app/test-eventdev/test_perf_common.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2017 Cavium, Inc
   3  */
   4
   5 #include "test_perf_common.h"
   6
   7 int
   8 perf_test_result(struct evt_test *test, struct evt_options *opt)
   9 {
  10         RTE_SET_USED(opt);
  11         struct test_perf *t = evt_test_priv(test);
  12
  13         return t->result;
  14 }
  15
  16 static inline int
  17 perf_producer(void *arg)
  18 {
  19         struct prod_data *p  = arg;
  20         struct test_perf *t = p->t;
  21         struct evt_options *opt = t->opt;
  22         const uint8_t dev_id = p->dev_id;
  23         const uint8_t port = p->port_id;
  24         struct rte_mempool *pool = t->pool;
  25         const uint64_t nb_pkts = t->nb_pkts;
  26         const uint32_t nb_flows = t->nb_flows;
  27         uint32_t flow_counter = 0;
  28         uint64_t count = 0;
  29         struct perf_elt *m;
  30         struct rte_event ev;
  31
  32         if (opt->verbose_level > 1)
  33                 printf("%s(): lcore %d dev_id %d port=%d queue %d\n", __func__,
  34                                 rte_lcore_id(), dev_id, port, p->queue_id);
  35
  36         ev.event = 0;
  37         ev.op = RTE_EVENT_OP_NEW;
  38         ev.queue_id = p->queue_id;
  39         ev.sched_type = t->opt->sched_type_list[0];
  40         ev.priority = RTE_EVENT_DEV_PRIORITY_NORMAL;
  41         ev.event_type =  RTE_EVENT_TYPE_CPU;
  42         ev.sub_event_type = 0; /* stage 0 */
  43
  44         while (count < nb_pkts && t->done == false) {
  45                 if (rte_mempool_get(pool, (void **)&m) < 0)
  46                         continue;
  47
  48                 ev.flow_id = flow_counter++ % nb_flows;
  49                 ev.event_ptr = m;
  50                 m->timestamp = rte_get_timer_cycles();
  51                 while (rte_event_enqueue_burst(dev_id, port, &ev, 1) != 1) {
  52                         if (t->done)
  53                                 break;
  54                         rte_pause();
  55                         m->timestamp = rte_get_timer_cycles();
  56                 }
  57                 count++;
  58         }
  59
  60         return 0;
  61 }
  62
  63 static int
  64 perf_producer_wrapper(void *arg)
  65 {
  66         struct prod_data *p  = arg;
  67         struct test_perf *t = p->t;
  68         /* Launch the producer function only in case of synthetic producer. */
  69         if (t->opt->prod_type == EVT_PROD_TYPE_SYNT)
  70                 return perf_producer(arg);
  71         return 0;
  72 }
  73
  74 static inline uint64_t
  75 processed_pkts(struct test_perf *t)
  76 {
  77         uint8_t i;
  78         uint64_t total = 0;
  79
  80         rte_smp_rmb();
  81         for (i = 0; i < t->nb_workers; i++)
  82                 total += t->worker[i].processed_pkts;
  83
  84         return total;
  85 }
  86
  87 static inline uint64_t
  88 total_latency(struct test_perf *t)
  89 {
  90         uint8_t i;
  91         uint64_t total = 0;
  92
  93         rte_smp_rmb();
  94         for (i = 0; i < t->nb_workers; i++)
  95                 total += t->worker[i].latency;
  96
  97         return total;
  98 }
  99
 100
 101 int
 102 perf_launch_lcores(struct evt_test *test, struct evt_options *opt,
 103                 int (*worker)(void *))
 104 {
 105         int ret, lcore_id;
 106         struct test_perf *t = evt_test_priv(test);
 107
 108         int port_idx = 0;
 109         /* launch workers */
 110         RTE_LCORE_FOREACH_SLAVE(lcore_id) {
 111                 if (!(opt->wlcores[lcore_id]))
 112                         continue;
 113
 114                 ret = rte_eal_remote_launch(worker,
 115                                  &t->worker[port_idx], lcore_id);
 116                 if (ret) {
 117                         evt_err("failed to launch worker %d", lcore_id);
 118                         return ret;
 119                 }
 120                 port_idx++;
 121         }
 122
 123         /* launch producers */
 124         RTE_LCORE_FOREACH_SLAVE(lcore_id) {
 125                 if (!(opt->plcores[lcore_id]))
 126                         continue;
 127
 128                 ret = rte_eal_remote_launch(perf_producer_wrapper,
 129                                 &t->prod[port_idx], lcore_id);
 130                 if (ret) {
 131                         evt_err("failed to launch perf_producer %d", lcore_id);
 132                         return ret;
 133                 }
 134                 port_idx++;
 135         }
 136
 137         const uint64_t total_pkts = opt->nb_pkts *
 138                         evt_nr_active_lcores(opt->plcores);
 139
 140         uint64_t dead_lock_cycles = rte_get_timer_cycles();
 141         int64_t dead_lock_remaining  =  total_pkts;
 142         const uint64_t dead_lock_sample = rte_get_timer_hz() * 5;
 143
 144         uint64_t perf_cycles = rte_get_timer_cycles();
 145         int64_t perf_remaining  = total_pkts;
 146         const uint64_t perf_sample = rte_get_timer_hz();
 147
 148         static float total_mpps;
 149         static uint64_t samples;
 150
 151         const uint64_t freq_mhz = rte_get_timer_hz() / 1000000;
 152         int64_t remaining = t->outstand_pkts - processed_pkts(t);
 153
 154         while (t->done == false) {
 155                 const uint64_t new_cycles = rte_get_timer_cycles();
 156
 157                 if ((new_cycles - perf_cycles) > perf_sample) {
 158                         const uint64_t latency = total_latency(t);
 159                         const uint64_t pkts = processed_pkts(t);
 160
 161                         remaining = t->outstand_pkts - pkts;
 162                         float mpps = (float)(perf_remaining-remaining)/1000000;
 163
 164                         perf_remaining = remaining;
 165                         perf_cycles = new_cycles;
 166                         total_mpps += mpps;
 167                         ++samples;
 168                         if (opt->fwd_latency && pkts > 0) {
 169                                 printf(CLGRN"\r%.3f mpps avg %.3f mpps [avg fwd latency %.3f us] "CLNRM,
 170                                         mpps, total_mpps/samples,
 171                                         (float)(latency/pkts)/freq_mhz);
 172                         } else {
 173                                 printf(CLGRN"\r%.3f mpps avg %.3f mpps"CLNRM,
 174                                         mpps, total_mpps/samples);
 175                         }
 176                         fflush(stdout);
 177
 178                         if (remaining <= 0) {
 179                                 t->result = EVT_TEST_SUCCESS;
 180                                 if (opt->prod_type == EVT_PROD_TYPE_SYNT) {
 181                                         t->done = true;
 182                                         rte_smp_wmb();
 183                                         break;
 184                                 }
 185                         }
 186                 }
 187
 188                 if (new_cycles - dead_lock_cycles > dead_lock_sample &&
 189                                 opt->prod_type == EVT_PROD_TYPE_SYNT) {
 190                         remaining = t->outstand_pkts - processed_pkts(t);
 191                         if (dead_lock_remaining == remaining) {
 192                                 rte_event_dev_dump(opt->dev_id, stdout);
 193                                 evt_err("No schedules for seconds, deadlock");
 194                                 t->done = true;
 195                                 rte_smp_wmb();
 196                                 break;
 197                         }
 198                         dead_lock_remaining = remaining;
 199                         dead_lock_cycles = new_cycles;
 200                 }
 201         }
 202         printf("\n");
 203         return 0;
 204 }
 205
 206 int
 207 perf_event_dev_port_setup(struct evt_test *test, struct evt_options *opt,
 208                                 uint8_t stride, uint8_t nb_queues)
 209 {
 210         struct test_perf *t = evt_test_priv(test);
 211         uint8_t port, prod;
 212         int ret = -1;
 213
 214         /* port configuration */
 215         const struct rte_event_port_conf wkr_p_conf = {
 216                         .dequeue_depth = opt->wkr_deq_dep,
 217                         .enqueue_depth = 64,
 218                         .new_event_threshold = 4096,
 219         };
 220
 221         /* setup one port per worker, linking to all queues */
 222         for (port = 0; port < evt_nr_active_lcores(opt->wlcores);
 223                                 port++) {
 224                 struct worker_data *w = &t->worker[port];
 225
 226                 w->dev_id = opt->dev_id;
 227                 w->port_id = port;
 228                 w->t = t;
 229                 w->processed_pkts = 0;
 230                 w->latency = 0;
 231
 232                 ret = rte_event_port_setup(opt->dev_id, port, &wkr_p_conf);
 233                 if (ret) {
 234                         evt_err("failed to setup port %d", port);
 235                         return ret;
 236                 }
 237
 238                 ret = rte_event_port_link(opt->dev_id, port, NULL, NULL, 0);
 239                 if (ret != nb_queues) {
 240                         evt_err("failed to link all queues to port %d", port);
 241                         return -EINVAL;
 242                 }
 243         }
 244
 245         /* port for producers, no links */
 246         const struct rte_event_port_conf prod_conf = {
 247                         .dequeue_depth = 8,
 248                         .enqueue_depth = 32,
 249                         .new_event_threshold = 1200,
 250         };
 251         prod = 0;
 252         for ( ; port < perf_nb_event_ports(opt); port++) {
 253                 struct prod_data *p = &t->prod[port];
 254
 255                 p->dev_id = opt->dev_id;
 256                 p->port_id = port;
 257                 p->queue_id = prod * stride;
 258                 p->t = t;
 259
 260                 ret = rte_event_port_setup(opt->dev_id, port, &prod_conf);
 261                 if (ret) {
 262                         evt_err("failed to setup port %d", port);
 263                         return ret;
 264                 }
 265                 prod++;
 266         }
 267
 268         return ret;
 269 }
 270
 271 int
 272 perf_opt_check(struct evt_options *opt, uint64_t nb_queues)
 273 {
 274         unsigned int lcores;
 275
 276         /* N producer + N worker + 1 master when producer cores are used
 277          * Else N worker + 1 master when Rx adapter is used
 278          */
 279         lcores = opt->prod_type == EVT_PROD_TYPE_SYNT ? 3 : 2;
 280
 281         if (rte_lcore_count() < lcores) {
 282                 evt_err("test need minimum %d lcores", lcores);
 283                 return -1;
 284         }
 285
 286         /* Validate worker lcores */
 287         if (evt_lcores_has_overlap(opt->wlcores, rte_get_master_lcore())) {
 288                 evt_err("worker lcores overlaps with master lcore");
 289                 return -1;
 290         }
 291         if (evt_lcores_has_overlap_multi(opt->wlcores, opt->plcores)) {
 292                 evt_err("worker lcores overlaps producer lcores");
 293                 return -1;
 294         }
 295         if (evt_has_disabled_lcore(opt->wlcores)) {
 296                 evt_err("one or more workers lcores are not enabled");
 297                 return -1;
 298         }
 299         if (!evt_has_active_lcore(opt->wlcores)) {
 300                 evt_err("minimum one worker is required");
 301                 return -1;
 302         }
 303
 304         if (opt->prod_type == EVT_PROD_TYPE_SYNT) {
 305                 /* Validate producer lcores */
 306                 if (evt_lcores_has_overlap(opt->plcores,
 307                                         rte_get_master_lcore())) {
 308                         evt_err("producer lcores overlaps with master lcore");
 309                         return -1;
 310                 }
 311                 if (evt_has_disabled_lcore(opt->plcores)) {
 312                         evt_err("one or more producer lcores are not enabled");
 313                         return -1;
 314                 }
 315                 if (!evt_has_active_lcore(opt->plcores)) {
 316                         evt_err("minimum one producer is required");
 317                         return -1;
 318                 }
 319         }
 320
 321         if (evt_has_invalid_stage(opt))
 322                 return -1;
 323
 324         if (evt_has_invalid_sched_type(opt))
 325                 return -1;
 326
 327         if (nb_queues > EVT_MAX_QUEUES) {
 328                 evt_err("number of queues exceeds %d", EVT_MAX_QUEUES);
 329                 return -1;
 330         }
 331         if (perf_nb_event_ports(opt) > EVT_MAX_PORTS) {
 332                 evt_err("number of ports exceeds %d", EVT_MAX_PORTS);
 333                 return -1;
 334         }
 335
 336         /* Fixups */
 337         if (opt->nb_stages == 1 && opt->fwd_latency) {
 338                 evt_info("fwd_latency is valid when nb_stages > 1, disabling");
 339                 opt->fwd_latency = 0;
 340         }
 341         if (opt->fwd_latency && !opt->q_priority) {
 342                 evt_info("enabled queue priority for latency measurement");
 343                 opt->q_priority = 1;
 344         }
 345         if (opt->nb_pkts == 0)
 346                 opt->nb_pkts = INT64_MAX/evt_nr_active_lcores(opt->plcores);
 347
 348         return 0;
 349 }
 350
 351 void
 352 perf_opt_dump(struct evt_options *opt, uint8_t nb_queues)
 353 {
 354         evt_dump("nb_prod_lcores", "%d", evt_nr_active_lcores(opt->plcores));
 355         evt_dump_producer_lcores(opt);
 356         evt_dump("nb_worker_lcores", "%d", evt_nr_active_lcores(opt->wlcores));
 357         evt_dump_worker_lcores(opt);
 358         evt_dump_nb_stages(opt);
 359         evt_dump("nb_evdev_ports", "%d", perf_nb_event_ports(opt));
 360         evt_dump("nb_evdev_queues", "%d", nb_queues);
 361         evt_dump_queue_priority(opt);
 362         evt_dump_sched_type_list(opt);
 363         evt_dump_producer_type(opt);
 364 }
 365
 366 void
 367 perf_eventdev_destroy(struct evt_test *test, struct evt_options *opt)
 368 {
 369         RTE_SET_USED(test);
 370
 371         rte_event_dev_stop(opt->dev_id);
 372         rte_event_dev_close(opt->dev_id);
 373 }
 374
 375 static inline void
 376 perf_elt_init(struct rte_mempool *mp, void *arg __rte_unused,
 377             void *obj, unsigned i __rte_unused)
 378 {
 379         memset(obj, 0, mp->elt_size);
 380 }
 381
 382 int
 383 perf_mempool_setup(struct evt_test *test, struct evt_options *opt)
 384 {
 385         struct test_perf *t = evt_test_priv(test);
 386
 387         t->pool = rte_mempool_create(test->name, /* mempool name */
 388                                 opt->pool_sz, /* number of elements*/
 389                                 sizeof(struct perf_elt), /* element size*/
 390                                 512, /* cache size*/
 391                                 0, NULL, NULL,
 392                                 perf_elt_init, /* obj constructor */
 393                                 NULL, opt->socket_id, 0); /* flags */
 394         if (t->pool == NULL) {
 395                 evt_err("failed to create mempool");
 396                 return -ENOMEM;
 397         }
 398
 399         return 0;
 400 }
 401
 402 void
 403 perf_mempool_destroy(struct evt_test *test, struct evt_options *opt)
 404 {
 405         RTE_SET_USED(opt);
 406         struct test_perf *t = evt_test_priv(test);
 407
 408         rte_mempool_free(t->pool);
 409 }
 410
 411 int
 412 perf_test_setup(struct evt_test *test, struct evt_options *opt)
 413 {
 414         void *test_perf;
 415
 416         test_perf = rte_zmalloc_socket(test->name, sizeof(struct test_perf),
 417                                 RTE_CACHE_LINE_SIZE, opt->socket_id);
 418         if (test_perf  == NULL) {
 419                 evt_err("failed to allocate test_perf memory");
 420                 goto nomem;
 421         }
 422         test->test_priv = test_perf;
 423
 424         struct test_perf *t = evt_test_priv(test);
 425
 426         t->outstand_pkts = opt->nb_pkts * evt_nr_active_lcores(opt->plcores);
 427         t->nb_workers = evt_nr_active_lcores(opt->wlcores);
 428         t->done = false;
 429         t->nb_pkts = opt->nb_pkts;
 430         t->nb_flows = opt->nb_flows;
 431         t->result = EVT_TEST_FAILED;
 432         t->opt = opt;
 433         memcpy(t->sched_type_list, opt->sched_type_list,
 434                         sizeof(opt->sched_type_list));
 435         return 0;
 436 nomem:
 437         return -ENOMEM;
 438 }
 439
 440 void
 441 perf_test_destroy(struct evt_test *test, struct evt_options *opt)
 442 {
 443         RTE_SET_USED(opt);
 444
 445         rte_free(test->test_priv);
 446 }