app/test-eventdev/test_perf_queue.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2017 Cavium, Inc
   3  */
   4
   5 #include "test_perf_common.h"
   6
   7 /* See http://doc.dpdk.org/guides/tools/testeventdev.html for test details */
   8
   9 static inline int
  10 perf_queue_nb_event_queues(struct evt_options *opt)
  11 {
  12         /* nb_queues = number of producers * number of stages */
  13         uint8_t nb_prod = opt->prod_type == EVT_PROD_TYPE_ETH_RX_ADPTR ?
  14                 rte_eth_dev_count_avail() : evt_nr_active_lcores(opt->plcores);
  15         return nb_prod * opt->nb_stages;
  16 }
  17
  18 static __rte_always_inline void
  19 mark_fwd_latency(struct rte_event *const ev,
  20                 const uint8_t nb_stages)
  21 {
  22         if (unlikely((ev->queue_id % nb_stages) == 0)) {
  23                 struct perf_elt *const m = ev->event_ptr;
  24
  25                 m->timestamp = rte_get_timer_cycles();
  26         }
  27 }
  28
  29 static __rte_always_inline void
  30 fwd_event(struct rte_event *const ev, uint8_t *const sched_type_list,
  31                 const uint8_t nb_stages)
  32 {
  33         ev->queue_id++;
  34         ev->sched_type = sched_type_list[ev->queue_id % nb_stages];
  35         ev->op = RTE_EVENT_OP_FORWARD;
  36         ev->event_type = RTE_EVENT_TYPE_CPU;
  37 }
  38
  39 static int
  40 perf_queue_worker(void *arg, const int enable_fwd_latency)
  41 {
  42         PERF_WORKER_INIT;
  43         struct rte_event ev;
  44
  45         while (t->done == false) {
  46                 uint16_t event = rte_event_dequeue_burst(dev, port, &ev, 1, 0);
  47
  48                 if (!event) {
  49                         rte_pause();
  50                         continue;
  51                 }
  52
  53                 if (prod_crypto_type &&
  54                     (ev.event_type == RTE_EVENT_TYPE_CRYPTODEV)) {
  55                         struct rte_crypto_op *op = ev.event_ptr;
  56
  57                         if (op->status == RTE_CRYPTO_OP_STATUS_SUCCESS) {
  58                                 if (op->sym->m_dst == NULL)
  59                                         ev.event_ptr = op->sym->m_src;
  60                                 else
  61                                         ev.event_ptr = op->sym->m_dst;
  62                                 rte_crypto_op_free(op);
  63                         } else {
  64                                 rte_crypto_op_free(op);
  65                                 continue;
  66                         }
  67                 }
  68
  69                 if (enable_fwd_latency && !prod_timer_type)
  70                 /* first q in pipeline, mark timestamp to compute fwd latency */
  71                         mark_fwd_latency(&ev, nb_stages);
  72
  73                 /* last stage in pipeline */
  74                 if (unlikely((ev.queue_id % nb_stages) == laststage)) {
  75                         if (enable_fwd_latency)
  76                                 cnt = perf_process_last_stage_latency(pool,
  77                                         &ev, w, bufs, sz, cnt);
  78                         else
  79                                 cnt = perf_process_last_stage(pool,
  80                                         &ev, w, bufs, sz, cnt);
  81                 } else {
  82                         fwd_event(&ev, sched_type_list, nb_stages);
  83                         while (rte_event_enqueue_burst(dev, port, &ev, 1) != 1)
  84                                 rte_pause();
  85                 }
  86         }
  87         return 0;
  88 }
  89
  90 static int
  91 perf_queue_worker_burst(void *arg, const int enable_fwd_latency)
  92 {
  93         PERF_WORKER_INIT;
  94         uint16_t i;
  95         /* +1 to avoid prefetch out of array check */
  96         struct rte_event ev[BURST_SIZE + 1];
  97
  98         while (t->done == false) {
  99                 uint16_t const nb_rx = rte_event_dequeue_burst(dev, port, ev,
 100                                 BURST_SIZE, 0);
 101
 102                 if (!nb_rx) {
 103                         rte_pause();
 104                         continue;
 105                 }
 106
 107                 for (i = 0; i < nb_rx; i++) {
 108                         if (prod_crypto_type &&
 109                             (ev[i].event_type == RTE_EVENT_TYPE_CRYPTODEV)) {
 110                                 struct rte_crypto_op *op = ev[i].event_ptr;
 111
 112                                 if (op->status ==
 113                                     RTE_CRYPTO_OP_STATUS_SUCCESS) {
 114                                         if (op->sym->m_dst == NULL)
 115                                                 ev[i].event_ptr =
 116                                                         op->sym->m_src;
 117                                         else
 118                                                 ev[i].event_ptr =
 119                                                         op->sym->m_dst;
 120                                         rte_crypto_op_free(op);
 121                                 } else {
 122                                         rte_crypto_op_free(op);
 123                                         continue;
 124                                 }
 125                         }
 126
 127                         if (enable_fwd_latency && !prod_timer_type) {
 128                                 rte_prefetch0(ev[i+1].event_ptr);
 129                                 /* first queue in pipeline.
 130                                  * mark time stamp to compute fwd latency
 131                                  */
 132                                 mark_fwd_latency(&ev[i], nb_stages);
 133                         }
 134                         /* last stage in pipeline */
 135                         if (unlikely((ev[i].queue_id % nb_stages) ==
 136                                                  laststage)) {
 137                                 if (enable_fwd_latency)
 138                                         cnt = perf_process_last_stage_latency(
 139                                                 pool, &ev[i], w, bufs, sz, cnt);
 140                                 else
 141                                         cnt = perf_process_last_stage(pool,
 142                                                 &ev[i], w, bufs, sz, cnt);
 143
 144                                 ev[i].op = RTE_EVENT_OP_RELEASE;
 145                         } else {
 146                                 fwd_event(&ev[i], sched_type_list, nb_stages);
 147                         }
 148                 }
 149
 150                 uint16_t enq;
 151
 152                 enq = rte_event_enqueue_burst(dev, port, ev, nb_rx);
 153                 while (enq < nb_rx) {
 154                         enq += rte_event_enqueue_burst(dev, port,
 155                                                         ev + enq, nb_rx - enq);
 156                 }
 157         }
 158         return 0;
 159 }
 160
 161 static int
 162 worker_wrapper(void *arg)
 163 {
 164         struct worker_data *w  = arg;
 165         struct evt_options *opt = w->t->opt;
 166
 167         const bool burst = evt_has_burst_mode(w->dev_id);
 168         const int fwd_latency = opt->fwd_latency;
 169
 170         /* allow compiler to optimize */
 171         if (!burst && !fwd_latency)
 172                 return perf_queue_worker(arg, 0);
 173         else if (!burst && fwd_latency)
 174                 return perf_queue_worker(arg, 1);
 175         else if (burst && !fwd_latency)
 176                 return perf_queue_worker_burst(arg, 0);
 177         else if (burst && fwd_latency)
 178                 return perf_queue_worker_burst(arg, 1);
 179
 180         rte_panic("invalid worker\n");
 181 }
 182
 183 static int
 184 perf_queue_launch_lcores(struct evt_test *test, struct evt_options *opt)
 185 {
 186         return perf_launch_lcores(test, opt, worker_wrapper);
 187 }
 188
 189 static int
 190 perf_queue_eventdev_setup(struct evt_test *test, struct evt_options *opt)
 191 {
 192         uint8_t queue;
 193         int nb_stages = opt->nb_stages;
 194         int ret;
 195         int nb_ports;
 196         int nb_queues;
 197         uint16_t prod;
 198         struct rte_event_dev_info dev_info;
 199         struct test_perf *t = evt_test_priv(test);
 200
 201         nb_ports = evt_nr_active_lcores(opt->wlcores);
 202         nb_ports += opt->prod_type == EVT_PROD_TYPE_ETH_RX_ADPTR ||
 203                 opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR ? 0 :
 204                 evt_nr_active_lcores(opt->plcores);
 205
 206         nb_queues = perf_queue_nb_event_queues(opt);
 207
 208         memset(&dev_info, 0, sizeof(struct rte_event_dev_info));
 209         ret = rte_event_dev_info_get(opt->dev_id, &dev_info);
 210         if (ret) {
 211                 evt_err("failed to get eventdev info %d", opt->dev_id);
 212                 return ret;
 213         }
 214
 215         ret = evt_configure_eventdev(opt, nb_queues, nb_ports);
 216         if (ret) {
 217                 evt_err("failed to configure eventdev %d", opt->dev_id);
 218                 return ret;
 219         }
 220
 221         struct rte_event_queue_conf q_conf = {
 222                         .priority = RTE_EVENT_DEV_PRIORITY_NORMAL,
 223                         .nb_atomic_flows = opt->nb_flows,
 224                         .nb_atomic_order_sequences = opt->nb_flows,
 225         };
 226         /* queue configurations */
 227         for (queue = 0; queue < nb_queues; queue++) {
 228                 q_conf.schedule_type =
 229                         (opt->sched_type_list[queue % nb_stages]);
 230
 231                 if (opt->q_priority) {
 232                         uint8_t stage_pos = queue % nb_stages;
 233                         /* Configure event queues(stage 0 to stage n) with
 234                          * RTE_EVENT_DEV_PRIORITY_LOWEST to
 235                          * RTE_EVENT_DEV_PRIORITY_HIGHEST.
 236                          */
 237                         uint8_t step = RTE_EVENT_DEV_PRIORITY_LOWEST /
 238                                         (nb_stages - 1);
 239                         /* Higher prio for the queues closer to last stage */
 240                         q_conf.priority = RTE_EVENT_DEV_PRIORITY_LOWEST -
 241                                         (step * stage_pos);
 242                 }
 243                 ret = rte_event_queue_setup(opt->dev_id, queue, &q_conf);
 244                 if (ret) {
 245                         evt_err("failed to setup queue=%d", queue);
 246                         return ret;
 247                 }
 248         }
 249
 250         if (opt->wkr_deq_dep > dev_info.max_event_port_dequeue_depth)
 251                 opt->wkr_deq_dep = dev_info.max_event_port_dequeue_depth;
 252
 253         /* port configuration */
 254         const struct rte_event_port_conf p_conf = {
 255                         .dequeue_depth = opt->wkr_deq_dep,
 256                         .enqueue_depth = dev_info.max_event_port_dequeue_depth,
 257                         .new_event_threshold = dev_info.max_num_events,
 258         };
 259
 260         ret = perf_event_dev_port_setup(test, opt, nb_stages /* stride */,
 261                                         nb_queues, &p_conf);
 262         if (ret)
 263                 return ret;
 264
 265         if (!evt_has_distributed_sched(opt->dev_id)) {
 266                 uint32_t service_id;
 267                 rte_event_dev_service_id_get(opt->dev_id, &service_id);
 268                 ret = evt_service_setup(service_id);
 269                 if (ret) {
 270                         evt_err("No service lcore found to run event dev.");
 271                         return ret;
 272                 }
 273         }
 274
 275         ret = rte_event_dev_start(opt->dev_id);
 276         if (ret) {
 277                 evt_err("failed to start eventdev %d", opt->dev_id);
 278                 return ret;
 279         }
 280
 281         if (opt->prod_type == EVT_PROD_TYPE_ETH_RX_ADPTR) {
 282                 RTE_ETH_FOREACH_DEV(prod) {
 283                         ret = rte_eth_dev_start(prod);
 284                         if (ret) {
 285                                 evt_err("Ethernet dev [%d] failed to start. Using synthetic producer",
 286                                                 prod);
 287                                 return ret;
 288                         }
 289
 290                         ret = rte_event_eth_rx_adapter_start(prod);
 291                         if (ret) {
 292                                 evt_err("Rx adapter[%d] start failed", prod);
 293                                 return ret;
 294                         }
 295                         printf("%s: Port[%d] using Rx adapter[%d] started\n",
 296                                         __func__, prod, prod);
 297                 }
 298         } else if (opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR) {
 299                 for (prod = 0; prod < opt->nb_timer_adptrs; prod++) {
 300                         ret = rte_event_timer_adapter_start(
 301                                         t->timer_adptr[prod]);
 302                         if (ret) {
 303                                 evt_err("failed to Start event timer adapter %d"
 304                                                 , prod);
 305                                 return ret;
 306                         }
 307                 }
 308         } else if (opt->prod_type == EVT_PROD_TYPE_EVENT_CRYPTO_ADPTR) {
 309                 uint8_t cdev_id, cdev_count;
 310
 311                 cdev_count = rte_cryptodev_count();
 312                 for (cdev_id = 0; cdev_id < cdev_count; cdev_id++) {
 313                         ret = rte_cryptodev_start(cdev_id);
 314                         if (ret) {
 315                                 evt_err("Failed to start cryptodev %u",
 316                                         cdev_id);
 317                                 return ret;
 318                         }
 319                 }
 320         }
 321
 322         return 0;
 323 }
 324
 325 static void
 326 perf_queue_opt_dump(struct evt_options *opt)
 327 {
 328         evt_dump_fwd_latency(opt);
 329         perf_opt_dump(opt, perf_queue_nb_event_queues(opt));
 330 }
 331
 332 static int
 333 perf_queue_opt_check(struct evt_options *opt)
 334 {
 335         return perf_opt_check(opt, perf_queue_nb_event_queues(opt));
 336 }
 337
 338 static bool
 339 perf_queue_capability_check(struct evt_options *opt)
 340 {
 341         struct rte_event_dev_info dev_info;
 342
 343         rte_event_dev_info_get(opt->dev_id, &dev_info);
 344         if (dev_info.max_event_queues < perf_queue_nb_event_queues(opt) ||
 345                         dev_info.max_event_ports < perf_nb_event_ports(opt)) {
 346                 evt_err("not enough eventdev queues=%d/%d or ports=%d/%d",
 347                         perf_queue_nb_event_queues(opt),
 348                         dev_info.max_event_queues,
 349                         perf_nb_event_ports(opt), dev_info.max_event_ports);
 350         }
 351
 352         return true;
 353 }
 354
 355 static const struct evt_test_ops perf_queue =  {
 356         .cap_check          = perf_queue_capability_check,
 357         .opt_check          = perf_queue_opt_check,
 358         .opt_dump           = perf_queue_opt_dump,
 359         .test_setup         = perf_test_setup,
 360         .mempool_setup      = perf_mempool_setup,
 361         .ethdev_setup       = perf_ethdev_setup,
 362         .cryptodev_setup    = perf_cryptodev_setup,
 363         .eventdev_setup     = perf_queue_eventdev_setup,
 364         .launch_lcores      = perf_queue_launch_lcores,
 365         .eventdev_destroy   = perf_eventdev_destroy,
 366         .mempool_destroy    = perf_mempool_destroy,
 367         .ethdev_destroy     = perf_ethdev_destroy,
 368         .cryptodev_destroy  = perf_cryptodev_destroy,
 369         .test_result        = perf_test_result,
 370         .test_destroy       = perf_test_destroy,
 371 };
 372
 373 EVT_TEST_REGISTER(perf_queue);