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://dpdk.org/doc/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         struct rte_event_dev_info dev_info;
 162
 163         nb_ports = evt_nr_active_lcores(opt->wlcores);
 164         nb_ports += opt->prod_type == EVT_PROD_TYPE_ETH_RX_ADPTR ||
 165                  opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR ? 0 :
 166                 evt_nr_active_lcores(opt->plcores);
 167
 168         nb_queues = perf_queue_nb_event_queues(opt);
 169
 170         memset(&dev_info, 0, sizeof(struct rte_event_dev_info));
 171         ret = rte_event_dev_info_get(opt->dev_id, &dev_info);
 172         if (ret) {
 173                 evt_err("failed to get eventdev info %d", opt->dev_id);
 174                 return ret;
 175         }
 176
 177         const struct rte_event_dev_config config = {
 178                         .nb_event_queues = nb_queues,
 179                         .nb_event_ports = nb_ports,
 180                         .nb_events_limit  = dev_info.max_num_events,
 181                         .nb_event_queue_flows = opt->nb_flows,
 182                         .nb_event_port_dequeue_depth =
 183                                 dev_info.max_event_port_dequeue_depth,
 184                         .nb_event_port_enqueue_depth =
 185                                 dev_info.max_event_port_enqueue_depth,
 186         };
 187
 188         ret = rte_event_dev_configure(opt->dev_id, &config);
 189         if (ret) {
 190                 evt_err("failed to configure eventdev %d", opt->dev_id);
 191                 return ret;
 192         }
 193
 194         struct rte_event_queue_conf q_conf = {
 195                         .priority = RTE_EVENT_DEV_PRIORITY_NORMAL,
 196                         .nb_atomic_flows = opt->nb_flows,
 197                         .nb_atomic_order_sequences = opt->nb_flows,
 198         };
 199         /* queue configurations */
 200         for (queue = 0; queue < nb_queues; queue++) {
 201                 q_conf.schedule_type =
 202                         (opt->sched_type_list[queue % nb_stages]);
 203
 204                 if (opt->q_priority) {
 205                         uint8_t stage_pos = queue % nb_stages;
 206                         /* Configure event queues(stage 0 to stage n) with
 207                          * RTE_EVENT_DEV_PRIORITY_LOWEST to
 208                          * RTE_EVENT_DEV_PRIORITY_HIGHEST.
 209                          */
 210                         uint8_t step = RTE_EVENT_DEV_PRIORITY_LOWEST /
 211                                         (nb_stages - 1);
 212                         /* Higher prio for the queues closer to last stage */
 213                         q_conf.priority = RTE_EVENT_DEV_PRIORITY_LOWEST -
 214                                         (step * stage_pos);
 215                 }
 216                 ret = rte_event_queue_setup(opt->dev_id, queue, &q_conf);
 217                 if (ret) {
 218                         evt_err("failed to setup queue=%d", queue);
 219                         return ret;
 220                 }
 221         }
 222
 223         if (opt->wkr_deq_dep > dev_info.max_event_port_dequeue_depth)
 224                 opt->wkr_deq_dep = dev_info.max_event_port_dequeue_depth;
 225
 226         /* port configuration */
 227         const struct rte_event_port_conf p_conf = {
 228                         .dequeue_depth = opt->wkr_deq_dep,
 229                         .enqueue_depth = dev_info.max_event_port_dequeue_depth,
 230                         .new_event_threshold = dev_info.max_num_events,
 231         };
 232
 233         ret = perf_event_dev_port_setup(test, opt, nb_stages /* stride */,
 234                                         nb_queues, &p_conf);
 235         if (ret)
 236                 return ret;
 237
 238         if (!evt_has_distributed_sched(opt->dev_id)) {
 239                 uint32_t service_id;
 240                 rte_event_dev_service_id_get(opt->dev_id, &service_id);
 241                 ret = evt_service_setup(service_id);
 242                 if (ret) {
 243                         evt_err("No service lcore found to run event dev.");
 244                         return ret;
 245                 }
 246         }
 247
 248         ret = rte_event_dev_start(opt->dev_id);
 249         if (ret) {
 250                 evt_err("failed to start eventdev %d", opt->dev_id);
 251                 return ret;
 252         }
 253
 254         return 0;
 255 }
 256
 257 static void
 258 perf_queue_opt_dump(struct evt_options *opt)
 259 {
 260         evt_dump_fwd_latency(opt);
 261         perf_opt_dump(opt, perf_queue_nb_event_queues(opt));
 262 }
 263
 264 static int
 265 perf_queue_opt_check(struct evt_options *opt)
 266 {
 267         return perf_opt_check(opt, perf_queue_nb_event_queues(opt));
 268 }
 269
 270 static bool
 271 perf_queue_capability_check(struct evt_options *opt)
 272 {
 273         struct rte_event_dev_info dev_info;
 274
 275         rte_event_dev_info_get(opt->dev_id, &dev_info);
 276         if (dev_info.max_event_queues < perf_queue_nb_event_queues(opt) ||
 277                         dev_info.max_event_ports < perf_nb_event_ports(opt)) {
 278                 evt_err("not enough eventdev queues=%d/%d or ports=%d/%d",
 279                         perf_queue_nb_event_queues(opt),
 280                         dev_info.max_event_queues,
 281                         perf_nb_event_ports(opt), dev_info.max_event_ports);
 282         }
 283
 284         return true;
 285 }
 286
 287 static const struct evt_test_ops perf_queue =  {
 288         .cap_check          = perf_queue_capability_check,
 289         .opt_check          = perf_queue_opt_check,
 290         .opt_dump           = perf_queue_opt_dump,
 291         .test_setup         = perf_test_setup,
 292         .mempool_setup      = perf_mempool_setup,
 293         .ethdev_setup       = perf_ethdev_setup,
 294         .eventdev_setup     = perf_queue_eventdev_setup,
 295         .launch_lcores      = perf_queue_launch_lcores,
 296         .eventdev_destroy   = perf_eventdev_destroy,
 297         .mempool_destroy    = perf_mempool_destroy,
 298         .ethdev_destroy     = perf_ethdev_destroy,
 299         .test_result        = perf_test_result,
 300         .test_destroy       = perf_test_destroy,
 301 };
 302
 303 EVT_TEST_REGISTER(perf_queue);