xref: /dpdk/examples/eventdev_pipeline/pipeline_worker_generic.c (revision daa02b5cddbb8e11b31d41e2bf7bb1ae64dcae2f)
1 /*
2  * SPDX-License-Identifier: BSD-3-Clause
3  * Copyright 2016 Intel Corporation.
4  * Copyright 2017 Cavium, Inc.
5  */
6 
7 #include "pipeline_common.h"
8 
9 static __rte_always_inline int
10 worker_generic(void *arg)
11 {
12 	struct rte_event ev;
13 
14 	struct worker_data *data = (struct worker_data *)arg;
15 	uint8_t dev_id = data->dev_id;
16 	uint8_t port_id = data->port_id;
17 	size_t sent = 0, received = 0;
18 	unsigned int lcore_id = rte_lcore_id();
19 
20 	while (!fdata->done) {
21 
22 		if (fdata->cap.scheduler)
23 			fdata->cap.scheduler(lcore_id);
24 
25 		if (!fdata->worker_core[lcore_id]) {
26 			rte_pause();
27 			continue;
28 		}
29 
30 		const uint16_t nb_rx = rte_event_dequeue_burst(dev_id, port_id,
31 				&ev, 1, 0);
32 
33 		if (nb_rx == 0) {
34 			rte_pause();
35 			continue;
36 		}
37 		received++;
38 
39 		/* The first worker stage does classification */
40 		if (ev.queue_id == cdata.qid[0])
41 			ev.flow_id = ev.mbuf->hash.rss
42 						% cdata.num_fids;
43 
44 		ev.queue_id = cdata.next_qid[ev.queue_id];
45 		ev.op = RTE_EVENT_OP_FORWARD;
46 		ev.sched_type = cdata.queue_type;
47 
48 		work();
49 
50 		while (rte_event_enqueue_burst(dev_id, port_id, &ev, 1) != 1)
51 			rte_pause();
52 		sent++;
53 	}
54 
55 	if (!cdata.quiet)
56 		printf("  worker %u thread done. RX=%zu TX=%zu\n",
57 				rte_lcore_id(), received, sent);
58 
59 	return 0;
60 }
61 
62 static int
63 worker_generic_burst(void *arg)
64 {
65 	struct rte_event events[BATCH_SIZE];
66 
67 	struct worker_data *data = (struct worker_data *)arg;
68 	uint8_t dev_id = data->dev_id;
69 	uint8_t port_id = data->port_id;
70 	size_t sent = 0, received = 0;
71 	unsigned int lcore_id = rte_lcore_id();
72 
73 	while (!fdata->done) {
74 		uint16_t i;
75 
76 		if (fdata->cap.scheduler)
77 			fdata->cap.scheduler(lcore_id);
78 
79 		if (!fdata->worker_core[lcore_id]) {
80 			rte_pause();
81 			continue;
82 		}
83 
84 		const uint16_t nb_rx = rte_event_dequeue_burst(dev_id, port_id,
85 				events, RTE_DIM(events), 0);
86 
87 		if (nb_rx == 0) {
88 			rte_pause();
89 			continue;
90 		}
91 		received += nb_rx;
92 
93 		for (i = 0; i < nb_rx; i++) {
94 
95 			/* The first worker stage does classification */
96 			if (events[i].queue_id == cdata.qid[0])
97 				events[i].flow_id = events[i].mbuf->hash.rss
98 							% cdata.num_fids;
99 
100 			events[i].queue_id = cdata.next_qid[events[i].queue_id];
101 			events[i].op = RTE_EVENT_OP_FORWARD;
102 			events[i].sched_type = cdata.queue_type;
103 
104 			work();
105 		}
106 		uint16_t nb_tx = rte_event_enqueue_burst(dev_id, port_id,
107 				events, nb_rx);
108 		while (nb_tx < nb_rx && !fdata->done)
109 			nb_tx += rte_event_enqueue_burst(dev_id, port_id,
110 							events + nb_tx,
111 							nb_rx - nb_tx);
112 		sent += nb_tx;
113 	}
114 
115 	if (!cdata.quiet)
116 		printf("  worker %u thread done. RX=%zu TX=%zu\n",
117 				rte_lcore_id(), received, sent);
118 
119 	return 0;
120 }
121 
122 static int
123 setup_eventdev_generic(struct worker_data *worker_data)
124 {
125 	const uint8_t dev_id = 0;
126 	/* +1 stages is for a SINGLE_LINK TX stage */
127 	const uint8_t nb_queues = cdata.num_stages + 1;
128 	const uint8_t nb_ports = cdata.num_workers;
129 	struct rte_event_dev_config config = {
130 			.nb_event_queues = nb_queues,
131 			.nb_event_ports = nb_ports,
132 			.nb_single_link_event_port_queues = 1,
133 			.nb_events_limit  = 4096,
134 			.nb_event_queue_flows = 1024,
135 			.nb_event_port_dequeue_depth = 128,
136 			.nb_event_port_enqueue_depth = 128,
137 	};
138 	struct rte_event_port_conf wkr_p_conf = {
139 			.dequeue_depth = cdata.worker_cq_depth,
140 			.enqueue_depth = 64,
141 			.new_event_threshold = 4096,
142 			.event_port_cfg = RTE_EVENT_PORT_CFG_HINT_WORKER,
143 	};
144 	struct rte_event_queue_conf wkr_q_conf = {
145 			.schedule_type = cdata.queue_type,
146 			.priority = RTE_EVENT_DEV_PRIORITY_NORMAL,
147 			.nb_atomic_flows = 1024,
148 			.nb_atomic_order_sequences = 1024,
149 	};
150 	struct rte_event_queue_conf tx_q_conf = {
151 			.priority = RTE_EVENT_DEV_PRIORITY_HIGHEST,
152 			.event_queue_cfg = RTE_EVENT_QUEUE_CFG_SINGLE_LINK,
153 	};
154 
155 	struct port_link worker_queues[MAX_NUM_STAGES];
156 	uint8_t disable_implicit_release;
157 	unsigned int i;
158 
159 	int ret, ndev = rte_event_dev_count();
160 	if (ndev < 1) {
161 		printf("%d: No Eventdev Devices Found\n", __LINE__);
162 		return -1;
163 	}
164 
165 	struct rte_event_dev_info dev_info;
166 	ret = rte_event_dev_info_get(dev_id, &dev_info);
167 	printf("\tEventdev %d: %s\n", dev_id, dev_info.driver_name);
168 
169 	disable_implicit_release = (dev_info.event_dev_cap &
170 			RTE_EVENT_DEV_CAP_IMPLICIT_RELEASE_DISABLE);
171 
172 	wkr_p_conf.event_port_cfg = disable_implicit_release ?
173 		RTE_EVENT_PORT_CFG_DISABLE_IMPL_REL : 0;
174 
175 	if (dev_info.max_num_events < config.nb_events_limit)
176 		config.nb_events_limit = dev_info.max_num_events;
177 	if (dev_info.max_event_port_dequeue_depth <
178 			config.nb_event_port_dequeue_depth)
179 		config.nb_event_port_dequeue_depth =
180 				dev_info.max_event_port_dequeue_depth;
181 	if (dev_info.max_event_port_enqueue_depth <
182 			config.nb_event_port_enqueue_depth)
183 		config.nb_event_port_enqueue_depth =
184 				dev_info.max_event_port_enqueue_depth;
185 
186 	ret = rte_event_dev_configure(dev_id, &config);
187 	if (ret < 0) {
188 		printf("%d: Error configuring device\n", __LINE__);
189 		return -1;
190 	}
191 
192 	/* Q creation - one load balanced per pipeline stage*/
193 	printf("  Stages:\n");
194 	for (i = 0; i < cdata.num_stages; i++) {
195 		if (rte_event_queue_setup(dev_id, i, &wkr_q_conf) < 0) {
196 			printf("%d: error creating qid %d\n", __LINE__, i);
197 			return -1;
198 		}
199 		cdata.qid[i] = i;
200 		cdata.next_qid[i] = i+1;
201 		worker_queues[i].queue_id = i;
202 		if (cdata.enable_queue_priorities) {
203 			/* calculate priority stepping for each stage, leaving
204 			 * headroom of 1 for the SINGLE_LINK TX below
205 			 */
206 			const uint32_t prio_delta =
207 				(RTE_EVENT_DEV_PRIORITY_LOWEST-1) /  nb_queues;
208 
209 			/* higher priority for queues closer to tx */
210 			wkr_q_conf.priority =
211 				RTE_EVENT_DEV_PRIORITY_LOWEST - prio_delta * i;
212 		}
213 
214 		const char *type_str = "Atomic";
215 		switch (wkr_q_conf.schedule_type) {
216 		case RTE_SCHED_TYPE_ORDERED:
217 			type_str = "Ordered";
218 			break;
219 		case RTE_SCHED_TYPE_PARALLEL:
220 			type_str = "Parallel";
221 			break;
222 		}
223 		printf("\tStage %d, Type %s\tPriority = %d\n", i, type_str,
224 				wkr_q_conf.priority);
225 	}
226 	printf("\n");
227 
228 	/* final queue for sending to TX core */
229 	if (rte_event_queue_setup(dev_id, i, &tx_q_conf) < 0) {
230 		printf("%d: error creating qid %d\n", __LINE__, i);
231 		return -1;
232 	}
233 	cdata.tx_queue_id = i;
234 
235 	if (wkr_p_conf.new_event_threshold > config.nb_events_limit)
236 		wkr_p_conf.new_event_threshold = config.nb_events_limit;
237 	if (wkr_p_conf.dequeue_depth > config.nb_event_port_dequeue_depth)
238 		wkr_p_conf.dequeue_depth = config.nb_event_port_dequeue_depth;
239 	if (wkr_p_conf.enqueue_depth > config.nb_event_port_enqueue_depth)
240 		wkr_p_conf.enqueue_depth = config.nb_event_port_enqueue_depth;
241 
242 	/* set up one port per worker, linking to all stage queues */
243 	for (i = 0; i < cdata.num_workers; i++) {
244 		struct worker_data *w = &worker_data[i];
245 		w->dev_id = dev_id;
246 		if (rte_event_port_setup(dev_id, i, &wkr_p_conf) < 0) {
247 			printf("Error setting up port %d\n", i);
248 			return -1;
249 		}
250 
251 		uint32_t s;
252 		for (s = 0; s < cdata.num_stages; s++) {
253 			if (rte_event_port_link(dev_id, i,
254 						&worker_queues[s].queue_id,
255 						&worker_queues[s].priority,
256 						1) != 1) {
257 				printf("%d: error creating link for port %d\n",
258 						__LINE__, i);
259 				return -1;
260 			}
261 		}
262 		w->port_id = i;
263 	}
264 
265 	ret = rte_event_dev_service_id_get(dev_id,
266 				&fdata->evdev_service_id);
267 	if (ret != -ESRCH && ret != 0) {
268 		printf("Error getting the service ID for sw eventdev\n");
269 		return -1;
270 	}
271 	rte_service_runstate_set(fdata->evdev_service_id, 1);
272 	rte_service_set_runstate_mapped_check(fdata->evdev_service_id, 0);
273 
274 	return dev_id;
275 }
276 
277 /*
278  * Initializes a given port using global settings and with the RX buffers
279  * coming from the mbuf_pool passed as a parameter.
280  */
281 static inline int
282 port_init(uint8_t port, struct rte_mempool *mbuf_pool)
283 {
284 	struct rte_eth_rxconf rx_conf;
285 	static const struct rte_eth_conf port_conf_default = {
286 		.rxmode = {
287 			.mq_mode = RTE_ETH_MQ_RX_RSS,
288 		},
289 		.rx_adv_conf = {
290 			.rss_conf = {
291 				.rss_hf = RTE_ETH_RSS_IP |
292 					  RTE_ETH_RSS_TCP |
293 					  RTE_ETH_RSS_UDP,
294 			}
295 		}
296 	};
297 	const uint16_t rx_rings = 1, tx_rings = 1;
298 	const uint16_t rx_ring_size = 512, tx_ring_size = 512;
299 	struct rte_eth_conf port_conf = port_conf_default;
300 	int retval;
301 	uint16_t q;
302 	struct rte_eth_dev_info dev_info;
303 	struct rte_eth_txconf txconf;
304 
305 	if (!rte_eth_dev_is_valid_port(port))
306 		return -1;
307 
308 	retval = rte_eth_dev_info_get(port, &dev_info);
309 	if (retval != 0) {
310 		printf("Error during getting device (port %u) info: %s\n",
311 				port, strerror(-retval));
312 		return retval;
313 	}
314 
315 	if (dev_info.tx_offload_capa & RTE_ETH_TX_OFFLOAD_MBUF_FAST_FREE)
316 		port_conf.txmode.offloads |=
317 			RTE_ETH_TX_OFFLOAD_MBUF_FAST_FREE;
318 
319 	if (dev_info.rx_offload_capa & RTE_ETH_RX_OFFLOAD_RSS_HASH)
320 		port_conf.rxmode.offloads |= RTE_ETH_RX_OFFLOAD_RSS_HASH;
321 
322 	rx_conf = dev_info.default_rxconf;
323 	rx_conf.offloads = port_conf.rxmode.offloads;
324 
325 	port_conf.rx_adv_conf.rss_conf.rss_hf &=
326 		dev_info.flow_type_rss_offloads;
327 	if (port_conf.rx_adv_conf.rss_conf.rss_hf !=
328 			port_conf_default.rx_adv_conf.rss_conf.rss_hf) {
329 		printf("Port %u modified RSS hash function based on hardware support,"
330 			"requested:%#"PRIx64" configured:%#"PRIx64"\n",
331 			port,
332 			port_conf_default.rx_adv_conf.rss_conf.rss_hf,
333 			port_conf.rx_adv_conf.rss_conf.rss_hf);
334 	}
335 
336 	/* Configure the Ethernet device. */
337 	retval = rte_eth_dev_configure(port, rx_rings, tx_rings, &port_conf);
338 	if (retval != 0)
339 		return retval;
340 
341 	/* Allocate and set up 1 RX queue per Ethernet port. */
342 	for (q = 0; q < rx_rings; q++) {
343 		retval = rte_eth_rx_queue_setup(port, q, rx_ring_size,
344 				rte_eth_dev_socket_id(port), &rx_conf,
345 				mbuf_pool);
346 		if (retval < 0)
347 			return retval;
348 	}
349 
350 	txconf = dev_info.default_txconf;
351 	txconf.offloads = port_conf_default.txmode.offloads;
352 	/* Allocate and set up 1 TX queue per Ethernet port. */
353 	for (q = 0; q < tx_rings; q++) {
354 		retval = rte_eth_tx_queue_setup(port, q, tx_ring_size,
355 				rte_eth_dev_socket_id(port), &txconf);
356 		if (retval < 0)
357 			return retval;
358 	}
359 
360 	/* Display the port MAC address. */
361 	struct rte_ether_addr addr;
362 	retval = rte_eth_macaddr_get(port, &addr);
363 	if (retval != 0) {
364 		printf("Failed to get MAC address (port %u): %s\n",
365 				port, rte_strerror(-retval));
366 		return retval;
367 	}
368 
369 	printf("Port %u MAC: %02" PRIx8 " %02" PRIx8 " %02" PRIx8
370 			" %02" PRIx8 " %02" PRIx8 " %02" PRIx8 "\n",
371 			(unsigned int)port, RTE_ETHER_ADDR_BYTES(&addr));
372 
373 	/* Enable RX in promiscuous mode for the Ethernet device. */
374 	retval = rte_eth_promiscuous_enable(port);
375 	if (retval != 0)
376 		return retval;
377 
378 	return 0;
379 }
380 
381 static int
382 init_ports(uint16_t num_ports)
383 {
384 	uint16_t portid;
385 
386 	if (!cdata.num_mbuf)
387 		cdata.num_mbuf = 16384 * num_ports;
388 
389 	struct rte_mempool *mp = rte_pktmbuf_pool_create("packet_pool",
390 			/* mbufs */ cdata.num_mbuf,
391 			/* cache_size */ 512,
392 			/* priv_size*/ 0,
393 			/* data_room_size */ RTE_MBUF_DEFAULT_BUF_SIZE,
394 			rte_socket_id());
395 
396 	RTE_ETH_FOREACH_DEV(portid)
397 		if (port_init(portid, mp) != 0)
398 			rte_exit(EXIT_FAILURE, "Cannot init port %"PRIu16 "\n",
399 					portid);
400 
401 	return 0;
402 }
403 
404 static void
405 init_adapters(uint16_t nb_ports)
406 {
407 	int i;
408 	int ret;
409 	uint8_t tx_port_id = 0;
410 	uint8_t evdev_id = 0;
411 	struct rte_event_dev_info dev_info;
412 
413 	ret = rte_event_dev_info_get(evdev_id, &dev_info);
414 
415 	struct rte_event_port_conf adptr_p_conf = {
416 		.dequeue_depth = cdata.worker_cq_depth,
417 		.enqueue_depth = 64,
418 		.new_event_threshold = 4096,
419 		.event_port_cfg = RTE_EVENT_PORT_CFG_HINT_PRODUCER,
420 	};
421 
422 	if (adptr_p_conf.new_event_threshold > dev_info.max_num_events)
423 		adptr_p_conf.new_event_threshold = dev_info.max_num_events;
424 	if (adptr_p_conf.dequeue_depth > dev_info.max_event_port_dequeue_depth)
425 		adptr_p_conf.dequeue_depth =
426 			dev_info.max_event_port_dequeue_depth;
427 	if (adptr_p_conf.enqueue_depth > dev_info.max_event_port_enqueue_depth)
428 		adptr_p_conf.enqueue_depth =
429 			dev_info.max_event_port_enqueue_depth;
430 
431 	init_ports(nb_ports);
432 	/* Create one adapter for all the ethernet ports. */
433 	ret = rte_event_eth_rx_adapter_create(cdata.rx_adapter_id, evdev_id,
434 			&adptr_p_conf);
435 	if (ret)
436 		rte_exit(EXIT_FAILURE, "failed to create rx adapter[%d]",
437 				cdata.rx_adapter_id);
438 
439 	ret = rte_event_eth_tx_adapter_create(cdata.tx_adapter_id, evdev_id,
440 			&adptr_p_conf);
441 	if (ret)
442 		rte_exit(EXIT_FAILURE, "failed to create tx adapter[%d]",
443 				cdata.tx_adapter_id);
444 
445 	struct rte_event_eth_rx_adapter_queue_conf queue_conf;
446 	memset(&queue_conf, 0, sizeof(queue_conf));
447 	queue_conf.ev.sched_type = cdata.queue_type;
448 	queue_conf.ev.queue_id = cdata.qid[0];
449 
450 	for (i = 0; i < nb_ports; i++) {
451 		ret = rte_event_eth_rx_adapter_queue_add(cdata.rx_adapter_id, i,
452 				-1, &queue_conf);
453 		if (ret)
454 			rte_exit(EXIT_FAILURE,
455 					"Failed to add queues to Rx adapter");
456 
457 		ret = rte_event_eth_tx_adapter_queue_add(cdata.tx_adapter_id, i,
458 				-1);
459 		if (ret)
460 			rte_exit(EXIT_FAILURE,
461 					"Failed to add queues to Tx adapter");
462 	}
463 
464 	ret = rte_event_eth_tx_adapter_event_port_get(cdata.tx_adapter_id,
465 			&tx_port_id);
466 	if (ret)
467 		rte_exit(EXIT_FAILURE,
468 				"Failed to get Tx adapter port id");
469 	ret = rte_event_port_link(evdev_id, tx_port_id, &cdata.tx_queue_id,
470 			NULL, 1);
471 	if (ret != 1)
472 		rte_exit(EXIT_FAILURE,
473 				"Unable to link Tx adapter port to Tx queue");
474 
475 	ret = rte_event_eth_rx_adapter_service_id_get(cdata.rx_adapter_id,
476 				&fdata->rxadptr_service_id);
477 	if (ret != -ESRCH && ret != 0) {
478 		rte_exit(EXIT_FAILURE,
479 			"Error getting the service ID for Rx adapter\n");
480 	}
481 	rte_service_runstate_set(fdata->rxadptr_service_id, 1);
482 	rte_service_set_runstate_mapped_check(fdata->rxadptr_service_id, 0);
483 
484 	ret = rte_event_eth_tx_adapter_service_id_get(cdata.tx_adapter_id,
485 				&fdata->txadptr_service_id);
486 	if (ret != -ESRCH && ret != 0) {
487 		rte_exit(EXIT_FAILURE,
488 			"Error getting the service ID for Tx adapter\n");
489 	}
490 	rte_service_runstate_set(fdata->txadptr_service_id, 1);
491 	rte_service_set_runstate_mapped_check(fdata->txadptr_service_id, 0);
492 
493 	ret = rte_event_eth_rx_adapter_start(cdata.rx_adapter_id);
494 	if (ret)
495 		rte_exit(EXIT_FAILURE, "Rx adapter[%d] start failed",
496 				cdata.rx_adapter_id);
497 
498 	ret = rte_event_eth_tx_adapter_start(cdata.tx_adapter_id);
499 	if (ret)
500 		rte_exit(EXIT_FAILURE, "Tx adapter[%d] start failed",
501 				cdata.tx_adapter_id);
502 
503 	if (rte_event_dev_start(evdev_id) < 0)
504 		rte_exit(EXIT_FAILURE, "Error starting eventdev");
505 }
506 
507 static void
508 generic_opt_check(void)
509 {
510 	int i;
511 	int ret;
512 	uint32_t cap = 0;
513 	uint8_t rx_needed = 0;
514 	uint8_t sched_needed = 0;
515 	struct rte_event_dev_info eventdev_info;
516 
517 	memset(&eventdev_info, 0, sizeof(struct rte_event_dev_info));
518 	rte_event_dev_info_get(0, &eventdev_info);
519 
520 	if (cdata.all_type_queues && !(eventdev_info.event_dev_cap &
521 				RTE_EVENT_DEV_CAP_QUEUE_ALL_TYPES))
522 		rte_exit(EXIT_FAILURE,
523 				"Event dev doesn't support all type queues\n");
524 	sched_needed = !(eventdev_info.event_dev_cap &
525 		RTE_EVENT_DEV_CAP_DISTRIBUTED_SCHED);
526 
527 	RTE_ETH_FOREACH_DEV(i) {
528 		ret = rte_event_eth_rx_adapter_caps_get(0, i, &cap);
529 		if (ret)
530 			rte_exit(EXIT_FAILURE,
531 				"failed to get event rx adapter capabilities");
532 		rx_needed |=
533 			!(cap & RTE_EVENT_ETH_RX_ADAPTER_CAP_INTERNAL_PORT);
534 	}
535 
536 	if (cdata.worker_lcore_mask == 0 ||
537 			(rx_needed && cdata.rx_lcore_mask == 0) ||
538 			(cdata.tx_lcore_mask == 0) ||
539 			(sched_needed && cdata.sched_lcore_mask == 0)) {
540 		printf("Core part of pipeline was not assigned any cores. "
541 			"This will stall the pipeline, please check core masks "
542 			"(use -h for details on setting core masks):\n"
543 			"\trx: %"PRIu64"\n\ttx: %"PRIu64"\n\tsched: %"PRIu64
544 			"\n\tworkers: %"PRIu64"\n",
545 			cdata.rx_lcore_mask, cdata.tx_lcore_mask,
546 			cdata.sched_lcore_mask,
547 			cdata.worker_lcore_mask);
548 		rte_exit(-1, "Fix core masks\n");
549 	}
550 
551 	if (!sched_needed)
552 		memset(fdata->sched_core, 0,
553 				sizeof(unsigned int) * MAX_NUM_CORE);
554 	if (!rx_needed)
555 		memset(fdata->rx_core, 0,
556 				sizeof(unsigned int) * MAX_NUM_CORE);
557 }
558 
559 void
560 set_worker_generic_setup_data(struct setup_data *caps, bool burst)
561 {
562 	if (burst) {
563 		caps->worker = worker_generic_burst;
564 	} else {
565 		caps->worker = worker_generic;
566 	}
567 
568 	caps->adptr_setup = init_adapters;
569 	caps->scheduler = schedule_devices;
570 	caps->evdev_setup = setup_eventdev_generic;
571 	caps->check_opt = generic_opt_check;
572 }
573