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