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