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