xref: /dpdk/examples/eventdev_pipeline/pipeline_worker_generic.c (revision 8f1d23ece06adff5eae9f1b4365bdbbd3abee2b2)
1 /*
2  * SPDX-License-Identifier: BSD-3-Clause
3  * Copyright 2016 Intel Corporation.
4  * Copyright 2017 Cavium, Inc.
5  */
6 
7 #include <stdlib.h>
8 
9 #include "pipeline_common.h"
10 
11 static __rte_always_inline int
12 worker_generic(void *arg)
13 {
14 	struct rte_event ev;
15 
16 	struct worker_data *data = (struct worker_data *)arg;
17 	uint8_t dev_id = data->dev_id;
18 	uint8_t port_id = data->port_id;
19 	size_t sent = 0, received = 0;
20 	unsigned int lcore_id = rte_lcore_id();
21 	uint16_t nb_rx = 0, nb_tx = 0;
22 
23 	while (!fdata->done) {
24 
25 		if (fdata->cap.scheduler)
26 			fdata->cap.scheduler(lcore_id);
27 
28 		if (!fdata->worker_core[lcore_id]) {
29 			rte_pause();
30 			continue;
31 		}
32 
33 		nb_rx = rte_event_dequeue_burst(dev_id, port_id, &ev, 1, 0);
34 
35 		if (nb_rx == 0) {
36 			rte_pause();
37 			continue;
38 		}
39 		received++;
40 
41 		/* The first worker stage does classification */
42 		if (ev.queue_id == cdata.qid[0])
43 			ev.flow_id = ev.mbuf->hash.rss
44 						% cdata.num_fids;
45 
46 		ev.queue_id = cdata.next_qid[ev.queue_id];
47 		ev.op = RTE_EVENT_OP_FORWARD;
48 		ev.sched_type = cdata.queue_type;
49 
50 		work();
51 
52 		do {
53 			nb_tx = rte_event_enqueue_burst(dev_id, port_id, &ev,
54 							1);
55 		} while (!nb_tx && !fdata->done);
56 		sent++;
57 	}
58 
59 	worker_cleanup(dev_id, port_id, &ev, nb_tx, nb_rx);
60 	if (!cdata.quiet)
61 		printf("  worker %u thread done. RX=%zu TX=%zu\n",
62 				rte_lcore_id(), received, sent);
63 
64 	return 0;
65 }
66 
67 static int
68 worker_generic_burst(void *arg)
69 {
70 	struct rte_event events[BATCH_SIZE];
71 
72 	struct worker_data *data = (struct worker_data *)arg;
73 	uint8_t dev_id = data->dev_id;
74 	uint8_t port_id = data->port_id;
75 	size_t sent = 0, received = 0;
76 	unsigned int lcore_id = rte_lcore_id();
77 	uint16_t i, nb_rx = 0, nb_tx = 0;
78 
79 	while (!fdata->done) {
80 		if (fdata->cap.scheduler)
81 			fdata->cap.scheduler(lcore_id);
82 
83 		if (!fdata->worker_core[lcore_id]) {
84 			rte_pause();
85 			continue;
86 		}
87 
88 		nb_rx = rte_event_dequeue_burst(dev_id, port_id, events,
89 						RTE_DIM(events), 0);
90 
91 		if (nb_rx == 0) {
92 			rte_pause();
93 			continue;
94 		}
95 		received += nb_rx;
96 
97 		for (i = 0; i < nb_rx; i++) {
98 
99 			/* The first worker stage does classification */
100 			if (events[i].queue_id == cdata.qid[0])
101 				events[i].flow_id = events[i].mbuf->hash.rss
102 							% cdata.num_fids;
103 
104 			events[i].queue_id = cdata.next_qid[events[i].queue_id];
105 			events[i].op = RTE_EVENT_OP_FORWARD;
106 			events[i].sched_type = cdata.queue_type;
107 
108 			work();
109 		}
110 		nb_tx = rte_event_enqueue_burst(dev_id, port_id, events, nb_rx);
111 		while (nb_tx < nb_rx && !fdata->done)
112 			nb_tx += rte_event_enqueue_burst(dev_id, port_id,
113 							events + nb_tx,
114 							nb_rx - nb_tx);
115 		sent += nb_tx;
116 	}
117 
118 	worker_cleanup(dev_id, port_id, events, nb_tx, nb_rx);
119 
120 	if (!cdata.quiet)
121 		printf("  worker %u thread done. RX=%zu TX=%zu\n",
122 				rte_lcore_id(), received, sent);
123 
124 	return 0;
125 }
126 
127 static int
128 setup_eventdev_generic(struct worker_data *worker_data)
129 {
130 	const uint8_t dev_id = 0;
131 	/* +1 stages is for a SINGLE_LINK TX stage */
132 	const uint8_t nb_queues = cdata.num_stages + 1;
133 	const uint8_t nb_ports = cdata.num_workers;
134 	struct rte_event_dev_config config = {
135 			.nb_event_queues = nb_queues,
136 			.nb_event_ports = nb_ports,
137 			.nb_single_link_event_port_queues = 1,
138 			.nb_events_limit  = 4096,
139 			.nb_event_queue_flows = 1024,
140 			.nb_event_port_dequeue_depth = 128,
141 			.nb_event_port_enqueue_depth = 128,
142 	};
143 	struct rte_event_port_conf wkr_p_conf = {
144 			.dequeue_depth = cdata.worker_cq_depth,
145 			.enqueue_depth = 64,
146 			.new_event_threshold = 4096,
147 			.event_port_cfg = RTE_EVENT_PORT_CFG_HINT_WORKER,
148 	};
149 	struct rte_event_queue_conf wkr_q_conf = {
150 			.schedule_type = cdata.queue_type,
151 			.priority = RTE_EVENT_DEV_PRIORITY_NORMAL,
152 			.nb_atomic_flows = 1024,
153 			.nb_atomic_order_sequences = 1024,
154 	};
155 	struct rte_event_queue_conf tx_q_conf = {
156 			.priority = RTE_EVENT_DEV_PRIORITY_HIGHEST,
157 			.event_queue_cfg = RTE_EVENT_QUEUE_CFG_SINGLE_LINK,
158 	};
159 
160 	struct port_link worker_queues[MAX_NUM_STAGES];
161 	uint8_t disable_implicit_release;
162 	unsigned int i;
163 
164 	int ret, ndev = rte_event_dev_count();
165 	if (ndev < 1) {
166 		printf("%d: No Eventdev Devices Found\n", __LINE__);
167 		return -1;
168 	}
169 
170 	struct rte_event_dev_info dev_info;
171 	ret = rte_event_dev_info_get(dev_id, &dev_info);
172 	printf("\tEventdev %d: %s\n", dev_id, dev_info.driver_name);
173 
174 	disable_implicit_release = (dev_info.event_dev_cap &
175 			RTE_EVENT_DEV_CAP_IMPLICIT_RELEASE_DISABLE);
176 
177 	wkr_p_conf.event_port_cfg = disable_implicit_release ?
178 		RTE_EVENT_PORT_CFG_DISABLE_IMPL_REL : 0;
179 
180 	if (dev_info.max_num_events < config.nb_events_limit)
181 		config.nb_events_limit = dev_info.max_num_events;
182 	if (dev_info.max_event_port_dequeue_depth <
183 			config.nb_event_port_dequeue_depth)
184 		config.nb_event_port_dequeue_depth =
185 				dev_info.max_event_port_dequeue_depth;
186 	if (dev_info.max_event_port_enqueue_depth <
187 			config.nb_event_port_enqueue_depth)
188 		config.nb_event_port_enqueue_depth =
189 				dev_info.max_event_port_enqueue_depth;
190 
191 	ret = rte_event_dev_configure(dev_id, &config);
192 	if (ret < 0) {
193 		printf("%d: Error configuring device\n", __LINE__);
194 		return -1;
195 	}
196 
197 	/* Q creation - one load balanced per pipeline stage*/
198 	printf("  Stages:\n");
199 	for (i = 0; i < cdata.num_stages; i++) {
200 		if (rte_event_queue_setup(dev_id, i, &wkr_q_conf) < 0) {
201 			printf("%d: error creating qid %d\n", __LINE__, i);
202 			return -1;
203 		}
204 		cdata.qid[i] = i;
205 		cdata.next_qid[i] = i+1;
206 		worker_queues[i].queue_id = i;
207 		if (cdata.enable_queue_priorities) {
208 			/* calculate priority stepping for each stage, leaving
209 			 * headroom of 1 for the SINGLE_LINK TX below
210 			 */
211 			const uint32_t prio_delta =
212 				(RTE_EVENT_DEV_PRIORITY_LOWEST-1) /  nb_queues;
213 
214 			/* higher priority for queues closer to tx */
215 			wkr_q_conf.priority =
216 				RTE_EVENT_DEV_PRIORITY_LOWEST - prio_delta * i;
217 		}
218 
219 		const char *type_str = "Atomic";
220 		switch (wkr_q_conf.schedule_type) {
221 		case RTE_SCHED_TYPE_ORDERED:
222 			type_str = "Ordered";
223 			break;
224 		case RTE_SCHED_TYPE_PARALLEL:
225 			type_str = "Parallel";
226 			break;
227 		}
228 		printf("\tStage %d, Type %s\tPriority = %d\n", i, type_str,
229 				wkr_q_conf.priority);
230 	}
231 	printf("\n");
232 
233 	/* final queue for sending to TX core */
234 	if (rte_event_queue_setup(dev_id, i, &tx_q_conf) < 0) {
235 		printf("%d: error creating qid %d\n", __LINE__, i);
236 		return -1;
237 	}
238 	cdata.tx_queue_id = i;
239 
240 	if (wkr_p_conf.new_event_threshold > config.nb_events_limit)
241 		wkr_p_conf.new_event_threshold = config.nb_events_limit;
242 	if (wkr_p_conf.dequeue_depth > config.nb_event_port_dequeue_depth)
243 		wkr_p_conf.dequeue_depth = config.nb_event_port_dequeue_depth;
244 	if (wkr_p_conf.enqueue_depth > config.nb_event_port_enqueue_depth)
245 		wkr_p_conf.enqueue_depth = config.nb_event_port_enqueue_depth;
246 
247 	/* set up one port per worker, linking to all stage queues */
248 	for (i = 0; i < cdata.num_workers; i++) {
249 		struct worker_data *w = &worker_data[i];
250 		w->dev_id = dev_id;
251 		if (rte_event_port_setup(dev_id, i, &wkr_p_conf) < 0) {
252 			printf("Error setting up port %d\n", i);
253 			return -1;
254 		}
255 
256 		uint32_t s;
257 		for (s = 0; s < cdata.num_stages; s++) {
258 			if (rte_event_port_link(dev_id, i,
259 						&worker_queues[s].queue_id,
260 						&worker_queues[s].priority,
261 						1) != 1) {
262 				printf("%d: error creating link for port %d\n",
263 						__LINE__, i);
264 				return -1;
265 			}
266 		}
267 		w->port_id = i;
268 	}
269 
270 	ret = rte_event_dev_service_id_get(dev_id,
271 				&fdata->evdev_service_id);
272 	if (ret != -ESRCH && ret != 0) {
273 		printf("Error getting the service ID for sw eventdev\n");
274 		return -1;
275 	}
276 	rte_service_runstate_set(fdata->evdev_service_id, 1);
277 	rte_service_set_runstate_mapped_check(fdata->evdev_service_id, 0);
278 
279 	return dev_id;
280 }
281 
282 /*
283  * Initializes a given port using global settings and with the RX buffers
284  * coming from the mbuf_pool passed as a parameter.
285  */
286 static inline int
287 port_init(uint8_t port, struct rte_mempool *mbuf_pool)
288 {
289 	struct rte_eth_rxconf rx_conf;
290 	static const struct rte_eth_conf port_conf_default = {
291 		.rxmode = {
292 			.mq_mode = RTE_ETH_MQ_RX_RSS,
293 		},
294 		.rx_adv_conf = {
295 			.rss_conf = {
296 				.rss_hf = RTE_ETH_RSS_IP |
297 					  RTE_ETH_RSS_TCP |
298 					  RTE_ETH_RSS_UDP,
299 			}
300 		}
301 	};
302 	const uint16_t rx_rings = 1, tx_rings = 1;
303 	const uint16_t rx_ring_size = 512, tx_ring_size = 512;
304 	struct rte_eth_conf port_conf = port_conf_default;
305 	int retval;
306 	uint16_t q;
307 	struct rte_eth_dev_info dev_info;
308 	struct rte_eth_txconf txconf;
309 
310 	if (!rte_eth_dev_is_valid_port(port))
311 		return -1;
312 
313 	retval = rte_eth_dev_info_get(port, &dev_info);
314 	if (retval != 0) {
315 		printf("Error during getting device (port %u) info: %s\n",
316 				port, strerror(-retval));
317 		return retval;
318 	}
319 
320 	if (dev_info.tx_offload_capa & RTE_ETH_TX_OFFLOAD_MBUF_FAST_FREE)
321 		port_conf.txmode.offloads |=
322 			RTE_ETH_TX_OFFLOAD_MBUF_FAST_FREE;
323 
324 	if (dev_info.rx_offload_capa & RTE_ETH_RX_OFFLOAD_RSS_HASH)
325 		port_conf.rxmode.offloads |= RTE_ETH_RX_OFFLOAD_RSS_HASH;
326 
327 	rx_conf = dev_info.default_rxconf;
328 	rx_conf.offloads = port_conf.rxmode.offloads;
329 
330 	port_conf.rx_adv_conf.rss_conf.rss_hf &=
331 		dev_info.flow_type_rss_offloads;
332 	if (port_conf.rx_adv_conf.rss_conf.rss_hf !=
333 			port_conf_default.rx_adv_conf.rss_conf.rss_hf) {
334 		printf("Port %u modified RSS hash function based on hardware support,"
335 			"requested:%#"PRIx64" configured:%#"PRIx64"\n",
336 			port,
337 			port_conf_default.rx_adv_conf.rss_conf.rss_hf,
338 			port_conf.rx_adv_conf.rss_conf.rss_hf);
339 	}
340 
341 	/* Configure the Ethernet device. */
342 	retval = rte_eth_dev_configure(port, rx_rings, tx_rings, &port_conf);
343 	if (retval != 0)
344 		return retval;
345 
346 	/* Allocate and set up 1 RX queue per Ethernet port. */
347 	for (q = 0; q < rx_rings; q++) {
348 		retval = rte_eth_rx_queue_setup(port, q, rx_ring_size,
349 				rte_eth_dev_socket_id(port), &rx_conf,
350 				mbuf_pool);
351 		if (retval < 0)
352 			return retval;
353 	}
354 
355 	txconf = dev_info.default_txconf;
356 	txconf.offloads = port_conf_default.txmode.offloads;
357 	/* Allocate and set up 1 TX queue per Ethernet port. */
358 	for (q = 0; q < tx_rings; q++) {
359 		retval = rte_eth_tx_queue_setup(port, q, tx_ring_size,
360 				rte_eth_dev_socket_id(port), &txconf);
361 		if (retval < 0)
362 			return retval;
363 	}
364 
365 	/* Display the port MAC address. */
366 	struct rte_ether_addr addr;
367 	retval = rte_eth_macaddr_get(port, &addr);
368 	if (retval != 0) {
369 		printf("Failed to get MAC address (port %u): %s\n",
370 				port, rte_strerror(-retval));
371 		return retval;
372 	}
373 
374 	printf("Port %u MAC: %02" PRIx8 " %02" PRIx8 " %02" PRIx8
375 			" %02" PRIx8 " %02" PRIx8 " %02" PRIx8 "\n",
376 			(unsigned int)port, RTE_ETHER_ADDR_BYTES(&addr));
377 
378 	/* Enable RX in promiscuous mode for the Ethernet device. */
379 	retval = rte_eth_promiscuous_enable(port);
380 	if (retval != 0)
381 		return retval;
382 
383 	return 0;
384 }
385 
386 static int
387 init_ports(uint16_t num_ports)
388 {
389 	uint16_t portid;
390 
391 	if (!cdata.num_mbuf)
392 		cdata.num_mbuf = 16384 * num_ports;
393 
394 	struct rte_mempool *mp = rte_pktmbuf_pool_create("packet_pool",
395 			/* mbufs */ cdata.num_mbuf,
396 			/* cache_size */ 512,
397 			/* priv_size*/ 0,
398 			/* data_room_size */ RTE_MBUF_DEFAULT_BUF_SIZE,
399 			rte_socket_id());
400 
401 	RTE_ETH_FOREACH_DEV(portid)
402 		if (port_init(portid, mp) != 0)
403 			rte_exit(EXIT_FAILURE, "Cannot init port %"PRIu16 "\n",
404 					portid);
405 
406 	return 0;
407 }
408 
409 static void
410 init_adapters(uint16_t nb_ports)
411 {
412 	int i;
413 	int ret;
414 	uint8_t tx_port_id = 0;
415 	uint8_t evdev_id = 0;
416 	struct rte_event_dev_info dev_info;
417 
418 	ret = rte_event_dev_info_get(evdev_id, &dev_info);
419 
420 	struct rte_event_port_conf adptr_p_conf = {
421 		.dequeue_depth = cdata.worker_cq_depth,
422 		.enqueue_depth = 64,
423 		.new_event_threshold = 4096,
424 		.event_port_cfg = RTE_EVENT_PORT_CFG_HINT_PRODUCER,
425 	};
426 
427 	if (adptr_p_conf.new_event_threshold > dev_info.max_num_events)
428 		adptr_p_conf.new_event_threshold = dev_info.max_num_events;
429 	if (adptr_p_conf.dequeue_depth > dev_info.max_event_port_dequeue_depth)
430 		adptr_p_conf.dequeue_depth =
431 			dev_info.max_event_port_dequeue_depth;
432 	if (adptr_p_conf.enqueue_depth > dev_info.max_event_port_enqueue_depth)
433 		adptr_p_conf.enqueue_depth =
434 			dev_info.max_event_port_enqueue_depth;
435 
436 	init_ports(nb_ports);
437 	/* Create one adapter for all the ethernet ports. */
438 	ret = rte_event_eth_rx_adapter_create(cdata.rx_adapter_id, evdev_id,
439 			&adptr_p_conf);
440 	if (ret)
441 		rte_exit(EXIT_FAILURE, "failed to create rx adapter[%d]",
442 				cdata.rx_adapter_id);
443 
444 	ret = rte_event_eth_tx_adapter_create(cdata.tx_adapter_id, evdev_id,
445 			&adptr_p_conf);
446 	if (ret)
447 		rte_exit(EXIT_FAILURE, "failed to create tx adapter[%d]",
448 				cdata.tx_adapter_id);
449 
450 	struct rte_event_eth_rx_adapter_queue_conf queue_conf;
451 	memset(&queue_conf, 0, sizeof(queue_conf));
452 	queue_conf.ev.sched_type = cdata.queue_type;
453 	queue_conf.ev.queue_id = cdata.qid[0];
454 
455 	for (i = 0; i < nb_ports; i++) {
456 		ret = rte_event_eth_rx_adapter_queue_add(cdata.rx_adapter_id, i,
457 				-1, &queue_conf);
458 		if (ret)
459 			rte_exit(EXIT_FAILURE,
460 					"Failed to add queues to Rx adapter");
461 
462 		ret = rte_event_eth_tx_adapter_queue_add(cdata.tx_adapter_id, i,
463 				-1);
464 		if (ret)
465 			rte_exit(EXIT_FAILURE,
466 					"Failed to add queues to Tx adapter");
467 	}
468 
469 	ret = rte_event_eth_tx_adapter_event_port_get(cdata.tx_adapter_id,
470 			&tx_port_id);
471 	if (ret)
472 		rte_exit(EXIT_FAILURE,
473 				"Failed to get Tx adapter port id");
474 	ret = rte_event_port_link(evdev_id, tx_port_id, &cdata.tx_queue_id,
475 			NULL, 1);
476 	if (ret != 1)
477 		rte_exit(EXIT_FAILURE,
478 				"Unable to link Tx adapter port to Tx queue");
479 
480 	ret = rte_event_eth_rx_adapter_service_id_get(cdata.rx_adapter_id,
481 				&fdata->rxadptr_service_id);
482 	if (ret != -ESRCH && ret != 0) {
483 		rte_exit(EXIT_FAILURE,
484 			"Error getting the service ID for Rx adapter\n");
485 	}
486 	rte_service_runstate_set(fdata->rxadptr_service_id, 1);
487 	rte_service_set_runstate_mapped_check(fdata->rxadptr_service_id, 0);
488 
489 	ret = rte_event_eth_tx_adapter_service_id_get(cdata.tx_adapter_id,
490 				&fdata->txadptr_service_id);
491 	if (ret != -ESRCH && ret != 0) {
492 		rte_exit(EXIT_FAILURE,
493 			"Error getting the service ID for Tx adapter\n");
494 	}
495 	rte_service_runstate_set(fdata->txadptr_service_id, 1);
496 	rte_service_set_runstate_mapped_check(fdata->txadptr_service_id, 0);
497 
498 	ret = rte_event_eth_rx_adapter_start(cdata.rx_adapter_id);
499 	if (ret)
500 		rte_exit(EXIT_FAILURE, "Rx adapter[%d] start failed",
501 				cdata.rx_adapter_id);
502 
503 	ret = rte_event_eth_tx_adapter_start(cdata.tx_adapter_id);
504 	if (ret)
505 		rte_exit(EXIT_FAILURE, "Tx adapter[%d] start failed",
506 				cdata.tx_adapter_id);
507 
508 	if (rte_event_dev_start(evdev_id) < 0)
509 		rte_exit(EXIT_FAILURE, "Error starting eventdev");
510 }
511 
512 static void
513 generic_opt_check(void)
514 {
515 	int i;
516 	int ret;
517 	uint32_t cap = 0;
518 	uint8_t rx_needed = 0;
519 	uint8_t sched_needed = 0;
520 	struct rte_event_dev_info eventdev_info;
521 
522 	memset(&eventdev_info, 0, sizeof(struct rte_event_dev_info));
523 	rte_event_dev_info_get(0, &eventdev_info);
524 
525 	if (cdata.all_type_queues && !(eventdev_info.event_dev_cap &
526 				RTE_EVENT_DEV_CAP_QUEUE_ALL_TYPES))
527 		rte_exit(EXIT_FAILURE,
528 				"Event dev doesn't support all type queues\n");
529 	sched_needed = !(eventdev_info.event_dev_cap &
530 		RTE_EVENT_DEV_CAP_DISTRIBUTED_SCHED);
531 
532 	RTE_ETH_FOREACH_DEV(i) {
533 		ret = rte_event_eth_rx_adapter_caps_get(0, i, &cap);
534 		if (ret)
535 			rte_exit(EXIT_FAILURE,
536 				"failed to get event rx adapter capabilities");
537 		rx_needed |=
538 			!(cap & RTE_EVENT_ETH_RX_ADAPTER_CAP_INTERNAL_PORT);
539 	}
540 
541 	if (cdata.worker_lcore_mask == 0 ||
542 			(rx_needed && cdata.rx_lcore_mask == 0) ||
543 			(cdata.tx_lcore_mask == 0) ||
544 			(sched_needed && cdata.sched_lcore_mask == 0)) {
545 		printf("Core part of pipeline was not assigned any cores. "
546 			"This will stall the pipeline, please check core masks "
547 			"(use -h for details on setting core masks):\n"
548 			"\trx: %"PRIu64"\n\ttx: %"PRIu64"\n\tsched: %"PRIu64
549 			"\n\tworkers: %"PRIu64"\n",
550 			cdata.rx_lcore_mask, cdata.tx_lcore_mask,
551 			cdata.sched_lcore_mask,
552 			cdata.worker_lcore_mask);
553 		rte_exit(-1, "Fix core masks\n");
554 	}
555 
556 	if (!sched_needed)
557 		memset(fdata->sched_core, 0,
558 				sizeof(unsigned int) * MAX_NUM_CORE);
559 	if (!rx_needed)
560 		memset(fdata->rx_core, 0,
561 				sizeof(unsigned int) * MAX_NUM_CORE);
562 }
563 
564 void
565 set_worker_generic_setup_data(struct setup_data *caps, bool burst)
566 {
567 	if (burst) {
568 		caps->worker = worker_generic_burst;
569 	} else {
570 		caps->worker = worker_generic;
571 	}
572 
573 	caps->adptr_setup = init_adapters;
574 	caps->scheduler = schedule_devices;
575 	caps->evdev_setup = setup_eventdev_generic;
576 	caps->check_opt = generic_opt_check;
577 }
578