ceph/src/spdk/dpdk/examples/eventdev_pipeline/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 /*
  30  * Parse the coremask given as argument (hexadecimal string) and fill
  31  * the global configuration (core role and core count) with the parsed
  32  * value.
  33  */
  34 static int xdigit2val(unsigned char c)
  35 {
  36         int val;
  37
  38         if (isdigit(c))
  39                 val = c - '0';
  40         else if (isupper(c))
  41                 val = c - 'A' + 10;
  42         else
  43                 val = c - 'a' + 10;
  44         return val;
  45 }
  46
  47 static uint64_t
  48 parse_coremask(const char *coremask)
  49 {
  50         int i, j, idx = 0;
  51         unsigned int count = 0;
  52         char c;
  53         int val;
  54         uint64_t mask = 0;
  55         const int32_t BITS_HEX = 4;
  56
  57         if (coremask == NULL)
  58                 return -1;
  59         /* Remove all blank characters ahead and after .
  60          * Remove 0x/0X if exists.
  61          */
  62         while (isblank(*coremask))
  63                 coremask++;
  64         if (coremask[0] == '0' && ((coremask[1] == 'x')
  65                 || (coremask[1] == 'X')))
  66                 coremask += 2;
  67         i = strlen(coremask);
  68         while ((i > 0) && isblank(coremask[i - 1]))
  69                 i--;
  70         if (i == 0)
  71                 return -1;
  72
  73         for (i = i - 1; i >= 0 && idx < MAX_NUM_CORE; i--) {
  74                 c = coremask[i];
  75                 if (isxdigit(c) == 0) {
  76                         /* invalid characters */
  77                         return -1;
  78                 }
  79                 val = xdigit2val(c);
  80                 for (j = 0; j < BITS_HEX && idx < MAX_NUM_CORE; j++, idx++) {
  81                         if ((1 << j) & val) {
  82                                 mask |= (1UL << idx);
  83                                 count++;
  84                         }
  85                 }
  86         }
  87         for (; i >= 0; i--)
  88                 if (coremask[i] != '0')
  89                         return -1;
  90         if (count == 0)
  91                 return -1;
  92         return mask;
  93 }
  94
  95 static struct option long_options[] = {
  96         {"workers", required_argument, 0, 'w'},
  97         {"packets", required_argument, 0, 'n'},
  98         {"atomic-flows", required_argument, 0, 'f'},
  99         {"num_stages", required_argument, 0, 's'},
 100         {"rx-mask", required_argument, 0, 'r'},
 101         {"tx-mask", required_argument, 0, 't'},
 102         {"sched-mask", required_argument, 0, 'e'},
 103         {"cq-depth", required_argument, 0, 'c'},
 104         {"work-cycles", required_argument, 0, 'W'},
 105         {"mempool-size", required_argument, 0, 'm'},
 106         {"queue-priority", no_argument, 0, 'P'},
 107         {"parallel", no_argument, 0, 'p'},
 108         {"ordered", no_argument, 0, 'o'},
 109         {"quiet", no_argument, 0, 'q'},
 110         {"use-atq", no_argument, 0, 'a'},
 111         {"dump", no_argument, 0, 'D'},
 112         {0, 0, 0, 0}
 113 };
 114
 115 static void
 116 usage(void)
 117 {
 118         const char *usage_str =
 119                 "  Usage: eventdev_demo [options]\n"
 120                 "  Options:\n"
 121                 "  -n, --packets=N              Send N packets (default ~32M), 0 implies no limit\n"
 122                 "  -f, --atomic-flows=N         Use N random flows from 1 to N (default 16)\n"
 123                 "  -s, --num_stages=N           Use N atomic stages (default 1)\n"
 124                 "  -r, --rx-mask=core mask      Run NIC rx on CPUs in core mask\n"
 125                 "  -w, --worker-mask=core mask  Run worker on CPUs in core mask\n"
 126                 "  -t, --tx-mask=core mask      Run NIC tx on CPUs in core mask\n"
 127                 "  -e  --sched-mask=core mask   Run scheduler on CPUs in core mask\n"
 128                 "  -c  --cq-depth=N             Worker CQ depth (default 16)\n"
 129                 "  -W  --work-cycles=N          Worker cycles (default 0)\n"
 130                 "  -P  --queue-priority         Enable scheduler queue prioritization\n"
 131                 "  -o, --ordered                Use ordered scheduling\n"
 132                 "  -p, --parallel               Use parallel scheduling\n"
 133                 "  -q, --quiet                  Minimize printed output\n"
 134                 "  -a, --use-atq                Use all type queues\n"
 135                 "  -m, --mempool-size=N         Dictate the mempool size\n"
 136                 "  -D, --dump                   Print detailed statistics before exit"
 137                 "\n";
 138         fprintf(stderr, "%s", usage_str);
 139         exit(1);
 140 }
 141
 142 static void
 143 parse_app_args(int argc, char **argv)
 144 {
 145         /* Parse cli options*/
 146         int option_index;
 147         int c;
 148         opterr = 0;
 149         uint64_t rx_lcore_mask = 0;
 150         uint64_t tx_lcore_mask = 0;
 151         uint64_t sched_lcore_mask = 0;
 152         uint64_t worker_lcore_mask = 0;
 153         int i;
 154
 155         for (;;) {
 156                 c = getopt_long(argc, argv, "r:t:e:c:w:n:f:s:m:paoPqDW:",
 157                                 long_options, &option_index);
 158                 if (c == -1)
 159                         break;
 160
 161                 int popcnt = 0;
 162                 switch (c) {
 163                 case 'n':
 164                         cdata.num_packets = (int64_t)atol(optarg);
 165                         if (cdata.num_packets == 0)
 166                                 cdata.num_packets = INT64_MAX;
 167                         break;
 168                 case 'f':
 169                         cdata.num_fids = (unsigned int)atoi(optarg);
 170                         break;
 171                 case 's':
 172                         cdata.num_stages = (unsigned int)atoi(optarg);
 173                         break;
 174                 case 'c':
 175                         cdata.worker_cq_depth = (unsigned int)atoi(optarg);
 176                         break;
 177                 case 'W':
 178                         cdata.worker_cycles = (unsigned int)atoi(optarg);
 179                         break;
 180                 case 'P':
 181                         cdata.enable_queue_priorities = 1;
 182                         break;
 183                 case 'o':
 184                         cdata.queue_type = RTE_SCHED_TYPE_ORDERED;
 185                         break;
 186                 case 'p':
 187                         cdata.queue_type = RTE_SCHED_TYPE_PARALLEL;
 188                         break;
 189                 case 'a':
 190                         cdata.all_type_queues = 1;
 191                         break;
 192                 case 'q':
 193                         cdata.quiet = 1;
 194                         break;
 195                 case 'D':
 196                         cdata.dump_dev = 1;
 197                         break;
 198                 case 'w':
 199                         worker_lcore_mask = parse_coremask(optarg);
 200                         break;
 201                 case 'r':
 202                         rx_lcore_mask = parse_coremask(optarg);
 203                         popcnt = __builtin_popcountll(rx_lcore_mask);
 204                         fdata->rx_single = (popcnt == 1);
 205                         break;
 206                 case 't':
 207                         tx_lcore_mask = parse_coremask(optarg);
 208                         popcnt = __builtin_popcountll(tx_lcore_mask);
 209                         fdata->tx_single = (popcnt == 1);
 210                         break;
 211                 case 'e':
 212                         sched_lcore_mask = parse_coremask(optarg);
 213                         popcnt = __builtin_popcountll(sched_lcore_mask);
 214                         fdata->sched_single = (popcnt == 1);
 215                         break;
 216                 case 'm':
 217                         cdata.num_mbuf = (uint64_t)atol(optarg);
 218                         break;
 219                 default:
 220                         usage();
 221                 }
 222         }
 223
 224         cdata.worker_lcore_mask = worker_lcore_mask;
 225         cdata.sched_lcore_mask = sched_lcore_mask;
 226         cdata.rx_lcore_mask = rx_lcore_mask;
 227         cdata.tx_lcore_mask = tx_lcore_mask;
 228
 229         if (cdata.num_stages == 0 || cdata.num_stages > MAX_NUM_STAGES)
 230                 usage();
 231
 232         for (i = 0; i < MAX_NUM_CORE; i++) {
 233                 fdata->rx_core[i] = !!(rx_lcore_mask & (1UL << i));
 234                 fdata->tx_core[i] = !!(tx_lcore_mask & (1UL << i));
 235                 fdata->sched_core[i] = !!(sched_lcore_mask & (1UL << i));
 236                 fdata->worker_core[i] = !!(worker_lcore_mask & (1UL << i));
 237
 238                 if (fdata->worker_core[i])
 239                         cdata.num_workers++;
 240                 if (core_in_use(i))
 241                         cdata.active_cores++;
 242         }
 243 }
 244
 245 /*
 246  * Initializes a given port using global settings and with the RX buffers
 247  * coming from the mbuf_pool passed as a parameter.
 248  */
 249 static inline int
 250 port_init(uint8_t port, struct rte_mempool *mbuf_pool)
 251 {
 252         struct rte_eth_rxconf rx_conf;
 253         static const struct rte_eth_conf port_conf_default = {
 254                 .rxmode = {
 255                         .mq_mode = ETH_MQ_RX_RSS,
 256                         .max_rx_pkt_len = ETHER_MAX_LEN,
 257                 },
 258                 .rx_adv_conf = {
 259                         .rss_conf = {
 260                                 .rss_hf = ETH_RSS_IP |
 261                                           ETH_RSS_TCP |
 262                                           ETH_RSS_UDP,
 263                         }
 264                 }
 265         };
 266         const uint16_t rx_rings = 1, tx_rings = 1;
 267         const uint16_t rx_ring_size = 512, tx_ring_size = 512;
 268         struct rte_eth_conf port_conf = port_conf_default;
 269         int retval;
 270         uint16_t q;
 271         struct rte_eth_dev_info dev_info;
 272         struct rte_eth_txconf txconf;
 273
 274         if (!rte_eth_dev_is_valid_port(port))
 275                 return -1;
 276
 277         rte_eth_dev_info_get(port, &dev_info);
 278         if (dev_info.tx_offload_capa & DEV_TX_OFFLOAD_MBUF_FAST_FREE)
 279                 port_conf.txmode.offloads |=
 280                         DEV_TX_OFFLOAD_MBUF_FAST_FREE;
 281         rx_conf = dev_info.default_rxconf;
 282         rx_conf.offloads = port_conf.rxmode.offloads;
 283
 284         port_conf.rx_adv_conf.rss_conf.rss_hf &=
 285                 dev_info.flow_type_rss_offloads;
 286         if (port_conf.rx_adv_conf.rss_conf.rss_hf !=
 287                         port_conf_default.rx_adv_conf.rss_conf.rss_hf) {
 288                 printf("Port %u modified RSS hash function based on hardware support,"
 289                         "requested:%#"PRIx64" configured:%#"PRIx64"\n",
 290                         port,
 291                         port_conf_default.rx_adv_conf.rss_conf.rss_hf,
 292                         port_conf.rx_adv_conf.rss_conf.rss_hf);
 293         }
 294
 295         /* Configure the Ethernet device. */
 296         retval = rte_eth_dev_configure(port, rx_rings, tx_rings, &port_conf);
 297         if (retval != 0)
 298                 return retval;
 299
 300         /* Allocate and set up 1 RX queue per Ethernet port. */
 301         for (q = 0; q < rx_rings; q++) {
 302                 retval = rte_eth_rx_queue_setup(port, q, rx_ring_size,
 303                                 rte_eth_dev_socket_id(port), &rx_conf,
 304                                 mbuf_pool);
 305                 if (retval < 0)
 306                         return retval;
 307         }
 308
 309         txconf = dev_info.default_txconf;
 310         txconf.offloads = port_conf_default.txmode.offloads;
 311         /* Allocate and set up 1 TX queue per Ethernet port. */
 312         for (q = 0; q < tx_rings; q++) {
 313                 retval = rte_eth_tx_queue_setup(port, q, tx_ring_size,
 314                                 rte_eth_dev_socket_id(port), &txconf);
 315                 if (retval < 0)
 316                         return retval;
 317         }
 318
 319         /* Display the port MAC address. */
 320         struct ether_addr addr;
 321         rte_eth_macaddr_get(port, &addr);
 322         printf("Port %u MAC: %02" PRIx8 " %02" PRIx8 " %02" PRIx8
 323                            " %02" PRIx8 " %02" PRIx8 " %02" PRIx8 "\n",
 324                         (unsigned int)port,
 325                         addr.addr_bytes[0], addr.addr_bytes[1],
 326                         addr.addr_bytes[2], addr.addr_bytes[3],
 327                         addr.addr_bytes[4], addr.addr_bytes[5]);
 328
 329         /* Enable RX in promiscuous mode for the Ethernet device. */
 330         rte_eth_promiscuous_enable(port);
 331
 332         return 0;
 333 }
 334
 335 static int
 336 init_ports(uint16_t num_ports)
 337 {
 338         uint16_t portid;
 339
 340         if (!cdata.num_mbuf)
 341                 cdata.num_mbuf = 16384 * num_ports;
 342
 343         struct rte_mempool *mp = rte_pktmbuf_pool_create("packet_pool",
 344                         /* mbufs */ cdata.num_mbuf,
 345                         /* cache_size */ 512,
 346                         /* priv_size*/ 0,
 347                         /* data_room_size */ RTE_MBUF_DEFAULT_BUF_SIZE,
 348                         rte_socket_id());
 349
 350         RTE_ETH_FOREACH_DEV(portid)
 351                 if (port_init(portid, mp) != 0)
 352                         rte_exit(EXIT_FAILURE, "Cannot init port %"PRIu16 "\n",
 353                                         portid);
 354
 355         return 0;
 356 }
 357
 358 static void
 359 do_capability_setup(uint8_t eventdev_id)
 360 {
 361         int ret;
 362         uint16_t i;
 363         uint8_t generic_pipeline = 0;
 364         uint8_t burst = 0;
 365
 366         RTE_ETH_FOREACH_DEV(i) {
 367                 uint32_t caps = 0;
 368
 369                 ret = rte_event_eth_tx_adapter_caps_get(eventdev_id, i, &caps);
 370                 if (ret)
 371                         rte_exit(EXIT_FAILURE,
 372                                 "Invalid capability for Tx adptr port %d\n", i);
 373                 generic_pipeline |= !(caps &
 374                                 RTE_EVENT_ETH_TX_ADAPTER_CAP_INTERNAL_PORT);
 375         }
 376
 377         struct rte_event_dev_info eventdev_info;
 378         memset(&eventdev_info, 0, sizeof(struct rte_event_dev_info));
 379
 380         rte_event_dev_info_get(eventdev_id, &eventdev_info);
 381         burst = eventdev_info.event_dev_cap & RTE_EVENT_DEV_CAP_BURST_MODE ? 1 :
 382                 0;
 383
 384         if (generic_pipeline)
 385                 set_worker_generic_setup_data(&fdata->cap, burst);
 386         else
 387                 set_worker_tx_enq_setup_data(&fdata->cap, burst);
 388 }
 389
 390 static void
 391 signal_handler(int signum)
 392 {
 393         static uint8_t once;
 394         uint16_t portid;
 395
 396         if (fdata->done)
 397                 rte_exit(1, "Exiting on signal %d\n", signum);
 398         if ((signum == SIGINT || signum == SIGTERM) && !once) {
 399                 printf("\n\nSignal %d received, preparing to exit...\n",
 400                                 signum);
 401                 if (cdata.dump_dev)
 402                         rte_event_dev_dump(0, stdout);
 403                 once = 1;
 404                 fdata->done = 1;
 405                 rte_smp_wmb();
 406
 407                 RTE_ETH_FOREACH_DEV(portid) {
 408                         rte_event_eth_rx_adapter_stop(portid);
 409                         rte_event_eth_tx_adapter_stop(portid);
 410                         rte_eth_dev_stop(portid);
 411                 }
 412
 413                 rte_eal_mp_wait_lcore();
 414
 415                 RTE_ETH_FOREACH_DEV(portid) {
 416                         rte_eth_dev_close(portid);
 417                 }
 418
 419                 rte_event_dev_stop(0);
 420                 rte_event_dev_close(0);
 421         }
 422         if (signum == SIGTSTP)
 423                 rte_event_dev_dump(0, stdout);
 424 }
 425
 426 static inline uint64_t
 427 port_stat(int dev_id, int32_t p)
 428 {
 429         char statname[64];
 430         snprintf(statname, sizeof(statname), "port_%u_rx", p);
 431         return rte_event_dev_xstats_by_name_get(dev_id, statname, NULL);
 432 }
 433
 434 int
 435 main(int argc, char **argv)
 436 {
 437         struct worker_data *worker_data;
 438         uint16_t num_ports;
 439         uint16_t portid;
 440         int lcore_id;
 441         int err;
 442
 443         signal(SIGINT, signal_handler);
 444         signal(SIGTERM, signal_handler);
 445         signal(SIGTSTP, signal_handler);
 446
 447         err = rte_eal_init(argc, argv);
 448         if (err < 0)
 449                 rte_panic("Invalid EAL arguments\n");
 450
 451         argc -= err;
 452         argv += err;
 453
 454         fdata = rte_malloc(NULL, sizeof(struct fastpath_data), 0);
 455         if (fdata == NULL)
 456                 rte_panic("Out of memory\n");
 457
 458         /* Parse cli options*/
 459         parse_app_args(argc, argv);
 460
 461         num_ports = rte_eth_dev_count_avail();
 462         if (num_ports == 0)
 463                 rte_panic("No ethernet ports found\n");
 464
 465         const unsigned int cores_needed = cdata.active_cores;
 466
 467         if (!cdata.quiet) {
 468                 printf("  Config:\n");
 469                 printf("\tports: %u\n", num_ports);
 470                 printf("\tworkers: %u\n", cdata.num_workers);
 471                 printf("\tpackets: %"PRIi64"\n", cdata.num_packets);
 472                 printf("\tQueue-prio: %u\n", cdata.enable_queue_priorities);
 473                 if (cdata.queue_type == RTE_SCHED_TYPE_ORDERED)
 474                         printf("\tqid0 type: ordered\n");
 475                 if (cdata.queue_type == RTE_SCHED_TYPE_ATOMIC)
 476                         printf("\tqid0 type: atomic\n");
 477                 printf("\tCores available: %u\n", rte_lcore_count());
 478                 printf("\tCores used: %u\n", cores_needed);
 479         }
 480
 481         if (rte_lcore_count() < cores_needed)
 482                 rte_panic("Too few cores (%d < %d)\n", rte_lcore_count(),
 483                                 cores_needed);
 484
 485         const unsigned int ndevs = rte_event_dev_count();
 486         if (ndevs == 0)
 487                 rte_panic("No dev_id devs found. Pasl in a --vdev eventdev.\n");
 488         if (ndevs > 1)
 489                 fprintf(stderr, "Warning: More than one eventdev, using idx 0");
 490
 491
 492         do_capability_setup(0);
 493         fdata->cap.check_opt();
 494
 495         worker_data = rte_calloc(0, cdata.num_workers,
 496                         sizeof(worker_data[0]), 0);
 497         if (worker_data == NULL)
 498                 rte_panic("rte_calloc failed\n");
 499
 500         int dev_id = fdata->cap.evdev_setup(worker_data);
 501         if (dev_id < 0)
 502                 rte_exit(EXIT_FAILURE, "Error setting up eventdev\n");
 503
 504         init_ports(num_ports);
 505         fdata->cap.adptr_setup(num_ports);
 506
 507         /* Start the Ethernet port. */
 508         RTE_ETH_FOREACH_DEV(portid) {
 509                 err = rte_eth_dev_start(portid);
 510                 if (err < 0)
 511                         rte_exit(EXIT_FAILURE, "Error starting ethdev %d\n",
 512                                         portid);
 513         }
 514
 515         int worker_idx = 0;
 516         RTE_LCORE_FOREACH_SLAVE(lcore_id) {
 517                 if (lcore_id >= MAX_NUM_CORE)
 518                         break;
 519
 520                 if (!fdata->rx_core[lcore_id] &&
 521                         !fdata->worker_core[lcore_id] &&
 522                         !fdata->tx_core[lcore_id] &&
 523                         !fdata->sched_core[lcore_id])
 524                         continue;
 525
 526                 if (fdata->rx_core[lcore_id])
 527                         printf(
 528                                 "[%s()] lcore %d executing NIC Rx\n",
 529                                 __func__, lcore_id);
 530
 531                 if (fdata->tx_core[lcore_id])
 532                         printf(
 533                                 "[%s()] lcore %d executing NIC Tx\n",
 534                                 __func__, lcore_id);
 535
 536                 if (fdata->sched_core[lcore_id])
 537                         printf("[%s()] lcore %d executing scheduler\n",
 538                                         __func__, lcore_id);
 539
 540                 if (fdata->worker_core[lcore_id])
 541                         printf(
 542                                 "[%s()] lcore %d executing worker, using eventdev port %u\n",
 543                                 __func__, lcore_id,
 544                                 worker_data[worker_idx].port_id);
 545
 546                 err = rte_eal_remote_launch(fdata->cap.worker,
 547                                 &worker_data[worker_idx], lcore_id);
 548                 if (err) {
 549                         rte_panic("Failed to launch worker on core %d\n",
 550                                         lcore_id);
 551                         continue;
 552                 }
 553                 if (fdata->worker_core[lcore_id])
 554                         worker_idx++;
 555         }
 556
 557         lcore_id = rte_lcore_id();
 558
 559         if (core_in_use(lcore_id))
 560                 fdata->cap.worker(&worker_data[worker_idx++]);
 561
 562         rte_eal_mp_wait_lcore();
 563
 564         if (!cdata.quiet && (port_stat(dev_id, worker_data[0].port_id) !=
 565                         (uint64_t)-ENOTSUP)) {
 566                 printf("\nPort Workload distribution:\n");
 567                 uint32_t i;
 568                 uint64_t tot_pkts = 0;
 569                 uint64_t pkts_per_wkr[RTE_MAX_LCORE] = {0};
 570                 for (i = 0; i < cdata.num_workers; i++) {
 571                         pkts_per_wkr[i] =
 572                                 port_stat(dev_id, worker_data[i].port_id);
 573                         tot_pkts += pkts_per_wkr[i];
 574                 }
 575                 for (i = 0; i < cdata.num_workers; i++) {
 576                         float pc = pkts_per_wkr[i]  * 100 /
 577                                 ((float)tot_pkts);
 578                         printf("worker %i :\t%.1f %% (%"PRIu64" pkts)\n",
 579                                         i, pc, pkts_per_wkr[i]);
 580                 }
 581
 582         }
 583
 584         return 0;
 585 }