examples/eventdev_pipeline_sw_pmd/main.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2016-2017 Intel Corporation
   3  */
   4
   5 #include <getopt.h>
   6 #include <stdint.h>
   7 #include <stdio.h>
   8 #include <signal.h>
   9 #include <sched.h>
  10
  11 #include "pipeline_common.h"
  12
  13 struct config_data cdata = {
  14         .num_packets = (1L << 25), /* do ~32M packets */
  15         .num_fids = 512,
  16         .queue_type = RTE_SCHED_TYPE_ATOMIC,
  17         .next_qid = {-1},
  18         .qid = {-1},
  19         .num_stages = 1,
  20         .worker_cq_depth = 16
  21 };
  22
  23 static bool
  24 core_in_use(unsigned int lcore_id) {
  25         return (fdata->rx_core[lcore_id] || fdata->sched_core[lcore_id] ||
  26                 fdata->tx_core[lcore_id] || fdata->worker_core[lcore_id]);
  27 }
  28
  29 static void
  30 eth_tx_buffer_retry(struct rte_mbuf **pkts, uint16_t unsent,
  31                         void *userdata)
  32 {
  33         int port_id = (uintptr_t) userdata;
  34         unsigned int _sent = 0;
  35
  36         do {
  37                 /* Note: hard-coded TX queue */
  38                 _sent += rte_eth_tx_burst(port_id, 0, &pkts[_sent],
  39                                           unsent - _sent);
  40         } while (_sent != unsent);
  41 }
  42
  43 /*
  44  * Parse the coremask given as argument (hexadecimal string) and fill
  45  * the global configuration (core role and core count) with the parsed
  46  * value.
  47  */
  48 static int xdigit2val(unsigned char c)
  49 {
  50         int val;
  51
  52         if (isdigit(c))
  53                 val = c - '0';
  54         else if (isupper(c))
  55                 val = c - 'A' + 10;
  56         else
  57                 val = c - 'a' + 10;
  58         return val;
  59 }
  60
  61 static uint64_t
  62 parse_coremask(const char *coremask)
  63 {
  64         int i, j, idx = 0;
  65         unsigned int count = 0;
  66         char c;
  67         int val;
  68         uint64_t mask = 0;
  69         const int32_t BITS_HEX = 4;
  70
  71         if (coremask == NULL)
  72                 return -1;
  73         /* Remove all blank characters ahead and after .
  74          * Remove 0x/0X if exists.
  75          */
  76         while (isblank(*coremask))
  77                 coremask++;
  78         if (coremask[0] == '0' && ((coremask[1] == 'x')
  79                 || (coremask[1] == 'X')))
  80                 coremask += 2;
  81         i = strlen(coremask);
  82         while ((i > 0) && isblank(coremask[i - 1]))
  83                 i--;
  84         if (i == 0)
  85                 return -1;
  86
  87         for (i = i - 1; i >= 0 && idx < MAX_NUM_CORE; i--) {
  88                 c = coremask[i];
  89                 if (isxdigit(c) == 0) {
  90                         /* invalid characters */
  91                         return -1;
  92                 }
  93                 val = xdigit2val(c);
  94                 for (j = 0; j < BITS_HEX && idx < MAX_NUM_CORE; j++, idx++) {
  95                         if ((1 << j) & val) {
  96                                 mask |= (1UL << idx);
  97                                 count++;
  98                         }
  99                 }
 100         }
 101         for (; i >= 0; i--)
 102                 if (coremask[i] != '0')
 103                         return -1;
 104         if (count == 0)
 105                 return -1;
 106         return mask;
 107 }
 108
 109 static struct option long_options[] = {
 110         {"workers", required_argument, 0, 'w'},
 111         {"packets", required_argument, 0, 'n'},
 112         {"atomic-flows", required_argument, 0, 'f'},
 113         {"num_stages", required_argument, 0, 's'},
 114         {"rx-mask", required_argument, 0, 'r'},
 115         {"tx-mask", required_argument, 0, 't'},
 116         {"sched-mask", required_argument, 0, 'e'},
 117         {"cq-depth", required_argument, 0, 'c'},
 118         {"work-cycles", required_argument, 0, 'W'},
 119         {"queue-priority", no_argument, 0, 'P'},
 120         {"parallel", no_argument, 0, 'p'},
 121         {"ordered", no_argument, 0, 'o'},
 122         {"quiet", no_argument, 0, 'q'},
 123         {"dump", no_argument, 0, 'D'},
 124         {0, 0, 0, 0}
 125 };
 126
 127 static void
 128 usage(void)
 129 {
 130         const char *usage_str =
 131                 "  Usage: eventdev_demo [options]\n"
 132                 "  Options:\n"
 133                 "  -n, --packets=N              Send N packets (default ~32M), 0 implies no limit\n"
 134                 "  -f, --atomic-flows=N         Use N random flows from 1 to N (default 16)\n"
 135                 "  -s, --num_stages=N           Use N atomic stages (default 1)\n"
 136                 "  -r, --rx-mask=core mask      Run NIC rx on CPUs in core mask\n"
 137                 "  -w, --worker-mask=core mask  Run worker on CPUs in core mask\n"
 138                 "  -t, --tx-mask=core mask      Run NIC tx on CPUs in core mask\n"
 139                 "  -e  --sched-mask=core mask   Run scheduler on CPUs in core mask\n"
 140                 "  -c  --cq-depth=N             Worker CQ depth (default 16)\n"
 141                 "  -W  --work-cycles=N          Worker cycles (default 0)\n"
 142                 "  -P  --queue-priority         Enable scheduler queue prioritization\n"
 143                 "  -o, --ordered                Use ordered scheduling\n"
 144                 "  -p, --parallel               Use parallel scheduling\n"
 145                 "  -q, --quiet                  Minimize printed output\n"
 146                 "  -D, --dump                   Print detailed statistics before exit"
 147                 "\n";
 148         fprintf(stderr, "%s", usage_str);
 149         exit(1);
 150 }
 151
 152 static void
 153 parse_app_args(int argc, char **argv)
 154 {
 155         /* Parse cli options*/
 156         int option_index;
 157         int c;
 158         opterr = 0;
 159         uint64_t rx_lcore_mask = 0;
 160         uint64_t tx_lcore_mask = 0;
 161         uint64_t sched_lcore_mask = 0;
 162         uint64_t worker_lcore_mask = 0;
 163         int i;
 164
 165         for (;;) {
 166                 c = getopt_long(argc, argv, "r:t:e:c:w:n:f:s:poPqDW:",
 167                                 long_options, &option_index);
 168                 if (c == -1)
 169                         break;
 170
 171                 int popcnt = 0;
 172                 switch (c) {
 173                 case 'n':
 174                         cdata.num_packets = (int64_t)atol(optarg);
 175                         if (cdata.num_packets == 0)
 176                                 cdata.num_packets = INT64_MAX;
 177                         break;
 178                 case 'f':
 179                         cdata.num_fids = (unsigned int)atoi(optarg);
 180                         break;
 181                 case 's':
 182                         cdata.num_stages = (unsigned int)atoi(optarg);
 183                         break;
 184                 case 'c':
 185                         cdata.worker_cq_depth = (unsigned int)atoi(optarg);
 186                         break;
 187                 case 'W':
 188                         cdata.worker_cycles = (unsigned int)atoi(optarg);
 189                         break;
 190                 case 'P':
 191                         cdata.enable_queue_priorities = 1;
 192                         break;
 193                 case 'o':
 194                         cdata.queue_type = RTE_SCHED_TYPE_ORDERED;
 195                         break;
 196                 case 'p':
 197                         cdata.queue_type = RTE_SCHED_TYPE_PARALLEL;
 198                         break;
 199                 case 'q':
 200                         cdata.quiet = 1;
 201                         break;
 202                 case 'D':
 203                         cdata.dump_dev = 1;
 204                         break;
 205                 case 'w':
 206                         worker_lcore_mask = parse_coremask(optarg);
 207                         break;
 208                 case 'r':
 209                         rx_lcore_mask = parse_coremask(optarg);
 210                         popcnt = __builtin_popcountll(rx_lcore_mask);
 211                         fdata->rx_single = (popcnt == 1);
 212                         break;
 213                 case 't':
 214                         tx_lcore_mask = parse_coremask(optarg);
 215                         popcnt = __builtin_popcountll(tx_lcore_mask);
 216                         fdata->tx_single = (popcnt == 1);
 217                         break;
 218                 case 'e':
 219                         sched_lcore_mask = parse_coremask(optarg);
 220                         popcnt = __builtin_popcountll(sched_lcore_mask);
 221                         fdata->sched_single = (popcnt == 1);
 222                         break;
 223                 default:
 224                         usage();
 225                 }
 226         }
 227
 228         if (worker_lcore_mask == 0 || rx_lcore_mask == 0 ||
 229             sched_lcore_mask == 0 || tx_lcore_mask == 0) {
 230                 printf("Core part of pipeline was not assigned any cores. "
 231                         "This will stall the pipeline, please check core masks "
 232                         "(use -h for details on setting core masks):\n"
 233                         "\trx: %"PRIu64"\n\ttx: %"PRIu64"\n\tsched: %"PRIu64
 234                         "\n\tworkers: %"PRIu64"\n",
 235                         rx_lcore_mask, tx_lcore_mask, sched_lcore_mask,
 236                         worker_lcore_mask);
 237                 rte_exit(-1, "Fix core masks\n");
 238         }
 239         if (cdata.num_stages == 0 || cdata.num_stages > MAX_NUM_STAGES)
 240                 usage();
 241
 242         for (i = 0; i < MAX_NUM_CORE; i++) {
 243                 fdata->rx_core[i] = !!(rx_lcore_mask & (1UL << i));
 244                 fdata->tx_core[i] = !!(tx_lcore_mask & (1UL << i));
 245                 fdata->sched_core[i] = !!(sched_lcore_mask & (1UL << i));
 246                 fdata->worker_core[i] = !!(worker_lcore_mask & (1UL << i));
 247
 248                 if (fdata->worker_core[i])
 249                         cdata.num_workers++;
 250                 if (core_in_use(i))
 251                         cdata.active_cores++;
 252         }
 253 }
 254
 255 /*
 256  * Initializes a given port using global settings and with the RX buffers
 257  * coming from the mbuf_pool passed as a parameter.
 258  */
 259 static inline int
 260 port_init(uint8_t port, struct rte_mempool *mbuf_pool)
 261 {
 262         static const struct rte_eth_conf port_conf_default = {
 263                 .rxmode = {
 264                         .mq_mode = ETH_MQ_RX_RSS,
 265                         .max_rx_pkt_len = ETHER_MAX_LEN,
 266                         .ignore_offload_bitfield = 1,
 267                 },
 268                 .rx_adv_conf = {
 269                         .rss_conf = {
 270                                 .rss_hf = ETH_RSS_IP |
 271                                           ETH_RSS_TCP |
 272                                           ETH_RSS_UDP,
 273                         }
 274                 }
 275         };
 276         const uint16_t rx_rings = 1, tx_rings = 1;
 277         const uint16_t rx_ring_size = 512, tx_ring_size = 512;
 278         struct rte_eth_conf port_conf = port_conf_default;
 279         int retval;
 280         uint16_t q;
 281         struct rte_eth_dev_info dev_info;
 282         struct rte_eth_txconf txconf;
 283
 284         if (port >= rte_eth_dev_count())
 285                 return -1;
 286
 287         rte_eth_dev_info_get(port, &dev_info);
 288         if (dev_info.tx_offload_capa & DEV_TX_OFFLOAD_MBUF_FAST_FREE)
 289                 port_conf.txmode.offloads |=
 290                         DEV_TX_OFFLOAD_MBUF_FAST_FREE;
 291
 292         /* Configure the Ethernet device. */
 293         retval = rte_eth_dev_configure(port, rx_rings, tx_rings, &port_conf);
 294         if (retval != 0)
 295                 return retval;
 296
 297         /* Allocate and set up 1 RX queue per Ethernet port. */
 298         for (q = 0; q < rx_rings; q++) {
 299                 retval = rte_eth_rx_queue_setup(port, q, rx_ring_size,
 300                                 rte_eth_dev_socket_id(port), NULL, mbuf_pool);
 301                 if (retval < 0)
 302                         return retval;
 303         }
 304
 305         txconf = dev_info.default_txconf;
 306         txconf.txq_flags = ETH_TXQ_FLAGS_IGNORE;
 307         txconf.offloads = port_conf_default.txmode.offloads;
 308         /* Allocate and set up 1 TX queue per Ethernet port. */
 309         for (q = 0; q < tx_rings; q++) {
 310                 retval = rte_eth_tx_queue_setup(port, q, tx_ring_size,
 311                                 rte_eth_dev_socket_id(port), &txconf);
 312                 if (retval < 0)
 313                         return retval;
 314         }
 315
 316         /* Start the Ethernet port. */
 317         retval = rte_eth_dev_start(port);
 318         if (retval < 0)
 319                 return retval;
 320
 321         /* Display the port MAC address. */
 322         struct ether_addr addr;
 323         rte_eth_macaddr_get(port, &addr);
 324         printf("Port %u MAC: %02" PRIx8 " %02" PRIx8 " %02" PRIx8
 325                            " %02" PRIx8 " %02" PRIx8 " %02" PRIx8 "\n",
 326                         (unsigned int)port,
 327                         addr.addr_bytes[0], addr.addr_bytes[1],
 328                         addr.addr_bytes[2], addr.addr_bytes[3],
 329                         addr.addr_bytes[4], addr.addr_bytes[5]);
 330
 331         /* Enable RX in promiscuous mode for the Ethernet device. */
 332         rte_eth_promiscuous_enable(port);
 333
 334         return 0;
 335 }
 336
 337 static int
 338 init_ports(unsigned int num_ports)
 339 {
 340         uint8_t portid;
 341         unsigned int i;
 342
 343         struct rte_mempool *mp = rte_pktmbuf_pool_create("packet_pool",
 344                         /* mbufs */ 16384 * num_ports,
 345                         /* cache_size */ 512,
 346                         /* priv_size*/ 0,
 347                         /* data_room_size */ RTE_MBUF_DEFAULT_BUF_SIZE,
 348                         rte_socket_id());
 349
 350         for (portid = 0; portid < num_ports; portid++)
 351                 if (port_init(portid, mp) != 0)
 352                         rte_exit(EXIT_FAILURE, "Cannot init port %"PRIu8 "\n",
 353                                         portid);
 354
 355         for (i = 0; i < num_ports; i++) {
 356                 void *userdata = (void *)(uintptr_t) i;
 357                 fdata->tx_buf[i] =
 358                         rte_malloc(NULL, RTE_ETH_TX_BUFFER_SIZE(32), 0);
 359                 if (fdata->tx_buf[i] == NULL)
 360                         rte_panic("Out of memory\n");
 361                 rte_eth_tx_buffer_init(fdata->tx_buf[i], 32);
 362                 rte_eth_tx_buffer_set_err_callback(fdata->tx_buf[i],
 363                                                    eth_tx_buffer_retry,
 364                                                    userdata);
 365         }
 366
 367         return 0;
 368 }
 369
 370 static void
 371 do_capability_setup(uint16_t nb_ethdev, uint8_t eventdev_id)
 372 {
 373         RTE_SET_USED(nb_ethdev);
 374         RTE_SET_USED(eventdev_id);
 375         set_worker_generic_setup_data(&fdata->cap, 1);
 376 }
 377
 378 static void
 379 signal_handler(int signum)
 380 {
 381         if (fdata->done)
 382                 rte_exit(1, "Exiting on signal %d\n", signum);
 383         if (signum == SIGINT || signum == SIGTERM) {
 384                 printf("\n\nSignal %d received, preparing to exit...\n",
 385                                 signum);
 386                 fdata->done = 1;
 387         }
 388         if (signum == SIGTSTP)
 389                 rte_event_dev_dump(0, stdout);
 390 }
 391
 392 static inline uint64_t
 393 port_stat(int dev_id, int32_t p)
 394 {
 395         char statname[64];
 396         snprintf(statname, sizeof(statname), "port_%u_rx", p);
 397         return rte_event_dev_xstats_by_name_get(dev_id, statname, NULL);
 398 }
 399
 400 int
 401 main(int argc, char **argv)
 402 {
 403         struct worker_data *worker_data;
 404         unsigned int num_ports;
 405         int lcore_id;
 406         int err;
 407
 408         signal(SIGINT, signal_handler);
 409         signal(SIGTERM, signal_handler);
 410         signal(SIGTSTP, signal_handler);
 411
 412         err = rte_eal_init(argc, argv);
 413         if (err < 0)
 414                 rte_panic("Invalid EAL arguments\n");
 415
 416         argc -= err;
 417         argv += err;
 418
 419         fdata = rte_malloc(NULL, sizeof(struct fastpath_data), 0);
 420         if (fdata == NULL)
 421                 rte_panic("Out of memory\n");
 422
 423         /* Parse cli options*/
 424         parse_app_args(argc, argv);
 425
 426         num_ports = rte_eth_dev_count();
 427         if (num_ports == 0)
 428                 rte_panic("No ethernet ports found\n");
 429
 430         const unsigned int cores_needed = cdata.active_cores;
 431
 432         if (!cdata.quiet) {
 433                 printf("  Config:\n");
 434                 printf("\tports: %u\n", num_ports);
 435                 printf("\tworkers: %u\n", cdata.num_workers);
 436                 printf("\tpackets: %"PRIi64"\n", cdata.num_packets);
 437                 printf("\tQueue-prio: %u\n", cdata.enable_queue_priorities);
 438                 if (cdata.queue_type == RTE_SCHED_TYPE_ORDERED)
 439                         printf("\tqid0 type: ordered\n");
 440                 if (cdata.queue_type == RTE_SCHED_TYPE_ATOMIC)
 441                         printf("\tqid0 type: atomic\n");
 442                 printf("\tCores available: %u\n", rte_lcore_count());
 443                 printf("\tCores used: %u\n", cores_needed);
 444         }
 445
 446         if (rte_lcore_count() < cores_needed)
 447                 rte_panic("Too few cores (%d < %d)\n", rte_lcore_count(),
 448                                 cores_needed);
 449
 450         const unsigned int ndevs = rte_event_dev_count();
 451         if (ndevs == 0)
 452                 rte_panic("No dev_id devs found. Pasl in a --vdev eventdev.\n");
 453         if (ndevs > 1)
 454                 fprintf(stderr, "Warning: More than one eventdev, using idx 0");
 455
 456
 457         do_capability_setup(num_ports, 0);
 458         fdata->cap.check_opt();
 459
 460         worker_data = rte_calloc(0, cdata.num_workers,
 461                         sizeof(worker_data[0]), 0);
 462         if (worker_data == NULL)
 463                 rte_panic("rte_calloc failed\n");
 464
 465         int dev_id = fdata->cap.evdev_setup(&cons_data, worker_data);
 466         if (dev_id < 0)
 467                 rte_exit(EXIT_FAILURE, "Error setting up eventdev\n");
 468
 469         init_ports(num_ports);
 470         fdata->cap.adptr_setup(num_ports);
 471
 472         int worker_idx = 0;
 473         RTE_LCORE_FOREACH_SLAVE(lcore_id) {
 474                 if (lcore_id >= MAX_NUM_CORE)
 475                         break;
 476
 477                 if (!fdata->rx_core[lcore_id] &&
 478                         !fdata->worker_core[lcore_id] &&
 479                         !fdata->tx_core[lcore_id] &&
 480                         !fdata->sched_core[lcore_id])
 481                         continue;
 482
 483                 if (fdata->rx_core[lcore_id])
 484                         printf(
 485                                 "[%s()] lcore %d executing NIC Rx\n",
 486                                 __func__, lcore_id);
 487
 488                 if (fdata->tx_core[lcore_id])
 489                         printf(
 490                                 "[%s()] lcore %d executing NIC Tx, and using eventdev port %u\n",
 491                                 __func__, lcore_id, cons_data.port_id);
 492
 493                 if (fdata->sched_core[lcore_id])
 494                         printf("[%s()] lcore %d executing scheduler\n",
 495                                         __func__, lcore_id);
 496
 497                 if (fdata->worker_core[lcore_id])
 498                         printf(
 499                                 "[%s()] lcore %d executing worker, using eventdev port %u\n",
 500                                 __func__, lcore_id,
 501                                 worker_data[worker_idx].port_id);
 502
 503                 err = rte_eal_remote_launch(fdata->cap.worker,
 504                                 &worker_data[worker_idx], lcore_id);
 505                 if (err) {
 506                         rte_panic("Failed to launch worker on core %d\n",
 507                                         lcore_id);
 508                         continue;
 509                 }
 510                 if (fdata->worker_core[lcore_id])
 511                         worker_idx++;
 512         }
 513
 514         lcore_id = rte_lcore_id();
 515
 516         if (core_in_use(lcore_id))
 517                 fdata->cap.worker(&worker_data[worker_idx++]);
 518
 519         rte_eal_mp_wait_lcore();
 520
 521         if (cdata.dump_dev)
 522                 rte_event_dev_dump(dev_id, stdout);
 523
 524         if (!cdata.quiet && (port_stat(dev_id, worker_data[0].port_id) !=
 525                         (uint64_t)-ENOTSUP)) {
 526                 printf("\nPort Workload distribution:\n");
 527                 uint32_t i;
 528                 uint64_t tot_pkts = 0;
 529                 uint64_t pkts_per_wkr[RTE_MAX_LCORE] = {0};
 530                 for (i = 0; i < cdata.num_workers; i++) {
 531                         pkts_per_wkr[i] =
 532                                 port_stat(dev_id, worker_data[i].port_id);
 533                         tot_pkts += pkts_per_wkr[i];
 534                 }
 535                 for (i = 0; i < cdata.num_workers; i++) {
 536                         float pc = pkts_per_wkr[i]  * 100 /
 537                                 ((float)tot_pkts);
 538                         printf("worker %i :\t%.1f %% (%"PRIu64" pkts)\n",
 539                                         i, pc, pkts_per_wkr[i]);
 540                 }
 541
 542         }
 543
 544         return 0;
 545 }