ceph/src/spdk/dpdk/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 inline __attribute__((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 inline __attribute__((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         PERF_WORKER_INIT;
  41         struct rte_event ev;
  42
  43         while (t->done == false) {
  44                 uint16_t event = rte_event_dequeue_burst(dev, port, &ev, 1, 0);
  45
  46                 if (!event) {
  47                         rte_pause();
  48                         continue;
  49                 }
  50
  51                 if (enable_fwd_latency && !prod_timer_type)
  52                 /* first stage in pipeline, mark ts to compute fwd latency */
  53                         atq_mark_fwd_latency(&ev);
  54
  55                 /* last stage in pipeline */
  56                 if (unlikely((ev.sub_event_type % nb_stages) == laststage)) {
  57                         if (enable_fwd_latency)
  58                                 cnt = perf_process_last_stage_latency(pool,
  59                                         &ev, w, bufs, sz, cnt);
  60                         else
  61                                 cnt = perf_process_last_stage(pool, &ev, w,
  62                                          bufs, sz, cnt);
  63                 } else {
  64                         atq_fwd_event(&ev, sched_type_list, nb_stages);
  65                         while (rte_event_enqueue_burst(dev, port, &ev, 1) != 1)
  66                                 rte_pause();
  67                 }
  68         }
  69         return 0;
  70 }
  71
  72 static int
  73 perf_atq_worker_burst(void *arg, const int enable_fwd_latency)
  74 {
  75         PERF_WORKER_INIT;
  76         uint16_t i;
  77         /* +1 to avoid prefetch out of array check */
  78         struct rte_event ev[BURST_SIZE + 1];
  79
  80         while (t->done == false) {
  81                 uint16_t const nb_rx = rte_event_dequeue_burst(dev, port, ev,
  82                                 BURST_SIZE, 0);
  83
  84                 if (!nb_rx) {
  85                         rte_pause();
  86                         continue;
  87                 }
  88
  89                 for (i = 0; i < nb_rx; i++) {
  90                         if (enable_fwd_latency && !prod_timer_type) {
  91                                 rte_prefetch0(ev[i+1].event_ptr);
  92                                 /* first stage in pipeline.
  93                                  * mark time stamp to compute fwd latency
  94                                  */
  95                                 atq_mark_fwd_latency(&ev[i]);
  96                         }
  97                         /* last stage in pipeline */
  98                         if (unlikely((ev[i].sub_event_type % nb_stages)
  99                                                 == laststage)) {
 100                                 if (enable_fwd_latency)
 101                                         cnt = perf_process_last_stage_latency(
 102                                                 pool, &ev[i], w, bufs, sz, cnt);
 103                                 else
 104                                         cnt = perf_process_last_stage(pool,
 105                                                 &ev[i], w, bufs, sz, cnt);
 106
 107                                 ev[i].op = RTE_EVENT_OP_RELEASE;
 108                         } else {
 109                                 atq_fwd_event(&ev[i], sched_type_list,
 110                                                 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_atq_worker(arg, 0);
 137         else if (!burst && fwd_latency)
 138                 return perf_atq_worker(arg, 1);
 139         else if (burst && !fwd_latency)
 140                 return perf_atq_worker_burst(arg, 0);
 141         else if (burst && fwd_latency)
 142                 return perf_atq_worker_burst(arg, 1);
 143
 144         rte_panic("invalid worker\n");
 145 }
 146
 147 static int
 148 perf_atq_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_atq_eventdev_setup(struct evt_test *test, struct evt_options *opt)
 155 {
 156         int ret;
 157         uint8_t queue;
 158         uint8_t nb_queues;
 159         uint8_t nb_ports;
 160         uint16_t prod;
 161         struct rte_event_dev_info dev_info;
 162         struct test_perf *t = evt_test_priv(test);
 163
 164         nb_ports = evt_nr_active_lcores(opt->wlcores);
 165         nb_ports += (opt->prod_type == EVT_PROD_TYPE_ETH_RX_ADPTR ||
 166                         opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR) ? 0 :
 167                 evt_nr_active_lcores(opt->plcores);
 168
 169         nb_queues = atq_nb_event_queues(opt);
 170
 171         memset(&dev_info, 0, sizeof(struct rte_event_dev_info));
 172         ret = rte_event_dev_info_get(opt->dev_id, &dev_info);
 173         if (ret) {
 174                 evt_err("failed to get eventdev info %d", opt->dev_id);
 175                 return ret;
 176         }
 177
 178         ret = evt_configure_eventdev(opt, nb_queues, nb_ports);
 179         if (ret) {
 180                 evt_err("failed to configure eventdev %d", opt->dev_id);
 181                 return ret;
 182         }
 183
 184         struct rte_event_queue_conf q_conf = {
 185                         .priority = RTE_EVENT_DEV_PRIORITY_NORMAL,
 186                         .event_queue_cfg = RTE_EVENT_QUEUE_CFG_ALL_TYPES,
 187                         .nb_atomic_flows = opt->nb_flows,
 188                         .nb_atomic_order_sequences = opt->nb_flows,
 189         };
 190         /* queue configurations */
 191         for (queue = 0; queue < nb_queues; queue++) {
 192                 ret = rte_event_queue_setup(opt->dev_id, queue, &q_conf);
 193                 if (ret) {
 194                         evt_err("failed to setup queue=%d", queue);
 195                         return ret;
 196                 }
 197         }
 198
 199         if (opt->wkr_deq_dep > dev_info.max_event_port_dequeue_depth)
 200                 opt->wkr_deq_dep = dev_info.max_event_port_dequeue_depth;
 201
 202         /* port configuration */
 203         const struct rte_event_port_conf p_conf = {
 204                         .dequeue_depth = opt->wkr_deq_dep,
 205                         .enqueue_depth = dev_info.max_event_port_dequeue_depth,
 206                         .new_event_threshold = dev_info.max_num_events,
 207         };
 208
 209         ret = perf_event_dev_port_setup(test, opt, 1 /* stride */, nb_queues,
 210                         &p_conf);
 211         if (ret)
 212                 return ret;
 213
 214         if (!evt_has_distributed_sched(opt->dev_id)) {
 215                 uint32_t service_id;
 216                 rte_event_dev_service_id_get(opt->dev_id, &service_id);
 217                 ret = evt_service_setup(service_id);
 218                 if (ret) {
 219                         evt_err("No service lcore found to run event dev.");
 220                         return ret;
 221                 }
 222         }
 223
 224         ret = rte_event_dev_start(opt->dev_id);
 225         if (ret) {
 226                 evt_err("failed to start eventdev %d", opt->dev_id);
 227                 return ret;
 228         }
 229
 230         if (opt->prod_type == EVT_PROD_TYPE_ETH_RX_ADPTR) {
 231                 RTE_ETH_FOREACH_DEV(prod) {
 232                         ret = rte_eth_dev_start(prod);
 233                         if (ret) {
 234                                 evt_err("Ethernet dev [%d] failed to start. Using synthetic producer",
 235                                                 prod);
 236                                 return ret;
 237                         }
 238
 239                         ret = rte_event_eth_rx_adapter_start(prod);
 240                         if (ret) {
 241                                 evt_err("Rx adapter[%d] start failed", prod);
 242                                 return ret;
 243                         }
 244                         printf("%s: Port[%d] using Rx adapter[%d] started\n",
 245                                         __func__, prod, prod);
 246                 }
 247         } else if (opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR) {
 248                 for (prod = 0; prod < opt->nb_timer_adptrs; prod++) {
 249                         ret = rte_event_timer_adapter_start(
 250                                         t->timer_adptr[prod]);
 251                         if (ret) {
 252                                 evt_err("failed to Start event timer adapter %d"
 253                                                 , prod);
 254                                 return ret;
 255                         }
 256                 }
 257         }
 258
 259         return 0;
 260 }
 261
 262 static void
 263 perf_atq_opt_dump(struct evt_options *opt)
 264 {
 265         perf_opt_dump(opt, atq_nb_event_queues(opt));
 266 }
 267
 268 static int
 269 perf_atq_opt_check(struct evt_options *opt)
 270 {
 271         return perf_opt_check(opt, atq_nb_event_queues(opt));
 272 }
 273
 274 static bool
 275 perf_atq_capability_check(struct evt_options *opt)
 276 {
 277         struct rte_event_dev_info dev_info;
 278
 279         rte_event_dev_info_get(opt->dev_id, &dev_info);
 280         if (dev_info.max_event_queues < atq_nb_event_queues(opt) ||
 281                         dev_info.max_event_ports < perf_nb_event_ports(opt)) {
 282                 evt_err("not enough eventdev queues=%d/%d or ports=%d/%d",
 283                         atq_nb_event_queues(opt), dev_info.max_event_queues,
 284                         perf_nb_event_ports(opt), dev_info.max_event_ports);
 285         }
 286         if (!evt_has_all_types_queue(opt->dev_id))
 287                 return false;
 288
 289         return true;
 290 }
 291
 292 static const struct evt_test_ops perf_atq =  {
 293         .cap_check          = perf_atq_capability_check,
 294         .opt_check          = perf_atq_opt_check,
 295         .opt_dump           = perf_atq_opt_dump,
 296         .test_setup         = perf_test_setup,
 297         .ethdev_setup       = perf_ethdev_setup,
 298         .mempool_setup      = perf_mempool_setup,
 299         .eventdev_setup     = perf_atq_eventdev_setup,
 300         .launch_lcores      = perf_atq_launch_lcores,
 301         .eventdev_destroy   = perf_eventdev_destroy,
 302         .mempool_destroy    = perf_mempool_destroy,
 303         .ethdev_destroy     = perf_ethdev_destroy,
 304         .test_result        = perf_test_result,
 305         .test_destroy       = perf_test_destroy,
 306 };
 307
 308 EVT_TEST_REGISTER(perf_atq);