ETH_FOREACH_DEV(port) {
1142                 rte_flow_flush(port, &error);
1143                 rte_eth_dev_stop(port);
1144                 rte_eth_dev_close(port);
1145         }
1146         return 0;
1147 }