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         uint16_t enq = 0, deq = 0;
  43         struct rte_event ev;
  44         PERF_WORKER_INIT;
  45
  46         while (t->done == false) {
  47                 deq = rte_event_dequeue_burst(dev, port, &ev, 1, 0);
  48
  49                 if (!deq) {
  50                         rte_pause();
  51                         continue;
  52                 }
  53
  54                 if (prod_crypto_type &&
  55                     (ev.event_type == RTE_EVENT_TYPE_CRYPTODEV)) {
  56                         struct rte_crypto_op *op = ev.event_ptr;
  57
  58                         if (op->status == RTE_CRYPTO_OP_STATUS_SUCCESS) {
  59                                 if (op->type == RTE_CRYPTO_OP_TYPE_SYMMETRIC) {
  60                                         if (op->sym->m_dst == NULL)
  61                                                 ev.event_ptr = op->sym->m_src;
  62                                         else
  63                                                 ev.event_ptr = op->sym->m_dst;
  64                                         rte_crypto_op_free(op);
  65                                 }
  66                         } else {
  67                                 rte_crypto_op_free(op);
  68                                 continue;
  69                         }
  70                 }
  71
  72                 if (enable_fwd_latency && !prod_timer_type)
  73                 /* first q in pipeline, mark timestamp to compute fwd latency */
  74                         mark_fwd_latency(&ev, nb_stages);
  75
  76                 /* last stage in pipeline */
  77                 if (unlikely((ev.queue_id % nb_stages) == laststage)) {
  78                         if (enable_fwd_latency)
  79                                 cnt = perf_process_last_stage_latency(pool,
  80                                         &ev, w, bufs, sz, cnt);
  81                         else
  82                                 cnt = perf_process_last_stage(pool,
  83                                         &ev, w, bufs, sz, cnt);
  84                 } else {
  85                         fwd_event(&ev, sched_type_list, nb_stages);
  86                         do {
  87                                 enq = rte_event_enqueue_burst(dev, port, &ev,
  88                                                               1);
  89                         } while (!enq && !t->done);
  90                 }
  91         }
  92
  93         perf_worker_cleanup(pool, dev, port, &ev, enq, deq);
  94
  95         return 0;
  96 }
  97
  98 static int
  99 perf_queue_worker_burst(void *arg, const int enable_fwd_latency)
 100 {
 101         /* +1 to avoid prefetch out of array check */
 102         struct rte_event ev[BURST_SIZE + 1];
 103         uint16_t enq = 0, nb_rx = 0;
 104         PERF_WORKER_INIT;
 105         uint16_t i;
 106
 107         while (t->done == false) {
 108                 nb_rx = rte_event_dequeue_burst(dev, port, ev, BURST_SIZE, 0);
 109
 110                 if (!nb_rx) {
 111                         rte_pause();
 112                         continue;
 113                 }
 114
 115                 for (i = 0; i < nb_rx; i++) {
 116                         if (prod_crypto_type &&
 117                             (ev[i].event_type == RTE_EVENT_TYPE_CRYPTODEV)) {
 118                                 struct rte_crypto_op *op = ev[i].event_ptr;
 119
 120                                 if (op->status ==
 121                                     RTE_CRYPTO_OP_STATUS_SUCCESS) {
 122                                         if (op->sym->m_dst == NULL)
 123                                                 ev[i].event_ptr =
 124                                                         op->sym->m_src;
 125                                         else
 126                                                 ev[i].event_ptr =
 127                                                         op->sym->m_dst;
 128                                         rte_crypto_op_free(op);
 129                                 } else {
 130                                         rte_crypto_op_free(op);
 131                                         continue;
 132                                 }
 133                         }
 134
 135                         if (enable_fwd_latency && !prod_timer_type) {
 136                                 rte_prefetch0(ev[i+1].event_ptr);
 137                                 /* first queue in pipeline.
 138                                  * mark time stamp to compute fwd latency
 139                                  */
 140                                 mark_fwd_latency(&ev[i], nb_stages);
 141                         }
 142                         /* last stage in pipeline */
 143                         if (unlikely((ev[i].queue_id % nb_stages) ==
 144                                                  laststage)) {
 145                                 if (enable_fwd_latency)
 146                                         cnt = perf_process_last_stage_latency(
 147                                                 pool, &ev[i], w, bufs, sz, cnt);
 148                                 else
 149                                         cnt = perf_process_last_stage(pool,
 150                                                 &ev[i], w, bufs, sz, cnt);
 151
 152                                 ev[i].op = RTE_EVENT_OP_RELEASE;
 153                         } else {
 154                                 fwd_event(&ev[i], sched_type_list, nb_stages);
 155                         }
 156                 }
 157
 158
 159                 enq = rte_event_enqueue_burst(dev, port, ev, nb_rx);
 160                 while (enq < nb_rx && !t->done) {
 161                         enq += rte_event_enqueue_burst(dev, port,
 162                                                         ev + enq, nb_rx - enq);
 163                 }
 164         }
 165
 166         perf_worker_cleanup(pool, dev, port, ev, enq, nb_rx);
 167
 168         return 0;
 169 }
 170
 171 static int
 172 worker_wrapper(void *arg)
 173 {
 174         struct worker_data *w  = arg;
 175         struct evt_options *opt = w->t->opt;
 176
 177         const bool burst = evt_has_burst_mode(w->dev_id);
 178         const int fwd_latency = opt->fwd_latency;
 179
 180         /* allow compiler to optimize */
 181         if (!burst && !fwd_latency)
 182                 return perf_queue_worker(arg, 0);
 183         else if (!burst && fwd_latency)
 184                 return perf_queue_worker(arg, 1);
 185         else if (burst && !fwd_latency)
 186                 return perf_queue_worker_burst(arg, 0);
 187         else if (burst && fwd_latency)
 188                 return perf_queue_worker_burst(arg, 1);
 189
 190         rte_panic("invalid worker\n");
 191 }
 192
 193 static int
 194 perf_queue_launch_lcores(struct evt_test *test, struct evt_options *opt)
 195 {
 196         return perf_launch_lcores(test, opt, worker_wrapper);
 197 }
 198
 199 static int
 200 perf_queue_eventdev_setup(struct evt_test *test, struct evt_options *opt)
 201 {
 202         uint8_t queue;
 203         int nb_stages = opt->nb_stages;
 204         int ret;
 205         int nb_ports;
 206         int nb_queues;
 207         uint16_t prod;
 208         struct rte_event_dev_info dev_info;
 209         struct test_perf *t = evt_test_priv(test);
 210
 211         nb_ports = evt_nr_active_lcores(opt->wlcores);
 212         nb_ports += opt->prod_type == EVT_PROD_TYPE_ETH_RX_ADPTR ||
 213                 opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR ? 0 :
 214                 evt_nr_active_lcores(opt->plcores);
 215
 216         nb_queues = perf_queue_nb_event_queues(opt);
 217
 218         memset(&dev_info, 0, sizeof(struct rte_event_dev_info));
 219         ret = rte_event_dev_info_get(opt->dev_id, &dev_info);
 220         if (ret) {
 221                 evt_err("failed to get eventdev info %d", opt->dev_id);
 222                 return ret;
 223         }
 224
 225         ret = evt_configure_eventdev(opt, nb_queues, nb_ports);
 226         if (ret) {
 227                 evt_err("failed to configure eventdev %d", opt->dev_id);
 228                 return ret;
 229         }
 230
 231         struct rte_event_queue_conf q_conf = {
 232                         .priority = RTE_EVENT_DEV_PRIORITY_NORMAL,
 233                         .nb_atomic_flows = opt->nb_flows,
 234                         .nb_atomic_order_sequences = opt->nb_flows,
 235         };
 236         /* queue configurations */
 237         for (queue = 0; queue < nb_queues; queue++) {
 238                 q_conf.schedule_type =
 239                         (opt->sched_type_list[queue % nb_stages]);
 240
 241                 if (opt->q_priority) {
 242                         uint8_t stage_pos = queue % nb_stages;
 243                         /* Configure event queues(stage 0 to stage n) with
 244                          * RTE_EVENT_DEV_PRIORITY_LOWEST to
 245                          * RTE_EVENT_DEV_PRIORITY_HIGHEST.
 246                          */
 247                         uint8_t step = RTE_EVENT_DEV_PRIORITY_LOWEST /
 248                                         (nb_stages - 1);
 249                         /* Higher prio for the queues closer to last stage */
 250                         q_conf.priority = RTE_EVENT_DEV_PRIORITY_LOWEST -
 251                                         (step * stage_pos);
 252                 }
 253                 ret = rte_event_queue_setup(opt->dev_id, queue, &q_conf);
 254                 if (ret) {
 255                         evt_err("failed to setup queue=%d", queue);
 256                         return ret;
 257                 }
 258         }
 259
 260         if (opt->wkr_deq_dep > dev_info.max_event_port_dequeue_depth)
 261                 opt->wkr_deq_dep = dev_info.max_event_port_dequeue_depth;
 262
 263         /* port configuration */
 264         const struct rte_event_port_conf p_conf = {
 265                         .dequeue_depth = opt->wkr_deq_dep,
 266                         .enqueue_depth = dev_info.max_event_port_dequeue_depth,
 267                         .new_event_threshold = dev_info.max_num_events,
 268         };
 269
 270         ret = perf_event_dev_port_setup(test, opt, nb_stages /* stride */,
 271                                         nb_queues, &p_conf);
 272         if (ret)
 273                 return ret;
 274
 275         if (!evt_has_distributed_sched(opt->dev_id)) {
 276                 uint32_t service_id;
 277                 rte_event_dev_service_id_get(opt->dev_id, &service_id);
 278                 ret = evt_service_setup(service_id);
 279                 if (ret) {
 280                         evt_err("No service lcore found to run event dev.");
 281                         return ret;
 282                 }
 283         }
 284
 285         ret = rte_event_dev_start(opt->dev_id);
 286         if (ret) {
 287                 evt_err("failed to start eventdev %d", opt->dev_id);
 288                 return ret;
 289         }
 290
 291         if (opt->prod_type == EVT_PROD_TYPE_ETH_RX_ADPTR) {
 292                 RTE_ETH_FOREACH_DEV(prod) {
 293                         ret = rte_eth_dev_start(prod);
 294                         if (ret) {
 295                                 evt_err("Ethernet dev [%d] failed to start. Using synthetic producer",
 296                                                 prod);
 297                                 return ret;
 298                         }
 299
 300                         ret = rte_event_eth_rx_adapter_start(prod);
 301                         if (ret) {
 302                                 evt_err("Rx adapter[%d] start failed", prod);
 303                                 return ret;
 304                         }
 305                         printf("%s: Port[%d] using Rx adapter[%d] started\n",
 306                                         __func__, prod, prod);
 307                 }
 308         } else if (opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR) {
 309                 for (prod = 0; prod < opt->nb_timer_adptrs; prod++) {
 310                         ret = rte_event_timer_adapter_start(
 311                                         t->timer_adptr[prod]);
 312                         if (ret) {
 313                                 evt_err("failed to Start event timer adapter %d"
 314                                                 , prod);
 315                                 return ret;
 316                         }
 317                 }
 318         } else if (opt->prod_type == EVT_PROD_TYPE_EVENT_CRYPTO_ADPTR) {
 319                 uint8_t cdev_id, cdev_count;
 320
 321                 cdev_count = rte_cryptodev_count();
 322                 for (cdev_id = 0; cdev_id < cdev_count; cdev_id++) {
 323                         ret = rte_cryptodev_start(cdev_id);
 324                         if (ret) {
 325                                 evt_err("Failed to start cryptodev %u",
 326                                         cdev_id);
 327                                 return ret;
 328                         }
 329                 }
 330         }
 331
 332         return 0;
 333 }
 334
 335 static void
 336 perf_queue_opt_dump(struct evt_options *opt)
 337 {
 338         evt_dump_fwd_latency(opt);
 339         perf_opt_dump(opt, perf_queue_nb_event_queues(opt));
 340 }
 341
 342 static int
 343 perf_queue_opt_check(struct evt_options *opt)
 344 {
 345         return perf_opt_check(opt, perf_queue_nb_event_queues(opt));
 346 }
 347
 348 static bool
 349 perf_queue_capability_check(struct evt_options *opt)
 350 {
 351         struct rte_event_dev_info dev_info;
 352
 353         rte_event_dev_info_get(opt->dev_id, &dev_info);
 354         if (dev_info.max_event_queues < perf_queue_nb_event_queues(opt) ||
 355                         dev_info.max_event_ports < perf_nb_event_ports(opt)) {
 356                 evt_err("not enough eventdev queues=%d/%d or ports=%d/%d",
 357                         perf_queue_nb_event_queues(opt),
 358                         dev_info.max_event_queues,
 359                         perf_nb_event_ports(opt), dev_info.max_event_ports);
 360         }
 361
 362         return true;
 363 }
 364
 365 static const struct evt_test_ops perf_queue =  {
 366         .cap_check          = perf_queue_capability_check,
 367         .opt_check          = perf_queue_opt_check,
 368         .opt_dump           = perf_queue_opt_dump,
 369         .test_setup         = perf_test_setup,
 370         .mempool_setup      = perf_mempool_setup,
 371         .ethdev_setup       = perf_ethdev_setup,
 372         .cryptodev_setup    = perf_cryptodev_setup,
 373         .ethdev_rx_stop     = perf_ethdev_rx_stop,
 374         .eventdev_setup     = perf_queue_eventdev_setup,
 375         .launch_lcores      = perf_queue_launch_lcores,
 376         .eventdev_destroy   = perf_eventdev_destroy,
 377         .mempool_destroy    = perf_mempool_destroy,
 378         .ethdev_destroy     = perf_ethdev_destroy,
 379         .cryptodev_destroy  = perf_cryptodev_destroy,
 380         .test_result        = perf_test_result,
 381         .test_destroy       = perf_test_destroy,
 382 };
 383
 384 EVT_TEST_REGISTER(perf_queue);