xref: /dpdk/app/test-eventdev/test_perf_common.c (revision 2808423a9ce42a748aed77a4b487be27d2b6acfa)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright(c) 2017 Cavium, Inc
3  */
4 
5 #include "test_perf_common.h"
6 
7 int
8 perf_test_result(struct evt_test *test, struct evt_options *opt)
9 {
10 	RTE_SET_USED(opt);
11 	int i;
12 	uint64_t total = 0;
13 	struct test_perf *t = evt_test_priv(test);
14 
15 	printf("Packet distribution across worker cores :\n");
16 	for (i = 0; i < t->nb_workers; i++)
17 		total += t->worker[i].processed_pkts;
18 	for (i = 0; i < t->nb_workers; i++)
19 		printf("Worker %d packets: "CLGRN"%"PRIx64" "CLNRM"percentage:"
20 				CLGRN" %3.2f\n"CLNRM, i,
21 				t->worker[i].processed_pkts,
22 				(((double)t->worker[i].processed_pkts)/total)
23 				* 100);
24 
25 	return t->result;
26 }
27 
28 static inline int
29 perf_producer(void *arg)
30 {
31 	struct prod_data *p  = arg;
32 	struct test_perf *t = p->t;
33 	struct evt_options *opt = t->opt;
34 	const uint8_t dev_id = p->dev_id;
35 	const uint8_t port = p->port_id;
36 	struct rte_mempool *pool = t->pool;
37 	const uint64_t nb_pkts = t->nb_pkts;
38 	const uint32_t nb_flows = t->nb_flows;
39 	uint32_t flow_counter = 0;
40 	uint64_t count = 0;
41 	struct perf_elt *m;
42 	struct rte_event ev;
43 
44 	if (opt->verbose_level > 1)
45 		printf("%s(): lcore %d dev_id %d port=%d queue %d\n", __func__,
46 				rte_lcore_id(), dev_id, port, p->queue_id);
47 
48 	ev.event = 0;
49 	ev.op = RTE_EVENT_OP_NEW;
50 	ev.queue_id = p->queue_id;
51 	ev.sched_type = t->opt->sched_type_list[0];
52 	ev.priority = RTE_EVENT_DEV_PRIORITY_NORMAL;
53 	ev.event_type =  RTE_EVENT_TYPE_CPU;
54 	ev.sub_event_type = 0; /* stage 0 */
55 
56 	while (count < nb_pkts && t->done == false) {
57 		if (rte_mempool_get(pool, (void **)&m) < 0)
58 			continue;
59 
60 		ev.flow_id = flow_counter++ % nb_flows;
61 		ev.event_ptr = m;
62 		m->timestamp = rte_get_timer_cycles();
63 		while (rte_event_enqueue_burst(dev_id, port, &ev, 1) != 1) {
64 			if (t->done)
65 				break;
66 			rte_pause();
67 			m->timestamp = rte_get_timer_cycles();
68 		}
69 		count++;
70 	}
71 
72 	return 0;
73 }
74 
75 static inline int
76 perf_event_timer_producer(void *arg)
77 {
78 	struct prod_data *p  = arg;
79 	struct test_perf *t = p->t;
80 	struct evt_options *opt = t->opt;
81 	uint32_t flow_counter = 0;
82 	uint64_t count = 0;
83 	uint64_t arm_latency = 0;
84 	const uint8_t nb_timer_adptrs = opt->nb_timer_adptrs;
85 	const uint32_t nb_flows = t->nb_flows;
86 	const uint64_t nb_timers = opt->nb_timers;
87 	struct rte_mempool *pool = t->pool;
88 	struct perf_elt *m;
89 	struct rte_event_timer_adapter **adptr = t->timer_adptr;
90 	struct rte_event_timer tim;
91 	uint64_t timeout_ticks = opt->expiry_nsec / opt->timer_tick_nsec;
92 
93 	memset(&tim, 0, sizeof(struct rte_event_timer));
94 	timeout_ticks = opt->optm_timer_tick_nsec ?
95 			(timeout_ticks * opt->timer_tick_nsec)
96 			/ opt->optm_timer_tick_nsec : timeout_ticks;
97 	timeout_ticks += timeout_ticks ? 0 : 1;
98 	tim.ev.event_type =  RTE_EVENT_TYPE_TIMER;
99 	tim.ev.op = RTE_EVENT_OP_NEW;
100 	tim.ev.sched_type = t->opt->sched_type_list[0];
101 	tim.ev.queue_id = p->queue_id;
102 	tim.ev.priority = RTE_EVENT_DEV_PRIORITY_NORMAL;
103 	tim.state = RTE_EVENT_TIMER_NOT_ARMED;
104 	tim.timeout_ticks = timeout_ticks;
105 
106 	if (opt->verbose_level > 1)
107 		printf("%s(): lcore %d\n", __func__, rte_lcore_id());
108 
109 	while (count < nb_timers && t->done == false) {
110 		if (rte_mempool_get(pool, (void **)&m) < 0)
111 			continue;
112 
113 		m->tim = tim;
114 		m->tim.ev.flow_id = flow_counter++ % nb_flows;
115 		m->tim.ev.event_ptr = m;
116 		m->timestamp = rte_get_timer_cycles();
117 		while (rte_event_timer_arm_burst(
118 				adptr[flow_counter % nb_timer_adptrs],
119 				(struct rte_event_timer **)&m, 1) != 1) {
120 			if (t->done)
121 				break;
122 			rte_pause();
123 			m->timestamp = rte_get_timer_cycles();
124 		}
125 		arm_latency += rte_get_timer_cycles() - m->timestamp;
126 		count++;
127 	}
128 	fflush(stdout);
129 	rte_delay_ms(1000);
130 	printf("%s(): lcore %d Average event timer arm latency = %.3f us\n",
131 			__func__, rte_lcore_id(), (float)(arm_latency / count) /
132 			(rte_get_timer_hz() / 1000000));
133 	return 0;
134 }
135 
136 static inline int
137 perf_event_timer_producer_burst(void *arg)
138 {
139 	int i;
140 	struct prod_data *p  = arg;
141 	struct test_perf *t = p->t;
142 	struct evt_options *opt = t->opt;
143 	uint32_t flow_counter = 0;
144 	uint64_t count = 0;
145 	uint64_t arm_latency = 0;
146 	const uint8_t nb_timer_adptrs = opt->nb_timer_adptrs;
147 	const uint32_t nb_flows = t->nb_flows;
148 	const uint64_t nb_timers = opt->nb_timers;
149 	struct rte_mempool *pool = t->pool;
150 	struct perf_elt *m[BURST_SIZE + 1] = {NULL};
151 	struct rte_event_timer_adapter **adptr = t->timer_adptr;
152 	struct rte_event_timer tim;
153 	uint64_t timeout_ticks = opt->expiry_nsec / opt->timer_tick_nsec;
154 
155 	memset(&tim, 0, sizeof(struct rte_event_timer));
156 	timeout_ticks = opt->optm_timer_tick_nsec ?
157 			(timeout_ticks * opt->timer_tick_nsec)
158 			/ opt->optm_timer_tick_nsec : timeout_ticks;
159 	timeout_ticks += timeout_ticks ? 0 : 1;
160 	tim.ev.event_type =  RTE_EVENT_TYPE_TIMER;
161 	tim.ev.op = RTE_EVENT_OP_NEW;
162 	tim.ev.sched_type = t->opt->sched_type_list[0];
163 	tim.ev.queue_id = p->queue_id;
164 	tim.ev.priority = RTE_EVENT_DEV_PRIORITY_NORMAL;
165 	tim.state = RTE_EVENT_TIMER_NOT_ARMED;
166 	tim.timeout_ticks = timeout_ticks;
167 
168 	if (opt->verbose_level > 1)
169 		printf("%s(): lcore %d\n", __func__, rte_lcore_id());
170 
171 	while (count < nb_timers && t->done == false) {
172 		if (rte_mempool_get_bulk(pool, (void **)m, BURST_SIZE) < 0)
173 			continue;
174 		for (i = 0; i < BURST_SIZE; i++) {
175 			rte_prefetch0(m[i + 1]);
176 			m[i]->tim = tim;
177 			m[i]->tim.ev.flow_id = flow_counter++ % nb_flows;
178 			m[i]->tim.ev.event_ptr = m[i];
179 			m[i]->timestamp = rte_get_timer_cycles();
180 		}
181 		rte_event_timer_arm_tmo_tick_burst(
182 				adptr[flow_counter % nb_timer_adptrs],
183 				(struct rte_event_timer **)m,
184 				tim.timeout_ticks,
185 				BURST_SIZE);
186 		arm_latency += rte_get_timer_cycles() - m[i - 1]->timestamp;
187 		count += BURST_SIZE;
188 	}
189 	fflush(stdout);
190 	rte_delay_ms(1000);
191 	printf("%s(): lcore %d Average event timer arm latency = %.3f us\n",
192 			__func__, rte_lcore_id(), (float)(arm_latency / count) /
193 			(rte_get_timer_hz() / 1000000));
194 	return 0;
195 }
196 
197 static int
198 perf_producer_wrapper(void *arg)
199 {
200 	struct prod_data *p  = arg;
201 	struct test_perf *t = p->t;
202 	/* Launch the producer function only in case of synthetic producer. */
203 	if (t->opt->prod_type == EVT_PROD_TYPE_SYNT)
204 		return perf_producer(arg);
205 	else if (t->opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR &&
206 			!t->opt->timdev_use_burst)
207 		return perf_event_timer_producer(arg);
208 	else if (t->opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR &&
209 			t->opt->timdev_use_burst)
210 		return perf_event_timer_producer_burst(arg);
211 	return 0;
212 }
213 
214 static inline uint64_t
215 processed_pkts(struct test_perf *t)
216 {
217 	uint8_t i;
218 	uint64_t total = 0;
219 
220 	rte_smp_rmb();
221 	for (i = 0; i < t->nb_workers; i++)
222 		total += t->worker[i].processed_pkts;
223 
224 	return total;
225 }
226 
227 static inline uint64_t
228 total_latency(struct test_perf *t)
229 {
230 	uint8_t i;
231 	uint64_t total = 0;
232 
233 	rte_smp_rmb();
234 	for (i = 0; i < t->nb_workers; i++)
235 		total += t->worker[i].latency;
236 
237 	return total;
238 }
239 
240 
241 int
242 perf_launch_lcores(struct evt_test *test, struct evt_options *opt,
243 		int (*worker)(void *))
244 {
245 	int ret, lcore_id;
246 	struct test_perf *t = evt_test_priv(test);
247 
248 	int port_idx = 0;
249 	/* launch workers */
250 	RTE_LCORE_FOREACH_SLAVE(lcore_id) {
251 		if (!(opt->wlcores[lcore_id]))
252 			continue;
253 
254 		ret = rte_eal_remote_launch(worker,
255 				 &t->worker[port_idx], lcore_id);
256 		if (ret) {
257 			evt_err("failed to launch worker %d", lcore_id);
258 			return ret;
259 		}
260 		port_idx++;
261 	}
262 
263 	/* launch producers */
264 	RTE_LCORE_FOREACH_SLAVE(lcore_id) {
265 		if (!(opt->plcores[lcore_id]))
266 			continue;
267 
268 		ret = rte_eal_remote_launch(perf_producer_wrapper,
269 				&t->prod[port_idx], lcore_id);
270 		if (ret) {
271 			evt_err("failed to launch perf_producer %d", lcore_id);
272 			return ret;
273 		}
274 		port_idx++;
275 	}
276 
277 	const uint64_t total_pkts = t->outstand_pkts;
278 
279 	uint64_t dead_lock_cycles = rte_get_timer_cycles();
280 	int64_t dead_lock_remaining  =  total_pkts;
281 	const uint64_t dead_lock_sample = rte_get_timer_hz() * 5;
282 
283 	uint64_t perf_cycles = rte_get_timer_cycles();
284 	int64_t perf_remaining  = total_pkts;
285 	const uint64_t perf_sample = rte_get_timer_hz();
286 
287 	static float total_mpps;
288 	static uint64_t samples;
289 
290 	const uint64_t freq_mhz = rte_get_timer_hz() / 1000000;
291 	int64_t remaining = t->outstand_pkts - processed_pkts(t);
292 
293 	while (t->done == false) {
294 		const uint64_t new_cycles = rte_get_timer_cycles();
295 
296 		if ((new_cycles - perf_cycles) > perf_sample) {
297 			const uint64_t latency = total_latency(t);
298 			const uint64_t pkts = processed_pkts(t);
299 
300 			remaining = t->outstand_pkts - pkts;
301 			float mpps = (float)(perf_remaining-remaining)/1000000;
302 
303 			perf_remaining = remaining;
304 			perf_cycles = new_cycles;
305 			total_mpps += mpps;
306 			++samples;
307 			if (opt->fwd_latency && pkts > 0) {
308 				printf(CLGRN"\r%.3f mpps avg %.3f mpps [avg fwd latency %.3f us] "CLNRM,
309 					mpps, total_mpps/samples,
310 					(float)(latency/pkts)/freq_mhz);
311 			} else {
312 				printf(CLGRN"\r%.3f mpps avg %.3f mpps"CLNRM,
313 					mpps, total_mpps/samples);
314 			}
315 			fflush(stdout);
316 
317 			if (remaining <= 0) {
318 				t->result = EVT_TEST_SUCCESS;
319 				if (opt->prod_type == EVT_PROD_TYPE_SYNT ||
320 					opt->prod_type ==
321 					EVT_PROD_TYPE_EVENT_TIMER_ADPTR) {
322 					t->done = true;
323 					rte_smp_wmb();
324 					break;
325 				}
326 			}
327 		}
328 
329 		if (new_cycles - dead_lock_cycles > dead_lock_sample &&
330 				opt->prod_type == EVT_PROD_TYPE_SYNT) {
331 			remaining = t->outstand_pkts - processed_pkts(t);
332 			if (dead_lock_remaining == remaining) {
333 				rte_event_dev_dump(opt->dev_id, stdout);
334 				evt_err("No schedules for seconds, deadlock");
335 				t->done = true;
336 				rte_smp_wmb();
337 				break;
338 			}
339 			dead_lock_remaining = remaining;
340 			dead_lock_cycles = new_cycles;
341 		}
342 	}
343 	printf("\n");
344 	return 0;
345 }
346 
347 static int
348 perf_event_rx_adapter_setup(struct evt_options *opt, uint8_t stride,
349 		struct rte_event_port_conf prod_conf)
350 {
351 	int ret = 0;
352 	uint16_t prod;
353 	struct rte_event_eth_rx_adapter_queue_conf queue_conf;
354 
355 	memset(&queue_conf, 0,
356 			sizeof(struct rte_event_eth_rx_adapter_queue_conf));
357 	queue_conf.ev.sched_type = opt->sched_type_list[0];
358 	RTE_ETH_FOREACH_DEV(prod) {
359 		uint32_t cap;
360 
361 		ret = rte_event_eth_rx_adapter_caps_get(opt->dev_id,
362 				prod, &cap);
363 		if (ret) {
364 			evt_err("failed to get event rx adapter[%d]"
365 					" capabilities",
366 					opt->dev_id);
367 			return ret;
368 		}
369 		queue_conf.ev.queue_id = prod * stride;
370 		ret = rte_event_eth_rx_adapter_create(prod, opt->dev_id,
371 				&prod_conf);
372 		if (ret) {
373 			evt_err("failed to create rx adapter[%d]", prod);
374 			return ret;
375 		}
376 		ret = rte_event_eth_rx_adapter_queue_add(prod, prod, -1,
377 				&queue_conf);
378 		if (ret) {
379 			evt_err("failed to add rx queues to adapter[%d]", prod);
380 			return ret;
381 		}
382 
383 		if (!(cap & RTE_EVENT_ETH_RX_ADAPTER_CAP_INTERNAL_PORT)) {
384 			uint32_t service_id;
385 
386 			rte_event_eth_rx_adapter_service_id_get(prod,
387 					&service_id);
388 			ret = evt_service_setup(service_id);
389 			if (ret) {
390 				evt_err("Failed to setup service core"
391 						" for Rx adapter\n");
392 				return ret;
393 			}
394 		}
395 
396 		ret = rte_eth_dev_start(prod);
397 		if (ret) {
398 			evt_err("Ethernet dev [%d] failed to start."
399 					" Using synthetic producer", prod);
400 			return ret;
401 		}
402 
403 		ret = rte_event_eth_rx_adapter_start(prod);
404 		if (ret) {
405 			evt_err("Rx adapter[%d] start failed", prod);
406 			return ret;
407 		}
408 		printf("%s: Port[%d] using Rx adapter[%d] started\n", __func__,
409 				prod, prod);
410 	}
411 
412 	return ret;
413 }
414 
415 static int
416 perf_event_timer_adapter_setup(struct test_perf *t)
417 {
418 	int i;
419 	int ret;
420 	struct rte_event_timer_adapter_info adapter_info;
421 	struct rte_event_timer_adapter *wl;
422 	uint8_t nb_producers = evt_nr_active_lcores(t->opt->plcores);
423 	uint8_t flags = RTE_EVENT_TIMER_ADAPTER_F_ADJUST_RES;
424 
425 	if (nb_producers == 1)
426 		flags |= RTE_EVENT_TIMER_ADAPTER_F_SP_PUT;
427 
428 	for (i = 0; i < t->opt->nb_timer_adptrs; i++) {
429 		struct rte_event_timer_adapter_conf config = {
430 			.event_dev_id = t->opt->dev_id,
431 			.timer_adapter_id = i,
432 			.timer_tick_ns = t->opt->timer_tick_nsec,
433 			.max_tmo_ns = t->opt->max_tmo_nsec,
434 			.nb_timers = 2 * 1024 * 1024,
435 			.flags = flags,
436 		};
437 
438 		wl = rte_event_timer_adapter_create(&config);
439 		if (wl == NULL) {
440 			evt_err("failed to create event timer ring %d", i);
441 			return rte_errno;
442 		}
443 
444 		memset(&adapter_info, 0,
445 				sizeof(struct rte_event_timer_adapter_info));
446 		rte_event_timer_adapter_get_info(wl, &adapter_info);
447 		t->opt->optm_timer_tick_nsec = adapter_info.min_resolution_ns;
448 
449 		if (!(adapter_info.caps &
450 				RTE_EVENT_TIMER_ADAPTER_CAP_INTERNAL_PORT)) {
451 			uint32_t service_id;
452 
453 			rte_event_timer_adapter_service_id_get(wl,
454 					&service_id);
455 			ret = evt_service_setup(service_id);
456 			if (ret) {
457 				evt_err("Failed to setup service core"
458 						" for timer adapter\n");
459 				return ret;
460 			}
461 			rte_service_runstate_set(service_id, 1);
462 		}
463 
464 		ret = rte_event_timer_adapter_start(wl);
465 		if (ret) {
466 			evt_err("failed to Start event timer adapter %d", i);
467 			return ret;
468 		}
469 		t->timer_adptr[i] = wl;
470 	}
471 	return 0;
472 }
473 
474 int
475 perf_event_dev_port_setup(struct evt_test *test, struct evt_options *opt,
476 				uint8_t stride, uint8_t nb_queues,
477 				const struct rte_event_port_conf *port_conf)
478 {
479 	struct test_perf *t = evt_test_priv(test);
480 	uint16_t port, prod;
481 	int ret = -1;
482 
483 	/* setup one port per worker, linking to all queues */
484 	for (port = 0; port < evt_nr_active_lcores(opt->wlcores);
485 				port++) {
486 		struct worker_data *w = &t->worker[port];
487 
488 		w->dev_id = opt->dev_id;
489 		w->port_id = port;
490 		w->t = t;
491 		w->processed_pkts = 0;
492 		w->latency = 0;
493 
494 		ret = rte_event_port_setup(opt->dev_id, port, port_conf);
495 		if (ret) {
496 			evt_err("failed to setup port %d", port);
497 			return ret;
498 		}
499 
500 		ret = rte_event_port_link(opt->dev_id, port, NULL, NULL, 0);
501 		if (ret != nb_queues) {
502 			evt_err("failed to link all queues to port %d", port);
503 			return -EINVAL;
504 		}
505 	}
506 
507 	/* port for producers, no links */
508 	if (opt->prod_type == EVT_PROD_TYPE_ETH_RX_ADPTR) {
509 		for ( ; port < perf_nb_event_ports(opt); port++) {
510 			struct prod_data *p = &t->prod[port];
511 			p->t = t;
512 		}
513 
514 		ret = perf_event_rx_adapter_setup(opt, stride, *port_conf);
515 		if (ret)
516 			return ret;
517 	} else if (opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR) {
518 		prod = 0;
519 		for ( ; port < perf_nb_event_ports(opt); port++) {
520 			struct prod_data *p = &t->prod[port];
521 			p->queue_id = prod * stride;
522 			p->t = t;
523 			prod++;
524 		}
525 
526 		ret = perf_event_timer_adapter_setup(t);
527 		if (ret)
528 			return ret;
529 	} else {
530 		prod = 0;
531 		for ( ; port < perf_nb_event_ports(opt); port++) {
532 			struct prod_data *p = &t->prod[port];
533 
534 			p->dev_id = opt->dev_id;
535 			p->port_id = port;
536 			p->queue_id = prod * stride;
537 			p->t = t;
538 
539 			ret = rte_event_port_setup(opt->dev_id, port,
540 					port_conf);
541 			if (ret) {
542 				evt_err("failed to setup port %d", port);
543 				return ret;
544 			}
545 			prod++;
546 		}
547 	}
548 
549 	return ret;
550 }
551 
552 int
553 perf_opt_check(struct evt_options *opt, uint64_t nb_queues)
554 {
555 	unsigned int lcores;
556 
557 	/* N producer + N worker + 1 master when producer cores are used
558 	 * Else N worker + 1 master when Rx adapter is used
559 	 */
560 	lcores = opt->prod_type == EVT_PROD_TYPE_SYNT ? 3 : 2;
561 
562 	if (rte_lcore_count() < lcores) {
563 		evt_err("test need minimum %d lcores", lcores);
564 		return -1;
565 	}
566 
567 	/* Validate worker lcores */
568 	if (evt_lcores_has_overlap(opt->wlcores, rte_get_master_lcore())) {
569 		evt_err("worker lcores overlaps with master lcore");
570 		return -1;
571 	}
572 	if (evt_lcores_has_overlap_multi(opt->wlcores, opt->plcores)) {
573 		evt_err("worker lcores overlaps producer lcores");
574 		return -1;
575 	}
576 	if (evt_has_disabled_lcore(opt->wlcores)) {
577 		evt_err("one or more workers lcores are not enabled");
578 		return -1;
579 	}
580 	if (!evt_has_active_lcore(opt->wlcores)) {
581 		evt_err("minimum one worker is required");
582 		return -1;
583 	}
584 
585 	if (opt->prod_type == EVT_PROD_TYPE_SYNT) {
586 		/* Validate producer lcores */
587 		if (evt_lcores_has_overlap(opt->plcores,
588 					rte_get_master_lcore())) {
589 			evt_err("producer lcores overlaps with master lcore");
590 			return -1;
591 		}
592 		if (evt_has_disabled_lcore(opt->plcores)) {
593 			evt_err("one or more producer lcores are not enabled");
594 			return -1;
595 		}
596 		if (!evt_has_active_lcore(opt->plcores)) {
597 			evt_err("minimum one producer is required");
598 			return -1;
599 		}
600 	}
601 
602 	if (evt_has_invalid_stage(opt))
603 		return -1;
604 
605 	if (evt_has_invalid_sched_type(opt))
606 		return -1;
607 
608 	if (nb_queues > EVT_MAX_QUEUES) {
609 		evt_err("number of queues exceeds %d", EVT_MAX_QUEUES);
610 		return -1;
611 	}
612 	if (perf_nb_event_ports(opt) > EVT_MAX_PORTS) {
613 		evt_err("number of ports exceeds %d", EVT_MAX_PORTS);
614 		return -1;
615 	}
616 
617 	/* Fixups */
618 	if ((opt->nb_stages == 1 &&
619 			opt->prod_type != EVT_PROD_TYPE_EVENT_TIMER_ADPTR) &&
620 			opt->fwd_latency) {
621 		evt_info("fwd_latency is valid when nb_stages > 1, disabling");
622 		opt->fwd_latency = 0;
623 	}
624 
625 	if (opt->fwd_latency && !opt->q_priority) {
626 		evt_info("enabled queue priority for latency measurement");
627 		opt->q_priority = 1;
628 	}
629 	if (opt->nb_pkts == 0)
630 		opt->nb_pkts = INT64_MAX/evt_nr_active_lcores(opt->plcores);
631 
632 	return 0;
633 }
634 
635 void
636 perf_opt_dump(struct evt_options *opt, uint8_t nb_queues)
637 {
638 	evt_dump("nb_prod_lcores", "%d", evt_nr_active_lcores(opt->plcores));
639 	evt_dump_producer_lcores(opt);
640 	evt_dump("nb_worker_lcores", "%d", evt_nr_active_lcores(opt->wlcores));
641 	evt_dump_worker_lcores(opt);
642 	evt_dump_nb_stages(opt);
643 	evt_dump("nb_evdev_ports", "%d", perf_nb_event_ports(opt));
644 	evt_dump("nb_evdev_queues", "%d", nb_queues);
645 	evt_dump_queue_priority(opt);
646 	evt_dump_sched_type_list(opt);
647 	evt_dump_producer_type(opt);
648 }
649 
650 void
651 perf_eventdev_destroy(struct evt_test *test, struct evt_options *opt)
652 {
653 	int i;
654 	struct test_perf *t = evt_test_priv(test);
655 
656 	if (opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR) {
657 		for (i = 0; i < opt->nb_timer_adptrs; i++)
658 			rte_event_timer_adapter_stop(t->timer_adptr[i]);
659 	}
660 	rte_event_dev_stop(opt->dev_id);
661 	rte_event_dev_close(opt->dev_id);
662 }
663 
664 static inline void
665 perf_elt_init(struct rte_mempool *mp, void *arg __rte_unused,
666 	    void *obj, unsigned i __rte_unused)
667 {
668 	memset(obj, 0, mp->elt_size);
669 }
670 
671 #define NB_RX_DESC			128
672 #define NB_TX_DESC			512
673 int
674 perf_ethdev_setup(struct evt_test *test, struct evt_options *opt)
675 {
676 	uint16_t i;
677 	struct test_perf *t = evt_test_priv(test);
678 	struct rte_eth_conf port_conf = {
679 		.rxmode = {
680 			.mq_mode = ETH_MQ_RX_RSS,
681 			.max_rx_pkt_len = ETHER_MAX_LEN,
682 			.split_hdr_size = 0,
683 			.offloads = DEV_RX_OFFLOAD_CRC_STRIP,
684 		},
685 		.rx_adv_conf = {
686 			.rss_conf = {
687 				.rss_key = NULL,
688 				.rss_hf = ETH_RSS_IP,
689 			},
690 		},
691 	};
692 
693 	if (opt->prod_type == EVT_PROD_TYPE_SYNT ||
694 			opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR)
695 		return 0;
696 
697 	if (!rte_eth_dev_count_avail()) {
698 		evt_err("No ethernet ports found.");
699 		return -ENODEV;
700 	}
701 
702 	RTE_ETH_FOREACH_DEV(i) {
703 		struct rte_eth_dev_info dev_info;
704 		struct rte_eth_conf local_port_conf = port_conf;
705 
706 		rte_eth_dev_info_get(i, &dev_info);
707 
708 		local_port_conf.rx_adv_conf.rss_conf.rss_hf &=
709 			dev_info.flow_type_rss_offloads;
710 		if (local_port_conf.rx_adv_conf.rss_conf.rss_hf !=
711 				port_conf.rx_adv_conf.rss_conf.rss_hf) {
712 			evt_info("Port %u modified RSS hash function based on hardware support,"
713 				"requested:%#"PRIx64" configured:%#"PRIx64"\n",
714 				i,
715 				port_conf.rx_adv_conf.rss_conf.rss_hf,
716 				local_port_conf.rx_adv_conf.rss_conf.rss_hf);
717 		}
718 
719 		if (rte_eth_dev_configure(i, 1, 1, &local_port_conf) < 0) {
720 			evt_err("Failed to configure eth port [%d]", i);
721 			return -EINVAL;
722 		}
723 
724 		if (rte_eth_rx_queue_setup(i, 0, NB_RX_DESC,
725 				rte_socket_id(), NULL, t->pool) < 0) {
726 			evt_err("Failed to setup eth port [%d] rx_queue: %d.",
727 					i, 0);
728 			return -EINVAL;
729 		}
730 
731 		if (rte_eth_tx_queue_setup(i, 0, NB_TX_DESC,
732 					rte_socket_id(), NULL) < 0) {
733 			evt_err("Failed to setup eth port [%d] tx_queue: %d.",
734 					i, 0);
735 			return -EINVAL;
736 		}
737 
738 		rte_eth_promiscuous_enable(i);
739 	}
740 
741 	return 0;
742 }
743 
744 void perf_ethdev_destroy(struct evt_test *test, struct evt_options *opt)
745 {
746 	uint16_t i;
747 	RTE_SET_USED(test);
748 
749 	if (opt->prod_type == EVT_PROD_TYPE_ETH_RX_ADPTR) {
750 		RTE_ETH_FOREACH_DEV(i) {
751 			rte_event_eth_rx_adapter_stop(i);
752 			rte_eth_dev_stop(i);
753 			rte_eth_dev_close(i);
754 		}
755 	}
756 }
757 
758 int
759 perf_mempool_setup(struct evt_test *test, struct evt_options *opt)
760 {
761 	struct test_perf *t = evt_test_priv(test);
762 
763 	if (opt->prod_type == EVT_PROD_TYPE_SYNT ||
764 			opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR) {
765 		t->pool = rte_mempool_create(test->name, /* mempool name */
766 				opt->pool_sz, /* number of elements*/
767 				sizeof(struct perf_elt), /* element size*/
768 				512, /* cache size*/
769 				0, NULL, NULL,
770 				perf_elt_init, /* obj constructor */
771 				NULL, opt->socket_id, 0); /* flags */
772 	} else {
773 		t->pool = rte_pktmbuf_pool_create(test->name, /* mempool name */
774 				opt->pool_sz, /* number of elements*/
775 				512, /* cache size*/
776 				0,
777 				RTE_MBUF_DEFAULT_BUF_SIZE,
778 				opt->socket_id); /* flags */
779 
780 	}
781 
782 	if (t->pool == NULL) {
783 		evt_err("failed to create mempool");
784 		return -ENOMEM;
785 	}
786 
787 	return 0;
788 }
789 
790 void
791 perf_mempool_destroy(struct evt_test *test, struct evt_options *opt)
792 {
793 	RTE_SET_USED(opt);
794 	struct test_perf *t = evt_test_priv(test);
795 
796 	rte_mempool_free(t->pool);
797 }
798 
799 int
800 perf_test_setup(struct evt_test *test, struct evt_options *opt)
801 {
802 	void *test_perf;
803 
804 	test_perf = rte_zmalloc_socket(test->name, sizeof(struct test_perf),
805 				RTE_CACHE_LINE_SIZE, opt->socket_id);
806 	if (test_perf  == NULL) {
807 		evt_err("failed to allocate test_perf memory");
808 		goto nomem;
809 	}
810 	test->test_priv = test_perf;
811 
812 	struct test_perf *t = evt_test_priv(test);
813 
814 	if (opt->prod_type == EVT_PROD_TYPE_EVENT_TIMER_ADPTR) {
815 		t->outstand_pkts = opt->nb_timers *
816 			evt_nr_active_lcores(opt->plcores);
817 		t->nb_pkts = opt->nb_timers;
818 	} else {
819 		t->outstand_pkts = opt->nb_pkts *
820 			evt_nr_active_lcores(opt->plcores);
821 		t->nb_pkts = opt->nb_pkts;
822 	}
823 
824 	t->nb_workers = evt_nr_active_lcores(opt->wlcores);
825 	t->done = false;
826 	t->nb_flows = opt->nb_flows;
827 	t->result = EVT_TEST_FAILED;
828 	t->opt = opt;
829 	memcpy(t->sched_type_list, opt->sched_type_list,
830 			sizeof(opt->sched_type_list));
831 	return 0;
832 nomem:
833 	return -ENOMEM;
834 }
835 
836 void
837 perf_test_destroy(struct evt_test *test, struct evt_options *opt)
838 {
839 	RTE_SET_USED(opt);
840 
841 	rte_free(test->test_priv);
842 }
843