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 __rte_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 __rte_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 uint16_t enq = 0, deq = 0; 43 struct rte_event ev; 44 PERF_WORKER_INIT; 45 46 while (t->done == false) { 47 deq = rte_event_dequeue_burst(dev, port, &ev, 1, 0); 48 49 if (!deq) { 50 rte_pause(); 51 continue; 52 } 53 54 if (prod_crypto_type && 55 (ev.event_type == RTE_EVENT_TYPE_CRYPTODEV)) { 56 struct rte_crypto_op *op = ev.event_ptr; 57 58 if (op->status == RTE_CRYPTO_OP_STATUS_SUCCESS) { 59 if (op->type == RTE_CRYPTO_OP_TYPE_SYMMETRIC) { 60 if (op->sym->m_dst == NULL) 61 ev.event_ptr = op->sym->m_src; 62 else 63 ev.event_ptr = op->sym->m_dst; 64 rte_crypto_op_free(op); 65 } 66 } else { 67 rte_crypto_op_free(op); 68 continue; 69 } 70 } 71 72 if (enable_fwd_latency && !prod_timer_type) 73 /* first q in pipeline, mark timestamp to compute fwd latency */ 74 mark_fwd_latency(&ev, nb_stages); 75 76 /* last stage in pipeline */ 77 if (unlikely((ev.queue_id % nb_stages) == laststage)) { 78 if (enable_fwd_latency) 79 cnt = perf_process_last_stage_latency(pool, 80 &ev, w, bufs, sz, cnt); 81 else 82 cnt = perf_process_last_stage(pool, 83 &ev, w, bufs, sz, cnt); 84 } else { 85 fwd_event(&ev, sched_type_list, nb_stages); 86 do { 87 enq = rte_event_enqueue_burst(dev, port, &ev, 88 1); 89 } while (!enq && !t->done); 90 } 91 } 92 93 perf_worker_cleanup(pool, dev, port, &ev, enq, deq); 94 95 return 0; 96 } 97 98 static int 99 perf_queue_worker_burst(void *arg, const int enable_fwd_latency) 100 { 101 /* +1 to avoid prefetch out of array check */ 102 struct rte_event ev[BURST_SIZE + 1]; 103 uint16_t enq = 0, nb_rx = 0; 104 PERF_WORKER_INIT; 105 uint16_t i; 106 107 while (t->done == false) { 108 nb_rx = rte_event_dequeue_burst(dev, port, ev, BURST_SIZE, 0); 109 110 if (!nb_rx) { 111 rte_pause(); 112 continue; 113 } 114 115 for (i = 0; i < nb_rx; i++) { 116 if (prod_crypto_type && 117 (ev[i].event_type == RTE_EVENT_TYPE_CRYPTODEV)) { 118 struct rte_crypto_op *op = ev[i].event_ptr; 119 120 if (op->status == 121 RTE_CRYPTO_OP_STATUS_SUCCESS) { 122 if (op->sym->m_dst == NULL) 123 ev[i].event_ptr = 124 op->sym->m_src; 125 else 126 ev[i].event_ptr = 127 op->sym->m_dst; 128 rte_crypto_op_free(op); 129 } else { 130 rte_crypto_op_free(op); 131 continue; 132 } 133 } 134 135 if (enable_fwd_latency && !prod_timer_type) { 136 rte_prefetch0(ev[i+1].event_ptr); 137 /* first queue in pipeline. 138 * mark time stamp to compute fwd latency 139 */ 140 mark_fwd_latency(&ev[i], nb_stages); 141 } 142 /* last stage in pipeline */ 143 if (unlikely((ev[i].queue_id % nb_stages) == 144 laststage)) { 145 if (enable_fwd_latency) 146 cnt = perf_process_last_stage_latency( 147 pool, &ev[i], w, bufs, sz, cnt); 148 else 149 cnt = perf_process_last_stage(pool, 150 &ev[i], w, bufs, sz, cnt); 151 152 ev[i].op = RTE_EVENT_OP_RELEASE; 153 } else { 154 fwd_event(&ev[i], sched_type_list, nb_stages); 155 } 156 } 157 158 159 enq = rte_event_enqueue_burst(dev, port, ev, nb_rx); 160 while (enq < nb_rx && !t->done) { 161 enq += rte_event_enqueue_burst(dev, port, 162 ev + enq, nb_rx - enq); 163 } 164 } 165 166 perf_worker_cleanup(pool, dev, port, ev, enq, nb_rx); 167 168 return 0; 169 } 170 171 static int 172 worker_wrapper(void *arg) 173 { 174 struct worker_data *w = arg; 175 struct evt_options *opt = w->t->opt; 176 177 const bool burst = evt_has_burst_mode(w->dev_id); 178 const int fwd_latency = opt->fwd_latency; 179 180 /* allow compiler to optimize */ 181 if (!burst && !fwd_latency) 182 return perf_queue_worker(arg, 0); 183 else if (!burst && fwd_latency) 184 return perf_queue_worker(arg, 1); 185 else if (burst && !fwd_latency) 186 return perf_queue_worker_burst(arg, 0); 187 else if (burst && fwd_latency) 188 return perf_queue_worker_burst(arg, 1); 189 190 rte_panic("invalid worker\n"); 191 } 192 193 static int 194 perf_queue_launch_lcores(struct evt_test *test, struct evt_options *opt) 195 { 196 return perf_launch_lcores(test, opt, worker_wrapper); 197 } 198 199 static int 200 perf_queue_eventdev_setup(struct evt_test *test, struct evt_options *opt) 201 { 202 uint8_t queue; 203 int nb_stages = opt->nb_stages; 204 int ret; 205 int nb_ports; 206 int nb_queues; 207 uint16_t prod; 208 struct rte_event_dev_info dev_info; 209 struct test_perf *t = evt_test_priv(test); 210 211 nb_ports = evt_nr_active_lcores(opt->wlcores); 212 nb_ports += opt->prod_type == EVT_PROD_TYPE_ETH_RX_ADPTR || 213 opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR ? 0 : 214 evt_nr_active_lcores(opt->plcores); 215 216 nb_queues = perf_queue_nb_event_queues(opt); 217 218 ret = rte_event_dev_info_get(opt->dev_id, &dev_info); 219 if (ret) { 220 evt_err("failed to get eventdev info %d", opt->dev_id); 221 return ret; 222 } 223 224 ret = evt_configure_eventdev(opt, nb_queues, nb_ports); 225 if (ret) { 226 evt_err("failed to configure eventdev %d", opt->dev_id); 227 return ret; 228 } 229 230 struct rte_event_queue_conf q_conf = { 231 .priority = RTE_EVENT_DEV_PRIORITY_NORMAL, 232 .nb_atomic_flows = opt->nb_flows, 233 .nb_atomic_order_sequences = opt->nb_flows, 234 }; 235 /* queue configurations */ 236 for (queue = 0; queue < nb_queues; queue++) { 237 q_conf.schedule_type = 238 (opt->sched_type_list[queue % nb_stages]); 239 240 if (opt->q_priority) { 241 uint8_t stage_pos = queue % nb_stages; 242 /* Configure event queues(stage 0 to stage n) with 243 * RTE_EVENT_DEV_PRIORITY_LOWEST to 244 * RTE_EVENT_DEV_PRIORITY_HIGHEST. 245 */ 246 uint8_t step = RTE_EVENT_DEV_PRIORITY_LOWEST / 247 (nb_stages - 1); 248 /* Higher prio for the queues closer to last stage */ 249 q_conf.priority = RTE_EVENT_DEV_PRIORITY_LOWEST - 250 (step * stage_pos); 251 } 252 ret = rte_event_queue_setup(opt->dev_id, queue, &q_conf); 253 if (ret) { 254 evt_err("failed to setup queue=%d", queue); 255 return ret; 256 } 257 } 258 259 if (opt->wkr_deq_dep > dev_info.max_event_port_dequeue_depth) 260 opt->wkr_deq_dep = dev_info.max_event_port_dequeue_depth; 261 262 /* port configuration */ 263 const struct rte_event_port_conf p_conf = { 264 .dequeue_depth = opt->wkr_deq_dep, 265 .enqueue_depth = dev_info.max_event_port_dequeue_depth, 266 .new_event_threshold = dev_info.max_num_events, 267 }; 268 269 ret = perf_event_dev_port_setup(test, opt, nb_stages /* stride */, 270 nb_queues, &p_conf); 271 if (ret) 272 return ret; 273 274 if (!evt_has_distributed_sched(opt->dev_id)) { 275 uint32_t service_id; 276 rte_event_dev_service_id_get(opt->dev_id, &service_id); 277 ret = evt_service_setup(service_id); 278 if (ret) { 279 evt_err("No service lcore found to run event dev."); 280 return ret; 281 } 282 } 283 284 ret = rte_event_dev_start(opt->dev_id); 285 if (ret) { 286 evt_err("failed to start eventdev %d", opt->dev_id); 287 return ret; 288 } 289 290 if (opt->prod_type == EVT_PROD_TYPE_ETH_RX_ADPTR) { 291 RTE_ETH_FOREACH_DEV(prod) { 292 ret = rte_eth_dev_start(prod); 293 if (ret) { 294 evt_err("Ethernet dev [%d] failed to start. Using synthetic producer", 295 prod); 296 return ret; 297 } 298 299 ret = rte_event_eth_rx_adapter_start(prod); 300 if (ret) { 301 evt_err("Rx adapter[%d] start failed", prod); 302 return ret; 303 } 304 printf("%s: Port[%d] using Rx adapter[%d] started\n", 305 __func__, prod, prod); 306 } 307 } else if (opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR) { 308 for (prod = 0; prod < opt->nb_timer_adptrs; prod++) { 309 ret = rte_event_timer_adapter_start( 310 t->timer_adptr[prod]); 311 if (ret) { 312 evt_err("failed to Start event timer adapter %d" 313 , prod); 314 return ret; 315 } 316 } 317 } else if (opt->prod_type == EVT_PROD_TYPE_EVENT_CRYPTO_ADPTR) { 318 uint8_t cdev_id, cdev_count; 319 320 cdev_count = rte_cryptodev_count(); 321 for (cdev_id = 0; cdev_id < cdev_count; cdev_id++) { 322 ret = rte_cryptodev_start(cdev_id); 323 if (ret) { 324 evt_err("Failed to start cryptodev %u", 325 cdev_id); 326 return ret; 327 } 328 } 329 } 330 331 return 0; 332 } 333 334 static void 335 perf_queue_opt_dump(struct evt_options *opt) 336 { 337 evt_dump_fwd_latency(opt); 338 perf_opt_dump(opt, perf_queue_nb_event_queues(opt)); 339 } 340 341 static int 342 perf_queue_opt_check(struct evt_options *opt) 343 { 344 return perf_opt_check(opt, perf_queue_nb_event_queues(opt)); 345 } 346 347 static bool 348 perf_queue_capability_check(struct evt_options *opt) 349 { 350 struct rte_event_dev_info dev_info; 351 352 rte_event_dev_info_get(opt->dev_id, &dev_info); 353 if (dev_info.max_event_queues < perf_queue_nb_event_queues(opt) || 354 dev_info.max_event_ports < perf_nb_event_ports(opt)) { 355 evt_err("not enough eventdev queues=%d/%d or ports=%d/%d", 356 perf_queue_nb_event_queues(opt), 357 dev_info.max_event_queues, 358 perf_nb_event_ports(opt), dev_info.max_event_ports); 359 } 360 361 return true; 362 } 363 364 static const struct evt_test_ops perf_queue = { 365 .cap_check = perf_queue_capability_check, 366 .opt_check = perf_queue_opt_check, 367 .opt_dump = perf_queue_opt_dump, 368 .test_setup = perf_test_setup, 369 .mempool_setup = perf_mempool_setup, 370 .ethdev_setup = perf_ethdev_setup, 371 .cryptodev_setup = perf_cryptodev_setup, 372 .ethdev_rx_stop = perf_ethdev_rx_stop, 373 .eventdev_setup = perf_queue_eventdev_setup, 374 .launch_lcores = perf_queue_launch_lcores, 375 .eventdev_destroy = perf_eventdev_destroy, 376 .mempool_destroy = perf_mempool_destroy, 377 .ethdev_destroy = perf_ethdev_destroy, 378 .cryptodev_destroy = perf_cryptodev_destroy, 379 .test_result = perf_test_result, 380 .test_destroy = perf_test_destroy, 381 }; 382 383 EVT_TEST_REGISTER(perf_queue); 384