xref: /dpdk/drivers/event/sw/sw_evdev.c (revision 89f0711f9ddfb5822da9d34f384b92f72a61c4dc)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright(c) 2016-2017 Intel Corporation
3  */
4 
5 #include <inttypes.h>
6 #include <string.h>
7 
8 #include <rte_bus_vdev.h>
9 #include <rte_kvargs.h>
10 #include <rte_ring.h>
11 #include <rte_errno.h>
12 #include <rte_event_ring.h>
13 #include <rte_service_component.h>
14 
15 #include "sw_evdev.h"
16 #include "iq_chunk.h"
17 
18 #define EVENTDEV_NAME_SW_PMD event_sw
19 #define NUMA_NODE_ARG "numa_node"
20 #define SCHED_QUANTA_ARG "sched_quanta"
21 #define CREDIT_QUANTA_ARG "credit_quanta"
22 
23 static void
24 sw_info_get(struct rte_eventdev *dev, struct rte_event_dev_info *info);
25 
26 static int
27 sw_port_link(struct rte_eventdev *dev, void *port, const uint8_t queues[],
28 		const uint8_t priorities[], uint16_t num)
29 {
30 	struct sw_port *p = port;
31 	struct sw_evdev *sw = sw_pmd_priv(dev);
32 	int i;
33 
34 	RTE_SET_USED(priorities);
35 	for (i = 0; i < num; i++) {
36 		struct sw_qid *q = &sw->qids[queues[i]];
37 		unsigned int j;
38 
39 		/* check for qid map overflow */
40 		if (q->cq_num_mapped_cqs >= RTE_DIM(q->cq_map)) {
41 			rte_errno = -EDQUOT;
42 			break;
43 		}
44 
45 		if (p->is_directed && p->num_qids_mapped > 0) {
46 			rte_errno = -EDQUOT;
47 			break;
48 		}
49 
50 		for (j = 0; j < q->cq_num_mapped_cqs; j++) {
51 			if (q->cq_map[j] == p->id)
52 				break;
53 		}
54 
55 		/* check if port is already linked */
56 		if (j < q->cq_num_mapped_cqs)
57 			continue;
58 
59 		if (q->type == SW_SCHED_TYPE_DIRECT) {
60 			/* check directed qids only map to one port */
61 			if (p->num_qids_mapped > 0) {
62 				rte_errno = -EDQUOT;
63 				break;
64 			}
65 			/* check port only takes a directed flow */
66 			if (num > 1) {
67 				rte_errno = -EDQUOT;
68 				break;
69 			}
70 
71 			p->is_directed = 1;
72 			p->num_qids_mapped = 1;
73 		} else if (q->type == RTE_SCHED_TYPE_ORDERED) {
74 			p->num_ordered_qids++;
75 			p->num_qids_mapped++;
76 		} else if (q->type == RTE_SCHED_TYPE_ATOMIC ||
77 				q->type == RTE_SCHED_TYPE_PARALLEL) {
78 			p->num_qids_mapped++;
79 		}
80 
81 		q->cq_map[q->cq_num_mapped_cqs] = p->id;
82 		rte_smp_wmb();
83 		q->cq_num_mapped_cqs++;
84 	}
85 	return i;
86 }
87 
88 static int
89 sw_port_unlink(struct rte_eventdev *dev, void *port, uint8_t queues[],
90 		uint16_t nb_unlinks)
91 {
92 	struct sw_port *p = port;
93 	struct sw_evdev *sw = sw_pmd_priv(dev);
94 	unsigned int i, j;
95 
96 	int unlinked = 0;
97 	for (i = 0; i < nb_unlinks; i++) {
98 		struct sw_qid *q = &sw->qids[queues[i]];
99 		for (j = 0; j < q->cq_num_mapped_cqs; j++) {
100 			if (q->cq_map[j] == p->id) {
101 				q->cq_map[j] =
102 					q->cq_map[q->cq_num_mapped_cqs - 1];
103 				rte_smp_wmb();
104 				q->cq_num_mapped_cqs--;
105 				unlinked++;
106 
107 				p->num_qids_mapped--;
108 
109 				if (q->type == RTE_SCHED_TYPE_ORDERED)
110 					p->num_ordered_qids--;
111 
112 				continue;
113 			}
114 		}
115 	}
116 	return unlinked;
117 }
118 
119 static int
120 sw_port_setup(struct rte_eventdev *dev, uint8_t port_id,
121 		const struct rte_event_port_conf *conf)
122 {
123 	struct sw_evdev *sw = sw_pmd_priv(dev);
124 	struct sw_port *p = &sw->ports[port_id];
125 	char buf[RTE_RING_NAMESIZE];
126 	unsigned int i;
127 
128 	struct rte_event_dev_info info;
129 	sw_info_get(dev, &info);
130 
131 	/* detect re-configuring and return credits to instance if needed */
132 	if (p->initialized) {
133 		/* taking credits from pool is done one quanta at a time, and
134 		 * credits may be spend (counted in p->inflights) or still
135 		 * available in the port (p->inflight_credits). We must return
136 		 * the sum to no leak credits
137 		 */
138 		int possible_inflights = p->inflight_credits + p->inflights;
139 		rte_atomic32_sub(&sw->inflights, possible_inflights);
140 	}
141 
142 	*p = (struct sw_port){0}; /* zero entire structure */
143 	p->id = port_id;
144 	p->sw = sw;
145 
146 	/* check to see if rings exists - port_setup() can be called multiple
147 	 * times legally (assuming device is stopped). If ring exists, free it
148 	 * to so it gets re-created with the correct size
149 	 */
150 	snprintf(buf, sizeof(buf), "sw%d_p%u_%s", dev->data->dev_id,
151 			port_id, "rx_worker_ring");
152 	struct rte_event_ring *existing_ring = rte_event_ring_lookup(buf);
153 	if (existing_ring)
154 		rte_event_ring_free(existing_ring);
155 
156 	p->rx_worker_ring = rte_event_ring_create(buf, MAX_SW_PROD_Q_DEPTH,
157 			dev->data->socket_id,
158 			RING_F_SP_ENQ | RING_F_SC_DEQ | RING_F_EXACT_SZ);
159 	if (p->rx_worker_ring == NULL) {
160 		SW_LOG_ERR("Error creating RX worker ring for port %d\n",
161 				port_id);
162 		return -1;
163 	}
164 
165 	p->inflight_max = conf->new_event_threshold;
166 	p->implicit_release = !conf->disable_implicit_release;
167 
168 	/* check if ring exists, same as rx_worker above */
169 	snprintf(buf, sizeof(buf), "sw%d_p%u, %s", dev->data->dev_id,
170 			port_id, "cq_worker_ring");
171 	existing_ring = rte_event_ring_lookup(buf);
172 	if (existing_ring)
173 		rte_event_ring_free(existing_ring);
174 
175 	p->cq_worker_ring = rte_event_ring_create(buf, conf->dequeue_depth,
176 			dev->data->socket_id,
177 			RING_F_SP_ENQ | RING_F_SC_DEQ | RING_F_EXACT_SZ);
178 	if (p->cq_worker_ring == NULL) {
179 		rte_event_ring_free(p->rx_worker_ring);
180 		SW_LOG_ERR("Error creating CQ worker ring for port %d\n",
181 				port_id);
182 		return -1;
183 	}
184 	sw->cq_ring_space[port_id] = conf->dequeue_depth;
185 
186 	/* set hist list contents to empty */
187 	for (i = 0; i < SW_PORT_HIST_LIST; i++) {
188 		p->hist_list[i].fid = -1;
189 		p->hist_list[i].qid = -1;
190 	}
191 	dev->data->ports[port_id] = p;
192 
193 	rte_smp_wmb();
194 	p->initialized = 1;
195 	return 0;
196 }
197 
198 static void
199 sw_port_release(void *port)
200 {
201 	struct sw_port *p = (void *)port;
202 	if (p == NULL)
203 		return;
204 
205 	rte_event_ring_free(p->rx_worker_ring);
206 	rte_event_ring_free(p->cq_worker_ring);
207 	memset(p, 0, sizeof(*p));
208 }
209 
210 static int32_t
211 qid_init(struct sw_evdev *sw, unsigned int idx, int type,
212 		const struct rte_event_queue_conf *queue_conf)
213 {
214 	unsigned int i;
215 	int dev_id = sw->data->dev_id;
216 	int socket_id = sw->data->socket_id;
217 	char buf[IQ_ROB_NAMESIZE];
218 	struct sw_qid *qid = &sw->qids[idx];
219 
220 	/* Initialize the FID structures to no pinning (-1), and zero packets */
221 	const struct sw_fid_t fid = {.cq = -1, .pcount = 0};
222 	for (i = 0; i < RTE_DIM(qid->fids); i++)
223 		qid->fids[i] = fid;
224 
225 	qid->id = idx;
226 	qid->type = type;
227 	qid->priority = queue_conf->priority;
228 
229 	if (qid->type == RTE_SCHED_TYPE_ORDERED) {
230 		char ring_name[RTE_RING_NAMESIZE];
231 		uint32_t window_size;
232 
233 		/* rte_ring and window_size_mask require require window_size to
234 		 * be a power-of-2.
235 		 */
236 		window_size = rte_align32pow2(
237 				queue_conf->nb_atomic_order_sequences);
238 
239 		qid->window_size = window_size - 1;
240 
241 		if (!window_size) {
242 			SW_LOG_DBG(
243 				"invalid reorder_window_size for ordered queue\n"
244 				);
245 			goto cleanup;
246 		}
247 
248 		snprintf(buf, sizeof(buf), "sw%d_iq_%d_rob", dev_id, i);
249 		qid->reorder_buffer = rte_zmalloc_socket(buf,
250 				window_size * sizeof(qid->reorder_buffer[0]),
251 				0, socket_id);
252 		if (!qid->reorder_buffer) {
253 			SW_LOG_DBG("reorder_buffer malloc failed\n");
254 			goto cleanup;
255 		}
256 
257 		memset(&qid->reorder_buffer[0],
258 		       0,
259 		       window_size * sizeof(qid->reorder_buffer[0]));
260 
261 		snprintf(ring_name, sizeof(ring_name), "sw%d_q%d_freelist",
262 				dev_id, idx);
263 
264 		/* lookup the ring, and if it already exists, free it */
265 		struct rte_ring *cleanup = rte_ring_lookup(ring_name);
266 		if (cleanup)
267 			rte_ring_free(cleanup);
268 
269 		qid->reorder_buffer_freelist = rte_ring_create(ring_name,
270 				window_size,
271 				socket_id,
272 				RING_F_SP_ENQ | RING_F_SC_DEQ);
273 		if (!qid->reorder_buffer_freelist) {
274 			SW_LOG_DBG("freelist ring create failed");
275 			goto cleanup;
276 		}
277 
278 		/* Populate the freelist with reorder buffer entries. Enqueue
279 		 * 'window_size - 1' entries because the rte_ring holds only
280 		 * that many.
281 		 */
282 		for (i = 0; i < window_size - 1; i++) {
283 			if (rte_ring_sp_enqueue(qid->reorder_buffer_freelist,
284 						&qid->reorder_buffer[i]) < 0)
285 				goto cleanup;
286 		}
287 
288 		qid->reorder_buffer_index = 0;
289 		qid->cq_next_tx = 0;
290 	}
291 
292 	qid->initialized = 1;
293 
294 	return 0;
295 
296 cleanup:
297 	if (qid->reorder_buffer) {
298 		rte_free(qid->reorder_buffer);
299 		qid->reorder_buffer = NULL;
300 	}
301 
302 	if (qid->reorder_buffer_freelist) {
303 		rte_ring_free(qid->reorder_buffer_freelist);
304 		qid->reorder_buffer_freelist = NULL;
305 	}
306 
307 	return -EINVAL;
308 }
309 
310 static void
311 sw_queue_release(struct rte_eventdev *dev, uint8_t id)
312 {
313 	struct sw_evdev *sw = sw_pmd_priv(dev);
314 	struct sw_qid *qid = &sw->qids[id];
315 
316 	if (qid->type == RTE_SCHED_TYPE_ORDERED) {
317 		rte_free(qid->reorder_buffer);
318 		rte_ring_free(qid->reorder_buffer_freelist);
319 	}
320 	memset(qid, 0, sizeof(*qid));
321 }
322 
323 static int
324 sw_queue_setup(struct rte_eventdev *dev, uint8_t queue_id,
325 		const struct rte_event_queue_conf *conf)
326 {
327 	int type;
328 
329 	type = conf->schedule_type;
330 
331 	if (RTE_EVENT_QUEUE_CFG_SINGLE_LINK & conf->event_queue_cfg) {
332 		type = SW_SCHED_TYPE_DIRECT;
333 	} else if (RTE_EVENT_QUEUE_CFG_ALL_TYPES
334 			& conf->event_queue_cfg) {
335 		SW_LOG_ERR("QUEUE_CFG_ALL_TYPES not supported\n");
336 		return -ENOTSUP;
337 	}
338 
339 	struct sw_evdev *sw = sw_pmd_priv(dev);
340 
341 	if (sw->qids[queue_id].initialized)
342 		sw_queue_release(dev, queue_id);
343 
344 	return qid_init(sw, queue_id, type, conf);
345 }
346 
347 static void
348 sw_init_qid_iqs(struct sw_evdev *sw)
349 {
350 	int i, j;
351 
352 	/* Initialize the IQ memory of all configured qids */
353 	for (i = 0; i < RTE_EVENT_MAX_QUEUES_PER_DEV; i++) {
354 		struct sw_qid *qid = &sw->qids[i];
355 
356 		if (!qid->initialized)
357 			continue;
358 
359 		for (j = 0; j < SW_IQS_MAX; j++)
360 			iq_init(sw, &qid->iq[j]);
361 	}
362 }
363 
364 static void
365 sw_clean_qid_iqs(struct sw_evdev *sw)
366 {
367 	int i, j;
368 
369 	/* Release the IQ memory of all configured qids */
370 	for (i = 0; i < RTE_EVENT_MAX_QUEUES_PER_DEV; i++) {
371 		struct sw_qid *qid = &sw->qids[i];
372 
373 		for (j = 0; j < SW_IQS_MAX; j++) {
374 			if (!qid->iq[j].head)
375 				continue;
376 			iq_free_chunk_list(sw, qid->iq[j].head);
377 			qid->iq[j].head = NULL;
378 		}
379 	}
380 }
381 
382 static void
383 sw_queue_def_conf(struct rte_eventdev *dev, uint8_t queue_id,
384 				 struct rte_event_queue_conf *conf)
385 {
386 	RTE_SET_USED(dev);
387 	RTE_SET_USED(queue_id);
388 
389 	static const struct rte_event_queue_conf default_conf = {
390 		.nb_atomic_flows = 4096,
391 		.nb_atomic_order_sequences = 1,
392 		.schedule_type = RTE_SCHED_TYPE_ATOMIC,
393 		.priority = RTE_EVENT_DEV_PRIORITY_NORMAL,
394 	};
395 
396 	*conf = default_conf;
397 }
398 
399 static void
400 sw_port_def_conf(struct rte_eventdev *dev, uint8_t port_id,
401 		 struct rte_event_port_conf *port_conf)
402 {
403 	RTE_SET_USED(dev);
404 	RTE_SET_USED(port_id);
405 
406 	port_conf->new_event_threshold = 1024;
407 	port_conf->dequeue_depth = 16;
408 	port_conf->enqueue_depth = 16;
409 	port_conf->disable_implicit_release = 0;
410 }
411 
412 static int
413 sw_dev_configure(const struct rte_eventdev *dev)
414 {
415 	struct sw_evdev *sw = sw_pmd_priv(dev);
416 	const struct rte_eventdev_data *data = dev->data;
417 	const struct rte_event_dev_config *conf = &data->dev_conf;
418 	int num_chunks, i;
419 
420 	sw->qid_count = conf->nb_event_queues;
421 	sw->port_count = conf->nb_event_ports;
422 	sw->nb_events_limit = conf->nb_events_limit;
423 	rte_atomic32_set(&sw->inflights, 0);
424 
425 	/* Number of chunks sized for worst-case spread of events across IQs */
426 	num_chunks = ((SW_INFLIGHT_EVENTS_TOTAL/SW_EVS_PER_Q_CHUNK)+1) +
427 			sw->qid_count*SW_IQS_MAX*2;
428 
429 	/* If this is a reconfiguration, free the previous IQ allocation. All
430 	 * IQ chunk references were cleaned out of the QIDs in sw_stop(), and
431 	 * will be reinitialized in sw_start().
432 	 */
433 	if (sw->chunks)
434 		rte_free(sw->chunks);
435 
436 	sw->chunks = rte_malloc_socket(NULL,
437 				       sizeof(struct sw_queue_chunk) *
438 				       num_chunks,
439 				       0,
440 				       sw->data->socket_id);
441 	if (!sw->chunks)
442 		return -ENOMEM;
443 
444 	sw->chunk_list_head = NULL;
445 	for (i = 0; i < num_chunks; i++)
446 		iq_free_chunk(sw, &sw->chunks[i]);
447 
448 	if (conf->event_dev_cfg & RTE_EVENT_DEV_CFG_PER_DEQUEUE_TIMEOUT)
449 		return -ENOTSUP;
450 
451 	return 0;
452 }
453 
454 struct rte_eth_dev;
455 
456 static int
457 sw_eth_rx_adapter_caps_get(const struct rte_eventdev *dev,
458 			const struct rte_eth_dev *eth_dev,
459 			uint32_t *caps)
460 {
461 	RTE_SET_USED(dev);
462 	RTE_SET_USED(eth_dev);
463 	*caps = RTE_EVENT_ETH_RX_ADAPTER_SW_CAP;
464 	return 0;
465 }
466 
467 static void
468 sw_info_get(struct rte_eventdev *dev, struct rte_event_dev_info *info)
469 {
470 	RTE_SET_USED(dev);
471 
472 	static const struct rte_event_dev_info evdev_sw_info = {
473 			.driver_name = SW_PMD_NAME,
474 			.max_event_queues = RTE_EVENT_MAX_QUEUES_PER_DEV,
475 			.max_event_queue_flows = SW_QID_NUM_FIDS,
476 			.max_event_queue_priority_levels = SW_Q_PRIORITY_MAX,
477 			.max_event_priority_levels = SW_IQS_MAX,
478 			.max_event_ports = SW_PORTS_MAX,
479 			.max_event_port_dequeue_depth = MAX_SW_CONS_Q_DEPTH,
480 			.max_event_port_enqueue_depth = MAX_SW_PROD_Q_DEPTH,
481 			.max_num_events = SW_INFLIGHT_EVENTS_TOTAL,
482 			.event_dev_cap = (
483 				RTE_EVENT_DEV_CAP_QUEUE_QOS |
484 				RTE_EVENT_DEV_CAP_BURST_MODE |
485 				RTE_EVENT_DEV_CAP_EVENT_QOS |
486 				RTE_EVENT_DEV_CAP_IMPLICIT_RELEASE_DISABLE|
487 				RTE_EVENT_DEV_CAP_RUNTIME_PORT_LINK |
488 				RTE_EVENT_DEV_CAP_MULTIPLE_QUEUE_PORT |
489 				RTE_EVENT_DEV_CAP_NONSEQ_MODE),
490 	};
491 
492 	*info = evdev_sw_info;
493 }
494 
495 static void
496 sw_dump(struct rte_eventdev *dev, FILE *f)
497 {
498 	const struct sw_evdev *sw = sw_pmd_priv(dev);
499 
500 	static const char * const q_type_strings[] = {
501 			"Ordered", "Atomic", "Parallel", "Directed"
502 	};
503 	uint32_t i;
504 	fprintf(f, "EventDev %s: ports %d, qids %d\n", "todo-fix-name",
505 			sw->port_count, sw->qid_count);
506 
507 	fprintf(f, "\trx   %"PRIu64"\n\tdrop %"PRIu64"\n\ttx   %"PRIu64"\n",
508 		sw->stats.rx_pkts, sw->stats.rx_dropped, sw->stats.tx_pkts);
509 	fprintf(f, "\tsched calls: %"PRIu64"\n", sw->sched_called);
510 	fprintf(f, "\tsched cq/qid call: %"PRIu64"\n", sw->sched_cq_qid_called);
511 	fprintf(f, "\tsched no IQ enq: %"PRIu64"\n", sw->sched_no_iq_enqueues);
512 	fprintf(f, "\tsched no CQ enq: %"PRIu64"\n", sw->sched_no_cq_enqueues);
513 	uint32_t inflights = rte_atomic32_read(&sw->inflights);
514 	uint32_t credits = sw->nb_events_limit - inflights;
515 	fprintf(f, "\tinflight %d, credits: %d\n", inflights, credits);
516 
517 #define COL_RED "\x1b[31m"
518 #define COL_RESET "\x1b[0m"
519 
520 	for (i = 0; i < sw->port_count; i++) {
521 		int max, j;
522 		const struct sw_port *p = &sw->ports[i];
523 		if (!p->initialized) {
524 			fprintf(f, "  %sPort %d not initialized.%s\n",
525 				COL_RED, i, COL_RESET);
526 			continue;
527 		}
528 		fprintf(f, "  Port %d %s\n", i,
529 			p->is_directed ? " (SingleCons)" : "");
530 		fprintf(f, "\trx   %"PRIu64"\tdrop %"PRIu64"\ttx   %"PRIu64
531 			"\t%sinflight %d%s\n", sw->ports[i].stats.rx_pkts,
532 			sw->ports[i].stats.rx_dropped,
533 			sw->ports[i].stats.tx_pkts,
534 			(p->inflights == p->inflight_max) ?
535 				COL_RED : COL_RESET,
536 			sw->ports[i].inflights, COL_RESET);
537 
538 		fprintf(f, "\tMax New: %u"
539 			"\tAvg cycles PP: %"PRIu64"\tCredits: %u\n",
540 			sw->ports[i].inflight_max,
541 			sw->ports[i].avg_pkt_ticks,
542 			sw->ports[i].inflight_credits);
543 		fprintf(f, "\tReceive burst distribution:\n");
544 		float zp_percent = p->zero_polls * 100.0 / p->total_polls;
545 		fprintf(f, zp_percent < 10 ? "\t\t0:%.02f%% " : "\t\t0:%.0f%% ",
546 				zp_percent);
547 		for (max = (int)RTE_DIM(p->poll_buckets); max-- > 0;)
548 			if (p->poll_buckets[max] != 0)
549 				break;
550 		for (j = 0; j <= max; j++) {
551 			if (p->poll_buckets[j] != 0) {
552 				float poll_pc = p->poll_buckets[j] * 100.0 /
553 					p->total_polls;
554 				fprintf(f, "%u-%u:%.02f%% ",
555 					((j << SW_DEQ_STAT_BUCKET_SHIFT) + 1),
556 					((j+1) << SW_DEQ_STAT_BUCKET_SHIFT),
557 					poll_pc);
558 			}
559 		}
560 		fprintf(f, "\n");
561 
562 		if (p->rx_worker_ring) {
563 			uint64_t used = rte_event_ring_count(p->rx_worker_ring);
564 			uint64_t space = rte_event_ring_free_count(
565 					p->rx_worker_ring);
566 			const char *col = (space == 0) ? COL_RED : COL_RESET;
567 			fprintf(f, "\t%srx ring used: %4"PRIu64"\tfree: %4"
568 					PRIu64 COL_RESET"\n", col, used, space);
569 		} else
570 			fprintf(f, "\trx ring not initialized.\n");
571 
572 		if (p->cq_worker_ring) {
573 			uint64_t used = rte_event_ring_count(p->cq_worker_ring);
574 			uint64_t space = rte_event_ring_free_count(
575 					p->cq_worker_ring);
576 			const char *col = (space == 0) ? COL_RED : COL_RESET;
577 			fprintf(f, "\t%scq ring used: %4"PRIu64"\tfree: %4"
578 					PRIu64 COL_RESET"\n", col, used, space);
579 		} else
580 			fprintf(f, "\tcq ring not initialized.\n");
581 	}
582 
583 	for (i = 0; i < sw->qid_count; i++) {
584 		const struct sw_qid *qid = &sw->qids[i];
585 		if (!qid->initialized) {
586 			fprintf(f, "  %sQueue %d not initialized.%s\n",
587 				COL_RED, i, COL_RESET);
588 			continue;
589 		}
590 		int affinities_per_port[SW_PORTS_MAX] = {0};
591 		uint32_t inflights = 0;
592 
593 		fprintf(f, "  Queue %d (%s)\n", i, q_type_strings[qid->type]);
594 		fprintf(f, "\trx   %"PRIu64"\tdrop %"PRIu64"\ttx   %"PRIu64"\n",
595 			qid->stats.rx_pkts, qid->stats.rx_dropped,
596 			qid->stats.tx_pkts);
597 		if (qid->type == RTE_SCHED_TYPE_ORDERED) {
598 			struct rte_ring *rob_buf_free =
599 				qid->reorder_buffer_freelist;
600 			if (rob_buf_free)
601 				fprintf(f, "\tReorder entries in use: %u\n",
602 					rte_ring_free_count(rob_buf_free));
603 			else
604 				fprintf(f,
605 					"\tReorder buffer not initialized\n");
606 		}
607 
608 		uint32_t flow;
609 		for (flow = 0; flow < RTE_DIM(qid->fids); flow++)
610 			if (qid->fids[flow].cq != -1) {
611 				affinities_per_port[qid->fids[flow].cq]++;
612 				inflights += qid->fids[flow].pcount;
613 			}
614 
615 		uint32_t port;
616 		fprintf(f, "\tPer Port Stats:\n");
617 		for (port = 0; port < sw->port_count; port++) {
618 			fprintf(f, "\t  Port %d: Pkts: %"PRIu64, port,
619 					qid->to_port[port]);
620 			fprintf(f, "\tFlows: %d\n", affinities_per_port[port]);
621 		}
622 
623 		uint32_t iq;
624 		uint32_t iq_printed = 0;
625 		for (iq = 0; iq < SW_IQS_MAX; iq++) {
626 			if (!qid->iq[iq].head) {
627 				fprintf(f, "\tiq %d is not initialized.\n", iq);
628 				iq_printed = 1;
629 				continue;
630 			}
631 			uint32_t used = iq_count(&qid->iq[iq]);
632 			const char *col = COL_RESET;
633 			if (used > 0) {
634 				fprintf(f, "\t%siq %d: Used %d"
635 					COL_RESET"\n", col, iq, used);
636 				iq_printed = 1;
637 			}
638 		}
639 		if (iq_printed == 0)
640 			fprintf(f, "\t-- iqs empty --\n");
641 	}
642 }
643 
644 static int
645 sw_start(struct rte_eventdev *dev)
646 {
647 	unsigned int i, j;
648 	struct sw_evdev *sw = sw_pmd_priv(dev);
649 
650 	rte_service_component_runstate_set(sw->service_id, 1);
651 
652 	/* check a service core is mapped to this service */
653 	if (!rte_service_runstate_get(sw->service_id)) {
654 		SW_LOG_ERR("Warning: No Service core enabled on service %s\n",
655 				sw->service_name);
656 		return -ENOENT;
657 	}
658 
659 	/* check all ports are set up */
660 	for (i = 0; i < sw->port_count; i++)
661 		if (sw->ports[i].rx_worker_ring == NULL) {
662 			SW_LOG_ERR("Port %d not configured\n", i);
663 			return -ESTALE;
664 		}
665 
666 	/* check all queues are configured and mapped to ports*/
667 	for (i = 0; i < sw->qid_count; i++)
668 		if (!sw->qids[i].initialized ||
669 		    sw->qids[i].cq_num_mapped_cqs == 0) {
670 			SW_LOG_ERR("Queue %d not configured\n", i);
671 			return -ENOLINK;
672 		}
673 
674 	/* build up our prioritized array of qids */
675 	/* We don't use qsort here, as if all/multiple entries have the same
676 	 * priority, the result is non-deterministic. From "man 3 qsort":
677 	 * "If two members compare as equal, their order in the sorted
678 	 * array is undefined."
679 	 */
680 	uint32_t qidx = 0;
681 	for (j = 0; j <= RTE_EVENT_DEV_PRIORITY_LOWEST; j++) {
682 		for (i = 0; i < sw->qid_count; i++) {
683 			if (sw->qids[i].priority == j) {
684 				sw->qids_prioritized[qidx] = &sw->qids[i];
685 				qidx++;
686 			}
687 		}
688 	}
689 
690 	sw_init_qid_iqs(sw);
691 
692 	if (sw_xstats_init(sw) < 0)
693 		return -EINVAL;
694 
695 	rte_smp_wmb();
696 	sw->started = 1;
697 
698 	return 0;
699 }
700 
701 static void
702 sw_stop(struct rte_eventdev *dev)
703 {
704 	struct sw_evdev *sw = sw_pmd_priv(dev);
705 	sw_clean_qid_iqs(sw);
706 	sw_xstats_uninit(sw);
707 	sw->started = 0;
708 	rte_smp_wmb();
709 }
710 
711 static int
712 sw_close(struct rte_eventdev *dev)
713 {
714 	struct sw_evdev *sw = sw_pmd_priv(dev);
715 	uint32_t i;
716 
717 	for (i = 0; i < sw->qid_count; i++)
718 		sw_queue_release(dev, i);
719 	sw->qid_count = 0;
720 
721 	for (i = 0; i < sw->port_count; i++)
722 		sw_port_release(&sw->ports[i]);
723 	sw->port_count = 0;
724 
725 	memset(&sw->stats, 0, sizeof(sw->stats));
726 	sw->sched_called = 0;
727 	sw->sched_no_iq_enqueues = 0;
728 	sw->sched_no_cq_enqueues = 0;
729 	sw->sched_cq_qid_called = 0;
730 
731 	return 0;
732 }
733 
734 static int
735 assign_numa_node(const char *key __rte_unused, const char *value, void *opaque)
736 {
737 	int *socket_id = opaque;
738 	*socket_id = atoi(value);
739 	if (*socket_id >= RTE_MAX_NUMA_NODES)
740 		return -1;
741 	return 0;
742 }
743 
744 static int
745 set_sched_quanta(const char *key __rte_unused, const char *value, void *opaque)
746 {
747 	int *quanta = opaque;
748 	*quanta = atoi(value);
749 	if (*quanta < 0 || *quanta >= 4096)
750 		return -1;
751 	return 0;
752 }
753 
754 static int
755 set_credit_quanta(const char *key __rte_unused, const char *value, void *opaque)
756 {
757 	int *credit = opaque;
758 	*credit = atoi(value);
759 	if (*credit < 0 || *credit >= 128)
760 		return -1;
761 	return 0;
762 }
763 
764 
765 static int32_t sw_sched_service_func(void *args)
766 {
767 	struct rte_eventdev *dev = args;
768 	sw_event_schedule(dev);
769 	return 0;
770 }
771 
772 static int
773 sw_probe(struct rte_vdev_device *vdev)
774 {
775 	static const struct rte_eventdev_ops evdev_sw_ops = {
776 			.dev_configure = sw_dev_configure,
777 			.dev_infos_get = sw_info_get,
778 			.dev_close = sw_close,
779 			.dev_start = sw_start,
780 			.dev_stop = sw_stop,
781 			.dump = sw_dump,
782 
783 			.queue_def_conf = sw_queue_def_conf,
784 			.queue_setup = sw_queue_setup,
785 			.queue_release = sw_queue_release,
786 			.port_def_conf = sw_port_def_conf,
787 			.port_setup = sw_port_setup,
788 			.port_release = sw_port_release,
789 			.port_link = sw_port_link,
790 			.port_unlink = sw_port_unlink,
791 
792 			.eth_rx_adapter_caps_get = sw_eth_rx_adapter_caps_get,
793 
794 			.xstats_get = sw_xstats_get,
795 			.xstats_get_names = sw_xstats_get_names,
796 			.xstats_get_by_name = sw_xstats_get_by_name,
797 			.xstats_reset = sw_xstats_reset,
798 
799 			.dev_selftest = test_sw_eventdev,
800 	};
801 
802 	static const char *const args[] = {
803 		NUMA_NODE_ARG,
804 		SCHED_QUANTA_ARG,
805 		CREDIT_QUANTA_ARG,
806 		NULL
807 	};
808 	const char *name;
809 	const char *params;
810 	struct rte_eventdev *dev;
811 	struct sw_evdev *sw;
812 	int socket_id = rte_socket_id();
813 	int sched_quanta  = SW_DEFAULT_SCHED_QUANTA;
814 	int credit_quanta = SW_DEFAULT_CREDIT_QUANTA;
815 
816 	name = rte_vdev_device_name(vdev);
817 	params = rte_vdev_device_args(vdev);
818 	if (params != NULL && params[0] != '\0') {
819 		struct rte_kvargs *kvlist = rte_kvargs_parse(params, args);
820 
821 		if (!kvlist) {
822 			SW_LOG_INFO(
823 				"Ignoring unsupported parameters when creating device '%s'\n",
824 				name);
825 		} else {
826 			int ret = rte_kvargs_process(kvlist, NUMA_NODE_ARG,
827 					assign_numa_node, &socket_id);
828 			if (ret != 0) {
829 				SW_LOG_ERR(
830 					"%s: Error parsing numa node parameter",
831 					name);
832 				rte_kvargs_free(kvlist);
833 				return ret;
834 			}
835 
836 			ret = rte_kvargs_process(kvlist, SCHED_QUANTA_ARG,
837 					set_sched_quanta, &sched_quanta);
838 			if (ret != 0) {
839 				SW_LOG_ERR(
840 					"%s: Error parsing sched quanta parameter",
841 					name);
842 				rte_kvargs_free(kvlist);
843 				return ret;
844 			}
845 
846 			ret = rte_kvargs_process(kvlist, CREDIT_QUANTA_ARG,
847 					set_credit_quanta, &credit_quanta);
848 			if (ret != 0) {
849 				SW_LOG_ERR(
850 					"%s: Error parsing credit quanta parameter",
851 					name);
852 				rte_kvargs_free(kvlist);
853 				return ret;
854 			}
855 
856 			rte_kvargs_free(kvlist);
857 		}
858 	}
859 
860 	SW_LOG_INFO(
861 			"Creating eventdev sw device %s, numa_node=%d, sched_quanta=%d, credit_quanta=%d\n",
862 			name, socket_id, sched_quanta, credit_quanta);
863 
864 	dev = rte_event_pmd_vdev_init(name,
865 			sizeof(struct sw_evdev), socket_id);
866 	if (dev == NULL) {
867 		SW_LOG_ERR("eventdev vdev init() failed");
868 		return -EFAULT;
869 	}
870 	dev->dev_ops = &evdev_sw_ops;
871 	dev->enqueue = sw_event_enqueue;
872 	dev->enqueue_burst = sw_event_enqueue_burst;
873 	dev->enqueue_new_burst = sw_event_enqueue_burst;
874 	dev->enqueue_forward_burst = sw_event_enqueue_burst;
875 	dev->dequeue = sw_event_dequeue;
876 	dev->dequeue_burst = sw_event_dequeue_burst;
877 
878 	if (rte_eal_process_type() != RTE_PROC_PRIMARY)
879 		return 0;
880 
881 	sw = dev->data->dev_private;
882 	sw->data = dev->data;
883 
884 	/* copy values passed from vdev command line to instance */
885 	sw->credit_update_quanta = credit_quanta;
886 	sw->sched_quanta = sched_quanta;
887 
888 	/* register service with EAL */
889 	struct rte_service_spec service;
890 	memset(&service, 0, sizeof(struct rte_service_spec));
891 	snprintf(service.name, sizeof(service.name), "%s_service", name);
892 	snprintf(sw->service_name, sizeof(sw->service_name), "%s_service",
893 			name);
894 	service.socket_id = socket_id;
895 	service.callback = sw_sched_service_func;
896 	service.callback_userdata = (void *)dev;
897 
898 	int32_t ret = rte_service_component_register(&service, &sw->service_id);
899 	if (ret) {
900 		SW_LOG_ERR("service register() failed");
901 		return -ENOEXEC;
902 	}
903 
904 	dev->data->service_inited = 1;
905 	dev->data->service_id = sw->service_id;
906 
907 	return 0;
908 }
909 
910 static int
911 sw_remove(struct rte_vdev_device *vdev)
912 {
913 	const char *name;
914 
915 	name = rte_vdev_device_name(vdev);
916 	if (name == NULL)
917 		return -EINVAL;
918 
919 	SW_LOG_INFO("Closing eventdev sw device %s\n", name);
920 
921 	return rte_event_pmd_vdev_uninit(name);
922 }
923 
924 static struct rte_vdev_driver evdev_sw_pmd_drv = {
925 	.probe = sw_probe,
926 	.remove = sw_remove
927 };
928 
929 RTE_PMD_REGISTER_VDEV(EVENTDEV_NAME_SW_PMD, evdev_sw_pmd_drv);
930 RTE_PMD_REGISTER_PARAM_STRING(event_sw, NUMA_NODE_ARG "=<int> "
931 		SCHED_QUANTA_ARG "=<int>" CREDIT_QUANTA_ARG "=<int>");
932